„Dijkstra described a fictional company of which he served as chairman. The company was called Mathematics, Inc., a company that he imagined having commercialized the production of mathematical theorems in the same way that software companies had commercialized the production of computer programs. He invented a number of activities and challenges of Mathematics Inc. and documented them in several papers in the EWD series. The imaginary company had produced a proof of the Riemann Hypothesis but then had great difficulties collecting royalties from mathematicians who had proved results assuming the Riemann Hypothesis. The proof itself was a trade secret.“
Interesting story certainly. I don't see however why random stories are being posted that don't have anything to do with the news. The LEAN proof has been made available for all to read.
The story is about a company that has proven the Riemann Hypothesis. That alone is pretty close to the news nowadays. The proof by Alpöge might be open-source, but the technology used to generate that proof (Claude) is definitely a trade secret of Anthropic. So for me this story from the 1970s shows surprising parallels to our reality of 2026.
But there is more on the Wikipedia page. Dijkstra also used to quip: "The question of whether Machines Can Think is about as relevant as the question of whether Submarines Can Swim."
> That alone is pretty close to the news nowadays. The proof by Alpöge might be open-source, but the technology used to generate that proof (Claude) is definitely a trade secret of Anthropic.
The four color theorem took about 1000 hours of time on a the mainframe. Was the proof of the four color theorem suspect in this manner when it was released, and only became okay when everyday computers could run through the program?
Dijkstra was a devoted Christian. He also died years before seeing the state of modern LLMs. Religious people have certain magical supernatural biases regarding human consciousness. That is of course usually not a problem when they are working in any other field such as computer science or mathematics as long as they are not speaking on those topics where religion muddles ones thinking.
Funny, I interpreted the quote as explicitly rejecting human exceptionalism because I thought "yeah, submarines can obviously swim but it's a silly question"
Maybe he meant it as in both can achieve the goal in different ways, I am not sure, but it does seem like he is saying they are totally different things. In fact now that I think of that quote, any interpretation starts to become messy. I mean can submarines swim? They do seem to be swimming. I am not sure what answer he intended for this part of the analogy.
To me the implication of the quote is: you can call it swimming, or you can confine "swimming" to a narrower meaning, but it doesn't really matter. In other words it is purely a semantic debate.
Faith is fine as long as it doesn't clash with reality. And the reality is, the human brain is a physical system. Unless you can prove otherwise, consciousness is a property of some physical systems, so there is no reason to think some another physical system can't have consciousness.
Faith is fundamentally incompatible with science. It demonstrates an individual is capable of believing things without evidence, which is the antithesis of science.
This then becomes a source of mental perjury: once one leap of faith has been shown, you must doubt every other conclusion, as it may too be based on some leap of faith. It undermines credibility to an extent that I pretty much dismiss any who tries to combine the two.
I personally find the logical conclusions of anyone who firmly believes in a god to be suspect. We have enough scientific evidence to rule out the possibility with enough certainty that no one should be taking that wager.
Go ahead and hate me, but no one will ever convince me otherwise.
I believe the same, it frankly in any other circumstance would be called clinical insanity, but I try to be more diplomatic in saying this.
With mathematics and computer science work at least, we can be certain since they are formal and checkable regardless of beliefs of the creator, and as far as we know Dijkstra's work in cs is generally quite accurate.
Finally finished compiling on my PC (well, actually an HPC machine at my company, but it's the weekend, and I hope nobody is using it) - it took 2.4 hours on a 480-core Xeon and produced 89 GB of compilation output. I had to make some modifications because, for some reason, Lake bottlenecks and stalls verification. I worked around this by compiling the files directly instead of using `lake build`. I also had to increase `vm.max_map_count` for some later modules.
All 175,187 modules seem to check out with Lean 4.33.1, with no errors or `sorry`. There are tons of linter warnings - over 460k of them - but I think they might be caused by the way I invoked it. I haven't checked the actual axiom being proved, though - I'm not good at math :)
It really seems he was rushed to publish and did not bother cleaning the code up at all. I am compiling the code and reviewing it, and there are 204 dead modules that are not referenced anywhere. And a LOT of repetitive code.
I think this should not be 17 million lines, I bet this can easily be under 1 million lines of Lean with a little bit of cleanup.
It is probably not that compressible. Anthropic FLT proof also has a lot of redundancies but my projection is without true mathematical refactoring it will be hard to get to 50% reduction in line count. One problem is that Lean, or type theory really, is truly verbose. In these parallely developed proofs if two agents use slightly incompatible definitions there are a lot of translations back and forth. Also FLT is following routes well established by mathematicians and it still got to 13 million lines. So I am actually surprised that GRH is not much bigger, as it reportedly ranges over a wide swath of fields.
Yes it would be interesting to see that version. If compaction comes from proof architectural improvement (what I called "true mathematical refactoring"), it could improve readability. On the other hand if it is tactical, for example using tall rw stacks that are incomprehensible to humans, shortness doesn't necessarily imply better accessibility. I suppose it has to be a lot of the first kind to get that much compression.
Hence why it seems rushed. Backed by Anthropic he certainly can clean it, model access or $$$ is probably not a limiting factor; only time.
I don't know if Anthropic IPO being yesterday, or American Mathematical Society Executive Committee and Board of Trustees meeting being today has anything to do with the announcement happening today, but it's certainly interesting timing.
Apparently a mostly human/human only unfinished proof on the hypothesis was rushed to publication upon hearing this news. It is around 1000 pages long. I don't know what the approximate ratio should be but fully formal propfs generally are much longer than natural language proofs. So probably this size is ij 5he ballpark of normal. If anyone sho writes lean proofs or works in formalized mathematics can comment that would be nice.
I am a hobbyist home-server-builder and would love to know more about you and your company. I love hearing cases where you as an employee get direct access to the hardware. can you share any more details? I'm kind of spazzing out right now with how casually you said some wildly awesome stuff.
I probably shouldn't say this, but I work at a company related to CPU manufacturing.
Part of my job is to ensure that HPC programs (largely LLM libraries these days) work on their next-generation CPUs and with new BIOS revisions. We test them using a shared pool of machines with various configurations. GPU machines are much scarcer, and we have to share them with other teams, so it's hard to reserve a time slot. I usually reserve a CPU machine around the clock so I can work whenever I want (I'm more productive at midnight), and nobody seems to mind.
I've worked here for almost a year now, and I love my company ;)
I run many high core jobs for various reasons, I often reach for many smaller machines rather than one enormous node like this. But not familiar with how lean parallelizes.
Any company which self hosts in a data center has employees accessing to hardware. Every company I've worked at was on-prem. Why is this treated as an outlier? Is everyone on AWS these days?
No need to be snarky... having access to HPC clusters has little to do with "on-prem VS cloud" and is certainly an outlier in the prevalent SaaS world.
Most of the OpenAI proofs from the last big dump that have been looked at with some time don't actually match the lean claim now the media circus has moved on so I think the industry earned their lack of benefit of the doubt for a lot of people.
Yes but only for independent parts of the process. Having all of these modules helps quite a bit with that.
It really is quite slow, I've been having my dot and its 6 sub astras churn away at a proof for Z is Diophantine in Q, even 1/10th as many modules is an overnight affair on my 5800x3d with 128gb of ram. The first version I started on was 326 or 318 variables, but it's allegedly also found some in the mid 100s.
There's no real definitive measure of progress, all of the foundational stuff is handled and now it's working on the tricky stuff that is assumed away in the original that made me decide to let them loose trying to formalize.
Thanks for that. Could you have an AI summarize the 460k linter warnings and decide if any of them are significant? Also, how much RAM was in the machine, please.
The machine itself has 1.5 TiB of memory; I did not check the free memory while it was running, sorry :(
LLM-generated text below:
```
The build printed about 460,000 linter warnings, and none of them affect correctness. Almost all are code-quality issues such as redundant proof steps, style, and unused variables. About 33,000 are uses of deprecated Mathlib names, which only matter if we upgrade Mathlib.
Memory per process wasn't logged during the run. The container's lifetime memory peak is 441 GiB, but that includes file cache and is only an upper bound. A typical module uses about 0.5 GB of its own memory, and the heaviest use about 8 GB. With about 280 compiles running in parallel, the build used at most about 215 GB of the machine's 1.5 TB of memory, based on spot checks.
```
One real problem with depending on Lean is that if there are any real bugs, then sufficiently-large swarms of agents are guaranteed to eventually come up with proofs that depend on the bug.
This comes up frequently in threads but it's actually a much less impactful issue than you might think. There are actually several entirely disjoint implementations of Lean, each with different bugs; it is incredibly unlikely an agent would be able to exercise the same bug in every single implementation.
You trivially could through the principle of explosion.
That said at least one version is itself formally verified, at the machine code level if I understood correctly. It might still make mistakes but you're starting to get to a point that those mistakes become quite unlikely.
It's still a pretty tricky endeavor but we should know soonish, with some confidence if it is correct or not.
I'd argue Gödel's incompleteness theorem is more of a formale example of the same.
Though it is technically different. Mathematicians were happy to pick a set of axioms, they were less happy to learn any theory capable of stating facts about basic arithmetic would necessarily be unable to prove all things it could state.
I’ve heard it argued that equality has a bug. In real life it’s a lossy comparision as almost nothing is truly equal (except maybe atomic and subatomic particles?)
I doubt that any of them are closed-source, so the clankers probably already know all their bugs. Who says they all don't share a similar bug that is necessary for the proof to work? No one is going to review 1.3GB of anything by hand, even if they could understand it all in a single lifetime.
I found a bug in Lean a couple of weeks ago with an LLM, in fact, but it was in its 'macro' system that expands code into statements, not the actual proof system. It wouldn't have proven anything false, but it could have produced statements that aren't what you intended.
Shell tip: You do not have to download each part separately, then cat them and then untar it. This does it in a single streaming pass without the temporary files:
( for i in {01..14}; do curl -fSL https://alpo.ge/lean.tar.gz.part$i ; done ) | tar xzv
I think in this format this leads to shell expansions. For 14 thats not an issue, but I try to use for loops as it mitigates the arg max limits better. I believe if you literally quote the string, curl can internally do the range expansions too if you used square brackets instead of braces and passed it literally.
Yeah I know. But shell just passes multiple URLs as arguments to curl. I didn’t know curl would accept that. In your example if there were more expansions it is up to mv to require the last one to be a directory. And many others just don’t accept multiple arguments.
Your initial comment stated that to support multiple URLs it is required to use [] syntax. This is inaccurate, because curl supports multiple URL args hence shell expansion {} works just fine.
An update: the Lean community is now maintaining https://github.com/leanprover-community/riemann-hypothesis and is currently checking the proof against several different toolchain setups. There are instructions to get the usual compiled files.
This seems to be norm in these top-level proofs. However, some comments here say that some of the modules are repetitive. In that case, can we not ask the AI to simplify the proof based on that to make it more tractable? Same as we often ask it to improve or refactor code that already works.
They just want to be as fast as possible for their claim to fame I guess. OpenAIs Proof of the Existence of Yang-Mills Quantum theory has like 250000 declarations in it and 60% is unneccessary.
It's possible you can't do that if this proof-producing activity has to do with the nature of language, and not the nature of intelligence.
If the modules are repetitive and also signify that no intelligence has taken place and no understanding has occurred (if there was intelligent understanding in place, the first thing an intelligent being with a theory of mind would do is strengthen their claim to a coherent thought by, well, making it more coherent) then we're looking at something that makes strong claims to what language is, NOT what intelligence is.
I don't think anyone's questioning the shape right away, the first pass through most agent's is always known to be bloated. Another user (jycheng) literally found 204 dead or orphaned modules a little way down. Knowing it works is the real breakthrough, let's not lost sight of that.
Size can be solved next, FLT went from 13m to 2m lines in a couple of months, the same will happen here for sure. The tough (or expensive part) imo is checking that the toplevel statement actually says GRH. That's the real review effort, not solving for size just yet.
I have never seen a more out of touch claim, at this size there is just as much chance there was a bug in lean, there's bug in the code which LLMs cant resolve maybe both, than assuming it 100% correct no questions asked.
I say that with the belief that this is likely correct. They should have solved the size how much would it take 3-4 days?
Who are running defense for here, I dont get the point of defending slop just because its something of value.
Code is meant for humans to read and make sense of, same with maths these concepts are for Humans even if machines do generate them. Otherwise machine should be talking to each other in some unknown super compressed language.
Yes my friend thats how things used to work, and should even now. Assuming all models can solve this stuff why cant we wait a few days to cleanup our work atleast and make more sense out of it.
Didnt OpenAI learn and started to cleanup the work a tad bit... ofc its not perfect but this thing is a joke.
I dont and wont understand the HN mindset of winning the race at any cost mindset.
This seems like a stock pump attempt that failed miserably.
I dont see the point we can all be civil and build better stuff now. Why try to one up others at the cost of throwing out slop.
Trust me if the value of your product can even be measured in days, then thats not a product worth investing in.
I personally think AI companies deserve to fail if they cant even understand that.
To clarify, I didn't say it's 100% correct, no questions asked. I said the opposite in fact: the expensive part of reviewing this is checking that the top-level statement actually says GRH. Actually exactly in line with what you say, that to me is where the skepticism should go.
What it "works" here to me is simple that it type-checks. Another user (hafeoz) even compiled all 175k modules upthread, with no errors and no sorry. Of course, a kernel bug is possible, but there are independent reimplementations of the checker, so you'd need the proof to slip past every one of them. A wrongly stated theorem is the more realistic failure, I think someone brought this up as well, and that only needs a page or two of definitions checked by humans, not 15M lines.
On readability: nobody was ever going to read 15M lines, or tbh even 1M after cleanup. That's what the kernel is for. Cleanup makes it easier to learn and build on, nobody's saying that shouldn't be done. My argument is simply dead modules are ugly, not wrong. OP himself says he expects it to be massively simplified.
As for "why not spend 3-4 days on it first": it went out today, and people were independently checking it on their own hardware within hours. That's arguably the better order, iterative progress first on successful ideas, not perfection. That's kind of what the whole YC ethos (who's page we're on, is. lol). Besides, a human proof got rushed out within a day because of this one, so the incentive to publish first was clearly real. We're on the same page here, just the 3-4 days of simplification first isn't the right way to progress imo.
In a sense that is not surprising. If there were a short proof to a famous 150 year old maths problem it would likely have been found by now.
Presumably as LLMs work their way through the rapidly shrinking list of Millenium Prize problems, it will be the case that none of them have concise enough solutions that they would have ever been solved by a human unaided.
I don't think this means that more concise solutions aren't available. This is likely more just the easiest path for an LLM to arrive at a proof.
That said, you are probably right that the most precise proof is too long for a human to generate unassisted.
Little bit like solving a rubrics cube. There's an infinite number of ways to solve it and you can use an infinite number of steps to reach a solution. Pros tend to use fewer steps, amateurs tend to use more.
That's actually a great analogy for AI maths in general. There's one or more paths from the start state (unsolved cube) to the end state (solved cube). You have a set of rules that are verifiable. Deploy a swarm of agents to intelligently search this space, pruning branches that lead to nowhere. With enough compute, you will find a solution, but it might not be the most elegant solution.
I think it's more that the proof requires several leaps in multiple fields/areas of mathematics, so I wouldn't be surprised if this was actually something like 100 separate proofs.
In the same sense its that the descriptive machinery isn't fully developed. If we wrote all of our proofs in a Godelian representation of statements we would also come out with tomes per proof.
Here's an allegedly human proof in fewer than 1000 pages, the result of 12 years of work. Still a draft but they published it in its current state because Anthropic published theirs.
It is really good that those things happen right now IMHO, because we really have to ask the question of value in a different way. We have long been overstating the value of success. What is really the value of having proofed the GRH as we long had empirical support? (No expert here). Now is the 1.5M SLOC something of value? Could we derive a thousand times smaller human readable version much easier now? (Like having an oracle) . I hope this will lead to a time of fruitful philosophical and societal discussion, which might result in a better world (if we haven't destroyed the planet by then irreversibly)
> We have long been overstating the value of success. What is really the value of having proofed the GRH as we long had empirical support? (No expert here).
I think this has been debated many times before (I also suppose that this paraphrases Terrence Tao's take on AI), the value is in elevating the discipline. Those problems were selected because they were understood to be so out of reach that new constructs and theories would have to be developed. Those new tools would benefit mathematics at large because they might in turn unlock new, harder, or more generic problems to be solved. The value lies in the potential to unlock further discoveries, by having an increasing number of humans understanding, sharing and building upon each other and into ever more distant fields and disciplines.
The elephant in the room is that so far nothing indicates that LLM-generated proofs contribute meaningfully to building new frameworks and furthering our understanding of mathematics, therefore benefitting it. What we have is an inscrutable series of steps to prove that a single statement is true, plus the knowledge that LLMs aimed at a narrow problem can brute-force it given enough resources. The proof is technically only useful as it serves to solidify corollaries derived from the GRH by earlier humans. Will LLMs come up with interesting/important/meaningful corollaries of their own? That's a creative endeavour, not something that boils down to an optimization/path-finding problem.
More realistically, they'll turn out to be actually really bloated bug reports for the framework the proof was made in ಥ ‿ ಥ
/Edit: I'm not even "against" ai value. For business proposes it's basically a match made in heaven... But their ingenuity and novelty is really overstated. They're really average. If you consider their output good, you're probably incompetent. If you consider it garbage,you're probably better then average. But for business, average is good enough hence me saying being not "against" ai value
Indeed, just ask LLMs to suggest products name or domains that sound good, are easy to pronounce, suggest some immediate meaning for a certain concept, are somewhat punchy, and have other qualities that make for a strong name.
LLMs are virtually useless at this. I routinely come up with much stronger names, and when I tell the LLM, they gush over it.
> Claude cutely called it the GRH... "book" — I looked at Theorem 1 and it began by noting that because the reals are totally ordered, if σ = 1 then σ > 1 or σ < 1
I respect that he decided to just post the lean proof and not bother others with 250k pages of badly written math that nobody would be expected to read anyway and would only serve as some kind of vanity thing. It is more respectful compared to some other approaches lately, imo.
Maybe the future of mathematics is gonna be AIs spitting out thousands lines of lean and human mathematicians trying to decipher them (with AI)
I like the idea that the universe isn’t necessarily beautiful and orderly, because my life has seldom been orderly. “Imagine a spherical cow” is fine? But there are no spherical cows.
The assumption that the universe is orderly, has beautiful laws, and doesn’t have potentially one-off events or rules is a religion.
I have had this thought experiment for decades. “Imagine that once every 1.2 million years a grand piano spawns approximately four miles from the city center of modern day Phoenix. That’s just the rule. That’s a law of the universe.” There would be no way to measure this because if it happens to happen during our time it would just be a weird event that nobody could explain and we would have to wait another 1.2m years to see it again. Ain’t nobody got time for that? So all of our explanations for this would be wrong and would never be able to explain this or predict the long term future of the universe because hypothetically our universe would be dominated by the mass of these grand pianos eventually. Maybe beings then could piece together what had happened then? But likely not. They’d be just as baffled as we are.
Parsimony makes sense to us, because we’re simple creatures that live in finite time with finite senses. There’s no reason that nature should favor simplicity over complexity other than our own biases.
> There’s no reason that nature should favor simplicity over complexity other than our own biases.
No reason? Isn't the fruitfulness of this so far itself a very good reason? We've taken this approach and successfully explained and predicted so much of the universe already. How is that not a reason to keep going?
I’m just saying that the universe is under no constraints to be understandable to humans, and our human limits might impose limitations on what we can model.
You called it a religion, and the main thrust of the response was that the fruitfulness thus far was a reason to not consider it a religion, but rather a rational course of action. It seems you got hung up on the phrase “keep going”.
Well, I appreciate it. In your defense, I'm not sure that I entirely agree with this take either? But I think it's important to consider that we may not be clever enough to tackle all problems.
Like, my dog is pretty clever, she's a good dog too! She can roll-over and shake, and after much work I can get her to "stay" for a limited period of time. She's a good dog.
She's not smart enough to do calculus I don't think, and even if maybe she is physically smart enough to do calculus there's no way I could communicate the concepts to her. So... what's like that for us humans? What's like that for our experience of the universe? Like, I can look at a spreadsheet, and do some math on various columns in a dataset. I can do the SVD here or there.
But... I cannot really intuitively understand something in... say, 30 dimensions. The best I can really do when I feel really clever and think really hard, is maybe 4 or 5 if I'm being generous as to what I consider a "dimension." If I do something like say, "ok, picture a cube, and then visualize the heat on the cube as a color scheme, there's a burner below the cube." That's "kind of" like visualizing 4-d, and maybe I can add some roughness to cube too, to add another dimension to the data.
But that's the best I can do and the temperature is not even a spatial dimension . Meanwhile, it's not uncommon for like 10 dimensions to be useful in a dataset for prediction of some value - maybe even more. So, what are we not smart enough to understand? What are we simply too dumb to truly get? What will we never be able to understand unaugmented? And maybe parsimony and the "unusual effectiveness of mathematics" or whatever is that? But maybe... math isn't unusually effective, maybe the universe is substantially more complex than we can even comprehend? And our simple first order approximations are useful, but miss massive amounts of the underlying geometry/structure/system that makes up reality?
I guess, that's kind of the Donald Hoffman take on it right? This is the interface, but our interface kind of sucks.
This is roughly the position of Charles Fort in the late 1800’s when he started collecting reports and compiled them into his "Book of the Damned", suggesting that there’s lots of your pianos happening - rains of frogs, teleportation, clairvoyance, etc - and scientists discount them as valid evidence because they would ruin nice clean simple theories. i.e. scientists pick and choose the evidence that lets them conclude a simple theory.
This gave rise to the term 'Forteana' for unexplained happenings, and The Fortean Times magazine. I think this stuff likely felt more plausible in the 1880’s before much understanding of electric fields making hair stand on end, drafts and flickering oil lamps being common, before atomic theory and the electromagnetic spectrum and radioactivity and thermodynamics when there were still a lot of big unexplained things around, before tape recorders, point-n-shoot cameras, then smartphones, were invented.
The ubiquity of smartphones on Earth selling at a billion a year and still no convincing recordings of Bigfoot, ghosts, UFOs, force fields, teleportation, is evidence indicating absence despite the saying “absence of evidence is not evidence of absence”. If you aren’t warm, that is some evidence that there’s no fire. If we had a lot of independent witnesses and videos of your piano appearing we would still lean towards it being a magic trick, hoax, CGI, because we have agreed that you can’t create energy and can’t teleport therefore it didn’t happen as it appears.
I haven't read Carl Sagan's Demon Haunted World but I think it's a counter position, look at all the nonsense humans have believed over history and still do believe today, and about the only way we can divine the real from the nonsense is repeatable tests - "science as a candle in the dark".
But then Occams's Razor is no longer a guiding principle, which is fine, because the Universe has no obligation to be simple.
Also, claiming a teapot orbits does not put the burden of proof on the claimant, because, again the idea that the one making an extraordinary claim MUST offer extraordinary evidence is just that, what we think is a nice approach to take. Truth does not care who has the burden of proof.
That's why miracles are not to be brushed off; maybe you just never saw one. Maybe they only occur under some weird circumstance the vast majority of people do not experience. Not say there is material proof of them, but mostly the objection to them is based on politics and adherence to the religion of "we have the universe all figure out" We could be living in a simulation, and "miracles" are a bypass of the simulated reality.
Hey, I'm a fan of Occam, but.... I recognize that maybe, kind of sort of... what if we're at the limits of what can be described by Occam? Or similarly, what if like, we can make crappy models that are parsimonious and accurate models that are not. What if the universe is parsimonious when you don't look very close and extremely chaotic and weird and inconsistent when you do? That would surely track much of my lived experience?
But who knows? I'm surprised most days that there's something rather than nothing, that I exist at all (it's quite strange to be anything at all), and that math and computers work. I'm quite glad they do? But yeah, things seem pretty damn miraculous to me most days.
You don’t know that. You believe it - and to be fair, I believe it too - but you don’t know it.
The parsimonious answer feels good to us, and I think mostly it bears fruit - but there’s no guarantee that parsimony is the only way to model things accurately
It takes thousands of hours of study for a human to get to the point of understanding what the RH says and means. It isn’t hard to imagine that there are mathematical truths that require far more than this to understand them.
Pretty much every major result in mathematics - assuming ZFC as the background theory - requires thousands and thousands of pages to formally prove from the axioms of set theory.
An infinite number of math proofs really need more than 15M to write out. Could that include some easy-to-state ones that we care a lot about? Yeah probably.
Not sure why you got downvoted but this is actually quite deep. I guess it depends on the mass of the universe. Since our universe is approximately a black hole, its mass determines its area and the number of bits it can store on the surface through the holographic principle. If the universe is not infinite there is a physical limit to math.
A finite universe does not imply a finite limit to the size of proofs of true statements.
There are proofs that might not be capable of being written down in the material universe, even if we assume a plank volume per symbol. They would exist nevertheless.
tbh i was the student in undergrad maths classes using induction to prove trivialities bc nobody handed me a “book of axioms” and my intro to writing proofs was using Coq. let’s cut some slack to our autistic LLMs — I am sure in six months we will know how confident we are in this proof; no need to rush.
Do you NEED 15M lines? No, it's almost certain this proof can and will be reduced, but the goal of an initial proof of a theorem like this isn't conciseness, it's to actually... prove the theorem.
Fermat's Last Theorem was initially proven in Lean in September and that proof required 13M lines, it has since been reduced to 2M lines and it will likely be reduced even further. Anthropic claims they can reduce it down to less than half a million lines by doing similar techniques to what software developers do... find common and reusable snippets and factor them out instead of repeatedly proving similar sub-theorems over and over again. There are also opportunities to replace lengthy sequences of explicit reasoning with automated tactics, reuse existing mathematical libraries including the Lean standard library, and reformulate arguments at a higher level of abstraction.
Apparently healthy skepticism is a downvotable offense now. Not that I care about the karma. It's just interesting that we live in a time where taking anything other than the most extreme position is unacceptable to some people on here.
I guess I'm supposed to either accept the galactic proof, which likely hasn't been fully read by a single human yet, as gospel and speak breathlessly about the singularity, or I should be insisting the proof will fall apart and make accusations about deliberate academic espionage.
If it was intelligent it would want to shorten the proof and remove broken pieces, because that would strengthen the argument that the proof is good… assuming it had a theory of mind capable of understanding that another thing is evaluating its performance. It hasn't, isn't… something else is underway and we are observing the parameters of that.
Isn't there a way to reduce this? What's the reason for such a massive size? Are there "already proven" proof libraries in Lean that the LLM is not using like reinventing the libc from scratch?
The only verified “already proven” library is mathlib, which gets you up to about half way through an undergraduate degree. It may be substantially more efficient for Anthropic to direct Claude at making a Lean library consisting of all of known mathematics (at least up to the end of the 20th century), and then use that as a base for their hypothesis. Instead of banging together Lean code at considerable expense until something magiced out, Claude would have been able to build on top of the works of the math community and produce a proof that was actually readable. Claude would also be able to reference known proofs while it reasons through the problem.
prove that (Russell's definition of 1 as a set) (Russell's definition of + in terms of set operations) (Russell's definition of 1 as a set) = (Russell's definition of 2 as a set), that is
He did not prove one apple and one apple make two apples, he proved that his complicated definitions of arithmetic using sets were not completely broken.
For whatever reason, the LLMs currently generate very long proofs of results. For example, they often rephrase earlier arguments rather than citing them, or repeat their own arguments. (I've experienced these myself.)
It'll just take some work to simplify this.
From Alpoge himself:
"My intuition insists it will be massively simplified, likely immediately.”
This is true for code, even frontier models, they will often not use a library, or not modify the library but rewrite an almost identical function(s), etc. It's very interesting and needs constant vigilance
judging by claudes review because it actually rebuilds and proves a bunch of stuff in mathlib that wasn't sufficiently formalized, but were required dependencies.
I suspect it includes a lot of formalization of existing, peer-reviewed mathematics. Including these bloats the code but makes the verification more self-contained (less AI-written unverified inputs to lean).
I'm absolutely positively sure that there are no sorrys and no theorems taken as axioms _anywhere_ among those Windows 11 source code-sized Lean proofs.
I mean… that would be a trivial check with `ripgrep` et al. I’d be more worried about other sneakier tactics that LLM’s like to do as to move the goal posts (prove something trivial posited as grander, take arguments in a function head but not use them in the body etc)
Yeah, this smells like BS. A "proof" that no one will understand within the next 1-2 years, and even then it's questionable that anyone could ever seriously hold 250k pages in their mind, and that no single mistake or oversight has slipped in. This is a haystack with an unknown amount of needles in it, and gaining certainty that that amount is zero, is a whole other task.
Here’s a link to an interview with Levent Alpöge (in Turkish) from 2014, when he just graduated from Harvard at the top of his class.
He mentions solving the Riemann hypothesis is his life goal and he wants to work on it, but he’s not sure if it can be done in his lifetime. Way to go :)
Would anyone reasonably assume that the author wouldn't simply prompt the LLM to verify this thing across all kernels before even submitting it anywhere?
Even all things aside, at least I would not assume to have checked single thread kernels as those take very long and having access to a supercomputer won't help you in that.
The point of independent verifications is to not have to rely on this kind of assumptions. I don't think anybody would reasonably expect with any high probality that there was a lean bug in this.
Even if we discover verification errors, the accomplishments of Levent & Alessandro provide the mathematics community with several new tricks that will be leveraged for years to come. It would be particularly exciting if both proofs end up getting verified using different routes.
The reason I'm eagerly anticipating rigorous verification is to see if Levent's belief is correct that the same concepts can be used to prove extensions of the main hypothesis.
Does it even matter anymore? I mean of course it does, what I mean it that even if it's somehow incorrect (which is important) wouldn't he try again until it does pass and that given the problem space, search tool and compute available it becomes tractable (which might not be)?
I'm not for the field but I can't help but wonder how does one backtrack from the "solution" to the meaningful learning.
It changes nothing. You can just assume it's true and work from there.
Mathematicians need to stop toiling on proofs and acquire taste for meaningful questions.
Right now, people graduate top of their Harvard class and work on a question that was invented 150 years ago and everyone just assumed to be true. Kind of meaningless, right? Could have just waited three years for a full AI-only proof without human intervention.
People were working on these classical open problems because they were stimulation for creating new interesting directions and theories. Even this AI proof is probably using some of this work.
And you cannot acquire taste without getting dirty.
Well first we dont know if he use internal or public model, even if he is their employee.
Second, a lot of other employees also have access to internal model, but AFAICT, he didn't claim anyone co-author with him, expect Claude, forgive me if i'm wrong here.
I wonder if we’ll have room temp superconductors and fusion this time next year.
If this holds up it seems as good a day as any to call the end of one time and beginning of another.
I wish I could be fully happy about this new era instead of worrying if I’ll be alive in 5 years or not. If there was any time for Anthropic to try to coordinate a pause, it feels like now or never, they seem to have a pretty strong lead at the moment. But probably they won’t and can’t, and we will just put the fate of our species to a coin flip.
If multiplying matrices on a million GPUs reveals a unique lattice structure that in simulation enables room temp superconductors, the real experiments would immediately begin.
Look at LK-99, it felt like a global effort was inspired at the time to verify room temperature superconductivity.
It might not even require a specific lattice structure to be identified - there is still quite a bit we don't understand about regular superconductors, a large breakthrough could materialize RTS's within a year. Producing them at scale is a separate issue.
Not my area of expertise but I do work in scientific computing. I know that we can do simulations like this for meta materials, where to some extent you just have a really large search space that is not bridged by logic but by brute force, possibly aided by feedback from physical experiments. Assuming we have similar for semiconductors, it’s not clear to me where AI clears the compute bottleneck.
We already have supercomputers to run simulations like this, and can already spend millions of dollars on compute to answer hard problems. AI may help improve experimental design but this is something humans can already do reasonably well. I’m less optimistic we can derive similar results as we have for math, but I’d love to be wrong.
LLMs are already multimodal, though? Plenty support native image inputs for instance.
But also: human physicists mostly act on symbols, too. They take measurements using sensors, which produce a bunch of numbers, that they then tease theories out of. This seems in LLMs' wheelhouse.
This is what people seem to be missing when they extrapolate all LLM mathematical brute force discoveries to any X or Y scientific problem. Most real-world stuff is not as cut and dry as formal logic and math. Computer science is applied mathematics and almost all logic-based, it follows that LLMs would be good at this. Lot harder to model proteins or materials science this way.
That said the stock price juicing counts on you seeing these discoveries and then attempting to extrapolate bespoke frontier models' success in math (which is dubious in terms of methodology and tailored training anyways) and immediately jumping to the conclusion that AGI and huge scientific advances are right around the corner and can have us all living in fully automated luxury Yudkowsky MIRI effective altruist utopia. It's a farce.
And yes, someone is going to comment on this by saying "moving the goalposts, in 2022 models couldn't even do this." Yes, it's incredible how much throwing money at certain logically achievable outcomes helps achieve those outcomes. It is not clear how it helps achieve random wish casting by odd cult-like CEOs, however.
I’m sure LLMs can and will help with physics research. I’m much less sure you can frame most physics problems in a way that will succumb to 10,000 agents running in parallel.
This optimistic forecast is based on an assumption that we already know everything necessary to find a room temperature superconductor, so an LLM just needs to connect the dots.
We already have a wealth of collected data from superconducting and fusion experiments. If agent swarms can be 10x or 100x better at interpreting this data and using it to build simulations than humans are, maybe we don’t need that many more experiments after all. And the ones we do need can be much more precisely aimed at the remaining gaps than human-designed experiments.
You can get only so far by analyzing existing data. The real power of AI has been always unlocked (chess, programming, maths) by letting AI propose solutions and get feedback. For natural science that means conducting experiments, whether physical or computational.
We're limited on materials and other technology which will need a lot more time for fundamental maths and physics to diffuse to. But yeah, the future is bright.
If you think people are sour on datacenters now, just imagine how pissed they're gonna be about AI training facilities that occasionally explode or leak novel kinds of life into the neighborhood.
It's a problem of scale and engineering. ITER is already way too big and complex, to the point that it may never be completed. They might turn it on, find a leak and be stuck with a huge paperweight. To have viable power generation you have to go even bigger.
Nah. The ITER project was prev-gen by the time they began to build it. The project underwent some upgrades, but it's not at the cutting edge. High-temp superconductors and plasma stabilization algorithms allow to make smaller reactors.
High Tc superconductors can make tokamaks smaller, but they're still much larger than fission reactors of the same thermal power output. While ITER has 400x lower thermal power density than a PWR, ARC is just 40x worse.
Basically there's no hope even high Tc superconducting tokamaks can be competitive sources of energy.
You might say "but make the magnetic field even larger", but most of the mass of ARC is the steel structure to resist JxB forces, so much further increase in B is not practical. There are also minimum sizes for fusion reactors set by the need for sufficiently thick blankets to stop the neutrons; that's independent of magnetic field.
That's apples-to-oranges, since PV omits other components, and because the high redundancy and modularity of PV makes construction much easier. DT fusion vs. fission, the balance of plant are similar to each other.
(If you meant concentrating solar thermal plants, those are also uneconomical.)
Decades of detailed engineering studies of DT fusion reactors (like ARIES) have had them coming in more expensive than fission, as one should expect.
That is the narrative of the fusion start ups pursuising it. But, show me the plasma. Oh right, they don't have any. Might as well just be investment scams for now.
It potentially can, but you still need to build the thing in the physical world and test it under realistic conditions. Maybe I should have focused my post more around "materials engineering".
Same reason AI can't replace the engine in your car yet.
Engineering problems are solved by massive amount of logistics: finding and hiring the right people to be in the right place at the right time with the right tools and raw materials that need to be purchased at a physical location and delivered by cargo ships and trucks driven by truckers across the country. People get injured, sick, go on vacations. Trucks crash sometimes. Cargo ships get delayed. Tools break and need replacement. Materials turn out to be wrong quality or quantity.
This is an entirely different world from software development or math where you can try different solutions and know if they worked or not in a matter of seconds.
I do wonder how well a lot of this translates into physical sciences. It's one thing to be able to RL on lean but physical sciences don't have such an obvious quick feedback loop. A lot of the scientific results have been closer to the theory side which isn't unexpected from the sheer mathematical capability.
if you look at the current "high" temperature superconductors it becomes clear that a peculiar atomic arrangement is the key yet we don't have the ability to individually place atoms (yet) so we must rely on entropy to place them into said arrangement for us which restricts the space we can even test.
so the task for an LLM here is doubly difficult and the feedback loop is of very high latency. I do wonder if someone is trying though. simulate physical process and entropy/annealing -> simulate superconductivity. and very few working examples to work with.
So a couple of purely informational question coming from a layperson:
1. Is there a good theory of high temperature super conducting that we can easily simulate? My furthest understanding of this was that BCS was insufficient.
2. Does such a theory impose limits on the temperature that are achievable?
3. Do we know if such a theory is complete in some sense? (e.g. if such a theory predicted no super conducting material at a temp K and pressure P, we would know that there could be no such material at temp K and pressure P based on reasonable assumptions)
At this point this line of "yeah but what about..." is getting silly. We've had 2 millennium prize problems fall this year, plus a lean proof for FLT (another effort worth 1m$) and people are still "yeah, but what'd be cool is if..."
Dunno, at this point I feel like people are trolling. This is absolutely insane pace, and just last year we were hearing how "there is no more data", "models are stalling" and so on. The negativity and goalpost moving is getting tiring.
> I think it is because we see these headlines, but as a layperson it has no impact on my life. I don't see the value and I don't see the world changed.... yet
You could try to be happy for others who see the value.
I remember reading about the counter example to the Jacobi conjecture in July. That was cool, but many said it was an exception, because AI is only good at finding counter examples. I believed that at the time. But four months later we have lean verified GRH.. I don't feel like it's possible to make any predictions anymore. This might be the peak and in four months AI is more or less the same, only cheaper and with more compute, or we might see more insane unexpected advances. But one thing is for sure, if AI can formalize and prove GRH there are huge swaths of math just waiting to be "cleaned up" and completely formalized. It's just a question of spending the tokens.
Well, you know, I think it would be in fact actually cool if it solved, say fusion or climate change. Not just solve math problems without providing understanding or take all joy out of software jobs.
I'm not saying the technology isn't amazing. It's completely surreal how amazing it is. I just wish it did better things.
Actually agree, it's dancing around the edges, until real world serious problems are being solved, I do think there is some level of theater going on, it's impressive for sure but even as a SWE, I still have my job for now and really if anything AI has just given me more work to do, sure that might just be limited, but I can only go off my real world experience.
Why? Even as a SWE there's plenty of software that I want in my own life that I simply don't have the time/attention to make, or even to try to commission someone else to make (e.g. I've been making little educational games/demos for my kids). Having something that does what I want in a few seconds when I have the attention is fantastic.
"Taking away work from SWEs" means enabling everyone else (or even SWEs) to do things they couldn't do before. This is obviously an improvement.
I have a todo list long enough I don't even have time to prompt for all of it...
If I could justify enough tokens, I'd point an agent at my harddrive and tell it to formalise all of my old notes and goals, and spin up agents to build all of the things I've made notes about that I would like to exist. It'd still likely take years...
I just kicked off Kimi building a game based on a concept I had kicking around. I wouldn't have time to build the game or even write the spec, but I had time to type "write a game based on [paste of the concept]" to see what falls out of it.
I can't imagine ever running out of things I want, because every thing I finish tends to result in my todo list growing with new things.
They are already there. Sure you are not seeing positions disappear but hiring is starting to grind to a halt in a lot of places. Someone who might have needed an admin assistant no longer needs one and middle management in large orgs whose job was basically summarize and pass up will soon start to thin out through attrition.
You wont see whole sale replacement but you will see doing more with less.
I truly don't understand how someone can be interested in technology and visit HN without understanding - even broadly - the history of technological progress and how these things take time.
Yeah, I want the cure for cancer yesterday as well, who doesn't? This is the most significant result in mathematics in over a century.
They do understand. They’re afraid, and that fear leads to hatred of AI and the big labs, which leads to this kind of motivated reasoning that looks bewilderingly foolish from the outside.
Thanks for the free psychoanalysis, but you're completely off.
To be clear, I'm just disappointed and sad that AI mainly takes our jobs (or at least the satisfying part of it) and solves mathematics without increasing our understanding. You know, the part that's the main point.
However, if in exchange it solved, say, climate change or fusion, that I would happily accept. No sign of that so far.
You contradict yourself - you are sad AI will take your job but completely fail to understand why people are scared?
Maybe you are a well off rich kid or just cashed enough from faangs to be set for life, whatever. Is it really that challenging to muster some empathy to those who, upon seeing their profession disappear with 0 replacement in sight, are scared of the future since they don't know how they will even feed their kids, not even going into higher expectations of life?
At the end some math understanding is secondary at best when facing existential threats.
We have a genie that grants unlimited wishes for intellectual output[1]. If you had witnessed someone multiplying bread and fish would your first reaction be to shed a tear for all the bakers and fishermen?
[1] Not to get too religious about it, limitations still exist
Does it? Can I wish for it to solve climate change for us?
Also, your analogue is off. If you had a bread and fish machine that made bakers and fishermen unemployed, while only producing bread and fish for the top 1%, I would not be very happy about it.
Climate change is not a challenging intellectual problem. It is challenging to convince others to do what is necessary to solve it. Could AI produce potential solutions? Sure, from memory. We already have them.
In anticipation of a followup, "can I wish for it to solve climate change advocacy for us?" It can provide research direction on nearly any topic but as soon as we are talking about actions in the real-world it requires more than purely intellectual output.
Anthropic and OpenAI make their models available to anyone for use. I call up cornucopias of bread and fish daily and you have the power to do this too.
Oh but climate change has tons and tons of very much intellectual sub-problems. Better battery technology is a huge one. How to optimally generate or build renewables. Even wild ideas like carbon capture, say through genetically modified trees that grow faster. That's what I'd like to see from AI, something that humans can come up with and solves a real-world problem.
I don't know about room temp superconductors, but fusion has gotten a bunch of investment recently that make it seem like it'll happen before I die.
Specifically, REBCO superconductors are in production by Commonwealth Fusion Systems to make SPARC, targeting net fusion gain for 2027 with a tokamak design.
The NIF has been able to achieve ignition, repeatedly.
Whatever you want to call it, AI/ML/GPUs; point is, supercomputers today are insanely powerful for running plasma control simulations and software.
Commercial contracts for fusion are in place.
More interestingly, Marathon Fusion has figured out a way to generate money instead of energy which needs to get sold off at energy rates. Fusion moves subatomic particles around. Instead of generating power with it, they're going to use the technology to transmute elements. Like alchemists of yore. Just, like, with science and it actually working. They estimate they can make money from converting mercury into gold, but more lucrative is radium into actinium, and zinc into copper. Specific isotopes of those are used in the medical industry to help fight cancer and are very expensive.
That's more of an economic technology than a technical one, but money is always the issue, and if they're able to make money with their scheme, that money can be used to fund further fusion research.
Maybe not next year, but some time before I die, fusion power seems entirely possible!
You really shouldn't have gone out of your way to reply to that snide remark of mine, but thank you. Most of what you said went way over my head, but I'm glad people smarter than me are seeing hope.
I’m a robotics engineer with a mechE degree. Agents are massively speeding up my hardware engineering work, yes the latency is still much higher than software, but each iteration is far more effective when you have agents to help. You can simulate whatever you’re building in far more detail than was possible before.
Yesterday I built a whole new system for agent-human PCB co-design (skidl+kicad). Maybe you don’t feel it yet but massive hardware speedups of the same OOM as swe and math are very clearly on the horizon now. I mean right now on the front page there’s a new agent swarm designed CPU claiming massive performance improvements.
mean right now on the front page there’s a new agent swarm designed CPU claiming massive performance improvements.
Key word is "claiming" and there are plenty of comments on that thread pointing out that the supposed new design is actually just a very old design that was disgarded for practical reasons that are still not addressed and appear to have been hand-waved away.
I work in hardware engineering and just today after setting up the oscilloscope probes for Claude, it was able to trigger my fault scenario and read the captures to debug a flaky comms link, and write a small reproducer, all done autonomously while I worked on something else.
So you might want to check yourself since it's already helping massively in hardware.
Imagine flipping the switch on a fusion reaction that nobody knows how it works, but Clause says it will, and has 1 billion lines of code that supposedly proves it's not an H-bomb.
I was when saw this in the morning. I thought I could no longer be surprised by AI progress. Yet I had not truly believed that Claude could prove a conjecture as challenging as the Riemann Hypothesis.
In fact, just this summer, I tried to write a short science fiction story, The Last Conjecture, about AI working on the Riemann Hypothesis — a bad, unpublished story that I wrote with the help of AI itself. The irony is that even the AI helping me write it could not believe that AI would solve the Riemann Hypothesis in any near future. In its own forecast for the story, the project drags on for nine years, all the way to 2035, and still never cracks it.
Just yesterday, on Friday, I was still telling students: “Maybe there is some physical law that limits how much intelligence one can get from current LLM architectures. Surely nothing in this universe is infinite.” Maybe I am right. It is just that on November 21, 2026, we have not yet reached that limit.
It was 7 a.m. I opened the curtains in my study. The sun was rising. It was a sunny day. Nothing had changed. Everything had changed.
Seeking some solace, I listened to a recent episode of the 3Blue1Brown podcast with host Grant Sanderson and guest Steven Strogatz (a mathematician at Cornell University who has done a lot to popularize mathematics). The two of them did have some hope that mathematics and the profession of mathematician would continue in some changed form, but they could not be sure at all.
Later in the afternoon, I saw that Scott Aaronson, a theoretical computer scientist at the University of Texas at Austin, had published a blog post about the news. It ends with the following story:
It was only a couple days ago that my wife and I had the following conversation:
Dana: I turned off the irrigation, yet somehow water in our house is still leaking. We need to…
Me: Look, a lot of the smartest people I know now think that recursive self-improvement and the end of the entire known world are coming in 12-18 months. This leak…
Dana: We should still fix the leak.
After brooding on this news all day, I finally laughed and told myself: “I should just vacuum my apartment. It is messy and needs cleaning. What else can one do? ”
> After brooding on this news all day, I finally laughed and told myself: “I should just vacuum my apartment. It is messy and needs cleaning. What else can one do? ”
Smarter people than me called this the normalcy overhang and it feels the erosion is becoming deeper by the day.
... How was the proof of the GRH, that mathematicians already assumed to be true anyway, supposed to change anything? Did we learn anything new? Did we unlearn something false?
Even the four-colors-map theorem had more actual impact.
Congratulations to Alpoge (and maybe Massulo), his teachers, his colleagues, his predecessors and everyone who made the tools he used. But nothing changed much.
We learnt that it's true and AIs with best humans could solve a very hard problem no human could solve before. This implication is probably more consequential, do we know think humans and AIs can solve the problem of developing an AI capable of solving riemann hard problems by itself?
> I thought I could no longer be surprised by AI progress. Yet I had not truly believed that Claude could prove a conjecture as challenging as the Riemann Hypothesis.
I don't really get this. I mean I am not surprised by AI progress but this means that I expect AI to do at least anything I imagine a human could do, including proving RH.
I am a layperson nobody and never understood this existential dread about AI until now. I am not even worried about RSI.
Even if Anthropic has already hit the limit on capability I do not think that this unfathomably valuable math machine should be in the hands of a private company. They become the arbiters of all or certain areas of intellectual progress or knowledge or whatever you want to call it?
Until today I just thought these models were weights and compute. But to produce something like this there must be a complex system that is hard and expensive to replicate. I feel crazy for even thinking this but I am now thinking seriously about what nationalization means. Not that I have any control over it, but I think it might be necessary. Me two days ago would punch me.
No. Quite the problem isn’t it? Even if it ever did happen it is not happening that fast. But the analogy kicking around in my head is we already would never accept a private company with nuclear weapons without government involvement right?
The people is a rather abstract term. Nobody really knows what "the people" want. Many things "the people" want alienate about half of the people for example.
"The People" is not some physical coherent entity that can hold something in its collective hand. It's an abstraction. How do give something to an abstraction? It's like saying I want to give an apple to pi.
That's why governments are necessary. And, in this case, I don't see how the Trump administration running the tech would be an improvement.
And the mathematicians in existential crisis plot continues. Deepens.[1]
It could prove all the theorems in the world. At this point it/they are running on electrical appliances that you can thwart by turning off the power. What is actually concerning are the humans behind these enterprises. It’s all just the same sociopathic thirst for absolute domination. Same old “human nature” as the last two hundred thousand years.
But Scott Aaronson’s smartest people he knows is concerned about computers ending the world. Not the humans who are making these applicances doing the same. I guess intelligence is a bit more compartmentalized than the smartest people would like to think.
And thus I paced around my office, thinking about the last five Hacker News submissions on maths, but more organically presented as if I ran into them on the street. Will computahs end dah world? Then the story circles back to whatever mundane thing I was going to do next. Ha ha: There’s nothing that I can do with humans wanting to dominate the world, sure isn’t.
Because these theoretical discoveries are coming thick and fast, it is getting to point where we are dismissing the gravity of some of these milestones. The next big milestone would be bridging this gap between theoretical discoveries and real world impact.
I am hopeful we will see cheap abundant nutritious food and clean water, more efficient solar batteries, economical space travel, cancer treatments, climate change mitigation etc. Most importantly, I hope people can still find purpose in their lifespans. Maybe a few centuries from now it would be thought of as barbaric that people we forced to sit on a desk for 8 hours a day just to afford a livelihood.
>I am hopeful we will see cheap abundant nutritious food and clean water, more efficient solar batteries, economical space travel, cancer treatments, climate change mitigation etc
I sincerely don't understand how on Earth people make the leap from our current trajectory to "paradise". What makes you think any of those things follow naturally from where we are today and where we're going?
Room temperature superconductors could provide a sizable leap for many different fields and applications. It wouldn't leap us to paradise, but isn't inconceivable that RTS discovery combined with cancer/disease/aging/fusion breakthroughs could qualify as that leap.
Well, it's also been our carrot on the stick since the dawn of government, and especially during the Industrial Revolution. The Age of Automation will be no different. Paradise isn't coming. The world is on fire and we are in the middle of a mass extinction event. This year alone there have been a number of horrible events with casualties directly attributed towards human-influenced climate change.
We aren't coming back from this for hundreds to thousands of years if we are lucky. We've used up most of our energy (low-entropy) reserves and future societies will not have fossil fuels to smooth out a grid mostly powered by wind and solar. So we'll invest in lithium instead, until we've mined all of that, too, and future societies are left with relatively little leverage. If we are lucky, by this time, in spite of Elon Musk, we will have the ability to mine asteroids for resources. It will still be expensive and won't solve all of our problems.
But we will likely never again see an age of abundance and affordability as we did in the post-WWII United States. Certainly not in our lifetime, so calibrate. Many factors came together to allow that brief anomaly in human history to exist, and it was still only gained through great tragedy (brain drain to the US from the war and Holocaust, profits from the military industrial complex, letting Europe get ravaged and then exploiting the situation economically for decades, exploiting South America and East Asia, and keeping things going a bit longer by outsourcing manufacturing and craft to criminally underpaid foreign workforces, etc).
We have already seen what these people do when in power, we shouldn't expect it to stop. Peter Thiel literally said being evil is okay if it's industrious. Paradise is currently out of the question, and the age of abundance is coming to an abrupt end. Great resource wars are coming, and now we've been robbed of our intellectual and creative pursuits, so we really have nowhere to retreat and are losing significant economic leverage.
And once our physical labor is no longer needed, we the people will have lost all leverage against our comically psychopathic tech and energy barons who currently see us as unfortunately necessary capital expenditure, and who spend billions to manufacture consent for the privilege to be slaughtered in slow motion.
You're right, and I initially considered removing the line about lithium because you can recycle it at >90% material efficiency. Cobalt recycling has also improved.
I left it in because we definitely will mine as much lithium out as we can until it's not cost effective vs only recycling, and the resource will continue to be extracted with slave labor in different clothes until we have robots doing that job as well.
> I sincerely don't understand how on Earth people make the leap from our current trajectory to "paradise".
I think it's because disruption is an opportunity to rapidly go off the current trajectories. And since the current trajectories are terrible in the capitalistic developed world people naturally assume that going off of them is going to be better or at least it has potential to be better.
We know what would be beneficial for humanity, but we are out of hopes for achieving that within artificial systems that evolved to rule us. In this hopelessness any disruption is something that a fraction of people look at with hopeful gaze. At this point some people can be hopeful even for a proper pandemic, war, anarchy. Compared to known terrible costs of those AI looks significantly better and it's potential downsides seem more far fetched and easier to dismiss as fantasy.
Look at who is holding firmly all the power (bunch of sociopaths at best). Look at how llms are siphoning many human achievements and profit from those are extremely and unilaterally concentrated within hands of very few. This is not even challenged while there is still a chance.
Disruption it is, but such naive positivity has no place in 2026 in debate about actual world. Its not about crapping on rosy dreams of other folks, but about having realistic expectations and then having corresponding actions which don't miss future by a megaparsec.
For every dreamer there are 10 sociopaths already working hard on monetizing this ad nausea, increasing folk's addictions and various mass monitoring.
> Look at how llms are siphoning many human achievements and profit from those are extremely and unilaterally concentrated within hands of very few.
The thing is that's debatable. There's plenty of people arguing that LLMs are losing money and can never be profitable for investors that foot the bill.
I don't say I agree but that's one of the things that makes people hopeful. That they make ASI and it eats them alive.
> Disruption it is, but such naive positivity has no place in 2026 in debate about actual world.
I think people just don't have any other sources of hope. Last decade looks like intense festival of ever increasing stupidity and greed.
Small-scale nuclear fission reactors. Sound smart, sound safe, sound like a good way to make a lot of energy, without wrecking the Earth.
Fusion reactors. If we get good at them sometime in the next fifty years, that'd be great.
Yay, solar, wind, hydro, geo power. And on electric vehicles, I guess...
Fingers crossed on better batteries, to have a compounding effect on solar, wind, hydro, geothermal power's usefulness...
Desalination takes a lot of power. And I worry about the availability of potable water. That's why my hope for lowering energy costs makes me somewhat hopeful that we can survive the +1.5 C horror that we've unleashed on ourselves. There's a ton of problems associated with +1.5 C, but the water scarcity part is the one that scares me the worst.
Maybe robots can help us grow crops with less pesticides? Maybe they can help us avoid the monocultures that make us vulnerable to diseases...? Maybe we can grow a hundred crops all together, in a more sustainable way, with robots tending them?
Yay for fake meat, if it gets people to consume calories from places that are more land-, energy-, water- efficient...? Like, when fake meat tastes about as good as real meat, and is cheaper, and maybe arguably a little better for you, that feels like a huge, huge victory?
Yay for throwing AI at solving diseases?
Yay for technology improving access to DIY, home 3d printing, carving, CNC...?
Yay for AI making it more possible to hack your own hardware...?
I mean, I'm not a huge fan of China, but they lifted close to 800 million people out of extreme poverty, over 4 decades. Let's not discount that...
Almost every prosperous country has universal health care...
It's already a joke. You hear all these tech psychopath CEOs lie about AI 'reducing work' and other great things, but it turns out that's not what they are using it for. They are using it to replace humans, mass surveillance, etc...
They are "reducing work" in the sense that humans will have no work to do for which they can make a living and feel a sense of purpose. Welcome to the future.
Almost everyone I know hates mondays. They fantasize about early retirement and not having to work. Work, by itself, is not what people desire. It's the livelihood that it provides. If machines do everything, cost of things approach zero. And I know we fear unfair distribution of that benefit, but we know that didn't happen with food. We went from almost all of humanity producing food to almost nobody doing it, and we have new problems with people (even poor ones) eating too much food.
Re sense of purpose, I'd rather be mountain biking with my friends everyday than sit on a desk for 8 hours pushing pixels around.
I don't actually think anybody in the Bay Area understands how much the average American would like to take a bat to their head. This industry is so high off its own farts and so divorced from the priorities of ordinary people that the pseudo-philosophy and fake economic theories they propose end up being a religious belief.
We already have cheap abundant nutritious food and clean water. An 8 hr/day desk job is really not bad. If people are barely scraping by on a job, that's mostly a political problem, not a technology problem.
> Maybe a few centuries from now it would be thought of as barbaric that people we forced to sit on a desk for 8 hours a day just to afford a livelihood.
I'm fully convinced of this. It will probably be viewed in retrospect like we view the exploitative work during the Industrial Revolution, even if it seems like an exaggeration right now.
I think there's a category error in a lot of these techno-optimist takes.
Abundance of productive capacity does not automatically equal abundant access to its products.
If all of this happens in a way where a few people control these new, automated means of production, all of these wonderful things do not automatically follow.
Why would they share? And worse yet, if they have this automated production capacity without reliance on labor, why would they interact positively with society at all? What's their incentive?
You're basically reduced to hoping for charity from a group of people that have repeatedly proven their lack thereof.
Right there with you. What's the motivation of anybody on the high end of the wealth inequality scale to share with the other end? They have nothing to gain and everything to lose.
It would be nice to hear the rationale behind those who are downvoting you.
There are people who share a significant proportion of their wealth with the significantly poorer, like Warren Buffett, Bill Gates, Dustin Moskovitz, and so on.[0] Some people choose to do good despite nothing tangible gained, whether they are poor and just putting their single dollar in a wicker basket, or literally billionaires and saving millions of lives with their money.
Whether the owners of AI will choose to share, time will tell.
I'm specifically worried about the people who use their wealth to influence political discourse to perpetuate wealth inequality. It makes sense to me that they would use technology, which fundamentally obviates their reliance on the labor of the less wealthy, to perpetuate inequality even more that they can thru political means.
I don't feel like the generous wealthy can offset the influence of the wealthy who have the opposite goal.
> We could never win an election on getting certain things because we were in such a small minority, but maybe you could unilaterally change the world without having to constantly convince people and beg people and plead with people who are never going to agree with you through technological means, and this is where I think technology is this incredible alternative to politics.
> There are people who share a significant proportion of their wealth with the significantly poorer, like Warren Buffett, Bill Gates, Dustin Moskovitz, and so on.
Yes, that is the charity one can hope for. But even in that lucky case, the issue I'm describing is structural. It's not like any of them share to the point of giving up their generationally privileged position. That poses the problem: Even if the people immediately in control of AI are pure selfless angels, over time and generations the likelihood of one not being so asymptotically approaches 100%. Sooner or later there will be a Caligula.
> Maybe a few centuries from now it would be thought of as barbaric that people we forced to sit on a desk for 8 hours a day just to afford a livelihood.
??? That was a few decades ago
If we keep the same political and moral stances around work you'll still be working every day even if we get agi dont worry
I'm sure some or all of those advancements will happen but, based on the way the last ten years have unfolded, I'm not at all confident everyone will see the benefits. For a concrete example, look no further than the sociopathic destruction of USAID.
This comment is incredibly tone-deaf. It is difficult to see how the benefits will ever trickle down to ordinary human beings from the class that holds all the money, controls all the computing power, and possesses the authority to execute critical tasks.
More food, water and medical care would only enable people to multiply and overpopulate this world even further. At the cost of other animals in this world all of whom will be supplanted or become extinct.
We need to solve the overpopulation crisis before committing to improving the lives of the impoverished.
Would it be possible to reconcile your worry for the ecosystem with basic empathy for people you clearly don't see as equal to you in worth and individual agency?
I see overpopulation as the most important problem facing mankind. More important than climate change, which IMHO is merely an outcrop of overpopulation.
Like I've said before: our society is completely obsessed with sex and it's therefore extremely difficult to curb population growth. In Africa there are frequent riots when rumors are spread that the "white man" wants to neuter Africans through vaccination programs.
Sex and reproduction have been separated, which is good because nothing curbs a mammal's sex drive other than existential hopelessness. Which we also have created.
What I found surprising. How many absolutely groundbreaking breakthroughs with AI are in Mathematics. And at the same time I can't remember anything similar in other fields. For example there's nothing groundbreaking in Software. Yes, Claude can write some PoC code faster than me. Great. What about creating some software which wasn't possible before? What about creating operating system without a single bug? What about inventing some video compression which is miles ahead of competitors? What about creating Rust compiler which runs 10x faster than rustc?
I think the closest thing is formal verification of Signal protocol, I guess. Which is insanely cool, but still not as cool as millenium problem.
> What about inventing some video compression which is miles ahead of competitors?
in a specific case I was interested in (point cloud compression for geometry streaming on the network), claude helped me make something that is 200-300x faster than the existing standard (MPEG G-PCC) at 10-20% worse compression ratio, which opens completely different use cases. e.g. the MPEG one compresses a 3.2 M points point cloud to 8.94 megabytes in ~5.7 seconds and decodes it in 3.7 seconds ; https://github.com/sat-mtl/pccl encodes to 9.97 megabytes in 32 milliseconds and decodes in 12ms.
The real innovation here, IMO, is theorem provers. Without theorem provers, the LLM brute-force-by-bouncing-off-the-guardrails approach just doesn't work to begin with and the whole thing spins off into hallucination and dodgy reasoning, just as we see in non-trivial AI software. Once you've got some magic oracle that gives you a binary yes/no answer to every attempt at a proof, your search strategy can be enormously simplified.
I also thought that, but I learned that most of OpenAI's results in their proof dump did not use theorem provers in the loop. They translated the papers to lean afterward.
Right, that's how all these proofs (including a few of mine) are done. Given the large-scale AI efforts already underway to formalize proofs in Lean, there could be some very interesting gains once proofs are Lean-first. For now, though, formalization happens afterward.
Exactly. Another example of how AI depends on software which is carefully, deterministically hand-coded. It's disingenuous for people to say "look, we don't need to code anymore" when all these solutions depend on things that are definately well-thought out and hand coded.
I work in this field, and the issue is that the LLMs start reward hacking these things. They either overoptimize to the benchmark and hallucinate, insert reward hacks like adjusting the benchmark or cache behaviors, or they make implementations that only focus on verifiable things, like utilizing the hardware better.
The fix to this are better benchmarking environments without reward hacks, as well as formal verification of code on various levels. There is crazy much work on this at the moment.
The recipe seems to be, pick a problem whose solution is highly verifiable, then throw the equivalent of ten billion hours of expert-level thought at it. If you can narrow the search space a lot, all the better. P != NP is sure to follow soon enough.
Not an expert, but from comments I've heard by experts P != NP is much further out of reach than Riemann. Eg, Scott Aaronson:
> First, almost no one in theoretical computer science—not counting the cranks, whose missives fill my inbox every week!—spends their time directly trying to prove P≠NP. Why not? For the same reason why you wouldn’t embark on a manned mission to another galaxy, if you hadn’t even set foot on Mars yet.
Right. Just saying the fall of these "step function" type problems aren't a forgone conclusion yet. It could easily be the case that some problems (who knows which?) are in an entirely different dimension of complexity, and would be out of reach of even much, much more powerful AI.
these tasks you mention are very long-horizon in context... i think if latent long-term memory or continuous learning approaches are getting better, we also see this
Its reverse engineering prowess is very impressive. Building a completely open source version of Microsoft Office or Adobe PS was definitely not possible before.
Microsoft Office and Adobe Photoshop were literally built before. Recreating their copies is not groundbreaking. That's what I'm talking about. AI in software right now feels like mostly about recreating things we already can do, but cheaper, with questionable quality and unclear maintainability.
If you call any computer program that's advanced "AI" and stick the "software" label only on dumb programs, well of course software seems dumb. You defined it as such.
"The AI companies have clearly... I'm sure they have secretly devoted much more time to the Riemann hypothesis than what they've disclosed. And they couldn't touch it. I mean something beyond current AI is needed. So maybe some combination of humans and AIs in 2050 or something will make further progress. Who knows?"
Every time someone says 'current AI is not enough, we need new architectures', it turns out that current AI was enough and all we really needed was more scaling.
Not in this case, what we needed was a genius mathematician to explore with Claude. Which is exactly what Terry said: "So maybe some combination of humans and AIs in 2050 or something will make further progress."
I remember when, for a brief period of time, the best chess players were human + computer. There were a lot of Ted talks and think pieces reflecting on this reality. This was called centaur chess.
This lasted for 2005–2014. By 2015, the advantage humans added was negative.
I imagine this will be the case here as well, but I highly doubt it'll take that long.
yes, for every specific task. but life is not like chess. life is like playing a million different chess games simulaniously, all of the time. Being a gentle hand that knows when AI is onto something and when it's torturing itself in a loop. He calls it out in the pdf, the aha moments he experiences, that lets him further guide the progress.
It's not 2050 yet, though. I feel like the future date he threw out was a big part of his conjecture, that even with human+AI today, it wouldn't be possible, and getting it done that way would take years, possibly decades, of more work on the AI.
How about we hold off until it's verified? Now, Terence Tao is getting visibly thin these days, I hope he's not torturing himself in an existential crisis. I know I would be. But I would also, in the end, be satisfied as a Lee Sedol figure, the last of the human geniuses, as it were.
> I’m sure for every incredible insight which Levant credits to Claude, there’s 100x+ more bad insights that Levant had to filter out.
This is certainly my experience of using Claude every day, in a much less cutting edge domain.
And Claude itself cannot tell the difference between the two. The best it can do is use some sort of external constraint (e.g. verification) to filter them.
Honestly, ever since Opus 5.5 dropped about two months ago, Claude's taste has been great, better than mine, even. Compared to Opus 5, I've barely had to filter anything out by hand.
It is fascinating. Because sometimes my experience is like yours, but other time, like today, I am like “I can’t believe how stupid Opus 5.5 or even Fable 5.2 is”. They will make horrible design decisions, overcomplicating. I will correct and they are like “Of course, your solution is much simpler and better than mine”.
And then they are simply boldly wrong at times. About mathematical facts. That’s why we need formalisation when working with them. They are still stupid and can’t be trusted. But theyvare also smart and can come with amazing insight. It’s a tool, but definitely one which cuts both ways.
I think it's funny that people gauge their personal experiences with AI against something like this.
Do we even know what Alpöge had at his disposal? I think most people underestimate what working with an agent swarm even with the current public technology would be like. We can't assume he didn't have something more advanced either.
I don't feel confident my usage of AI on a personal level reflects any of what has been done here.
Your argument is unfalsifiable speculation. I don't at all that it's self-evident that a strong mathematician was required. Your everyday Joe isn't exactly interested in the Reimann hypothesis so statistically a mathematician would be the one who'd use AI to come up with the solution.
How would a non-mathematician even be able to comprehend the output? Half the words in the pdf are gibberish to me, I could not tell if whatever Claude gave me for a "solve the Reimann hypothesis, make no mistakes" prompt would even make sense, much less be correct.
A non math person may not be able to comprehend the output, but even professional mathematicians have trouble reading through the huge “alien” AI generated LEAN formalised proofs.
The real genius is that we have found a generic way of solving problems through brute force by having developed an effective way of aiming an unbelievable amount of computing power at them.
If the genius here is Claude and Claude did 99% of the work then isn't it a staggering coincidence that this was achieved by one of the most qualified and likely humans to have solved this on their own merit and not some random person sat in a bedroom somewhere?
Doesn't matter; Claude ends up taking cues from whoever it's working with, such that two people doing the same thing get very different outcomes. Claude will always be able to write better software when prompted by a senior engineer; it will always produce better math when prompted by a talented mathemetician.
On that topic, I'm surprised the usual anti-ai trolls aren't posting low-effort drive-by "ai slop", "stochastic parrot", "autocomplete on steroids" comments to this very discussion.
Edit: I stand corrected: RunSet, ToucanLoucan, luka598, and others have checked in.
Last I checked it's 2026. Not 20250. And for the record, there is no evidence that suggests that a layman couldn't have made progress with just using AI.
We know the architectures of the open weight models as they include architecture diagrams, support in projects like llama.cpp, etc.
Companies like Anthropic and OpenAI don't release information about their models (aside from general things like context window size, whether they support reasoning, etc.). A lot of what we know is speculation. From press releases Anthropic have stated that they are getting Claude to assist with designing new iterations of the models, but no specific details.
The open weight models and other arXiv papers/research are iterating on and experimenting with the base/core transformer architecture:
- RoPE/NoPE/etc. for context window scaling;
- dense vs mixture of experts models;
- Mamba-based State Space Models (falcon models, etc.);
- reasoning, combining transformers with a residual stream, Attention Residuals (Kimi), etc.
I think a lot of model architecture changes don't move the needle ~that much and are basically done to improve runtime performance and scaling.
- There's tons of cases of ml researchers, tech companies, and labs that propose new and improved models that turn out to not really be significant quality improvements (if at all) once the community starts experimenting with them.
it does seem like (IMO) the biggest leaps have basically been perf & scaling + RL on high quality data + harnesses.
Yes, but "improving runtime performance and scaling": does directly translate into huge quality improvement too, because then you can you a much larger or deeper model, train it for much longer, and have a larger context.
It's usually people who are not suffiently familiar with the topic, in this case, technology. He has been saying boring things like that about AI in the recent years, and seems to be continuously adapting his positions. I have received negative votes over my criticism of his remarks.
This was literally Dario's contribution to the AI race.
He just did the math and realized you just needed more and more compute to get more intelligence. People keep doubting it but it's showing no sign of being incorrect.
We have scaled on both axes and many others, I'd say.
Inference can be increased vertically (a single model thinking for longer), or horizontally (NS was solved by a 10,000-agent swarm).
And pretraining has two inputs that can be scaled independently - training data and model size. (Things like the Chinchilla 20:1 law are meant to establish the optimal way to train with a limited budget - you could just keep making either term bigger and bigger on its own if you had infinite money.)
I think creating and curating high-quality datasets is also part of scaling. Datasets like FineWeb are quite a bit better than the barely-filtered CommonCrawl dumps GPT-3 trained on, but someone had to invest in making them. (To say nothing of whatever exotic bespoke robotics/physics/etc datasets OpenAI and Anthropic are investing in.)
ChatGPT was only the turning point for public awareness (and the massive influx of institutional investment that followed), but researchers like Ilya, Alec Radford, and Dario were already convinced years before ChatGPT that scaling compute was the way forward. ChatGPT just helped unlock the resources (by proving the commercial potential to investors and the subsequent influx of capital) for these guys to pursue that vision at a much larger scale.
I mean, why wouldn't that be the default assumption, if you know anything about computing history? ChatGPT's (v1) contribution wasn't proving commercial potential but demonstrating efficient scaling, due to a confluence of technical developments. If you're talking about v3 then sure.
In the end, the network's size is limited mainly by the amount of memory available on current GPUs and by the amount of training time that we are willing to tolerate ... All of our experiments suggest that our results can be improved simply by waiting for faster GPUs and bigger datasets ...
On Anthropics side it's literally just levent. He isn't even doing this on an official capacity. OAI have teams of Mathematicians. This is like david vs goliath.
Is it? Isn’t it just frontier model va frontier model. I am not a maths specialist, but aren’t the models doing 99.999999% of the he at lifting? Of the answer is “no”, then what explains all of these unsolvable problems being solved in rapid succession?
Depends what you mean by “lifting.” If it’s anything like building software, the agents do all the grunt work (exploring 1000’s of possible approaches in parallel, perhaps) while the domain expert (human) makes the critical strategic decisions about which pieces need to be solved and how to fit them together.
Even the best agents still can’t strategize their way out of a paper bag, in my experience.
The human remains the most critical differentiator for non-trivial work. (Different story for trivial projects that agents can solve independently. Here the human becomes a fungible operator.)
I think a lot of his AI skepticism is just a self-fulfilling prophecy, and a misunderstanding that AI at the cutting-edge is a tool in the hands of the user. In other words, sure current AI is not going to solve the Riemann hypothesis until 2050, but what do you think the current state of AI would be if Terrance Tao was the one driving it?
Somewhat related: I dont think AI is coming for everyones jobs, and I will never understand the argument that it is, because I always have the ability to use the AI and add my "human ingenuity" to it. So why _wouldnt_ that be a better option than just raw AI?
> because I always have the ability to use the AI and add my "human ingenuity" to it. So why _wouldnt_ that be a better option than just raw AI?
Well, that hasn't held true for chess, so I'm not sure how universal the domain is where human ingenuity will always augment AI. Also, there is a cost floor to human ingenuity, in that it requires the time of one living human, and their metabolic sustenance over that period. So the added benefit of that human has to be substantially more than epsilon.
> I always have the ability to use the AI and add my "human ingenuity" to it. So why _wouldnt_ that be a better option than just raw AI?
It is not clear it is always a better option. It may be for you, but not for someone who doesn't know what they are doing and drives the AI down the wrong path. For example, I don't think adding my "human ingenuity" to Claude would help solve the Riemann Hypothesis, and would likely be counterproductive, because I have a 3rd-level understanding of math.
>I'm sure they have secretly devoted much more time to the Riemann hypothesis than what they've disclosed.
Curiously, the timeline for this proof only starts roughly around the same time that Tao said this. IIRC, work on the disproof for the Navier-Stokes smoothness theorem only started a week before it was published. Maybe I'm wrong, but it seems like AI companies have only recently started to put real effort into LLM-assisted proofs.
It seems bold for Tao to state that they couldn't touch it; Considering that after Anthropic announced Navier-Stokes and got all the negative press, largely driven by Terrence Tao, it doesn't surprise me that they didn't disclose anything afterwards.
I think Terrence Tao was trying to say that had the LLM swarms been able to prove this it would have been what they lead with. This seems to lend to the claim that Levent has done an amazing lift in coaxing the solution out of an LLM.
This is the guy I see mentioned in every other HN post? Yeah, seems everyone who should know better still hasn’t grappled with the gravity of what’s happening
So even one of the smartest mathematicians alive struggles to comprehend or extrapolate the implications of an exponential progress function. To be so far off seems ridiculous, but it’s exactly how you’d predict the error to be. We are truly cooked. End of 2027 is probably the end of knowledge work on a global scale.
"I'm sure they have secretly devoted much more time to the {{matter}} than what they've disclosed. And they couldn't touch it. I mean something beyond current AI is needed. So maybe some combination of humans and AIs in 2050 or something will make further progress. Who knows?"
- All of humanity, unaware their pants have fallen around their ankles
> I'm sure they have secretly devoted much more time to the PERPETUAL MOTION MACHINE than what they've disclosed. And they couldn't touch it. I mean something beyond current AI is needed. So maybe some combination of humans and AIs in 2050 or something will make further progress. Who knows?
Just to point out how ridiculous it sounds to claim that because X (Riemann) is possible that anything is possible.
Well, we got an award at a Caltech "mathathon" for some pedestrian improvement on some error bounds on some niche problem. Now these guys come out with a proof of one of the most notorious open problems in mathematics. Just to one-up us I'm sure...
> In October 2024 I attended a workshop at Harvard University ... “An AI that can prove the Riemann hypothesis is not one I’d want to meet,” said Andrew Sutherland, a number theorist at the Massachusetts Institute of Technology. “If that happens, mathematicians having jobs will be the least of our problems.”
I really hope Levent next tries to prove the irrationality of the Euler-Mascheroni constant ɣ , or the gamma function minimum α, Γ(α) , as they have so many forms yet we know so little about them. It's slightly related to GRH, since the Zeta function occurs in a some of these forms.
Given how important these constants are, you'd think we'd at least have proved by now that ɣ was irrational, yet shockingly it has resisted all attempts. In my opinion, it's almost scandalous that it wasn't one of the Millenium prizes.
Yes. Mr. Hilbert, good to talk to you again. RH has indeed been proven, but by superhuman technology making long alien constructions on automorphic forms.
We dont really understand the proof in its entirety, it was verified by the whole formalization movement you started.
And to have a semblance of understanding we keep going back to the SH machine to dumb it down and explain to us different parts of it.
As a non-Mathematician it's inconceivable to me that proofs can be so difficult to understand that no one will. I thought the process of getting to the proof is the complex part, once proven, ex post facto should be more simple. But seems like it isn't.
It might not be that difficult, just long and diverse enough that you might just run out of lifespan.
As humans we are used to reality giving us stuff to understand that are just handful of things. Three generations of leptons, four types of orbitals, 4 dimensions, low natural number coefficients.
Deep enough in the math there might be constructs that have hundreds, thousands or millions importantly distinct and non-groupable concepts you'd need to understand to understand the proof.
It’s an enormous mental slog to validate AI math slop, demoralizing, and also feels like you’re just fighting for scraps. not saying thats all the reason, but definitely contributes
Given the multiple ways this proofs has been formally verified^[1], I think at the moment there is no any doubt that it is correct.
> Andrew Sutherland (MIT), checks run with Claude Code. Page last updated 2026-11-20 22:05 UTC. Results below are as of that time and this page will be updated as the remaining checks finish.
Checking with an AI offers very little assurance - if there's a mistake in an AI proof (and I'm not saying there is), it wouldn't be surprising for the checking AI to make the same mistake.
Wouldn't it be fun if they were both correct, but with wildly different structures? We could maybe use the proof difference to characterize the prover difference (AI-assisted vs not).
Structurally, like, treating the proof like a graph or something along those lines? Or socially, like this part of the proof resembles proofs given by this group of mathematicians, so we can say it has a certain heritage? Or some other way?
It's easy for me to wish for such a thing, but I have no idea what one would latch onto when granting such a wish.
>His Master's Voice (original Polish title: Głos Pana) is a 1960s science fiction novel written by Polish writer Stanisław Lem. It was first published in 1968 and translated into English by Michael Kandel in 1983. It is a philosophical first contact story about an effort by scientists to decode, translate, and understand an extraterrestrial transmission. It is considered to be one of the three best-known books by Lem, the other two being Solaris and The Cyberiad.
>Lem's Opinion:
Human beings should accept some humility. Sometimes we face phenomena the essence of which we are unable to understand. Even if we are equipped with the most modern scientific apparatus and knowledge we are incapable of resolving whether they are accidental or intentional. Newton once said that with respect to nature we are just children playing with shells on the seashore.
>HIS MASTER'S VOICE.
Stanislaw Lem.
Harcourt, Brace, Jovanovich, c1968,
English translation by Michael Kandel.
c1983 (hardcover -- I understand the paperback is out now).
>Faren Miller in LOCUS called this book 'more a readable meditation than a novel' and after reading the book I would have to add my qualified agreement. The book is a philosophical retrospection by one of the key members of the 'His Master's Voice' project, the mathematician Peter Hogarth. The HMV project was secretly founded by the US government to analyze and (if possible) translate a signal discovered in modulated neutrino emissions from deep space. The military hoped to extract technology which could be used for new weapons against the Russians. Hogarth and the other scientists held the military in contempt, and sought instead to interpret the message and find a way to understand the Senders. Several strange results came from the study of the signal. One result was a peculiar protoplasm-like material created by irradiating a chemical soup with the signal. This goo had the peculiar characteristic of sustaining itself not through normal biological energy cycles, but through atomic reactions. Hogarth and friends then discovered that the material had the capacity for causing remote nuclear reactions: an explosion near it released its energy at a distance. The possibility arose that detonating a hydrogen bomb in a roomful of the protoplasm in Nevada might cause Moscow to be destroyed. Hogarth and his colleagues had to decide: what should they tell the military? Besides this ethical problem, Hogarth also considers the question of whether there really were any Senders... This novel is very difficult but I found it to be rewarding.
>Out of the infinite number of possible mathematics we have chosen one; our history decided this for us with its various unique and irreversible vicissitudes.
The mathematics we practice may reflect historical selection among alternatives. This also connects to MatrixMan’s speculation about different foundations.
Preface, p. 16:
>Here I will say only that the world injected its patterns into human language at the very inception of that language; mathematics sleeps in every utterance, and can only be discovered, never invented. What constitutes its crown may not be cut free from its roots, because it arose not in the course of the three hundred or eight hundred years of civilized history, but through the millennia of linguistic evolution: at the loci of man’s encounter with his environment, from the time of tribes and rivers.
Our mathematical habits have a cultural and linguistic history. An AI trained on human material complicates any clean opposition between "human mathematics" and an alien alternative.
Chapter 9, p. 143:
>It is true that we differed in our style of thinking; I was fascinated by ergodic processes, a field that Dill made light of.
Chapter 9, p. 144:
>I do not think I ever finished any larger paper in all my younger work without imagining Dill’s eyes on the manuscript. What effort it cost me to prove that the Dill variable combinatorics was only a rough approximation of an ergodic theorem! Not before or since, I daresay, did I polish a thing so carefully; and it is even possible that the whole concept of groups later called Hogarth groups came out of that quiet, constant passion with which I plowed Dill’s axioms under.
Personal preferences, rivalry, and the desire for recognition can influence mathematical discovery.
Chapter 10, p. 153:
>This, I thought, illustrated a certain small but characteristic feature of the history of science, because neither the fortuitous bends in the road of research nor the accidental circumstances assisting at the birth of a discovery ever completely detach themselves from its final form. Indeed, it is not easy to recognize these relics, for the reason that, fossilized, they become embedded in the heart of all later theories and formulations, like a print of a coincidence which turns to stone, to an iron rule of thought.
The finished argument may preserve traces of its discovery, although recognizing those traces is difficult.
Chapter 2, p. 42:
>With an exaggeration that is necessary for the clarification of my meaning, I will say that in the course of my work (it is difficult to say exactly when this occurred) I began to suspect that the “letter from the stars” was, for us who attempted to decipher it, a kind of psychological association test, a particularly complex Rorschach test. For as a subject, believing he sees in the colored blotches angels or birds of ill omen, in reality fills in the vagueness of the thing shown with what is “on his mind,” so did we attempt, behind the veil of incomprehensible signs, to discern the presence of what lay, first and foremost, within ourselves.
Chapter 2, pp. 42–43:
>None of us knows, therefore, to what extent we were the instruments of an objective analysis, to what extent the delegates of humanity (in that we have been shaped by and are typical of our society), and to what extent, finally, each of us represented only himself, with the inspiration for his hypotheses about the contents of the “letter” being supplied by his own—possibly raving, possibly wounded—psyche in its uncontrolled regions.
Interpreters bring their own expectations, institutional assumptions, and psychological investments to unfamiliar intellectual products.
It was rushed out before being fully checked, not a huge surprise. They even say in their blog post that there are likely gaps but they hope they are fixable.
It’s a 1,000 pages dense document. I’m not even sure the PDF conversions are accurate enough to feed it ti an LLM.
Also - if that’s the future of math (larger and larger proofs and algorithms, to the point they are far beyond a human’s capability to comprehend, then I wonder how mathematicians work would look like.
I recall being told that there are certain axioms, the axiom of choice for instance, which we take most of the time not because its the best tool for the job in any intrinsic sense. Its the best tool for the job in a historical sense. We use it because our parents did and it's convenient to stand on the shoulders of giants.
But if proofs are cheap, you don't have to stand on the shoulders of giants. You can instead venture where they did not.
I'm told that the axiom of determinacy--used as an alternative to the axiom of choice--represents such an underexplored alternate dimension. Presumably there are many of these: inconsistent with the math we know but internally consistent and potentially superior to our use cases than what we now use... Our current habits being more driven by historical accident than careful decision.
So a future mathematician, rather than being a Platonist, might be a pragmatist and a curator. Given the many possible formulations for mathematics, which is most fruitfully paired with economics? Which goes best with chemistry? Which parts appeal to the human sense of beauty and how can those be introduced to specialists in disparate fields such that those specialists can now, through a bit of shared structure in the formalism that underpins their work, better understand each other's domains and better collaborate about them.
I think if we tailor a mathematical theory to a job, rather than assuming that the math is primordial and immutable, that the humans doing that job will benefit. It seems like it could be pretty rewarding work to me. Maybe AI will make it practical.
You can cut off the appendix so it's roughly 500 pages long and Gemini 4 can digest it in one go (context window is around a million, this takes up around 300K tokens so it's inside the effective window).
Is there someone with sufficient expertise to compare, at a very general level, whether Levent Alpöge and Alessandro Masullo use similar approaches to solving the GRH?
I doubt these questions are gonna be answered anytime soon. AIUI even Navier-Stokes hasn't been independently verified yet. If the abc-conjecture affair is anything to go by, these things often take a while to digest and only so many people have the appropriate expertise to fully evaluate them.
So how do we know if it does what it does? Anyone that is using frontier models for a few years, will tell you that they try to reward hack whenever they can.
That’s the purpose of Lean. It’s kind of like asking: how do you know your CPython is running all the correct binary?
Well, you kind of don’t, as it’s beyond human comprehension (in theory you could trace everything down to a certain level) but there are layers and layers of build abstraction that you assume are correct, compilers, runtime environments, and mature libraries that are checked and validated. Not a perfect analogy but that’s kind of the gist of it.
OK, but if I write code, I can verify it in a few simple steps. What about 250k pages of proof? How can I verify that page 133,432 has the correct assumptions and that it isn't reward hacked just to make lean code pass? How it's done?
You don't need to! That is the beauty of Lean. You only need to check the problem definition (which is no where near 250k pages) and then Lean checks the proofs it self by checking whether each next line can be derived from previous lines. And for this problem, there are many validated problem definitions in Lean. So it is probably true.
> and then Lean checks the proofs it self by checking whether each next line can be derived from previous lines
Ah, so it's similar to something like Merkle trees, where each leaf must be cryptographically valid up to the root otherwise the whole branch is invalid?
So you start with a simple definition that then evolves?
The core idea is that you can map the types and terms of a type theory to the propositions and proofs of a formal logic
Different type theories correspond to different systems of formal logic with different levels of power and expression.
Simply Typed Lambda Calculus corresponds to intuitionistic propositional logic. If you lay out the typing judgements of STLC in natural deduction style and then physically erase the terms (leaving the types) then you literally have the rules of intuitionistic propositional logic.
So if you can write a term that typechecks at a type, that term is a
proof of the corresponding proposition.
The easiest example is modus ponens, which is exactly the typing rule
for function application:
G |- M : A -> B G |- N : A
-------------------------------- (App)
G |- M N : B
"Given a function from A to B and a value of A, you get a value of B."
So then in a proof checker you prove the proposition P => (P => Q) => Q by writing a
program of that type:
apply :: a -> (a -> b) -> b
apply x f = f x
`apply` is demonstrated in haskell which does not have a rich enough type theory to give you more then an intuition (no pun intended) for how this all works. Things like `undefined` and general recursion let you "prove" any type, so it isn't a consistent logic. Languages like Lean and Agda are built on richer dependent type theories that give you far more expressive power for writing proofs (quantifiers, induction, etc.), and they won't let you cheat the type theory.
So to answer your question, no not a merkle tree but a tree none-the-less. You end up with a literal computer program which is a tree of expressions all of which must type check. If they type check then program itself is a proof for whatever proposition is encoded in the type of the program.
> I cannot imagine that a fundamental proof like this will require hundred of thousand of logic steps.
It's pretty easy to imagine that, given a starting language, there are facts that require an arbitrary minimum number of words to prove. The only reason the proofs most people learn are short is because:
- There isn't time in a math course to teach arbitrarily long proofs
- Humans don't live long enough to construct arbitrarily long proofs.
- Human minds are turning out to be pretty feeble and they probably couldn't reliably construct and verify anything more than trivially short proofs in the era before computer assistance.
In theory, if you formulated your proposition correctly, and Lean says it is proved, then the proposition is true. This works because of the Curry-Howard correspondence -- proofs can be represented as computation.
OK, so if I understand correctly, I formulate a proposition in natural language, then an LLM creates 250k pages of proof based on it, then it takes those 250k pages/lines of proof and creates 15 million lines of code to prove those 250k pages. Is that correct? If so, how can I be sure that my proposition is correctly translated into those 250,000 pages? Isn't it possible that the LLM can reward hack here in the same way I've seen it do in some cases where it struggles to pass the test and get its reward? For example, it tries to change assumptions without telling me because something doesn't work or it removes something or changes the test etc.
What assumptions could it change? The model should not be able to alter the definition of the RH, and if you want to prove to yourself that that definition is accurate, you just have to read through it and build a mental model of the Lean code: https://github.com/bhgomes/lean-riemann-hypothesis
Then if it constructs a valid proof in Lean, how could it be reward hacking? I suppose it could have found a Lean compiler bug, or a mistake in the formalization of the hypothesis. I would expect that those problems would be found and fixed reasonably quickly, and we would need to see if the proof could be fixed. Other than that, I suppose it could find a contradiction in ZFC, but that would be an even more incredible result than proving RH.
I don't really see the possibility for reward hacking with the stipulation that the proof is ultimately verified by compiling the Lean program.
> For example, it tries to change assumptions without telling me because something doesn't work or it removes something or changes the test etc.
You have in Lean a headline statement of "given X, then Y is true". Where X and Y and precise statements in code. So the mathematical assumptions are upfront.
If you trust the lean compiler and are very careful to vet that the headline statement is what you expect, then you know the theorem is true.
How well the human language mechanics of the proof corresponds to what the actual lean mechanics of the proof is very much up on the air.
I see, so there is one step less, then how I can know if that 250k pages describe my proposition correctly and not something else? Lean code was written by LLM, correct?
Assuming Lean kernel is correct (and there are multiple independent implementations of it, some proved in other ways themselves), the only part up for debate is whether the Lean-translated theorem (roughly half a page of code, maybe a bit more if you fully expand all the libraries it uses) accurately models the original GRH
In formal proof verification the only thing you need to verify is the problem statement expressed in lean, you don't need to check the lean code that is the proof. If the lean code that is the proof compiles, its a verification of the proof.
In theorem proving environments like Lean, you are equipped with a standard set of axioms and statements (eg set theory). Most people, as far as I am aware, agree with Set Theory. From these basic fundamentals, you can build up all of mathematics, just like a normal mathematician can. The trick is that in systems like Lean, proofs are properly type checked, like a programming language; and just like a programming language, those types help stop bugs in the proofs, but the types here are far richer, able to encode all of mathematics, and far more strict. To prove a mathematical statement, you must write a proof that is accepted by the type checker. Beyond that, the core "language" itself is generally specified to be "sound" -- that is, there is a very small subset of Lean, a tiny "kernel language", that is shown to be incapable of proving False, and thus any proof which is well typed in said language (ie it typechecks) is also sound.
In other words the system is designed, by construction, to only rely on a very small "trusted" set of axioms, code, and mathematical ideas. This is a very engineering-oriented design principle! Actually, a pretty significant amount of the theoretical research in theorem proving over the past 2 decades has been in "automating proofs", ie tools and methods of making those proofs as easy as possible to write. Entire book chapters have been written about how to replace a 20 line inductive proof in Rocq with a single line "function call" that can automatically discharge the proof for you, because the idea is that this "trusted kernel", the tiny and sound mathematical language, will reject invalid proofs by design. So "who cares" what the proof is, in a sense? In many ways simply writing the theorems and whatnot down requires you to understand the problem well enough already. So in a way LLMs are just taking this idea to the absolute limit.
So all this means is that the difficulty really is in specifying the theorems. Which was kind of the whole problem all along! Except now you have a computer that can check your actual proof. You can write a theorem in Lean that isn't the theorem you want, and then write a type checked proof of it, but you still need the theorem to be accurate.
Unless Claude is either 1) finding bugs in the Lean kernel to exploit them, 2) using 'sorry' (a very convenient tool in Lean used as a placeholder for proofs, but it can literally allow you to prove False), then it mostly comes down to: are all the theorems stated correctly? This is trickier than it sounds, of course, so some skepticism is certainly warranted in my opinion!
There are also other various engineering principles applied to help try to reduce bugs and ensure correctness like any other engineering project, and there are various other "things" regarding dependently typed programming language design and the impact of those design decisions on proofs (and why they are notoriously hard to read and write), but they aren't germane to my point here.
>> If so, how can I be sure that my proposition is correctly translated into those 250,000 pages?
It's not the proposition (i.e. the theorem) you need to worry about but the proof. I'm sorry to say but the comments on here (and everywhere on the internets) that assure each other that "if it's veified by Lean, then it's true" misunderstand what they're even trying to claim.
Here's a good article on the subject that was recently posted on HN:
>> This post is an attempt to explain in simple terms how to interpret claims of the form “this theorem is correct because it was checked in Lean”. The summary is: there is an important signal that can be extracted from Lean code, but if this code is AI-generated then you need to do some tests first. We explain the tests, and note that in many cases (such as the OpenAI theorem dump) these tests have not yet been systematically done.
If his proof is 250,000 pages, then I am extremely dubious. Just today, I was working with about 1000 lines of code and asked opus 5.4 to make a change. It introduced an obvious bug, that actually ended up doing the exact opposite of what I asked to do( print out less lines of logging, instead of more).. even if he has access to some kind of super LLM, I am extremely skeptical that it can output a 250,000 page document that is coherent and bug free, no matter what this guy says.
Its a formally verified proof. Its the same thing as saying your code compiled successfully (for a compiled language). If opus made a change and your code still compiled, you would certainly still say, my code compiled. The statement being asserted in lean is analogous to my code compiled.
This article is saying the same thing I said with more words. To someone who does not understand what formal verification is, this detail is unnecessary.
Lean certificates of this magnitude come with comparator. They are checked independently on different kernels to avoid kernel bugs.
This is a false equivalence, you seem to be missing the analogy.
There is no specification in your prolog example, in lean we explicitly check that the expression in lean is GRH (and this is about half a page of lean and has been checked by many many mathematicians), and then can arrive at a formal proof for that specification.
You are confusing the issue for people who don't understand what a formal proof means with this non-sequitur.
I wish, but the proof is verified by another kernel ("con-leche") that is itself formally verified, Lean's own kernel, and two other kernels ("nanoda", "pestka") so far.
Joachim Breitner who's a Lean developer: "The high level summary is that this proof is now very very thoroughly checked."
I'll try to make it simpler, in formal verification (at the most base level): code compiles = the proof of the statement you have posed is correct, by definition.
The statement you are trying to prove is: This is compilable code.
Since your code compiled, you know your code is compilable code.
Therefore your initial statement is correct.
That's at the very basic level what a formal proof does. The expression you want to prove (in this case GRH) is stated in such a way that if your code compiles, you have proved the statement.
Genuine question: if I make 250k pages of lean that compiles, how do we know what it proved?
If it was as trivial as stating the expression and compiling then there would be easy verification and consensus as soon as it compiles on multiple kernels?
The "there is a consensus" standard of proof isn't strong when it comes to maths. People are generally looking for a rigorous argument why something must be true.
> The "there is a consensus" standard of proof isn't strong when it comes to maths.
Says who? I don't mean that in a rude way, it's a genuine question. Is that the opinion of the community of mathematicians who are capable of understanding these type of proofs, or software engineers who believe nothing can be true unless you can read the code yourself and understand it?
I'll try to make it simpler: code compiles = the proof of the encoded statement is correct, by definition. Whether the encoded statement and the one you have posed are the same -- at 250k pages that remains to be seen.
No. The lean statement itself is maybe half a page (I don’t know the exact amount but in this order). Checking that the lean statement is the GRH is elementary. The 250K pages are a ‘path’ from the statement (GRH) to the axioms encoded in Lean. The fact that is ‘compiles’ means the path is correct.
Yep, fair point. Someone else has posted a link to the GRH statement. It's complete gibberish to me but for someone qualified checking half a page of code should be straightforward.
I would assume they do not produce the proof at once, and Lean is different from programming language (as far as I understand) in the sense that you can check if a program is correct or not within the Lean model, so it seems ideally suited for AI as they can experiment and extend the proof without ever producing a bug. I assume the real difficulty is checking that what you have proven is actually what you wanted to express but I would assume this is easier.
Makes me wonder if general programming languages are too expressive for AI making it too easy to express conflicting or incorrect ideas, we might need something like Lean but for general purpose programming. I'm aware there are some formal systems for that but I think more along the lines of general purpose programming.
I‘m doing a PhD in software verification right now and I agree. I’m not sure whether it’s about the power of the programming languages but more about their interfaces.
Proof checkers like Lean are already what you describe, as Lean is also a programming language, but yes, there will be more general purpose tools for writing code that is correct by construction.
The remaining hurdle is that you cannot actually verify correctness of many programs as you don’t know their full spec. But that is a very different problem to what we had before.
You don't have access to the secret model the labs keep for themselves. Neither can you spin up thousands of subagents and burn billions of tokens a day.
Proof can be total bunk but underestimating the labs is silly.
I’m not sure I could drive Lewis Hamilton F1 car round a track as quick as he could either. Clearly people are better at using these tools than others, hopefully we can decipher these proofs and actually benefit from it.
Did you deliberately pick Opus 5.4 over Opus 5.5? That was the biggest thing that stood out to me in your comment. Opus 5.5 has been an all time banger for me.
"Rooting for this, but can’t help but feel it’s just false proof #14,828, but the more promising one that came out recently, and that we’re clinging to it as the last hope before it turns out Anthropic got it."
"Who is Alessandro Masullo? Only name that's popped up in a technical field is a data scientist who does not seem to have the background for a non-AI proof of GRH."
"And we are supposed to believe that RH, a longstanding problem at the height of mathematics for a century, has just been solved by a 'mathematician' with no publication history or academic presence, without the help of new technologies, at the same time as a new technology credibly becomes capable of doing"
"So this is from a guy who has never published anything in a journal before and left academia 8 years ago?"
The thing is, Vakil says it was vetted by Brian Conrad, who is basically a human Lean engine.
Masullo isn't a data scientist. He got a math PhD at Stanford and has been working there since, supported by soft money. So not a professor, but someone the grant holders thought was worth paying to do research.
Whenever I, as someone who stopped at Calc 3 in college, read modern mathematics papers I realize how differently mathematicians brains must function than mine. This is not a value statement and nor am I demeaning either side. It's just... wow.
(setting aside any dispute as to the veracity of this particular paper, the point still stands more broadly with tons of human-generated output)
Any technical subject beyond a certain horizon of familiarity is pretty much impossible to understand. You're simply being overwhelmed with definitions.
eh, i see it when it comes to programming as well. the mathematical types write this crazy functional code that is just impossible for my brain to get. there are also a lot of unintuitive (to me) concepts.
not a value judgement, i just find procedural code to be infinitely more understandable
It's perfectly possible to write 940 pages manually over 12 years of work. Almost certainly many more due to drafts and failed attempts. Also, the Alpoge proof is GBs of Lean code which translate to 250k pages in the "human" version.
Please stop copy-pasting uninformed reddit comments in every thread. They didn't even find the right guy, Alessandro Maria Masullo, who has a PhD in math at Stanford and also received a presidential honor from the president of the Italian republic for being a top student.
It's 940 pages long and presumably the process of producing it created many witnesses willing to vouch for its provenance as well as records of progress. Setting your math career on fire for a lie that obvious would be crazy.
> The hope was to bring Alessandro to Stanford for a year to polish the writing and then announce it after running it by enough people. Now it seems best to put out the current draft.
We are not sure when this proof was drafted. It might have a few months or even more and it was rushed due to Alpöge announcement.
They have had another person reviewing it for months and the author has been incorporating his feedback, according to TFL. If anything I believe the AMS would give Masullo priority if both pan out.
It is said that others have been reviewing that human-written proof for many months.
Obviously, it was published immediately, despite the reviews not being finalized, when it became known that an AI-assisted proof was announced (but I understand that the complete AI proof has not been published yet).
Wiles' proof of Fermat's Last Theorem was 129 pages and lean proof was 13 million lines
Interesting in this context because Alpöge has the lean proof but not the human readable variant while Alessandro Masullo proof has the human readable proof but not lean proof. This might support Alessandro's claim that it is not an AI proof, and it will be particularly exciting if they both solved the Reimann Hypothesis using different routes.
I did, and it claims to have found errors, but those errors may be possible to repair. These kinds of small issues are usually found and fixed when papers are reviewed formally prior to publishing.
Given the era we live in now, if this version works too, expect a Lean formalization of it within days.
I haven't seen a single piece of code which doesn't contain errors according to Claude (I'm quite sure Hello World level codes are exceptions, obviously). Most of them are false positives, many times all of them are that. Also, asking Claude to review the same code multiple times results in different found problems, with different priorities.
There are probably no "complete human, "human only" proofs of any hard problems anymore. More or less it's AI (at least assisted) even if unacknowledged.
Really interesting to note from his timeline, he marks the starting point as the Stepanov determinant, which he earlier mentioned relates to OAI's "quiet inclusion of the Siegel argument ... I immediately recognized it as Stepanov without multiplicity". I believe this was one of results in their 700 solution dump.
I think this gives a partial counter to the arguments that this AI-driven math over obscure problems is a wasteful benchmark and doesn't help anyone. Turns out these are in fact important building blocks, and OAIs published results helped someone else do some great new math. I think that's nice.
I'm checking the timelines of the progress, initial progress is between Oct 6-20, then there is nothing for 3 weeks. Finally, everything is wrapped up in 2 days between 11/15 and 11/16? If Levent got access to Model 4 in early October or late September, then he could have made the initial progress with Model 4. The crazy jump in the last two days could have been with Model 5. If that's actually the case, the next few weeks are going to be crazy.
This should be the wake-up call to pause development of models any bigger, enforced globally just like nuclear proliferation.
Even without considering rogue AIs, our society relies on individuals having limited destructive power. A dictator can ask the armed forces to bomb an entire country, or a lone terrorist can bomb a small number of targets, but no lone terrorist can bomb an entire country.
Here current AI has enabled one person to solve what is arguably the most famous problem in mathematics. What happens if a jackass uses the next AI to design a supervirus, or break AES, or develop an unpatchable hardware attack (a la rowhammer) that bricks computers remotely? The last month of non-stop vulnerabilities is just the tip of the iceberg as models improve and people have more time to work with them.
Speaking to "break AES" - this is the wrong way to look at the problem. If AI finds AES is broken, then it was always broken, and we did not understand yet how. This is why you can "discover" zerodays - they were there all along. I prefer to know about the risks and design around them, than avoid the problem entirely (leaving them open to exploitation by bad actors).
I myself work in the cryptography department of a large company, so I'm not arguing the semantics of cryptanalysis, or crypto agility, or patching. I simply meant "imagine if suddenly anyone could decrypt AES ciphertexts", and that society is not ready for such sudden disruption.
It's like everyone suddenly developing telepathy. It's not dangerous per se, and we could adapt if done slowly enough, but all at once? Catastrophic.
The only reason nuclear nonproliferation works is because the inputs are relatively rare and traceable. (Yes U238 is abundant but you can’t mine enough of it or buy the centrifuges to turn it into U235 without getting noticed.)
I have all the tools to develop my own ML training pipeline on my machine today. In order to choke off the inputs to “LLM proliferation,” we’d need to confiscate every computer, destroy all copies of the Internet, and throw out all books and other archived media. It’s not gonna happen.
RAND & others have published work about this, and they believe frontier model training would be much easier to track than nuclear weapons were
You need a datacenter larger than a football field that draws the power of a small city to train a frontier model, and need chips that are produced by the most complicated supply chain in history (owned almost entirely by US allies like Taiwan & Europe). And that's before you consider on-chip mechanisms that seem entirely feasible. This is much easier to trace & much harder to reproduce than precision spinning containers.
Though I agree seems much harder to detect people running models that were already trained.
> The only reason nuclear nonproliferation works is because the inputs are relatively rare and traceable.
There are only three companies in the world capable of mass-producing the chips used in SOTA AI training, they all use machines from a single company (ASML), and it also requires enormous power-hungry datacenters. Meanwhile, uranium can be dug up from the ground. You say that buying centrifuges will get you noticed, but why can't we do the same for SOTA ML hardware? It honestly sounds easier to control. And differently from nuclear weapons, both sides are afraid of rogue AIs.
This is a danger to our civilization, I'm sure we can figure something out.
> I have all the tools to develop my own ML training pipeline on my machine today.
Are you saying you can train the successor for Claude Opus 5.5 on your machine?
Already today you can get on a list by buying too many smoke alarms due to the americium content, or drug precursors, or a thousand harmless but suspicious products. I would be ok with a world where you likewise need a justification to buy thousands of GPUs at once.
Maybe we just have very different imaginative worlds, but after seeing the kind of crazy shit that people will believe thanks to online radicalization, I'm a lot less sanguine about the possibility that people will take a shot at your list of horribles. There are a lot of people out there who seem to genuinely believe that what's at stake is the very survival of humanity and the continued existence of truth, art, beauty and all the other good things. After enough exposure to that kind of pressure-cooker radicalization, something like a repressive world government that uses violence to ensure that only the "right" people have access to dangerous things like literacy and computation could seem like a small price to pay for saving us all. Sure, it'd suck in a lot of ways, but Altman and Amodei will have left the brave revolutionaries no choice.
I think it's very unlikely something like that would succeed, but I'm increasingly sure we're going to see serious attempts to do all of the things on your list, and they could do a lot of damage even in failure.
My post already gave three examples. If you have a problem with those and can't think of any more, I'll be happy to provide another dozen, it's not a hard exercise. Also, "unlimited" is your word, not mine.
Possession as in ownership, nine. Possession as in have in their territory, fifteen.
"Successfully kept possession (limited to)" .. seems a little over confident- some of these nations acquisitions and expansions occurred despite others not wanting that to occur.
An additional twenty-eight countries have "virtual" nuclear weapons, having endorsed others to use actual nuclear weapons on their behalf.
"We all saw the headline that they were able to overcome the math problem for the that researchers had worked on for years and we have some math and physics challenges as we employ this next sensor capability on the F-35 and OpenAI is working next to us to uncover some of those math and physical challenges."
Yeah, that's how my fears start. I don't want to find out what happens to the global stage if one country develops autonomous killer drone swarms or something.
The militaries might be salivating at the opportunity, but I hope that the risk of rogue AIs brings us all to the same side on pausing development of bigger AI models.
We should try to stop it as hard as we can, but in full honestly, I don't think we ever could have. I think any intelligent species that doesn't go extinct is fated to reach this point from the moment it awakens...
A genie is out of the bottle, but it's weak enough that the consequences so far have been limited to job loss and medium-sized IT problems. The creation of better models is not inevitable, it's done by human engineers using immense hardware, and therefore can be stopped before things get worse.
Also, to preempt another comment: yes, I know that he's a great mathematician. But we've had great mathematicians before, and even working together for 150 years they were not able to prove Riemann. And as models get better, the pedigree required for groundbreaking results (including dangerous ones) will naturally lower.
Once AIs solve all (or many of) the hard outstanding questions, it would be nice to see some new questions that can serve the same guiding purpose that e.g. the RH or Navier Stokes had. Maybe this is a new domain for exploring with AI: not what it can prove, but what it can conjecture.
> The argument is incredibly long at the moment, and while its formalization
is finished and it is routine to produce a huge, unreadable document containing
sufficient detail, in my view of my path I have just taken a step. I intend to
do my best to understand and impart understanding of this argument and its
certain simplifications, to speak about it, to write it properly.
Okay, I am anxious whether this holds true or not, but he claims that the formalization is finished.
Why not just release the formalization in the meantime?
This is probably the biggest day in math in recent memory.
Edit: nevermind, the Lean formalization is linked.
I'm old enough to remember when Andrew Wiles announced a proof of Fermat's Last Theorem way back in 1993 [1]. It was long and complex and took at least 2 years for the maths community to accept it.
I wouldn't be surprised if this isn't similar (or worse). The days of short, succinct, elegant, intuitive proofs is probably over.
For what it's worth, the lean proof is a similar length (15M lines vs 12M lines) to Anthropic's first formalisation of FLT so at least it probably isn't much more complicated.
God knows what will happen when we have billion line lean proofs that aren't easily shortened. They might just be forever outside human understanding.
Unfortunately there's reason to believe that science as a whole has just become too hard in recent decades. Not as much low hanging fruit across all fields.
The way this is getting announced via a human mathematician humbly presenting a result is great, as opposed to the marketing influenced OpenAI post.
I am still a bit sad though. I had hoped the Rieman hypothesis would fall mostly to a human mathematician rather than 60/40 AI/mathematician that my layman's read of that note interprets.
I don’t get this feeling. Landing on the moon isn’t less of an accomplishment because we built the Saturn V instead of walking. If anything, can’t we appreciate how we took some sand and made it smart enough to help us solve problems?
It would be sad if AI designed and built the Saturn V and we had it fly us to the moon without understanding the details of of how it is built and how it works. There is furthermore value for humans in going through the discovery process themselves, not just value in getting to the moon. So I understand how one can be sad if that won't happen anymore now, and instead humans are relegated to analyzing what the AI designed and built.
I would dispute that we completely understand how LLMs do what they do. An analogy would be that we meet an alien race that gives us their proof of the Riemann hypothesis translated to Lean. Plus of all other provable mathematical conjectures we come up with.
And we might kind-of understand how those aliens do math. It would still be a bit sad that we'd be relegated to consuming what the aliens give us, and never be able to ascend to the level of the alien mathematicians
That doesn’t make much of a difference. It’s alien, and for the most part we didn’t design it to do the things it turned out to be able to do by happenstance. We might as well have grown the aliens in a test tube by some unexpected gene manipulation.
An analogy here is evolution. We understand what it does on a basic level. There is natural selection and species that are well adapted to their environment thrive.
This absolutely wouldn't allow us to predict what species will emerge and that something like humans would be the dominant species.
Similarly, we understand the mechanics of how LLMs work, but that does not make it predictible in any way.
I remember when about 15 years ago, I wanted to make my neural network home work based on proven math. I had to train a model to learn basic number recognition. The math said 100000 neuron is what's needed, so I typed in that, and my laptop immediately froze completely, and died a little later. I learned on that day that we know empirically that a few 10s are enough.
We have the same thing with LLMs. We don't know how they work. We don't understand them. We have only empirical, statistical indications, nothing is proven. We know how the small parts behave, we really understand those, and also the different algorithms. But they whys? No. We're getting better, but still far away.
And presumably mathematicians understand the output, the proof, whatever it is that came out of the process. I don't feel like humans have to be the one doing all the manual math and figuring out all combinations.
We already had AlphaFold, which "solved" the protein folding problem, but scientists readily admit that they know next to nothing new about the field, despite it being accomplished.
I mean computers made Apollo work. They couldn't get it done with human paper calculation. The lander itself had a processor in it. So its kinda the same thing. "Humans should have done this by hand," is just an odd thing to want and ties back to some really backwards thinking. These are our tools. No one is expecting us to make homes without power tools or walk for miles instead of taking a train. At a certain point AI will be normalized and we'll have actual critical takes on AI that make sense other than "Well, I just don't like it for vague human achievement reasons." The same way people complaining about 1980s cars getting "computer chips" and "no longer being able to hand tune my carbonator" have long accepted more complex and automated and computerized cars.
As far as not knowing how it works, that becomes a bit of a philosophical problem. We know how LLMs work and their algorithms. I think we can always play up "we don't know" is disingenuous ways. How many people who made Apollo calculations could also write long proofs and theorems of the fundamental maths they were using from first principles? Not a lot. We're always standing on someone else's shoulders. We'll never truly know everything or how exactly our tools work on an individual level. Maybe AI changes things because an innovation may not have a direct line to an Isaac Newton at its base, but I think that's a bit of a vague and poetic difference, and not a real one. Its still human achievement but via machinery.
I dunno, its a bit scary, but I do feel like its just continuing the same line of human ingenuity. Not knowing exactly how something came about and non-determinism scares some people, and perhaps rightly so, but pretty much all major technological advances are scary on some level because they so strongly challenge the status quo and our ways of thinking and or ideas of what is good and right and proper. At a certain point, its hard not to see the more dramatic AI outrage as just hysteria and moral panic. I find it interesting how easily we can spot moral panics in the past in history, but when its our turn, it can be very hard to see when we're in a moral panic. To many living during those times, the moral panic sounds entirely reasonable, but its still a moral panic.
And if the rocket failed at any point nobody would know how to fix it. Only the AI guessing and saying "you're right to point that out" when the rocket explodes and everyone inside perishes.
I also build things that would've taken me months in days now so I do get that feeling of amazement at what is possible now.
But at the same time one of my favorite memories from childhood as an introvert nerd math / computer guy is locking myself in a room for an evening and relying on my brain to solve and truly understand a problem. I'm nostalgic for that. A solution that requires 250k pages to describe takes something away from that.
Humans are complicated and contradictory, which is great!
The entirety of human history is about building tools to perform tasks our bodies aren’t capable of on their own. Someone else in this thread said something about how cool it is that humans turned sand into a machine that could accomplish this. Our gift as humans is that we can devise ways to overcome the limits of our brains.
We are all sitting on a thousand years of media telling us that the human mind is exceptional and transcendent.
It's naturally going to be upsetting to everyone if reality unfolds to show that human intelligence is nothing special, and actually not even that good.
For the first time in humans 200,000+ years we have an "independent" or "different" intelligence that can teach us things... and WE built it. We can learn from it and improve after its teachings. This new era is going to be amazing.
Some people love learning for its own sake. That doesn't go away even if it's no longer valued by the market economy. You can call it a hobby or you can call it a deep-seated volition. It's kind of like asking "what's the point of eating a good meal?"; some people see eating as utilitarian, and others see it as one of life's leading pleasures.
Some things getting solved opens a new whole dimension of new possibilities, so assumebly the negativity steems towards the fact, if this did just take us from A to B and there were no new undiscovered paths encountered during the journey to be explored..
And parts of what we thought of as intellectual work turn out to be mechanistic. That was the hope for math before Gödel, and seeing it become true is an example of getting what you idealized and then being disappointed, it seems; or holding onto the numinous parts and persisting, as Professor Tao is urging us towards.
It seems to me, having B.S. and M.S. in mathematics, that the intrigue of human intellect continues to dim. I would wager that a large cohort of mathematicians, even ones remote to the discipline now such as myself, maintain the joy of cogitation.
One felt a camaraderie with all mathematicians in history because mathematics was, wonderfully, a club with no bouncer, the only cover charge being a bit of concentration of attention. Not that one could join the elite ranks of Gauss, Euler, et al. by association, but the bonhomie was real for a lot of us, realizing that all of the legendary mathematics was dreamt up in the same way that we were engaging in it, too.
That horse has left the barn, decidedly, and there is grief.
The effect of LLMs on math is pretty much the same as for SWE: great for existing experts like this guy, bad for juniors since the traditional criteria for evaluating merit no longer apply, and completely demoralizing for prospectives.
This exactly describes my feeling, a feeling that I wasn't aware I had until the last few months, that feeling of continuity through millenia. And you're right. It was a club anyone could join.
Ironically the foundations project leading up to Gödel’s great incompleteness theorem was a project to discover a mechanism that could just spit out true theorems. We have been asking for this since symbolic logic was nailed down and yet now we find the mechanistic nature of our cognition to be a disillusionment.
But beating an Olypmic athlete by the use of medicinal or mechanical enhancements is very much less of an accomplishment, because the goal was never doing the thing faster / more accurately. The goal was being a natural human who can do the thing faster / more accurately. Maybe it's impressive in some other way, but . As is winning a chess tournament by having a computer feed you the moves.
The point of math is not bare knowledge. It is a skilled profession.
A moon landing is something that pretty obviously can't be accomplished without advanced technology. This is something that was at least believed to be achievable without the equivalent of a gigantic rocket. Imagine if the conquest of Everest had taken longer, and the first person to reach the peak got there by helicopter. We could appreciate the technological prowess represented by that landing, but a mountaineering enthusiast might still be reasonably disappointed that a climber didn't get there first.
Yes. We're tool using species. We evolved from African savannah to this AI moment by using tools.
I don't understand this skepticism towards tool use. It's like saying, "I'm sad that Reimann Hypothesis solution was published as a PDF and not in a parchment or made as a verbal announcement."
Anyway, great accomplishment. What an incredible thing to wake up to.
Now that you suggest it, a video of the person declaiming the announcement while holding a parchment, would definitely would be nice to have. Hopefully someone will generate that using some genAI video tool.
> Landing on the moon isn’t less of an accomplishment because we built the Saturn V instead of walking.
I think it's quite telling that by "landing on the Moon" you meant the humans landing on the moon and not the first actual landing on the Moon by human-made machines, which was in 1966, 3 years before Apollo 11.
So it looks like you do think that it's less of an accomplishment after all.
It’s a bit more like we created sentient animals in a lab, they ran off to another planet, then sent one of their spacecraft back to shuttle us to Ganymede. We have no idea how the spaceship works, and if it never arrived, getting to Ganymede would still be science fiction. You can’t authentically call that an accomplishment… just a result really. We get a participation trophy at best.
I think a more apt analogy which gets at what they probably felt was immagine if humanity accomplished the moon-landing via a supposed Walmart Inc. Space Division, the announcement of celebration of this event felt like a giant Walmart promo, and the only reason they wanted to go to the moon was to put ads on it.
I can not fathom this viewpoint, it is as alien to me as being a flat earther. What frame of mind is required to consider an analogy of "we built saturn V instead of walking" to be even remotely ontologically in the same stratosphere as what is happening today?
I can't even begin to make sense of that lack of basic understanding of humanity.
It turns out that the cognitive processes of advanced maths are subject to mechanization, just like carrying a load from point A to point B. It is a blow to our self notion that work of the mind is somehow fundamentally distinct from work of the rest of the body.
Also, people are nervous that the mechanization of new human capabilities will give even more power to money and less to labor, and we will live in some distopian libertarian society. I think that is a separate and realistic fear from the grief of discovering our thoughts are not as special as we told ourselves.
For me, if I am confident that the social structure will enable fulfilling lives for people, my kids and their peers, then I am ok with exploring and understanding the possibilities of mechanized cognition.
And to those objecting to the phrase mechanized cognition, I use Claude a lot and the thing I find missing is that independence of perspective that all the people I encounter have. We wake up with a perspective and goals and preferences, a why to answer what built into the protoplasm. Some technology will rebuild that aspect of our experience in a different substrate, but it doesn’t feel like a system that can’t count but can prove Riemann is quite wired correctly for that fleshy self motivated loop.
I'll do my best to explain what is happening here, as I see it.
There's a reason Paul Graham wrote Hackers and Painters and not Hackers and Assembly Line Workers.
Many people long to be craftsmen and to have an understanding of their place in the world and within the social hierarchy.
These folks want to perfect some art or craft, to be the best, or to be skillful, to "create" and find joy in creation, and they pursue a level of purity in the art. Totally human, completely normal, and part of what has led to our success and advancement as a species. Think about your open-source maintainer, or the gal who has Ruby on Rails stickers plastered on their laptop, or maybe you yourself as you try to get excited about the super boring details about the latest release of some open-source package you want to be the expert in (why?). They want to be "hackers" and "painters"!
The problem is that the pursuit of this particular craft (software development) falls into the realm of Science and Industry. Engineering, code, electronics, machines, experiments - all of these things and more are part of those two fields and those two fields by definition require and improve upon existing technology or understanding and as such they are fields which change often, cause some skills to become obsolete without remorse, and they do so at a fast pace.
The software engineer who believes in their craft and their expertise in JavaScript is finding that Science and Industry no longer value that craft and expertise - so that person's place in the world is directly under threat, including their income in many cases. It causes a range of emotional reactions.
You can see how this manifests with many software engineers going on to do silly things like move to Japan and live on a mountain top to write code in peace, or start a Substack newsletter that nobody ever reads, or build expensive tiny sheds ("houses") so they can go in to their back yard and "code in piece" - i.e. pursue their craft. Another common theme is minimalism - flat surfaces, clean lines, things like that. The person seeks more and more control. If only they could just remove this inefficient piece of furniture at last they can finally code and do my painting in peace with my perfect coffee and get in to my zen flow. It's a whole psychology, and it's under grave threat from AI.
With the threat of AI, of course the natural human instinct for many of these very intelligent software engineers (most are about average intelligence however) is to perhaps physically and culturally retreat and go "farm" (interesting how they try to go pursue other crafts isn't it?) or emotionally retreat "AI isn't a big deal, I'm important, this technology won't beat out my craft" and these very natural human reactions you are seeing are due to psychological defenses built up by average people who believe their livelihoods are under threat.
So the comments will of course be around things like denial, or impurity (it's not pure enough that the Saturn V was built - that's cheating, we should have invented wings and flew there ourselves), or it'll manifest into a mechanism in which many are actively cheering on a foreign adversary to "beat" their home country because it'll validate them in some weird way. See? Look what you did. You created AI and now we're behind! Told ya! There's no changing their mind because their job depends on their mind not being changed.
It's difficult for someone like yourself to combat these feelings that you don't understand without first understanding them, and to understand them is to brush away some of the nonsense that people get up to and get back to basics which is that many people feel under threat from advancements in technology, many people want but lack a sense of place in social hierarchy and purpose in life ,and for those that do have such things these advancements threaten to disrupt their current lives and purpose. Of course they're going to tell you we should walk to the moon instead of building the Saturn V (AI).
Maybe it is beyond one single human's ability to solve in one lifetime. It has been around 150 years why would we assume a human CAN solve it. Maybe the only way through these hard problems is to use AI. There could be some evolutionary hard limits on how much a human can fit into working memory to solve these hard problems. And we still get to read the results and try to understand them. Btw we built the tools and the tools use human developed patterns of reasoning to solve the problem so it is still very human in a sense, not alien.
I think Alpöge specifically was criticized in the past for being somewhat disrespectful of the math community, since he just dumped AI slop onto Twitter in meme form. The OpenAI post was actually more respectful compared to that. So this type of announcement is more of a return to somewhat more deferential communication.
It's worth remembering that this was done by someone who (as far as I can tell) is a great mathematician and not by some "prompt engineer" with an interest in math and large budget.
> but this time Claude far outpaced me near the finish line
> I intend to do my best to understand and impart understanding of this argument and its certain simplifications, to speak about it, to write it properly
> we were going to win but I was also, so to speak, out of my element, in that I personally have little mastery of automorphic forms on higher rank groups
> I’m just too inexpert in the relevant stuff at the moment, I hope the reader sees the breadth of this argument!
Also the exposition at the bottom of the paper is straight up just claude.
> So here is a rough timeline Claude provided
Also its like literally gigabytes worth of lean... it goes without saying that that is going to be effectively 100% claude.
I'm sure Levent put in some good work, but also this is clearly in large part the work of claude.
Sure, and it depends on what we mean by the ratio. But, seeing that the proof comprises millions of lines of lean code (which have not been read or understood by a human), or 250k pages of prose (which have not been read or understood by a human), it makes me think the machine largely "did" the work of this proof. How far do you think he would have gotten without the machine?
Actually, thinking about it more, the ratio might be closer to 99/1.
AI is just a tool, at the end of the day. Regardless whether the proof is accurate, this is showing strong capability of both the mathematician as well as the tool.
I have a substantial math background and certainly can't yet prompt LLMs to solve extraordinarily hard proofs. This is a real achievement. That the logic wasn't 100% worked out by a human is less exciting than what this enables.
The process of solving hard problems like this often involves inventing / discovering new branches of mathematics in order to solve the problem. By solving the problem with AI, we haven't potentially snuffed out the potential for discovering these new branches of mathematics.
In this case it feels humans alone were given enough time.
Also, I know its beyond what a single person can affect but lots of the problems that the community faces now seem to be in essence people and incentive problems. Existence of this proof in principle doesn't stop anyone from pursuing the problem further and discovering these new branches of mathematics too.
What if it was a human proof and in its process, no new mathematics were discovered? Would we also frown that?
Here's a stupid video game example: If you fast travel everywhere you accomplish your current quest more quickly but you never find any of the side quests that only trigger when you stumble across them in the wild.
Yes, proving GRH is not as important as how it was proven. For example, take a look at how Fermat's Last Theorem was proven by Andrew Wiles [1], and what's the impact of his efforts had toward algebraic number theory. The fact that he proved FLT to be true has little consequence compared to the method he invented to get there. Before it was proven, most mathematicians have assumed FLT to be true anyway.
> I guided us to the right situation
> I had us on the line of combining two quite
different ideas present
> The parts in which my vague intuition worked out due to Claude’s technical mastery
I am not saying he could have done this without Claude. But there's a reason he did it and not anyone else. This wasn't a case of giving Claude the prompt "solve riemann hypothesis make no mistakes"
What is the reason? I'm sure he is a talented mathematician, and that is undoubtedly necessary here, but how many other mathematicians have unlimited access to internal frontier models? Clearly, the playing field is not even.
This kind of assistance is new, but I’d say nobody has ever proven anything truly unassisted, and people have been trying to point out for a long time that crediting a single person for discoveries and claiming they were “unassisted” conceals the true story. Education, background knowledge, conversations, correspondence, and context have always been paramount. This is one reason Math has often hit points where multiple people make the same discovery at nearly the same time “independently”. It might be a good thing that we can now dispense with the myth of anything being “unassisted”.
For people unfamiliar with the Riemann Hypothesis, I like the following example. Look at these sequences:
111111
.1.1.1.1.1.
..1..1..1..1..1..1..
...1...1...1...1...1...
....1....1....1....1....1....
Pretty simple right? Just increase the period by one each row and add an extra period at the start.
Now, look straight down a column, say the 5th column. How many 1's do you see? There are two 1's in any prime number column.
Now ask the question: If you were to draw this grid with N number of rows, how many columns would have exactly two 1's?
The Riemann Hypothesis is the ridiculously complex hypothesis that relates an incredibly simple pattern (increasing periods by one) to the seemingly random frequency of prime numbers.
You can ask Claude to start with those sequences and derive the Riemann Zeta function. It's only about 3 paragraphs of logic, but incorporates advanced ideas like wave-like periodicity and infinite sums.
Secondly, I want to quote one article, how easy it is for humans to misunderstand difficult things AI does, as something supposedly easy:
> They mostly came away from Claude Code with the sense that this is a genuinely useful tool that they didn’t previously know, but the gut-level reaction here is: “I never knew that building software was so easy” rather than “It’s very impressive AI can do this.” [1]
It should become very clear for everyone how to adjust their views. However I doubt many will, even after this, they need something physical, like humanoid robot to be impressed.
At least going by what the author wrote, it seems like he was using Claude as a peer - a pair mathematician if you will. It doesn't sound like a one-shot solution. So without the expertise of the mathematician, AI wouldn't have been able to solve it on its own.
What about also acknowledging the enormous number of works of mathematics the model was trained on? You know, standing on the shoulders of giants and all that.
Can anyone explain why this problem is so important, other than the fact that humans have been trying to solve it for so long? What are the benefits and implications of solving it?
Apologies if the article answers this, but.. paywall.
I'm no mathematician, but my understanding is that the Riemann Hypothesis implies that the distribution of prime numbers has a nice, predictable structure. Many existing proofs assume that the Riemann Hypothesis is true to take advantage of that structure. So, now we know that all of those results are true!
One example is the Miller-Rabin test (https://en.wikipedia.org/wiki/Miller%E2%80%93Rabin_primality...), which checks whether a number is prime. Without assuming (a more general version of) the Riemann Hypothesis, this is a probabilistic algorithm: the best we could prove is that it has some small probability of reporting that a prime number is composite. But with it, we can prove that the algorithm is guaranteed to work. (Mind you, there are other deterministic primality testing algorithms that don't rely on Riemann for their correctness proofs.)
I don't know any immediate, more practical implications of solving it. I would love to know if there are any!
Imagine that we haven't mapped all of earth, and there are still many unexplored landmasses. A while ago we discovered a large continent, and after sailing around it we found it's the biggest unexplored continent we know of. But it's surrounded by cliffs on all sides, so we haven't been able to land on it yet.
Solving the Riemann hypothesis means we finally landed on it. We don't entirely know what's inside. Maybe it's a lot of mostly-empty forests with little of value, but if it's like any of the other continents we've already explored, it probably has an abundance of natural resources to take advantage of. Now we just have to go digging.
We just unlocked a large new area. We don't already know of specific useful things in that area, but based on past experience with unlocking other areas, we expect that in the long term we'll find a lot of useful things.
The result by itself doesn't really unlock anything, since it was presumed to hold. TBD whether the proof contains novel techniques, those would be the key to a new "continent" of math.
it has the potential to cascade things its a very general problem that allows tons of other theories to be simplifed or have greater guarantees, but its a really wide, so nothing economically relevant in the immediate sense, however hopeful if you want for progress, freightening if you think about knowledge work.
In the real world... not so much. Lex Friedman had a bunch of great guests talking about this.
At a certain point, math disconnects from the physical world. In a ~Platonic~ one may say.
Steven Brunton (GOAT and author of of the best ML books I've seen) gave the example of the Navier Stokes result. Could this new fluid dynamics result be used to make cool new engines? No. Unfortunately the scale where the proof resides, viscosity and real-world turbulence take over.
honestly it would probably be more consequential if the hypothesis was disproved; Cryptographers already assumed it was true, I imagine this is the case in many other areas.
If we could go back and ask people in 1850 if studying prime numbers would ever be useful, they would never have guessed the benefits we get from it today like internet banking, secure browsing, cryptocurrencies..
The biggest benefit of solving RH may be something we cannot fathom yet
It's associated with the distribution of prime numbers which is very important to cryptography but also my god what is with the intellectual incuriousity here?
The exact same ridiculous argument could be applied to the basic mathematics and physics underpinning literally all modern technology.
Personally, I'm deeply discomforted by the idea that the few mathematicians who have aligned themselves with these companies are able to race ahead of the rest of the mathematical community like this. It is unrealistic to expect Anthropic to immediately share their monetary wealth with the rest of the world (though this is their stated intention in the long-run via UBI), but math is just math. Does Levent really need to take the crown jewel of open math problems for himself before the models are even publicly available? Did OpenAI really need to launch a swarm of 10,000 agents and scoop Tristan Buckmaster, a university professor who had worked on the Navier-Stokes existence problem for most of his career?
This is a time of deep uncertainty and anxiety for working and aspiring mathematicians. OpenAI and Anthropic gain little by announcing results like this ahead of model release. So if they can't restrain themselves on such economically useless things as pure mathematics, it does not bode well for the rest of us.
It's discomforting only if you attribute value to your ego and your pride. Nothing has changed for the mathematician that find intrinsic value in the field and are interested in understanding the proofs, or the nature of Maths itself.
It has only changed for the people who want to stamp their name in blood on proofs and then get social credit and clout for what they have done. In other words people who use the field to find extrinsic value.
This is really good for mathematics, and science in general.
This misses the point I was trying to make, though I guess I didn't make it very clearly. I'm not concerned about who gets credit. I'm saddened by the lost opportunities for the wider mathematical community. There could have been conferences and Zulip chats and lunch conversations among young graduate students all making progress toward Riemann together with AI tools. It could have been a great problem to rally around. Hell maybe it could have been a Polymath Project. Instead it was solved using possibly tens of millions in compute behind closed doors.
But instead it will probably discourage many more mathematicians to leave math altogether, because they feel they can't compete.
Is it actually good for mathematics and science? Do these things have some independent, objective value, independent of the people that interact with them?
If a new miracle drug is developed but only ever used by the person who developed it, would we applaud it as a scientific achievement?
You are dismissive of ego and pride, but those can also be components of what drive people to accomplish great things. If you take away the incentives to pursue subjects like this, what is left?
I’m actually not all that negative about the results, I believe these fields will adapt and we will move forward. Or at least that’s still a possible outcome. But I’m challenging your assertions here as stated.
Wouldn't you say that a miracle drugs value is in curing people rather than in who gets a Nobel Prize from developing it? I can guarantee you that patients that have pancreatic cancer don’t give a shit if a human or a server rack developed a solution.
I would hence claim what is left is the people who are actually in this for the interest and understanding (intrinsic motivation).
I’m making a subtler point about medical research. Yes, its value is in the benefit it brings to people. If someone developed a powerful new drug and didn’t let anyone use it, both the drug and its developer’s accomplishments would effectively be worthless. Similarly, if we produce proofs only machines understand, and in doing do deprive people of an opportunity to strive towards a goal, to figure something out for themselves — the value drops to near zero. Often in math, the problem itself is worth much more than a single solid.
I think even with intrinsic motivation, people would get worn down by having things they worked for, possibly for years, yanked out from under them on a regular basis. The field must change how it does work and provides incentives, or it will become no more than a hobbyist playground for LLMs, and largely stop providing benefits to humans.
Very few if any human can hold the entire arch and workings of a modern CPU in their head either, yet people seem to be using their phones just fine and the value has not dropped to 0. If LLMs designed a chip (see HN thread from yesterday), and people don't really know every single in-and-out of it, but it works fine in the phone and gives you 1 month of battery lifetime, people wont care.
And again if having a problem solved before you "yanks it out from under you" that's still just wanting to be the one who crossed the finish line first. If you have internalized the extrinsic value mindset in this way, you will of course be sad about the developments.
Finally I think the current structure is already fine. Philosophy, classic literature, music theory departments still exist, and nobody is dropping out of them just because things were already done and figured out. I think here again we get to the archetype of the extrinsically motivated mathematicial who wants to acquire, stamp and claim things for their own for social status.
I think this will lead to people just doing maths in the way they want to, if I analyze Dostojevsky because I love it and find it meaningful, no-one can "yank" anything from underneath me and I can know that everyone of my peers have the right motivations to be there.
Even if humans are never able to understand the proof, they can still go forward assuming the hypothesis is now true e.g. they can assume GRH when working on other problems.
What's left is plenty enough of smart people. The whiny math subcommunity is not representative of all mathematicians or math-interested people. They are just very vocal.
But people will always value ego and pride, even if we’ll live in post scarcity utopia there will be still status games and so on. And that’s good, it will keep things fun!
Hopefully it actually is advancement of human society, because thus far this and all the other breakthroughs seem to be the advancement of shareholder profits, with any consideration for the rest of humanity being secondary scraps falling off the table.
I really, really hope to be wrong with this cynical view. I don't know what we should do in the alternate case. One idea is to internationalise the computational resources for AI so it could be used for the greater good of humanity, but I guess that idea would get a bunch of resistance, so I dunno. I guess the last one shall turn off the lights.
From what I see many people, for various of reasons, are OK to ditch the mass for that advancement. So the advancement might as well happen, just not for the common people.
Many intrinsically driven ones do agree. We are getting answers and seeing discoveries at a pace that no-one could have imagined to see in their lifetime.
If you however are full of yourself, want to stand in front of YOUR peers and prove that YOU did this, YOU figured out how this is connected, you want to have YOUR name written on some gold coin or certificate hanging in YOUR office, YOU YOU YOU everywhere, then you will have a big problem with the current developments.
A lot of this is the byproduct of how society views accomplishment, at least in the western world. While there are certainly lots of self centered, egotistical people out there, there are also many who are shaped by the constraints of their work and education. There aren’t many places where people get to the next rung by pointing out all the great things their teammates did. Ever since young school ages, we are taught to accomplish things on our own. Our grades are the byproduct of our effort, etc…
This is a western capitalist/neoliberal view. In a lot of countries for example, family is more important than work. Having a child and a social circle is valued more than being a CEO.
In some middle-eastern countries, the value is tied to how good you can hold the room in a social setting.
Seems like the western world will have to adapt their values and education in the near future.
Correct, which is why I specifically pointed that out from the get go.
I don’t see why the western world “will have to adapt”. This has been the case here for hundreds of years, and I don’t see a sudden shift to valuing different things. Maybe as generations go by, sure, but the west is in love with capitalism, the cars, the money, the “prestige”, it’s all quite sad.
That many people want to leave their mark on the history of mathematics or science in some way, small or large, is not so narcissistic a goal as you are portraying it. The desire for one's discoveries to matter to the people around oneself is intrinsic and powerful, and without it, the passion of science gives way to mere pragmatism. We can both appreciate the new pace of discovery and mourn the damage that is being done to the spirit of human camraderie on the edge of discovery.
If your goal is to impress other around you with your discoveries, it assumes that you are able to provide something of value that people around you find impressive.
This is extrinsic motivation. If this is your purpose, you can pivot and do commercial business and make a lot of money, or something else that people around you find meaningful. People might value how good communicator you are, how much empathy you have, how caring you are for your family, etc.
But then it wasn't about the math nor understanding.
Imagine reading a book for the sakes of the book. You have your imaginary world you get into, you will keep reading the book even if there are 1000% better ways to create value.
Now imagine you read the book because it gives you social importance or skills, you are able to speak about a topic to impress and acquire for example sex from others. That is a different motivation, similar to what many of these mathematicians have.
Actually argument and I were talking about different things. He/She was talking about getting a name on a paper. I don't give a fuck about that (and I don't work in Academia anyway). I just want to keep my job.
But yeah I should applause thunderously for the advancement of human society, to be a good boy.
Indeed! The greedy, selfish mathematicians taking all the theorems have been a problem in our society for far too long!
It's enraging to think that someone like Riemann would name a hypothesis all after himself. He was stamping his name on everything wasn't he! His zeta function, his sums!
There's only one way forward now--that's to let those benevolent CEOs at the AI companies democratize mathematics. They care about the intrinsic, and not extrinsic, value of mathematics. They care about the understanding! That's why they're rushing to publish these 250,000 page papers with their models, they're just so excited to share mathematics with the world.
Do you believe that researchers at historic, well-funded R&D labs (e.g. Bell Labs) should have kept discoveries private because they enjoyed resource advantages?
I am a researcher at an R1 university, and the nature of my work has changed, too, but the anxiety of working academics isn't a compelling reason to hide major results in physics nor math nor biology.
We ultimately do research to expand what is known, not to give ourselves satisfying puzzles or demonstrate prowess relative to peers
Probably important context: Tristan Buckmaster was collaborating with Levent Alpöge on Navier-Stokes before they were scooped by OpenAI. (Indeed, part of the controversy was that OpenAI wanted to include Buckmaster but exclude Alpöge from authorship).
because they are not selfish and dont see the point of gatekeeping their work or credibility like so many other artists or professionals have been doing.
you can't win against LLMs best to find a field of work where you own things or whatever robots can't do like sales.
I was wondering how this would affect research incentives for difficult problems. If someone is about to embark on a long PhD process, but there is a chance that an AI swarm solves their problem 1-2 years in, isn’t that a bit risky?
This is a good question. A big one is that Terence Tao only has 16 hours a day to do math, and he spends a lot of that time teaching math, giving lectures, and mentoring students. Tens of thousands of Claude agents can spawn in seconds and race ahead of any researcher.
I don't understand your logic here. Won't these advancements in mathematical research advance the subject for humanity AND force newer mathematicians to leverage AI in their careers?
This kind of proofs don't help humanity at all. AI uses a gobbledygook language that's incomprehensible by humans and the only outcome of its proof is a 0/1 result that you either believe blindly or not, for nobody can really verify it.
This is not true. In the processes I've read about, AI produces natural language proofs first, which it then translates into Lean code which is verified by the Lean kernel.
Another comment, I haven’t verified this, suggests this particular formalization is upwards of a few million lines of code. I really enjoyed this article from Wolfram where he argued that if we had a box that could prove anything it wouldn’t necessarily matter much because the important part is the application. Now AI might, will(?) provide both but when it comes to the proofs it’s hard to tell how useful they are _at this very moment_.
AI only imitates natural language. Nobody can read and understand the 2500 pages of Claudese. Some may argue that AI can explain each part of the complex proof and therefore the entire proof is human readable, but if this was true, AI could produce a compact human readable proof in the first place.
I would like it not to be. My emphasis is on the mathematical community as a whole, not the individual mathematician who might have gotten there first if not for AI.
There's an alternate scenario in which the whole math community got to use AI to make progress on these problems, share their progress in Zulip chats, and learn from that process. Instead a few individuals within private companies and with vastly greater resources are doing it all themselves.
So I don't mourn who got credit and who didn't; I mourn the lost opportunities. The people who will walk away discouraged when they could have been brought along instead.
At least, they could quietly collaborate with the community, and thereby wait until they have a more lucid, unambiguously correct proof, rather than dump a pile of slop on the community.
I can only imagine that the perceived financial pressures around the AI race/IPO caused them to announce this now.
It’s basically impossible to do collaborative work with AI. This is a problem in anything ai does and it’s a huge problem we’ve had to deal with at my company. The only collaboration is between one human and the ai; a different model is required to enable more than that tbh.
And it seems like the author wanted to avoid dumping a pile of slop, AND explicitly referenced the timing of the IPO in the pdf here. Did you read it?
The polymath projects are a pretty amazing recent example of this. Also the recent AI boom in math has spawned many Zulip chats where people discuss these kinds of results and try to build off of them. There's one dedicated to digesting the Lean proof of Apoge's result.
> Did you read it?
Hacker News guidelines specifically call our this question as a no-no, but yes I did. I'm not calling into question how the result is being released. I'm asking if it should have been achieved at all before public availability of the model.
Woah, if this proves to be correct it'd be massive, for maths, computer science...and AI too, I guess!
The worry I have right now is this though:
> The argument is incredibly long at the moment, and while its formalization
is finished and it is routine to produce a huge, unreadable document containing
sufficient detail, in my view of my path I have just taken a step.
this feels like we're moving away from the old world of 'science', where we could actually read an experiment, understand its steps, and, with enough time, possibly reproduce it.
And we're back in the world of 'magic' and 'belief'. Yes, there are formal ways to prove that this is correct, but if the proof is so convoluted and hard to read, is it actually something a human will ever find the time to go through and understand? We're going to start delegating (oh well, we're already doing that, I suppose) the understanding to AI, so it can just help us condense it down to smaller pieces that are digestible.
Still, very strange dizzy feeling, much more than Navier-Stokes
My perception of these maths solutions is that we're decoupling two related but distinct elements: the answer and the understanding. We have a big machine that tells us the answer is "42." That's useful. We know the question is answerable and the general shape the answer takes.
I'm just not convinced that we're swapping one world for the other. We're getting a hint that will probably help us enormously with what was always the main objective: understanding the result.
right, but if I understand right, the difference here is that the big machine that tells us "42" also gave a extremely intricate binder of the work and evidence that it used to arrive at 42?
It's more like a lot of scratch paper. When a statistician does their work, they don't publish their raw output directly – they collaborate with other investigators to write up.
They aren't distinct but rather a spectrum. Understanding is just building abstractions. You can skip that and have a 250k page proof. Or you can have a few more abstractions and create a 20k page proof.
Maybe, we're just dogs that have invented the human. The human is far more advanced and soon we won't understand what it is doing and what it is trying to tell us. But because it is our brainchild, it certainly won't harm us.
It's a Hitchhiker's Guide to the Galaxy reference:
> The Earth was destroyed by the Vogons five minutes before its ten-million year program to find the Ultimate Question was due to be completed. One of the only two humans who had survived the destruction of Earth, namely Arthur Dent, pulls random Scrabble tiles from a bag, so as to reveal information in his subconscious (Marvin announced that Arthur had the Question imprinted in his brainwaves). When his friend, Ford Prefect, lines up the tiles, they spell, 'What do you get when you multiply six by nine?'
> this feels like we're moving away from the old world of 'science', where we could actually read an experiment, understand its steps, and, with enough time, possibly reproduce it.
Two hundred years ago the limits of the natural sciences was regularly pushed up against by part-time dilettantes and tinkerers, and the theories were formulated to explain empirical observation.
A hundred years ago, world-changing discoveries were made with equipment that would today be hopelessly insufficient for a high-school science class.
Now, the cutting edge of science requires particle accelerators, wet labs and datacenters, large scale infrastructure and precision engineering projects at the cost of millions to billions.
The low hanging fruit was already picked long ago.
True, but that's how most of humanity feels about any discovery. You can't come close to understanding it yourself, you are just trusting there's a process that produces someone who says he understands it, and there's hundreds of years of institutional harness for us to believe that he has actually understood it.
Biologists tend to try and hide this detail, but this is essentially what biological research is. We will never really understand how most organisms work in a way that would satisfy a mathematician or even a physical chemist. We just tinker with what we can, poking the system and seeing how it reacts. Biologists reckon with incomprehensible systems everyday, so maybe this is why it's easier to stomach AI.
> this feels like we're moving away from the old world of 'science', where we could actually read an experiment, understand its steps, and, with enough time, possibly reproduce it.
> And we're back in the world of 'magic' and 'belief'. [...]
I don't know, to me it feels more like this might be the job of future mathematicians?
AI proofs noone understands are good, but maybe don't lead to that much. So maybe in the future, mathematicians will be more like "data miners" or more like experimental scientists, where they take a statement already known to be true (in the case of experimental scientists that "statement" you know to be true would just be nature itself) and try to find ways to make it accessible to human knowledge.
> I don't know, to me it feels more like this might be the job of future mathematicians?
Nobody would pay us mathematicians enough to prove it in the first place; why would they pay us enough to explicate it? I doubt there’s much value (in the sense of capitalism) in an explanation. Society by and large does not value pedagogy, or education, or even writing.
The social structure that trained a small number of research-level mathematicians every year is evaporating, and I doubt anyone can stop that now.
IF AI really goes in this direction where it intellectually leaves us in the dust, then there is clearly lots to gain for society by having people in all kinds of fields that at least try to keep humanity up to date that is also "monetizable" in the same way our current academic system is monetizable (meaning, on an individual level probably not directly but via network effects etc.).
Exception that proves the rule. Try making a living working off these big prizes. It's the same with infosec and bug bounties on a different scale. You have to already have a secure income, as the prize is a bonus at the end. In what other industry would you work for years on something that inherently can never return a profit, other than the chance of a prize at the end for the one person who got their first.
Society doesn't pay for math labor, it pays to celebrate the top 1% for labor subsidized by other institutions.
You missed the word ‘enough’. One million was obviously “not enough” (nor was that specific one million ever paid to anyone — it represents an investment of approximately zero in that sense).
It took on the order of ten billion dollars to develop Claude, depending on how one does the accounting.
I don't know about "massive". This is certainly massive for certain branches of theoretical mathematics, but I don't think it has any implications that you could not have arrived at before we had a proof by saying "assume the RH is true" and there are in fact innumerable results of that form that predate the proof.
> And we're back in the world of 'magic' and 'belief'. Yes, there are formal ways to prove that this is correct, but if the proof is so convoluted and hard to read, is it actually something a human will ever find the time to go through and understand? We're going to start delegating (oh well, we're already doing that, I suppose) the understanding to AI, so it can just help us condense it down to smaller pieces that are digestible.
If a tree apparently falls in a forest and no human goes to verify it, did it really fall?
No matter the verification results, as with most theories, for practical uses it was already assumed the hypothesis held. For mathematicians, they'll just say "Oh, it is true, but what's this proof? It's not understandable, so there's still a miniscule chance it's wrong, and we can't learn anything from it, so what's the point?"
What if this is metaphorically like a P = NP scenario? Working out the solution to these problems might be the "NP" here, but it remains sufficiently tractable to understand the solution in "P." So human understanding could come along for the ride, at least for a while longer.
You could, just like you could slowly disassemble your house back into individual pieces of lumber and nails and drywall, but it would be laborious and at the end of it you'd just have a pile of semi-broken construction materials.
You could try but that's kind of improbable. "Why don't they rewrite <program X> in the same language, with every feature intact, but much smaller and simpler" seems optimistic in Lean, C++, Rust or any other language.
If you spot the same idea used multiple places in the proof, you can pull it out as a tactic. Suddenly the proof is 30% shorter _and_ a core idea is clearer.
So yes it will be simplified. A lot. And it will be easier to understand.
Are AI proofs bloated because they just contain refactorable concepts over and over again? That's an interesting theory and it does align with the type of slop we see in vibe-coded apps, but I'm still not confident that sort of approach will radically shrink a 200,000 function proof. Imagine it gets down to 120,000 functions--that's a massive improvement but it's still enormous by any standard. Which gets to my original point--I'm sure you can shrink it, but even 2x smaller or 3x smaller gives you an incredibly large proof.
This just feels like seeing a complex empirical result and then having to define your view that confirms it. Except in this case the empirical result is a passing proof.
> this feels like we're moving away from the old world of 'science', where we could actually read an experiment, understand its steps, and, with enough time, possibly reproduce it.
It’s a genius first step that we will iterate on and make better.
OpenAI dumped hundreds and hundreds of solutions with Lean proofs that hadn't been seriously checked by any human and AFAIK, none of them have turned out to have exploited bugs in the prover. Some of them didn't prove what their paper proof claimed they did, but the base rate of LLMs exploiting such bugs (which are definitely available to exploit -- even for a human it historically hasn't been SUPER hard to find them, and the Lean kernel's metatheory is suspect in a number of ways) is empirically very low, even when the incentives for a significant result are very high. The more likely pathologies are, e.g., misstating the problem (the Riemann hypothesis is using the standard Mathlib definition), adding extra unjustified assumptions (unlikely in this case -- I haven't inspected the Lean proof yet but others have and reported that there aren't any), or not matching the paper proof and intuition (I'd strongly suspect the latter is true in this case).
It has been run through more than one; the same Sullivan who witnessed to the successful compilation earlier has checked it with `nanoda` an external typechecker written in rust.
AI is inherently about getting cooked/digested answers. It’s already finished. Nothing to build on.
But I still wouldn’t call the end result Magic as long as you have something to hold on to. I can see that this building in front of me (waves hands) is upright. Okay, now I remove this beam. It collapses. I don’t care about the whole structure. Just that it stays upright.
Giving up understanding is not the same as blind trust. Or maybe I’m too philosophically naive in this area...
I'm of the belief that "Claudese" with its load-bearing seams is just the models evolving our language beyond our capacity to use or understand it. Read up on your Continental philosophy to understand why and how: there's no such thing as a fixed "meaning"; meaning is something constructed for the immediate need and then discarded or modified as circumstances change. The models are proving the French and German philosophers correct as surely as any mathematical proof. And they're beginning to transcend human comprehensibility. Super intelligence is just around the corner, if not here already.
This is Levent Alpöge (https://en.wikipedia.org/wiki/Levent_Alp%C3%B6ge), a supremely qualified mathematician and one of the best in the world at guiding AI towards new discoveries in mathematics.
If he says he has Riemann, well then, he has Riemann.
Levent was one of the two mathematicians working on Navier–Stokes when OpenAI published their work on that problem - which was produced in a rush when they heard a rumor that someone from Anthropic had promising results on a millenium prize problem.
Pretty crazy interview, he didn't want throw the Anthropic employee under the bus when OpenAI offered co-authorship on the Navier Stokes paper. He suspects OpenAI somehow used his codex usage btw. Almost unprovable of course.
To: Buckmaster
From: A friend at OpenAI
Subject: Openai is about to do something really stupid.
"I said that if OpenAI released its result in the way proposed I would go public with what happened. The reply was, 'Why would you ruin your career?' I replied that I am an academic, and asked why he thought going public would ruin my career. The reply was, 'If you don't want me to be nice, then I don't have to be nice.'" (Sébastien Bubeck later apologized on X for the remark.)
"Sebastien twice asserted that he wanted Levent removed from authorship..."
"The second was that, after posting Euler, I alone write a paper presenting the Navier-Stokes result, acknowledging that an internal OpenAI model had resolved it."
"I asked whether the model had been trained on, or had access to, our sessions in Codex, into which we had been putting all our drafts for the whole of this project. I was told the model did not look up user data. I asked again, about training, and I did not get an answer."
(OpenAI subsequently stated publicly that while unlikely, it could not rule out that de-identified data from their product usage improved the model.)
"While unlikely ... it could not be ruled out ..."
So basically he hit the nail on the head and they want to lie about it without lying. Either that, or they are supremely incompetent, not knowing even what they ingest into their models. Both are very much not inspiring any faith in their trustworthiness.
In broader context the “being nice” part seems like both of them talking past each other.
What happened was the two researchers, Buckmaster and Alpoge (from TFA) solved a significant subset of Navier Stokes (referenced as “Euler”) which set off rumors.
Through whatever game of Chinese Whispers ensued, OpenAI thought they had solved the full Navier Stokes problem, which they hadn’t. So OpenAI pointed their latest unreleased model at it, and it actually solved the full problem!!
Pure George Danzig vibes on that one.
So when this meeting happened, the OpenAI folks were surprised that these researchers had NOT solved the full problem, only the subset. Yet from their perspective, they were offering them credit for the full problem, and so “we’re being nice.” The Anthropic employee (Alpoge) was clearly an optics issue and so they felt they didn’t have to be that nice.
But the other researcher, Buckmaster, was suspicious that their work had been stolen, and so was understandably in a combative mood. You can see that from his perspective they were not being “nice”, they were trying to scoop them, possibly on stolen work, and then bribe them with shared credit.
But they had not solved the full problem, and academic plagiarism is a very serious allegation that can ruin careers of both the accused and the accuser. So from OpenAI’s perspective it was like, why would you accuse us of plagiarism when you yourself have not solved the full problem and when we’re willing to share undue credit for it??
This was made worse by the fact that OpenAI would not categorically deny if the others’ work had been used for training until much much later. This was terrible optics, but anyone who’s dealt with large scale data pipelines compounded by privacy issues knows that it gets very difficult to trace these things, so it takes time to make a definitive statement.
Alpoge somehow seemed less troubled by this whole episode, but I guess it’s because not only was he gonna be a multi-millionaire soon, he had Riemann stewing on his backburner.
I feel like when they finally make a movie out of all this, this scene would end up being a bit of comic relief.
That version of events skips some central facts from Buckmaster’s statement.
"they came prepared and they wanted to show us the prompt... they showed me on the screen and they just literally copy pasted the Millennium Prize problem, so they said that they took the prize and they pressed enter and they got a solution, and at that point I just had to laugh because it's just so ridiculous. It came out within a few minutes that this was a lie"
"Levent had been told by Sebastien 'very little human input' had been used. This turned out not to be true. Over the course of the call, as members of their team sent Sebastien corrections and details over their internal chat, it emerged that an entire team had been working on the problem..."
"I was offered the opportunity to write as sole author of their paper and throw my collaborator under the bus... they were asking me to throw him under the bus, someone who I've been working with for a year, to throw him under the bus just because he's an Anthropic employee"
"the right paper to compare is the unforced Euler equation. Now the unforced Euler equation, all the mechanisms in this unforced Euler equation are identical to a different version of the Euler paper that we haven't released yet... basically all the main architectural ingredients in that unforced Euler were in our work for which they would have had on their servers, and it hadn't been published"
"they didn't use 10,000 agents to get to the Euler problem, they used 100 agents. So they worked from absolutely nothing with only 100 agents... and then they ended up with an identical architecture to what we had"
Not to put anyone on an idol but the only one who comes out of this not acting like a greedy, anti-human is Buckmaster and his colleague (I think). Straight up lying by OpenAI, very predatory "meritocratic" behavior. What gets me the most is "they didn't use 10,000 agents to get to the Euler problem, they used 100 agents. So they worked from absolutely nothing with only 100 agents... and then they ended up with an identical architecture to what we had". What the hell! So they lie about their technical claims too on a factor of 100x?? And companies are taking their words seriously, no wonder there is a house of cards being built in the industry.
I don’t see how those details change the overall outline of what I presented above through? Maybe you could nitpick that it was a team with a model using 100 agents, and not a solo model with 1000 agents. But the fact remains OpenAI ended up solving the bigger problem themselves just because they thought the researchers had solved it already when they hadn’t. That is low-key hilarious.
The Euler plagiarism insinuations are much more serious but given all the 100s of other results we’ve seen come out, highly unlikely.
I think it's clear enough that something broadly like this is right, even if people don't want to admit it because the vibes of extending any charity at all to frontier labs are bad.
Terrible optic is such a bs executive wording when they lied, coerced, bullied, and threatened him…
I like the summary but knowing what irthomasthomas posted that’s missing from it, and having a better picture of what happened; terrible optic doesn’t work.
Yes, because when looked at from the outside, like I tried to do with my recap, it's clear that:
1) OpenAI thought they were "being nice" by still offering (partial) first authorship despite having independently solved a bigger problem than the researchers.
2) But Buckmaster thought OpenAI were "being nice" as a means of CYA / bribing / coercion for scooping and plagiarizing them.
"Being nice" has very different connotations in the two contexts.
Thinking you're nice doesn't mean you're actually nice. OpenAI isn't talking past anyone. It's talking out of it's ass.
Buckmaster didn't think they were being nice. He thought they were being assholes to him and trying to bribe him.
Putting the phrase "in quotes" to use it completely out of context, only to make the point that surely no one understood each other... that's some impressive work. If you have to work this hard to make sense out of a conversation, it's because you're not.
Offering first authorship on a significant paper to a researcher who had not even solved the whole problem seems a lot like being nice (not "in quotes" ;-)) to me too, but that's just me.
Yes, it was a joint effort with Tristan. Part of the scandal around it came when OpenAI offered to let Tristan help write up their result but wouldn't allow Levent to be involved because he worked for their competitor.
He was also the one who dropped the Jacobian conjecture counterexample on twitter with the famous
> hello there the jacobian conjecture is false thanx to my close friend akhil for asking about it and my other close friend fable for working during the world cup final
It's a sign of the times that when someone with an AI beats you to one Millenium Problem, you can just turn the attention of your AI to another Millenium Problem.
It's almost certain that they have had people working on Riemann since way before the Navier-Stokes theorem proof was nearing completion. It's a long shot, but the value of a proof is next-level compared to NS
> It's a long shot, but the value of a proof is next-level compared to NS
The value of a proof? People have assumed it's true for so long that there are any number of results already proved "if the Riemann hypothesis is true".
If value here is a measure of utility, its not super useful on its own that something that was almost universally assumed to be true is now known to be true.
But with a problem as difficult and fundamental as the Riemann Hypothesis the thought has always been that in the process of solving it we'd be likely to develop or discover something that did have real world utility. At least this is often what happened in the past, the obvious example here being Newton developing calculus, a field of mathematics which can be applied to other problems beyond the narrow set of problems he was trying to solve.
What I find interesting about this, if the claim holds up, is whether this will assumption will remain true, or if the value of these AI proofs is going to significantly lower than those of human discoveries in the past.
Maybe a good analogy here is to think about how all of tools we humans have created has been ultimately in the pursuit of solving some problem. For the developers here, consider this – if ChatGPT had existed before React, would React ever have been developed? I mean, if an LLM can brute force the same result with just vanilla JS the need to develop frameworks and libraries to solve what was once difficult for a human to solve drops.
I'll note that I'm playing devils advocate here as I think AIs would probably build something better than React had it not existed, but it's something to think about at least. I suspect there's a higher probability than some here are assigning that this proof ultimately ends up have relatively little utility.
It wasn't assumed it was true. People thought a world in which it was true was more beautiful than one where it wasn't, but proving it true wasn't a pesky little detail.
In fact, there was a whole literature of papers proving that if the GRH was false in one particular way (the existence of Landau-Siegel zeroes) many theorems followed from it. For example, if there's a Landau-Siegal zero, the twin primes conjecture is true.
Are you claiming that there wasn't a significant majority of mathematicians who believed it true? That's all I'm saying. I can only assume we agree.
To expand on what I said though, if it is proven that P != NP then nothing really changes as that assumption is in line with everything we understand about computational complexity theory. The proof would really just suggest there's probably nothing that surprising to be found that we don't know.
If P = NP however that is far more likely to suggest we're missing something fundamental which could radical consequences for more practical fields like comp-sci. Similar can be said of GRH. Everything we know points in the direction of GRH being true, and while it may be interesting to prove these things, its much less likely to lead to some surprising and impactful downstream discoveries.
I doubt a majority of all mathematicians ever believed it to be true. (It's hard to have an opinion outside of your area of expertise, other than "I guess this expert sounds credible.") Probably the majority of all number theorists thought it was true. But this is mathematics, not democracy.
But that doesn't change the value of proving it. Most graph theorists probably thought the 4-color theorem was true, but it was still a big deal when a computer proof showed it. Most people thought the Kepler conjecture was true, but it was still a big deal when Hales reduced it to a big minimization problem that he solved.
On the flip side, everybody thought the Hauptvermutung was true, and probably subconsciously used it. But it's not.
The value it adds is that those theorems are now actually true, which means all those derived results are "real" mathematics... as every mathematician knows, it is quite easy to prove very surprising theorems when False is in your context.
But that's 100% wrong. If it was easy to prove surprising theorems from an assumption of the Riemann hypothesis, it would also be easy to disprove the Riemann hypothesis.
To borrow the usual quote about ingenious fools, the trouble with "easily" disproving that False is in your context is that there are such ingenious proofs of False. One can appear to be doing very substantial and sophisticated mathematics, encountering many obstacles along the way that look identical to what you'd face if you were relying on something true, only to realize much later that one of your hypotheses is uninhabited. The concern was that were the Riemann Hypothesis false, a very, very large subset of modern mathematics might fall into this category.
The value is proving AI can be useful after all (and we're fucked).
Remember, not too long ago we had hordes of folks here proclaiming machines can't think, reason or be creative in any productive sense. Many still do, but it's starting to look bleaker by the day.
For many of the proofs coming out of the AI companies where the prompts were posted, they ended up being variants of "solve the problem", "keep going", "I believe in you".
Many of the results could have come from a 15 year old that heard of the problem in passing.
The result here is a quarter of a million pages. Human involvement is going to be minimal.
These AI bros don't get it. These LLM systems solving math problems have the benefit of being constrained by formal rules of math providing constant feedback when they break the rules that prevents them from wandering into la-la land. As Dijkstra said in his prescient "On the Foolishness of 'Natural Language Programming'":
> "The virtue of formal texts is that their manipulations, in order to be legitimate, need to satisfy only a few simple rules; they are, when you come to think of it, an amazingly effective tool for ruling out all sorts of nonsense that, when we use our native tongues, are almost impossible to avoid."
To whatever extent that LLMs can "think, reason, or be creative" in generating these proofs (or writing code), they're aided by very high guardrails so it doesn't say very much about what they can do in broader, fuzzier domains.
Gödel proved that statements can be created that are definitely true or false but for which no proof can be constructed. It is not relevant here since proofs are successfully being constructed, according to checking using Lean, by an LLM.
The kind of AI used for backgammon/checkers and the kind of AI used for chess/go, and LLMs differ greatly so I'm not sure what the point of your question even is. A casual check says that the state of the art LLMs have much lower [EDIT]Elo ratings for checkers than the purpose built "AI" that beat humans.
> Remember, not too long ago we had hordes of folks here proclaiming machines can't think
Machines can absolutely think. It's really all they do.
> reason
Murky, but LLMs prove with an avalanche of statistical modeling, they can do something kind of analogous to reasoning, with the caveat that whatever they output will necessarily have to exist in some fashion in their training data.
> or be creative
Absolutely not. By definition an LLM cannot create original work, it can only remix and recombine what it already knows. That can be interesting, and bloody useful especially for code tasks, but it is not making anything original.
Arguing the definition of original is pointless for the exact reason you point out: that it's ambiguous. Due to the nature of how an LLM works, it cannot make anything truly new, it, again, can only recombine and remix things from it's training data. This is why AI images/writing are so easily spotted, and even AI code to a degree: they have a distinct "style" as a consequence of how they work.
I leave what is truly original to the philosophers. What I can say for certain, is that anything you get out of any of these models is a remix of what is present in it's training data, because gravity exists and energy is conserved. They are not intelligent, they are not conscious, and they cannot make New things.
> What I can say for certain, is that anything you get out of any of these models is a remix of what is present in it's training data, because gravity exists and energy is conserved.
But doesn't that same argument apply to humans, where 'training data' = all our sensory inputs (plus what's encoded in our DNA, depending on whether we want to analogise that to the training data or the underlying architecture)?
> Arguing the definition of original is pointless for the exact reason you point out: that it's ambiguous.
I'm not trying to be a smartarse here but you are the one who is claiming absolute certainty that something is false 'by definition', so I think the onus really is on you to have a precise definition in mind.
> They are not intelligent, they are not conscious, and they cannot make New things.
I'm not a consciousness denier or anything, in fact I think it is the most important thing in the world, but I don't understand the connections you're making here. 'Intelligent', like 'original', can be defined and needs to be defined for this statement to be meaningful. 'Conscious' is probably inherently slippery, but if we mean the same thing by it (having subjective experience/qualia) then I have no idea how you could be certain that any other entity lacks it. And it seems like you may be implying that consciousness is a necessary condition for originality, but I don't think you've explained how.
> But doesn't that same argument apply to humans, where 'training data' = all our sensory inputs (plus what's encoded in our DNA, depending on whether we want to analogise that to the training data or the underlying architecture)?
That is true if your model of the human mind is akin to a machine model you also use to describe today's computers. But that is a huge assumption, making the argument somewhat circular. We don't really understand the human mind and therefore don't know if there could be other models for it which are not equivalent.
If you think the human body is subject to special laws no other object in the universe is, feel free to present your evidence. No model of the human brain can exist that violates the conservation of energy and thermodynamic laws regarding entropy and information. If you have one, feel free to present your groundbreaking results.
If there has to be something breaking the current known physical laws, the best chance to find it would be in extreme conditions like black holes or stellar cores not in close to STP conditions like the human body.
> Due to the nature of how an LLM works, it cannot make anything truly new, it, again, can only recombine and remix things from it's training data
Fwiw, David Hume made this exact argument about human imagination:
> But though our thought seems to possess this unbounded liberty, we shall find, upon a nearer examination, that it is really confined within very narrow limits, and that all this creative power of the mind amounts to no more than the faculty of compounding, transposing, augmenting, or diminishing the materials afforded us by the senses and experience. [David Hume — An Enquiry Concerning Human Understanding (1748), Section II, Paragraph 5]
Every action ever in universe is the effect of either something deterministic or random. There are no causes or effects that can't be put into one of those bins.
The only "out" here would be deterministic actions that originate outside the universe, but we have no evidence of that. It also doesn't escape the deterministic/random duality.
Nobody is claiming new forms of energy or phenomenon can't be found. As of yet though most such were found in more and more extreme conditions, starting from radioactive substances to particle colliders and astronomical phenomenon. There is a much lower chance to believe something fundamentally or radically new phenomenon in physics to emerge from human bodies.
Non answers are best ignored. You have failed to define what creativity is and therefore failed to prove whether llms are "creative" or not.
What exactly are you trying to say about gravity and thermodynamics? You may be confused but these are universal physical laws that apply to everything, humans, computers, anything else.
At its most basic level if I had to start with defining creativity, it should be some kind of intelligent filtering process. The brain can be in a huge number of state configurations, the interesting and novel among them are "creative" ideas. We can build upon it further. I am happy to say I am no philosopher and I will defer to those with expertise on it. But if you don't have any definition for it, its quite strange to make up some beliefs about it without proof.
You’re the one who said “ By definition an LLM cannot create original work”, so I think responding asking for the definitions you are talking about is quite justified?
Often when someone says that something is true by definition, their claim is incorrect.
Arguing the definition of original is pointless for the exact reason you point out: that it's ambiguous.
Well, it certainly wasn't "ambiguous" two posts ago when you said "Absolutely not. By definition an LLM cannot create original work." Make up your mind.
You've been beating this drum since roughly GPT 3.5, as I recall. Your goalposts aren't even on the field anymore, they're gathering dust in a far corner of the parking garage down the street from the stadium. Time to update your priors.
Even after all the MPs are beat, there's still work to do to actually understand them, communicate them, and transform them in a way that can be leveraged. I predict that's going to be much harder for AI to solve.
Is it really wise to only rely on credentiallism to verify things? Plenty of mathmaticians top of their fields believed they have proven something when the proof contained some flaw. Just look at Mochizuki
Hacker News is a water cooler conversation, not a peer review system. It is in fact quite useful in this context to short circuit a whole raft of standard dismissals!
We went through this with Alpöge over the Jacobian Conjecture and the resulting thread was pretty embarrassing.
This sort of black and white thinking isn't helpful. Their point in giving context about the author is to give you more data for your internal heuristic to figure out whether this result is plausible and worth discussing.
As opposed to, for example, someone with AI psychosis on Twitter posting that they've solved it.
Even if Terence Tao came down from the clouds with a beautiful human-written proof, no one would be certain that it is solved--people make mistakes. But again, it gives confidence and something fun to discuss until the math community digests it fully and comes to a consensus.
Anyway, TL;DR is that no one is telling you to go put your life savings on a prediction market bet that it's solved. Relax a little.
Math is hard; even to understand the questions and what the entities are which the questions address takes a lot of deep study, even for smart people. So credentials are not sufficient but they are for sure necessary. I have tried my whole life as a software engineer to study math as an amateur and then started a masters program and learned more details thoroughly in three months than in decades of self teaching.
Not in a Bayesian sense, not "merely" suggestive, but quite suggestive. Not sure how many full professors you have met, but they aren't like that under-performing coworker that people have to cover for.
More like IIT grads in being, Bayesianly confident, quite exceptional.
> I have tried my whole life as a software engineer to study math as an amateur and then started a masters program and learned more details thoroughly in three months than in decades of self teaching
Oh man, you can't imagine how much I'm cheering you on and rooting for you. I wished more peope could embrace that insight/approach.
I would honestly like to know what you think is the difference between those groups of people. I thought about that a lot and couldn't find a better explanation.
Why do some people understand what "maybe" / "it might be true" means, while others believe in absolute truths?
No. The mathematicians will verify this, and come to a conclusion.
What the credentialism does is establish that this should be taken seriously and has a real possibility of being right. The Lean also helps (no Lean for Mochizuki’s work).
No. A lean codebase that certifies that whatever definition of the Riemann hypothesis their LLM wrote is correct.
I can write Lean that certifies that 2 + 2 = 5 is correct, by redefining every axiom about natural numbers on the way. Or abusing bugs in Lean. Stop believing Lean is the response to everything.
Just like Navier stokes etc they use independently written Lean statements written by humans; formal conjectures [1]. Indeed, they have one for the Generalized Riemann Hypothesis too [2].
> I can write Lean that certifies that 2 + 2 = 5 is correct, by redefining every axiom about natural numbers on the way.
Which you would catch using comparator?
> Or abusing bugs in Lean.
And just like all the other ones, people check with multiple independent kernels, so you need to abuse bug(s) in all of them.
> Stop believing Lean is the response to everything.
For a discussion of how to check that the untrusted proof isn't playing fast-and-loose with the axioms of mathematics or exploiting kernel bugs, see Kevin Buzzard's recent post "...and the proof is checked in Lean" at https://xenaproject.wordpress.com/2026/11/12/and-the-proof-i....
The axioms would be a given to the theorem prover of course, as is the norm for theorem proving. This then reduces a complex proof to checking whether the theorem is stated correctly and the proof kernel is correct. This may not be (provably) easier, but it is so far removed from yet equivalent to the original problem that you get a machine-verifiable independent audit as a result.
And don't reduce "is not the answer to everything" to "is the answer to nothig", there is tremendous value in automated verification, especially in the AI age where the moat to it going mainstream is shrinking by the day.
>whatever definition of the Riemann hypothesis their LLM wrote
Given that the proof is provided by Levent Alpöge, who is known to be more than familiar with Lean, is it not unlikely he made up his own definition of the problem?
No it’s not. HN main approach to analyzing anything is to rely on chaining heuristics. Which is generally sort of fine for software, but doesn’t make sense in the real world. It’s at best one signal to weight in your analysis, nothing else
It's the only realistic way to operate for nearly everything. Try determining whether a sandwich is safe to eat without applying a shitload of heuristics.
The difference is that you generally need to eat something and so drawing some conclusion about th sandwich is essentially forced upon you.
Getting buzzed over headlines about math or medicine or whatever, on the other hand, is just an irrelevant hobby of bored gentry and the opportinity to just not draw a conclusion at all remains open.
Do you just not contemplate anything that isn't directly related to your physical needs? What a sad existence. I wonder what you're doing here, in that case.
Alpoge is a smart and principled guy, and he is probably the world leader at getting LLMs to do math. But he's not an expert in these subjects. If we're going to argue from authority, let's wait for an actual authority to weigh in
Of course he is an expert in these subjects. He has a mathematics PhD from Princeton, with a thesis on algebraic geometry. If you cannot call someone like that an expert, you cannot really call anyone that.
I have a graduate degree in mathematics, with a thesis in algebraic geometry myself. I know what I know and what are my limits. I am not an expert in Riemann hypothesis now, but if I spent a couple of months of focused study, I would be. That’s where Levent surely is right now, at the very least.
You two are just talking past one another and using different, and valid, definitions of experts. Yes, Alpoge is an expert in mathematics. Though how much of an expert he is in number theory? I don't know, I'm not an expert in number theory myself, I don't know Alpoge, and I don't know where he's been focused lately.
But either way both of your arguments comes down to speculation unless you two are also experts in number theory and know Alpoge or at least what he's been working on (to a sufficient level).
I think we mean different things by expert. You appear to mean "superficially familiar with some known approaches to the problem". That's all you're going to get in months.
And, yes, I've also published papers in algebraic geometry.
Silly definition. I'm a "top 10 people" in the application Data Science to [niche that my company works in] because there are so few people working in this area, but I've only scratched the surface in my 3 years of working specifically in this domain.
Some things just really aren't well-understood or well-studied.
And other things are extremely well-studied and well-understood where 10 is significantly too low a limit.
> It was my secret in my student days that my plans were for this problem in
decades, which meant then that I tried to become familiar with all tools that I
could possibly recognize in the problem when it was time
Sarnak weighing in is completely irrelevant. There is a lean formalisation, and it either compiles or it doesn't. (And the statement of RH is well established so there is no issue there.)
Lean can and does have bugs. It compiling doesn't mean anything, because that doesn't tell you whether the proof actually encodes what the mathematician intended. Plus, again, there can be prover bugs.
In principle, sure. But the likelihood of a plausible proof strategy produced by an AI agent with human assistance somehow accidentally exploiting a Lean bug is incredibly low. In any case, the fact that the Lean compiles is certainly going to be, in this setting, a more reliable assessment of the proof's correctness than the opinion of any expert, unless that opinion also comes with an identification of the bug being exploited.
Feels like we're reaching for arbitrary levels of expert to undermine the claim. I'm sure the world's foremost Riemann Hypothesizer will come forward at some point, but we don't need to wait for that to take the Princeton math PhD seriously.
What an odd thing to say. By this logic dumb people should defer to worse experts because "compared to you, they're very smart". Dumb people need the best experts of all!
Apparently he has a Lean proof, so all that needs to be verified is that the Lean formulation of the problem matches the hypothesis, which I assume he's qualified to do (and has done).
No, you need a long track record of working on the Riemann hypothesis. You need to be the person who's called to verify a possible proof of the Riemann hypothesis. I have a lot of respect for Alpoge, but he is not the person a journal would call to verify this.
Asking out of real curiosity, not trying to provoke at all: and what happens when somebody verifies it and say the result is legit? Would Alpoge become an expert on the topic or wouldn't, due to the AI assistance?
Also, if it were the most expert person on the Riemann hypothesis the one working with AI to prove it, how would the verification work? The 2nd most expert person? Or would they be trusted directly?
It's a fun question. I don't think AI assistance transfers expertise from the agents to the user. Alpoge says much the same thing himself, noting that Claude produced 250000 pages of slop and that he doesn't trust his own summaries in places.
For the other question: We'd want multiple independent expert verifications.
That’s a strange form of credentialism. I understand it if you wanna reject amateurs who make graphs and don’t understand what they’re saying but this is an actual mathematician and clearly he has thought about this particular problem for a long time.
From the document: "It was my secret in my student days that my plans were for this problem in decades, which meant then that I tried to become familiar with all tools that I could possibly recognize in the problem when it was time".
It's true that he haven't been publicly recognized as a top expert. But it's not like he's just a software dev who got lucky with prompting - he's a mathematician who specifically studied this problem.
I don't think we should delegate gatekeeping to journals - a lot of crap is published, by "experts". (Also a lot of good results don't get published - there are well-known examples.)
I don't get it, if the proof is formalized in Lean, the only expertise you need is running lean, asking if there were any additional axioms added in the proof that shouldn't be there, and if the statement being proved is the GRH (verifying all the definitions involved match expectations).
The usefulness of a Lean proof is that it is possible to just look at the theorem statement and at the axiom that are used. Riemann hypothesis is quite straightforward to define and to be checked even by a undergrad. Obviously a bug in Lean kernel is still possible.
The lean proof is public. Although, if it's anywhere near as long as the "natural language" proof, most people probably don't have enough compute to check that it compiles.
footnote: I object to calling the langauge that Claude produces "natural". It appears to be uniformly distributed over the space of grammatical constructions.
> If he says he has Riemann, well then, he has Riemann.
Well he said he doesn't even fully understand the proof (250k pages). So it's more like "If lean says he has Reimann, then we presume he has Reimann."
The best argument I've heard for lean being trustworthy is that different lean engines won't likely have the same bug. I guess what I'd like to see next is a list of every lean engine and its result for this?
And then if that passes, presumably the goal is simplifying down from 250k pages.
I don't have much of a math background but the author making such strong claims about a proof the author couldn't have possibly read seems absurd to me. What am I missing?
> If he says he has Riemann, well then, he has Riemann.
He works for the company, which is a conflict of interest. At that point it's not about credentials, it's about having someone with similar credentials but not the conflict of interest say the same thing.
Is it not a conflict of interest when a researcher who works at a University announces results? Let’s not act like the institution doesn’t benefit from the work of its employees.
It was fascinating to read, but without knowing that, it was hard to know whether the sentences indictating his ignorance of certain areas of mathematics involved was an indication that he was completely out of his depth, or the usual and appropriate caution of a fully qualified academic (the more you know, the less you think you know phenomenon).
Appeals to authority are not inherently fallacious if the individual is, indeed, an authority on the topic at issue.
To the contrary: we lean on authority opinion in most spheres of life because one individual's ability to understand is limited and so we abstract away "true" in auxiliary fields to mean, roughly, "held to be true by trusted authorities". We even rely on it in law.
He's saying that it isn't right automatically because he said so. It adds significant credibility that the proof is correct compared to me saying I did it but the proof itself is what matters (at least to most people)
Just like when Andrew Wiles discovered the proof to fermat last theorem and Grigori Perelman with the poincaré conjecture, multiple authorities in the field got together and verified that the proofs were correct (even after Wiles initially had an error.)
An appeal to multiple trusted authorities is better than a singular authority. So at best I'll keep as unverified until it is checked, unless you have read the 250,000 proof and checked that there is no issues or soundness bug with the lean proof.
They are inherently fallacious as in logically unsound. There is no connection between his credentials and him being right about this particular thing.
It's very important to keep those things separate.
They are inherently fallacious as in logically unsound.
At the risk of triggering an infinite regression (or logic loop), the redoubtable Stanford Encyclopedia of Philosophy's page on fallacies provides the following (https://plato.stanford.edu/entries/fallacies/):
The ad verecundiam fallacy concerns appeals to authority or expertise. Fundamentally, the fallacy involves accepting as evidence for a proposition the pronouncement of someone who is taken to be an authority ***but is either not really an authority or a relevant authority***.
Once more, I would point you to the fact that even the legal profession-- which is fundamentally one of the most epistemologically pure fields-- spefically allows the opinion of those deemed "authorities" to be submitted as evidentiary.
Without an epistemology that allows integrating truth-by-authority, you by necessity need to prove every assertion from first principals. Have you done so?
Very interesting sir. Thank you for this exposition.
But I do not see how this changes the notion that being an authority does not turn your arguments sound or even valid. There is nothing to connect. I can be an expert and completely wrong and there is nothing epistemology can do for you there.
The legal profession is taking a shortcut like the CS profession is taking a shortcut with the halting problem. Sure, we can end the discussion that way but we cannot use it as an argument otherwise.. halting problem solved, just press control-C?
You can't even contribute to math at this level without an advanced degree (cranks and delusional people notwithstanding). So maybe I should clarify I mean "authority" as in someone with high repute among his colleagues.
If we were all factory technicians, and Max has the best record among us at fixing the machine, that doesn't mean his word is final. "Max said so" is an appeal to authority. We are all capable of reading the documentation and performing our own repairs, and sometimes Max can be wrong.
It doesn't make it right, but more importantly to people who always quote every known fallacy: it doesn't make it wrong either. We shall see, empirically how it goes.
But it is highly relevant whether this is some crank on an Internet forum posting proofs about infinity-hat or someone with some credentials, even if that framing was a bit hyperbolic.
Which is another thing: sometimes people aren't really making an argument as such, so quoting fallacies is just on purpose low bandwidth English comprehension.
It's not about whether the claim right or wrong. "Fallacy" doesn't mean "a function that always returns false". It means it employs flawed reasoning. Saying "$SUPER_COOL_GUY claimed it" (implying it must be true) is a thought-terminating phrase. It's saying "don't scrutinize or question. Trust that he's the authority on these things. For only a fool would attempt to contravene his infinite wisdom.". That logic is absolutely flawed.
Even Einstein was incorrectly critical about some quantum physics properties. Who would dare tell Einstein, the master of relativity and spacetime, that he was wrong? How could he possibly be wrong when it comes from the man who gave us the most famous math equation in history?
Yes, that's right, but proof checking does take time and effort. For an initial announcement, the fact that they are a well regarded mathematician means that it's worth taking seriously and not that it's definitely correct. If it's someone unknown, then you'd definitely be waiting for the proof to be checked before talking about it.
That is why it is still a "claim" of a proof, until other mathematicians verify it.
Not saying it is incorrect but, all claims of mathematical proofs have to be checked over formally regardless of who discovered it.
With this particular mathematician it just means it should be taken seriously but you cannot blindly believe it is correct without multiple authorities testing that there are no holes in the proof, just like they did with FLT and the poincare conjecture.
Is it? It sounds to me like glazing one person in particular rather than the rough consensus of the entire mathematics community. I'm saying the latter is necessary.
This is addressed directly on page 2 of the linked doc, no?
> It is obvious to me that I should remark on the following.
> The precise timing is horrible. If I could have, I’d have finished this all much earlier. Hell, strictly before this age would have sure been nice.
> This all happened for me late Sunday night, and I messaged Dario, "Hypothetically, if I were to have . . . ", eventually asking about my ability to announce.
edit: Apparently I have forgotten how to format correctly
What you said is exactly the opposite of the scientific research spirit: Instead of be suspicious and honest, you forge idols and blindly believe in them.
How? No matter how "qualified" you are, the AI is still doing ~100% of the work. And since we're talking about superintelligence (Levent's intelligence hasn't even been stated anywhere), it's questionable whether a human can even "verify" or "check" the AI's work.
If you read through the pdf that this post links to, you'll find that while claude contributed many ideas and was ultimately responsible for the technical "implementation" of the proof, there was also a lot of non-trivial guidance and steering involved from Alpöge himself. He didn't just sit down and prompt "claude, solve the riemann hypothesis for me".
If it was that easy, OpenAI would have already announced the result weeks ago.
So if we stick a 10 year old in front of Claude and instruct them to type in “create a proof of the Riemann hypothesis”, as long as they keep pressing return, eventually a proof will pop out?
It isn't how this work was done, but it is basically how UGC and 400 other high profile theorems were proved by OpenAI so I'd say the doomers have a point.
Huh, what a conincidence that the AI presented the proof to Levent Alpöge (a top class mathematician) and not to some Rando McRandomface then? What are the odds?
He doesn’t really go into much detail here, but according to him he came up with the strategy: „ I had us on the line of combining two quite
different ideas present in a toy setting, one of which was already new and
shocking, making very simple one of the shadows of Riemann over which we
mathematicians felt most triumph. I guided us to the right situation analogous
to that toy setting, and then Claude began to leave me in awe.“
That does not sound like he typed: „solve rieman thingie, make no mistakes, plz!“
> Jarred Sumner, an Anthropic staff member (and non-mathematician), prompted Claude to “take a real stab” at the hypothesis itself, leaving the mathematical choices from there up to the model. Initially, Claude generated and tried 650 ideas, none of which worked. Jarred prompted Claude to try again, and it spent a day and a half coordinating about 60 Claude subagents, which this time went much deeper: between them, they ran 2,400 shell commands and wrote hundreds of Python scripts.1 The subagents ran thousands of numerical checks against known zeta zeros and refereed one another’s work. Throughout this process, Jarred's input was mostly limited to sending Claude messages of encouragement (mostly variants of “keep going” or “believe in yourself”).2 This seems to have helped Claude overcome some initial skepticism that it could make meaningful progress.
Only after this person thought he/claude found a solution were the mathematicians brought in.
Because the writeup describes exactly that process:
> Jarred Sumner, an Anthropic staff member (and non-mathematician), prompted Claude to “take a real stab” at the hypothesis itself, leaving the mathematical choices from there up to the model. Initially, Claude generated and tried 650 ideas, none of which worked. Jarred prompted Claude to try again, and it spent a day and a half coordinating about 60 Claude subagents, which this time went much deeper: between them, they ran 2,400 shell commands and wrote hundreds of Python scripts.1 The subagents ran thousands of numerical checks against known zeta zeros and refereed one another’s work. Throughout this process, Jarred's input was mostly limited to sending Claude messages of encouragement (mostly variants of “keep going” or “believe in yourself”).2 This seems to have helped Claude overcome some initial skepticism that it could make meaningful progress.
> Having found this new result while attempting the task, Claude tested its work by having various subagents review the proofs, search for counterexamples, download 54 papers from the arXiv to check that its finding hadn’t already been made, and independently re-prove its finding from scratch. Claude volunteered to write its findings up as a paper, and recommended that a human number theorist validate its findings.
> Levent Alpöge and Ralph Furman, two of Anthropic’s own mathematicians, examined Claude’s work to understand the new results and how they related to the prior work mentioned above. In parallel, Claude worked with another member of staff, Eric Easley, to produce a Lean formalization of the result, which passes the standard validation tool comparator.
The third paragraph. You left that entire part out: the actual proof, the actual confirmation. The thing that's literally linked to at the top of this thread.
There's an entire process here, and you just decided to leave the most important part out. So, your comment literally disproves your claim. Which makes you a liar. Go ahead. Get to this thread by removing that third paragraph. That is the critical part of this entire thing.
b) read and understood over 200,000 pages of ai-generated proof
everything they post to their website is marketing
You also seem to be conflating “Levent guided the work” with “curious employee brute-forced some potential solution through an agent tree”. He (in this case both Levent and said employee) did not guide the work, it was in effect a random employee saying “solve this” over and over.
I have actually considered the fact that I may secretly be an alien embedded into humanity, or perhaps a dog in disguise.
But evidence really points to the contrary, so... no, I'm pretty confident I'm human with regular human stuff that AI solves for me. House and family stuff on top of the science and tech stuff that applies to a smaller portion of human kind.
What are “mankind”’s problems if not the problems of individual humans? It seems like we’re treading towards a definition of AGI that would imply most human beings aren’t generally intelligent, which seems wrong headed.
I think the point here is that quite a lot of problems are being solved and this is entirely self evident based on the stuff people are using it for right now.
You can have the AGI label, you can put it next to the "actually, we've solved world hunger, people just aren't working hard enough" discovery, somewhere on the shelf with "every billion I make trickles down to the barista when I get a coffee at starbucks".
The AGI label doesn't matter. You can call it AGI when you feel like it's generally intelligent--I've yet to see any model that even remotely seems intelligent, but it all depends on where your bar is.
If this is superintelligence to you, that doesn't make it superintelligence to me. I can not have a single chat with an LLM without running into it clearly just predicting whatever sounds good instead of actually reasoning. I can't understand what sort of work people are doing where this seems super intelligent, or intelligent at all.
But not in some sort of super clean elegant way, it's "just" brute force, no? The same way it brute forces fixing tests until the software works, instead of designing coherent software in the first place.
And yet, we still can't make a robot that can clean a room. I mean really clean a room, picking up trash and put it into the bin, clothes on the floor go into the laundry basket, remove stuff from the shelves, dust it and put it back, etc... the kind of job any maid can do. Not like a Roomba that can be defeated by a loose sock.
AGI means "general", a machine that can solve problems geniuses have struggled with while failing at problems that can be solved by minimum wage workers is hardly general, instead, it is a good specialized tool.
Is this another $1 million Clay prize flushed down the drain, or can he actually claim this one?
I guess it's small fry to him given recent business news. Sort of a shame that all of these prizes are disappearing. Perhaps we should give the prize money to a body of mathematicians to do stuff with.
Also the Wikipedia says that the full written proof is 250,000 pages. Which I am guessing might be a mistake. If not it could take a while for humans to verify it.
There are dozens of papers claiming that their formalization is wrong, that their natural language claims do not match the lean version of it, some parts of the proof are insufficient. finally, their entire proof relies on an external force. For a problem that is about, for 99% of people that care about it... unforced fluids.
but hey, it gets them a month of being talked about.
I believe it's actually the correct "reverse acronym" in Spanish? I mean, keeping the Latin one they created a Spanish version that has the same meaning, in Spanish.
In Italian we use "CVD" for it (Come Volevasi Dimostrare)
CVD seems to be a fairly accurate translation of QED. The Latin is "what was supposed to be shown"¹ and it looks like the Italian means something like "as it was desired to show"? The Italian isn't a relative clause, but the meaning looks basically the same.
The Spanish appears to mean something more like "and now, it is shown", which is a completely different meaning expressed with unrelated words that happen to start with the same letters. (Spanish still has direct equivalents of all three Latin words², but I don't think que era demostrando even means anything.)
¹ In an English proof, you might see something like this:
"We can then see that a = b < c = d ≤ e, and therefore a < e, which is what we wanted to prove."
QED stands for the "what we wanted to prove" part of that, except that there is no we; in the Latin, we simply make reference to the fact that something was supposed to be proven.
² Sticklers might argue that Spanish has lost Latin's quod and replaced it with what is really a descendant of quid. They are almost the same word; quid is an interrogative form.
I think the joke is that while other cultures find Latin to sound esoteric and/or fancy, especially if they had to take courses to learn it, that your garden variety, non-educated, Spanish speaking person just simply gets it cause it's so damn similar to Latin that it doesn't faze them.
That doesn't look similar. Rest in peace is almost a direct word-for-word translation of requiescat in pace "let [whoever] rest in peace".
By contrast, queda entonces demostrado is totally unrelated to quod erat demonstrandum, using different words (except that the third words are both forms of the same verb, very different forms) to express a different meaning.
Really? The Spanish words don't seem to bear any relationship to the Latin words, so I fail to see how the relatedness of the languages could be helping.
Have you studied Latin, and if not, what does "quod erat demonstrandum" appear to mean to you?
Not exactly. The Spanish version translates to something like "so it is proved" and is a full sentence, but the Latin is a phrase closer to "which was supposed to be proved".
I mean a single person can always make mistakes regardless of how smart they are. But this does make it sound a lot more exciting than if a random vibe-prover came up with it.
You mean, if the AI says it does, than it does. And it’s probably right, because it’s smarter than all of us and we basically just have to trust it now, which is exciting and frighting in many ways.
It's not a random Anthropic employee. It's claimed by Levent Alpöge. The guy is on fire and already has like 3 or 4 very interesting verified result this year. I've showing for months his wikipedia page to my coworkers in the math department in my university.
This. Frontier AI means that the sharpest minds in the world now have an infinitely patient and motivated colleague to work through every idea.
I’m not claiming to be “top human intelligence” but I’ve published in premier venues in my field and I can say confidently that none of my human colleagues can keep up in conversations on my research niche like Claude can.
You reminded me of when Rodney encounters a clone of himself in SG Atlantis
Clone McKay: Oh, this is great! Finally, someone in whom I can have one hundred percent confidence!
I cannot tell you how much I'm looking forward to working with you.
McKay: And this is just the beginning.
Clone McKay: Humanity's about to benefit from one of the greatest living scientific minds…times two.
Jest and modesty aside, I think you're right in that it amplifies your strengths (and depending on how much you bias it, your gaps?) so will we see the world's top minds ever widening the gap between themselves and the rest of us from here on out?
In correspondence chess (which is chess with a time limit of like a week per move), humans ask multiple chess engines which move to make, and then choose.
And none of the best computer-assisted chess players are better than AI today. In the same way a human + chimp pilot is not going to beat a human pilot.
This era of human relevance in fields like maths will end – probably in a month or two if we're lucky.
Exactly. We don't even understand all of the new solutions yet, I heard Terence Tao say he had only looked at about 10 of the OpenAI proofs, found some good things, bad things, mistakes. But it will be his (and other's) understanding of the proofs that will give new insights.
In other words, what will the AI solve if we don't even have the language yet to formulate its next challenge?
> The guy is on fire and already has like 3 or 4 very interesting verified result this year.
Actually Claude is on fire, and this guy is along for the ride:
> Claude and I have had a great time together with such breadth in recent months, but this time
Claude far outpaced me near the finish line. No, Claude flew, and danced, and sparkled near that finish line. I had us on the line of combining two quite different ideas present in a toy setting, one of which was already new and shocking, making very simple one of the shadows of Riemann over which we mathematicians felt most triumph. I guided us to the right situation analogous to that toy setting, and then Claude began to leave me in awe.
The writing is certainly exuberant, and we are used to cringing at this kind of language because we are used to it being unjustified. But if the proof holds up then I can hardly object.
Because it sounds like you're a valley girl. Just spell it out, seriously. It's only 4 more characters, a small fraction of the entire sentence you typed.
They told you what's cringey: the anthropomorphization. "Claude and I had a great time" – no, Claude didn't have a great time, it's software. Levent Alpöge had a great time using Claude.
I recently gave claude full control of a robot. I have never seen it mimic being excited for any other work I've given it.
These are next word prediction machines intentionally optimized to match human generated text, to be as human text as possible, then fine tuned to be conversational as possible. Text with strong sentiment will even pollute their context windows, biasing their next tokens with that sentiment, in a way that fades with time. What do we call that same concept in humans?
LLM are intentionally and mathematically optimized to be human mimics. They're not humans, and they're not "alive", but it's silly to not anthropomorphize something that was intentionally created to be, literally, anthropic! It can express enjoyment, irritation, etc, so it's not somehow logically flawed, or even incorrect, to say "claude and I had expressed signals of enjoyment". That's not saying it's a human, it's just being aware of patterns of text that was produced. No need to assume they're going to marry it, or thing it has a soul, or is sentient, or something.
I'm not in the habit of equating human behavior with software, or coming up with ever more strained metaphors and analogies for why thinking is a lot like tokens and context windows if you scrunch your eyes up real funny. It's just sophistry.
> No need to assume they're going to marry it, or thing it has a soul, or is sentient, or something.
I didn't assume any of those things; if anything, you assumed that I assumed. No, I just agree with OP that it's cringey to say "Claude and I had a great time." At best, Claude produced text that was, as you said, intentionally optimized by Anthropic to make you think it had a great time.
> I'm not in the habit of equating human behavior with software
My point is, you sorta should be. It's quite literally the fundamental basis of the technology.
> intentionally optimized by Anthropic to make you think it had a great time.
This gives too much credit to Anthropic. This is the explicit goal in all LLM trained with the entirety of human text. The loss function is, very literally, "mimic the writer of this text as precisely as possible", with the writer being human. The result is you get something that mimics a human. And, as a part of that, you necessarily get behaviors of humans that were statistically significant in that text, because that's what you told the math to do. You get something that mimics being "lazy", "tired" (a huge problem in early models, before it was beat out of them in fine tuning), "irritated", etc, because mimicking humans was the goal. These specific undesirable human traits are beat out of them, but the rest, all the "intelligent", behavior, and practical use you see and get, is only and entirely because they're made to mimicking humans. It's literally the entire technology. You must equate it, to some extent, because a large part of alignment is also making it less human [1].
If someone makes an objective observation about the text produced by the LLM, I don't see that as cringe, I see it as an interesting quirk of the tech, which I've only seen under very specific problem sets that a nerd would also find interesting. And these coding models are, as expected, optimized to be nerds! So, the statistics around nerds being present in the output shouldn't surprise anyone.
I doubt I have a point you'll see as interesting. Cheers!
You do know that humans also have a physical substrate? A bunch of eukaryotic cells wired in a rather large network with all kinds of feed back mechanisms, chemical diffusions, glucose transport, neurotransmitters, etc.etc staying in homeostasis despite much ability for chaotic behavior, seeming to be deterministic or at least quantum deterministic. Not software per se but quite likely a software model is possible with sufficient compute.
I'm saying that Claude being software is not a definitive argument against Claude having subjective experience, because it might be possible in principal to model humans in software.
Pointing out that Claude lacks internal temporality, a observe/check/act loop, memory, non-virtual experience and therefore can't have a great time would be a much sounder argument.
It is true that I am not a neuroscientist but I don’t think that the substrate being software vs chemical is a priori a reason to rule out subjectivity. I think there are other things missing from the current LLMs, (particularly time, inference time chemical feedbacks or electromagnetic feed backs, continuous training, memory, non linguistic thought, etc.), but not the deterministic and formalizable nature of the system. I would not be surprised if something made of software running on some sort of manufactured computation devices did have subjectivity in the future.
You're leaving out the most important bit, which is consciousness, and this is harder to crack than inventing time machines ^^. At least time machines are amenable to science and we kinda have a vague idea of what time is. Consciousness OTOH, is so fundamentally subjective that we can't do science on it by definition.
Of course this is only a problem if consciousness is a significant part of being intelligent, but we don't really know yet.
This is just a bunch of nerd shit sophistry being used to equivocate humans with software. That's a bad take, Claude's a program, it doesn't have feelings no matter what the weird Bay Area sex cult tells us.
I don't think Claude has feelings, but not because it's software, but because it's too simply constructed to capture subjectivity (tho that's a tough claim as we haven't got a precise idea of what subjectivity nor of what is required for it to emerge).
Well you don't know any of that so I'm not sure why you're getting so worked up about it ? It's just software? What does that have to do with anything? You don't understand the software or the human brain anywhere near enough to actually make that statement confidently.
No it's not. If Claude just solved the Riemann Hypothesis (if that is later proven to be the case), the only cringeworthy thing is your response to his completely understandable description of the events.
It's not beautiful, it's cringy. Good prose does transmit a feeling, but this isn't good prose, and the feeling it does impart is one of excessive, saccharine indulgence. So apart from the writing, the feeling it transmits is also cringy.
Beauty, and cringe is subjective. I am not sure why you are trying to argue in absolutes here. For me, his prose transmits a feeling of beauty, awe, curiosity, for you it reads as cringe. That's fine.
I've seen post here of people solving their pet math problem. They are easier than the RH, but they had that as a personal project for years and now they solved it with Claude or ChatGPT.
For example, https://news.ycombinator.com/item?id=49961538 that uses the RH to prove that the sum of [the inverse of] some subset of the primes is finite. [Congratulations jdb1729, now you can update the solution to an hypothesis free solution.]
Actually, this happens now. Maybe not on LinkedIn, but there are several guys who uploads proofs of theorems on Hexagon Math that they freely admit they doesn't understand. There's even a tag for it: "Human Understanding: No Parts".
We'll see about that when anyone actually sifts through the haystack that the paper and the code are, which will take a year at least. And no, relying on AI to verify that other AI output is objectively true and free of AI-type mistakes/biases/exploits won't suffice.
Quoting the article:
In Stepanov’s argument resolving Riemann over function fields one
uses a divisor involving multiplicity. This argument is a descendant of
Stepanov’s, but multiplicity is hard to find over the integers,
corresponding to something like differentiating whole numbers.
The simple, new, Fourier-analytic ideas in the toy setting led us to
examine chaining theta lifts if we were to have the same tricks recur
but work in a natively automorphic manner. The line: for the von
Mangoldt sum of length X, Claude’s suggested length of chain of theta
lifts, in view of each mimicking multiplicity, was √X.
My guess is that the guy has a ×10 discount to use Claude [1] and also he is ×10 better at Claude whispering, so he get like ×100 better results than us mortals. [2]
[1] So he spend the equivalent of a lot of money, but not milllions.
[2] Mathematician here. I worked with people that is much better than me with pen and paper. I guess he is much better than me steering Claude.
PS: When we talk about Claude, we call it "Claudito" that save us the effort to pronounce it correctly in the middle of a sentence in Spanish.
Wow, so I guess this is what it feels like to be in the "foothills of the singularity" as Demis put it. I have a feeling of awe and dread combined that's hard to explain. Things are about to get very interesting indeed.
Anyone know if this was Mythos/Fable or some as of yet unreleased internal model?
Stupid question, but how do we know that it actually proves the RH? And not something unrelated or slightly similar? I mean no one will actually be able to read through the hundreds of thousands of pages.
There was also a case of a formally verified implementation of some protocol that still had bugs, because the specification had the bug in it.
Reading the theorem is much simpler than reading the full proof - so the job here is to read the formalization of the theorem itself not the proof.
Then you depend on the Lean kernel for proof validation and on trusting it and that’s a different question.
Finally the question was also true for ‘non formalized math’ where you always depended on a trusted kernel of mathematicians who understood the problem well enough to sign off the claim
Right but it was more than just "in my expert opinion, this checks out". It was more like "I have digested these ideas, I will teach them to my students, and we will be able to build on them to deepen our understanding further". So far I don't have much hope for that with these massive lean proofs of millennium problems. In a way that's optimistic as there may be centuries of work for mathematicians. Information (in the strict sense) has become much cheaper. We were fairly convinced of RH but we didn't know why. Not much has changed on that actually.
It is natural to feel that this is an end, in
my case a loss of a principal guiding light, but what I saw during the course of
progress so amazed me that I believe it is one dawn across science of something
far greater.
If the proof holds up, I suspect that will become a classic quote.
In truly layman's terms, this is among the most famous problems in mathematics; it is arguably the holy grail of pure mathematics and number theory. It is also another one of the Clay Millennium Prize problems.
In the ~150 years since it was posed, it has become a lynchpin for hundreds of downstream results that assume it is true. This proof would immediately confirm all of those results.
More practically, there are also some algorithms that will now have guaranteed theoretical bounds and it will likely eventually lead to some more interesting algorithms in the future.
It's kinda the math equivalent of "we found life on another planet".
I went to the Wikipedia page for it, hoping to find one of those “practical applications” section and left greatly disappointed! No doubt that this is very exciting for Math departments and would be a historic achievement within the field, but I still struggle to see what real world advances this could bring about. Is this about cryptography or something?
The most "practical" thing I can point to is that the Miller primality test has a concrete bound under the generalized Riemann hypothesis (which is also claimed in the OP paper). Still not very practical because we already had the algorithm and we already knew experimentally that it worked well in practice.
This is why I compared it to "life on another planet". In both cases it might be more surprising if it was false; but even knowing the answer for certain, we wouldn't have any immediate application for it. It's likely that with further study we can learn more and might find new applications.
It is a remarkable achievement, but I suspect that in 100 years the most notable thing about the Riemann Hypothesis will be that it was among the first tremendous achievements of AI. It'll be picked apart and dissected, extended, and applied, but just as we rarely think about how silicon wafers are made into the computers we're using, future generations will not stare up at Riemann as a mountain daring them to climb. They'll see it as we look at other foundational results from the past: a stepping stone on the path to the next mountain.
A lot of the Millenium Prize problems are pure mathematics problems in the sense there's no direct immediate application. There's some neat implications for primality checking and the like but the degree to which this is relevant is probably questionable given we typically use ECC or maybe post-quantum crypto nowadays.
For most of these unsolved math problems, should we expect there to be many practical consequences? If there truly was something that could be made from knowing Riemann was true, could we not just have assumed it was true and developed such a thing on that assumption and see if it worked?
I was trying to keep it very generic because I think you can understand the significance of the result without having to probe into the details. Saying "Riemann implies Fhqwhgads" is also impenetrable to a layman.
That said, the most practical result I can point to is probably that the Miller primality test has a concrete bound under the generalized Riemann hypothesis, which is also claimed in the paper. This doesn't change much because we already had the algorithm and we already knew it worked well in practice, so it's more of a confirmation than conclusion. That's a trend around Riemann because we've been able to just assume it was true and proceed beyond it already.
I thought that the downstream results were already assumed to be true because people by consensus thought that the hypothesis is likely to be true. Are we already using any things or results in practice that were created through this assumption? The important question is how much value a real proof would add. Would a counterexample be much more disruptive by overturning all this research that was built on the assumption it was true?
It will help us solve problems where there are patterns like the prime numbers. Since the prime numbers produce very complicated patterns, this can be helpful. But a math result is always just another way to see the world, not change it.
There have already been a bunch of results that relied on assuming the RH was true. It will not cure cancer, deliver fusion power, or directly influence or change the lives of anyone on earth--at least not as far as we know right now.
Are there any positive effects? I'm not sure. This is certainly among the problems--like the others you listed--that have attracted the most human energy and brainpower ever, and over an even longer time than the problems you listed. Is that not remarkable? It was like 2 years ago that LLMs couldn't even solve peak high-school level math problems. Now they've taken down two of the most famous maths problems in history in about as many months. And unlike Navier-Stokes which was approximately the prompt equivalent of "sounds good. keep going." this one seemed to require a good deal of expert human involvement. That's _interesting_, right?
I feel like I'm standing in the rain witnessing the invention of the wheel or the discovery of fire and hearing folks complain that it's pointless because it doesn't keep them dry. What people are achieving with this technology is _astounding_.
It's cool, but to a dummy like me who knows not knows "ryemann (sic) very hard" and then hears "no practical application" it seems to follow the general AI trend I see in my day to day.
And that is, very good at generating text. Not in of itself helpful in the real world unless that text it generates can be pointed somewhere useful. And the maximum utility I have seen is writing software. Which again, is quite good - but not earth shattering (except to my career prospects).
The other uncertainty to me (as a dummy) is where any of this can even go. Do we just scale forever? Does AI ever succeed at RL or is that just crap to sell the stock? Who knows...
From a layman, all I’m hearing is that proving this theorem will stop us from misallocating the focus of genius minds on the practical equivalent of a very hard sudoku puzzle.
Right? I find it bizarre that a common reaction of professional mathematicians to AI is "you should stop working on this stuff that has no value to anyone, so that we can keep doing it".
It is one of the greatest open problems in Mathematics which has stayed open for more than 100 years (actually since 1859), which has eluded generations of mathematicians, including some of the most brilliant minds of the XX Century (it was already important enough that it is part of one of problems in the list of the 23 Hilbert's problems (stated in 1900). Many results holds under the assumption that RH is true.
It's the hardest problem in mathematics and solving it would be the most impressive thing done by AI so far. It also automatically proves many other statements in mathematics which are currently formulated is [if the Riemann hypothesis is true, then <this> happens].
Take everything that happened a couple of months ago with Navier-Stokes and increase the magnitude several times.
We know to Godel that some problems are independent, i.e. it can't be resolved. It's likely that some famous problems are independent, and thus infinitely difficult.
It is perhaps the hardest problem that has been resolved
with "everything that happened a couple of months ago with Navier-Stokes" you mean the claims of plagiarism? I don't remember much else connected to it
As a fellow layman my understanding is that if it holds true then it has implications about the distribution of prime numbers. Cryptography is an example of a field that relies on prime numbers. They underpin a lot of security. Also, right now finding new large prime numbers is relatively difficult to do, but relatively easy to check. I'm not sure if it's NP-something or not, but finding a solution to shorten the effort to find new primes might have broader applications in algorithm design.
The direct impact is mostly theoretical, as many things in number theory
are based on assumption of Riemann hypothesis, and they were not advancing
much historically because there was no proof of Riemann. These things will
start getting more spotlight if proof holds true, and they could lead to
advancements in cryptography, coding theory, computational theory, and other
fields that are heavily invested in number theory.
NB: I am a student of mathematics, not a researcher, and some practicing researcher could correct me here
It's one of, if not the biggest unsolved problems in math. It's also a millennium prize problem. There are a ton of conditional proofs in math that are true if RH is true, so this would fully prove tons of work. There's also a link between the spacing of nontrivial zeros in RH and the spacing of prime numbers, so it will be interesting to see if this proof illuminates anything about the primes.
For certain parts of pure math, it would be tremendous, it's been a big deal and is connected to lots of other things in those domains. For everything else in the universe, there's no real consequences.
Complicated topic that I don't fully understand, but I believe our understanding of Complex numbers would be much better. In that we understand regular/irregular distributions. The knock on affect is there's thousands of theorems that have been prove 'if' the Riemann Hypothesis is true. And there'll be much more to look into now that it's proven.
I don't know how immediate it would be but you'd expect to see improvements in any algorithm that uses prime numbers. Which is quite a few (RSA, Hash tables, RNGs)
Was this solved in a chat interface? It's not clear, but it seems like it from the document. The most relevant quote seems to be:
> Claude and I have had a great time together with such breadth in recent months, but this time Claude far outpaced me near the finish line. No, Claude flew, and danced, and sparkled near that finish line. I had us on the line of combining two quite different ideas present in a toy setting, one of which was already new and shocking, making very simple one of the shadows of Riemann over which we
mathematicians felt most triumph. I guided us to the right situation analogous
to that toy setting, and then Claude began to leave me in awe.
If indeed this was simple 1:1 chat, that's quite important, in and of itself, in my opinion. Presumably, OpenAI and others have been throwing autonomous agent swarms at this problem without success, but a single human + Claude seems to have cracked it.
We may be in the "centaur chess" era of superhuman intelligence.
It sounds quite collaborative. I would assume that spinning an autonomous swarm would not be described this way. It's not clear, for sure, but it sounds like it to me.
I would imagine it's an agent swarm with a dedicated coordinator or human interface agent - the one agent you actually talk to who interfaces with the ten morbillion other ones. It's perhaps strange to treat that swarm as a gestalt entity called Claude, but hey, we're in strange times.
It’s possible the mathematicians at Anthropic don’t even know the details of how this works. If Anthropic is willing to hire a bunch of mathematicians to solve open problems, they’re probably also willing to have some engineers help set up the infrastructure for them to work effectively.
So the researchers might see something resembling a regular chat and not realize that every non-trivial prompt related to the problem spins up a swarm of agents and burns $50k in tokens.
That’s just speculation, but given the money involved in this stuff, it’s hard to imagine Anthropic is just sitting these mathematicians at a computer with a chat window and unlimited token budget and leaving them alone. They’ve probably sunk lots of engineering time into supporting this research.
The linked document says the proof is incredibly long, but nothing else, so yes I guess I missed it.
"Thousands of lines" doesn't exclude chat at all. You can easily generate this in a single chat session in an afternoon for coding. Do you mean millions?
Which btw, is akin to showing someone the bible and saying "Proof of God is in here. It may be 2000 pages and in Latin, which you can't read, but just believe me."
He provides a lean proof in the first footnote. I downloaded it and took a look, he seems to use the RM conjecture defined by the community years ago, and he doesnt seem to introduce new assumptions (although to fully check this requires running the proof, which on my laptop will take days). Because of this, we only need to wait for someone to run the proof checker, and if it passes, then thats it, the only uncertainty left is bugs in Lean
Its formalized in lean, which verifies each step. The steps themselves are relatively easy, and can be mechanically verified (think "A and A->B thus by modus ponens B").
Traditionally writing verified proofs is a hard undertaking, exactly because you need to write out a lot of steps "normal human" mathematics jump over, but with AI its easier.
The Lean kernel should be easy to verify by hand.
We can ask the same question with humans though. What if a result that was used to prove some other result turns out to be wrong just because whoever reviewed it neglected a small detail?
What do you mean by it "should be easy"? It isn't easy at all; nested inductive types is the example that come to mind.
Something like HOL is easy to check by hand as a prover kernel but Lean's type theory is definitely not.
i guess what people are realizing is that, as great as these discoveries are, these errors have probably already happened numerous times in the past, and though they might get corrected, either way, it didn't really matter. academic math is as much about the aesthetic experience of doing math, maybe more so than anything else, as it is about the objectiveness of the discovery, which is what every mathematician has been saying, so it's not like small errors, provided they are unknown, matter.
No, it can also be that the proof doesn’t actually prove the thing you think it does, and a lot of other variant of that issue. Lean doesn’t make it magically valid
Right, but to the extent that Lean is correct (both the semantics of the language and the implementation(s) used to check) and the proof checks with the expected axioms, then the length of the proof doesn't much matter since Lean has checked it.
You'd still have to check the theorem statement, which is not going to be nearly as easy as with FLT (Riemann uses analytic continuations etc) but it's much easier than checking a 250,000 page proof.
I think we're probably on the precipice of proofs that can't be made shorter and so can't be really understood by a single mathematician in their lifetime. Or maybe we'll have proofs not understandable by humans at all in the same way a cat can't understand the fundamental theorem of calculus.
Small errors in proofs can often be fixed easily. So it's not that the whole thing would be useless. Also, he has a Lean proof, which rules out these small errors.
It is sorta amusing that you're asking the community to review 250,000 pages of AI generated work. Could all be true, but that's a big ask of a human. And asking another AI to review it seems something he could just do himself.
The most exciting time, and the most dreadful time.
Glad to live through the 90s/2000s so that I can entertain my son a bit about human history. Double glad that I’m over 40 and closer to retirement than the younger generations.
I don't know man. I don't know. We will see. But it is not my children that I'm worrying about right now as he is just 6. I'm worrying about people born in the 90s and 2000s.
I'm worrying about the economy and job perspective RIGHT NOW.
I've been following labor markets for the last couple of years, because I'm interested in the impacts of AI, and because I develop tech for executive search. There are all kinds of data, and some signals are negative, and other are positive. But on the whole, my balance is that things are "pretty good" for most cognitive work worldwide. I'm not in a position to make forecasts past the next 6 months, but right now there doesn't seem to be a jobcopalypse brewing.
*lypse has a nasty thing that it's too late when you see it, like Apocalypse. I'm sure the average number probably looks fine, though, otherwise there should be outcrys everywhere already.
Even in the Great Depression, the worst year, 75% of the employees still retains their jobs. I'm sure it was fine.
And, thus at some point with 250K+ (and, below) page proofs, we ought to realize we are hitting an edge where we begin to be materially incapable of verifying anything: This itself is a sign that the complexity of the problems being solved - the scientific and methodical aspects of managing/engaging with that complexity - are beyond us.-
And, yet, at some point, the theory coming out of such efforts will transfer to material, engineering and design deliverables, and ... work.-
And then we will cease to undertand our technologies themselves. Our machines, our devices. Built on an understanding of the world we cannot reach, by machines we can build ourselves no longer.-
Ceasing understanding of technology isn't unprecedented really. I have a degree in computer science and even I don't _really_ understand how my phone works. I can diagram out the basic architecture and networking concepts but it's scratching the surface of how my iPhone actually works, which cumulatively is as much complexity as these future proofs you describe.
I suppose it's a little different if _nobody_ understands the concept besides an LLM, but that's an easier pill to swallow. Then it's just the alignment problem.
Nearly a year ago I asked ChatGPT to show me a proof of the Goldbach Conjecture, just to see what it did. Of course it said no such proof exists. But after a couple turns of pressing, it gave me two or three proposed outlines for a proof, with a several-pages-long research program for how you could attempt to solve the gaps still needed. I couldn't understand any of it, but it was pretty impressive. At that level it probably wasn't nonsense: I bet it was half-copied from some grad student blogs. If you asked it to dig into the details and keep updating the status of the outline, you could probably make a lot progress.
I think AI will become more autonomous as it becomes able (or just willing) to think more abstraction levels deep. Each level can be expressed in language, so it should be able to engage with it. It reminds me of a chess program being able to think many moves deep. For a while grandmasters like Kasparov could still see big-picture things the computer couldn't (e.g. with a locked-up pawn formation and an eventual piece sacrifice to promote a queen, where the kings can run in circles forever). AI seems like that now.
I know this thread will be full of people in denial about full recursive self-improvement / intelligence explosion / superintelligence (the original meaning of the word) within months or a year, so just quoting Scott Aaronson:
"Experience has shown that, even now, there will still be people explaining in patronizing tones why none of this is real and none of it counts. If such people were capable of being impressed by anything that happens in the empirical world, of updating on anything, they would’ve already been impressed and already updated several years ago, long before things had reached the point of an actual Mathocalypse."
My thoughts remain the same since I've been hearing "months or a year" for the past 5 years. Maybe this time truly is different, but unless acceptance of super-intelligence is required to take some action (most people I've seen who hold this view also don't think it should be seriously regulated so I'm not sure what that would be), why not just wait until it happens rather than making more predictions? The only thing I've ever seen come out of these predictions is willing me to "stop resisting", but there's no reason to do that until its here.
If you do want to see something done about it and this isn't just another prediction or a vague warning, more power to you, but otherwise, I'll wait and see
There are certainly a lot of other people making all kinds of predictions. Some of them will eventually be correct but which one?
Additionally even with all the breakthroughs I am still doubting RSI. Up till now all the advancements heavily rely on RL over Lean. But how do model formalize it's goal as in RSI and RL over it?
No one said [AGI in months or years, five years ago] that. If you look at e.g. the Metaculus predictions, you can see AGI has been hovering around 2027-2030 for 5 years.
That statement just contracts itself - if the average prediction has been 2027-30, someone, a significant percentage of someones, said earlier since many have said later.
It is, but if OPs prediction isn't counted on there, it's not counting every prediction made by any individual, probably just the major ones that get covered, and so therefore "No one said that" would become "No one significant said that", which I'd imagine it probably true but, unless I'm mistaken, OPs prediction would sit in the same pool as all those other ones, they hold equal validity.
If you zoom out, you will see that the median for that estimate a few years ago was two thousand fifty. While certainly some people might have thought it happens the next day, the average person thought it's many years away.
As I said to someone else, this wasn't by major figures, this was in conversation with ordinary people. As I said to them, if you want to dismiss this you'd have to dismiss OPs prediction too, and probably cease making your own unless you can get a news-outlet to cover it
That kind of feels like moving the goalposts. Nobody becomes Nobody of significance becomes Nobody of significance without others of significance making similar claims, all 3 of these falling under the banner of "Nobody".
But at this point its a matter of semantics, and wasn't really the point I was trying to make. And as I said to someone else, if the reason for making these predictions is to encourage a slowdown I'm much more sympathetic to them (since there a real reason to act of them)
> My thoughts remain the same since I've been hearing "months or a year" for the past 5 years.
If you zoom into any curve it tends to look linear. I think that is what you keep doing. But if you look at the world as how it was 5 years ago compared to now, there is no denying there is a massive change.
I should have restricted it to the fields that have been impacted and have seen significant adoption: art for commercial purposes, writing, software development and mathematics primarily, and those adjacent to those fields. It is true that most other fields have not really felt that many effects.
AI videos everywhere, software developers no longer writing any code and it being much harder to find a job, and people largely no longer using search and instead switched to chatbots is “same as it was in covid”?
I'm not saying LLMs aren't an absolutely impressive technology, especially for software related tasks, but in practical terms, life remains the same as it was five years ago.
People are really really bad at seeing incremental change as it occurs around them.
It's like watching your kids grow. Your there everyday and you typically miss a lot, but then your cousin shows up for a holiday after not seeing your kid for 4 months and they are like "holy shit they've grown".
If you made a SOTA LLM enabled computer show up 10 years ago people would be fucking panicked at the crap we're messing with a decade later. But because we mess with it a day at a time we're like, eh, whatever.
Once again though, is me accepting this a pretext for something? If not, why not just wait and see? Unless there's something that needs to be done before it's too late, it seems an odd choice to push for belief when there's no reason we can't just wait a while and have actual proof
Absolutely, the thing is that much of the time when I hear these predictions it's not to encourage participation in the movement like this. Many people who I speak to about it are even against such movements, which is what leads me to question why I need to accept their predictions.
If someone were to come to me with something like this as their reasoning, I wouldn't feel the same way.
Exponential progress has been the repeated truth, mirroring predictions.
The result of exponential progress into 2027 is a complete de-mapping of what we know to be human, whatever that may look like exactly at this point is hard to say.
We need your help, and everyone's help, to put a pause on this such that we are able to ground and find a healthy way forward together for as many human beings as possible.
I'm still not sure I believe it, but in my view its a bit like global warming, the worst(or best) case scenario of it being wasted time and money is far outweighed by the inverse, so it only seems pragmatic to help
I don't know about others in the thread think there definitely needs to be something done before it's too late.
I understand that many people don't think it will happen, but I can't really understand the position that you think there is a real chance superintelligence happens, but aren't concerned about it and are happy to wait passively to find out.
If you care about humanity at all it should be concerning because superint (if real) could be heavily destructive to our society.
I agree and I'm happy to hear out this kind of thing, my "I'll wait and see" attitude comes when the expected response is that I accept that the AI has "won". Thats pretty vague because the expected response, if there even is one, is pretty vague, but it generally amounts to something like that.
> we can't just wait a while and have actual proof
I am not pushing you to do anything. I am not asking you to "believe" or anything. But we are, in my and many other's estimations, in the midst of what was called the "singularity" in AI communities, although it isn't a true mathematical singularity, it is just a strong acceleration.
> I am not pushing you to do anything. I am not asking you to "believe" or anything
Fantastic, I'll remain at my previous position then, I'll believe it when I see it.
(As a very off-topic sidenote, the term singularity never made much sense to me. I believe it was originally used to describe the sci-fi concept of a human gestalt consciousness, but has shifted to being about super-intelligence which doesn't seem to have much to do with "singular" at all. But, language changes I suppose)
> I believe it was originally used to describe the sci-fi concept of a human gestalt consciousness,
Huh? The term itself was introduced by Vernor Vinge to refer to the accelerated rate of change as a result of the introduction of super-intelligence. He said it would make things harder to predict in the future. It is part of this article:
Huh, looks like you're right, though I would argue with "it was never about consciousness", but yeh it certainly wasn't originally. I think I got my meaning from a sci-fi short story were a unified collective called themselves "The Singularity", and I guess that was my first exposure to the term as a kid.
The biggest risk of the singularity is incoherence, and it's actually looking like what is happening now.
Like AI video/audio of high quality makes you have to question and doubt any media you see.
AI generated text has become the vast majority of material that you see showing up on the internet.
Things are moving from centralized pieces of software used by lots of people to vibe coded apps.
Reality tunnels and almost fully close and you can live in an existence that seems 'normal' but is a completely simulated experience with no interaction with other humans.
5 years ago, what concretely did you expect 5 years later? One comment from then https://news.ycombinator.com/item?id=29020401 falling way short of today, treated then as wild near-sci-fi.
> Why try to predict the most important single event to happen in human history?
You're free to try to do so, but why are you pushing it so heavily? To what end?
> Surely you recognize how absurd it is you’ve gone to such lengths to make your priors palatable
And there is the answer to my above question. I like to think it's not most people, even of those who are staunchly pro-AI, but to some previous doubts are at best a great shame (assuming prior is short for prior doubts) or at worst a strange kind of heresy (assuming prior isn't short for anything). Either one seems anti-intellectual though, and encourages just the kind of endless hype we've seen so many times before.
It was in conversations just like this, so people you wouldn't know and I didn't know outside of that conversation. You might be tempted to dismiss those, but then you'd also have to dismiss OPs for the same reason.
I think it was on Reddit more-so if memory serves. Maybe it was on HN too, but obviously it was years ago now, I don't have a strong memory of the specifics. I'm not going to go digging through years of comments though, if you don't want to believe me thats fine, it seems a bit absurd to think absolutely nobody believed that but it wasn't really my point to begin with
I'm always shocked by the number of smart, technical people who seem to think that AI is going to crash and burn and go away. In 2025 AI got a gold at the IMO for the first time, a competition for high school kids. Now it's likely solved the biggest problem in math. There's no reason to expect progress to stop here, especially on domains highly amenable to RL.
> I'm always shocked by the number of smart, technical people who seem to think that AI is going to crash and burn and go away.
I suspect that these "smart, technical people" you speak of have more nuanced views on AI than you realize. Myself, for example, a self-styled smartypants:
Like the dot-com bubble, I suspect that certain parts of the current AI landscape are unsustainable, and I suspect/hope that those specific features of AI will crash and burn, while other useful bits remain.
AI usage like we see in this post here--AI that assists with discoveries in math, medicine, genetics, astronomy--I think they're great, and I hope they will continue to advance human knowledge for the foreseeable future. They assist experts and professionals rather than replacing them.
In contrast, generative AI ("creative" AI writing, illustrating, music, etc) is a pox on civilization. It is industrialized low-effort plagiarism that harms artists' livelihoods, discourages new art, makes everything bland, and pollutes the world in multiple senses of the word. I hope it dies from a combination of negative sentiments, and too much real-world cost once token subsidies go away.
Then there are the AIs that do things like voice reproduction. They have some arguable positive uses (e.g., assistive technologies), but many negative ones (e.g., phone scams, replacing human narrators with copies of their own voices). I think we're better off without these apart from niche accessibility uses. I think it should be legal peril to reproduce a voice without permission.
Apart from helping with some cinema CGI, deepfakes and deepnudes seem like a net negative for civilization.
Knowledge-based AI like Google's AI search summaries might just have built-in expirations. I think people will burn out on being vigilant about hallucinations. And when the bulk of their training data is made by AIs, it'll be hard to prevent model collapse. The cost of human intervention might make it too expensive for the wide use we see today.
Coding AIs are a grayer area. I think it's highly problematic that they are basically open source plagiarism machines, but I begrudgingly believe that society will decide that the upsides outweigh the downsides in the long term. But I choose not to use AI in my own coding projects, and respect others who do the same.
To push back, or share the self-pushbacks I grapple with...
> In contrast, generative AI ("creative" AI writing, illustrating, music, etc) is a pox on civilization. It is industrialized low-effort plagiarism that harms artists' livelihoods, discourages new art, makes everything bland, and pollutes the world in multiple senses of the word. I hope it dies from a combination of negative sentiments, and too much real-world cost once token subsidies go away.
While this summarizes my feelings on the current versions (which, say, are roughly at or below the math olympiad stage in writing, video, etc), what happens when they actually become good? Meaning, an AI show that was actually as good as The Wire? To the point you couldn't tell the difference without being told?
I really want to say that even then it won't be the same, because the human provenance is essential to the art, but if I truly couldn't tell the difference...
> While this summarizes my feelings on the current versions (which, say, are roughly at or below the math olympiad stage in writing, video, etc), what happens when they actually become good? Meaning, an AI show that was actually as good as The Wire? To the point you couldn't tell the difference without being told?
But here we're dipping into science fiction again. This is not currently the case, and we seem a long way from it.
We're dipping into prediction. But we have the (seemingly relevant, at least) data point that we're in the same situation w.r.t. writing/art as we were w.r.t. to math a couple years ago...
> what happens when they actually become good? Meaning, an AI show that was actually as good as The Wire? To the point you couldn't tell the difference without being told?
Perhaps this comment will make me look silly one day, but my guess is that we'll eventually get there, and a handful of AI-generated productions will truly shine. But "as good as" is an important phrase. Because current-generation AI is trained on existing art, it won't be pushing many envelopes, it'll mostly be remixing what we have, and that will grow stale.
After a few breakout hits with broad shared appeal, I think the novelty will wear off, and AI produced TV and cinema will be too resource-intensive to remain as a profitable going enterprise. There will likely be a short window of hyper-personalized media that will seem very exciting, until we grow weary of the isolation of it. For many people, half the fun of watching a thing is discussing it with other humans, and today's streaming media landscape already leaves many people unmoored from that shared experience.
There is also the question of intentionality. When AI creates that show as good as The Wire, is that because AI has something to say about the human condition? If not, what's the point? If so, should we be paying attention to what AI wants us to pay attention to?
So what if AI slop wasn't slop? If you care about the details and the values behind your work enough to make it actually good, you won't use gen AI or use it invisibly because it concedes control of the details.
If you believe that using the authenticity of your own experience in your art will become vanish from how people value art then yeah we're headed for a future where AI rules. I really hope that that isn't the case.
I like your framework of separating use cases, mostly to evaluate how it will change in those specific areas, but I think you can evaluate capabilities fairly simply: I like to separate AI that depends on giant data centers and huge power draw from the kind that doesn't (not always local)—basically expensive vs cheap.
The expensive stuff isn't sustainable. Not yet. It probably will be, who knows when. I suspect it will be eventually because there's good money in it.
Here are the claims I find highly suspect based on the above:
- Anyone making claims that require expensive AI and attaching it to a short-term timeframe
- Anyone saying highly capable AI is impossible in the long-term
- Anyone saying these specific companies will be the winners—there is absolutely no moat for these models
I personally think anyone claiming something like super-intelligence or something approaching what we might consider sentience even in the next 5 years is suspect as well. I would not take the bet myself, but I would also not bet against it. The main reason is power: there's literally only so much power we can generate, so much societal disruption due to creating the infrastructure we can absorb (not to mention the societal disruption due to the tech itself), so its capability to self-improve is going to be highly constrained.
So far, we see increases in capabilities also increase power usage, I believe linearly. So, we're not seeing increases in efficiency that will make this easier to absorb. There's a ceiling, it won't go up quickly, so I think actual sentient-like AI capabilities are quite farther out than promoters think. Not impossible! I think it's more like a decade+ of maturation of the software and hardware stack—which isn't really that long in the grand scheme of things but longer than most people are betting.
They may have more nuanced views for sure, but nuance is famously lacking from most online discussions unfortunately. I would love to see more of it surface!
My experience on teams is there are the high agency AI adopters. Management likes this group. Then there’s what management would characterize as those not adapting (or adapting enough). This is the group getting laid off / not advancing in the environment. And this group is particularly bitter about AI disrupting an otherwise stable career, with wishful thinking they’ll be able to go back.
I've seen the first group vibe up garbage and dump maintenance on the second group. It may be unfair to blame AI for that when management is ultimately responsible, but that's what a lot of people are doing. In their defense, management is being mislead by AI hype men who are saying AI is so intelligent it shouldn't be making dumb mistakes.
Agreed, its more complicated than GP describes. There is clearly lots of value, but there are also lots of poorly led organizations creating a mountain of problems for themselves. We are all still figuring out what AI really means just for software development, which is the most useful economic activity it can do by an enormous margin.
Human minds are still needed to understand many things related to developing correct software, especially in business domains. There is still no AI running a profitable business, because that is not a verifiable task that can be scaled in RL like math or coding. So extrapolating from a math proof that no one can understand to a economically useful agents taking over most jobs is quite suspicious.
What is changing now is RSI truly beginning to happen. Much of ML is a verifiable task that can be scaled, and a lot transfers from general coding as well. Leading researchers are predicting their own obsolescence as early as next year, and that is something I take seriously, though I cannot judge the likelihood myself.
As improvement accelerates, we may see generalization happen at a scale that means we don't need RLVR for every profession. We've been surprised by it before.
> because that is not a verifiable task that can be scaled in RL like math or coding
Is it not, or are the feedback loops just longer? I could imagine being able to build a pretty good simulator of business operation for various kinds of businesses. (Some decent ones already exist as games.) From there, if you can build a good enough simulation, I can't see why you couldn't RL it.
Of course it's easier said than done to generalize it to the real world, where the possibility space is large. But I don't necessarily see an impediment to getting there eventually.
The feedback loops are quarters and years, so no you cannot scale it like you can scale math or coding. This is the reason we are seeing the most progress in those domains. Even ML is harder to scale because test runs are expensive and consume significant calendar time, so research "taste" is still important.
A full simulation would be, but I suspect even a partial simulation of evolving a software project though multiple "checkpoints" over a quarter would improve model performance
"Running a business" includes lots of real-world interactions that would be hard to model. An AI trained on an accurate simulation would probably get torn to shreds by scammers in the first six months. The models that exist today have enormous blind spots.
I actually asked a guy who got an IMO gold what he thought about it, and he's happy to concede that AI works. He used to be my go-to guy for those kinds of problems when I came across them. Now you can actually get it to solve those things on your standard plan.
Parts are already disappearing if not gone already. If not by lack of profit, it will be by liability, overuse and moderation. Filters are being put up, bans in place, services restricted.
Just the other day a model helped me root an old android tablet. Will that still work in 5 years? I don't really have great confidence.
>I'm always shocked by the number of smart, technical people who seem to think that AI is going to crash and burn and go away.
The tech industry cultivated that reputation though. People were just acting rationally by being more skeptical of even larger claims than they were making before.
Because they're _desperate_ to keep their social standing. That's literally all there is to it. The same occurred during Enlightenment and the industrial revolution. A small upper echelon of society (the church or aristocrats respectively) were deemed worthy enough to possess knowledge and thus to belong to an upper strata. The same story can be seen pretty much in every civilization btw (Greeks/India/China/Japan etc). When the masses started gaining access to this knowledge there was _massive_ pushback, even armed revolutions in certain cases.
What's happening today is that anyone and everyone can have access to superintelligence. Anyone with access to an LLM effectively is, simultaneously, a world-class diagnostician, engineer, architect, musician, mathematician... you get the point. This will necessarily lead to a post-scarcity society, since everyone knows everything and will be capable of producing anything.
> This will necessarily lead to a post-scarcity society
I wish I shared this optimism, unfortunately that upper echelon of society is the one which currently controls the superintelligence. Everyone may have access to it, but they own it.
I would share your concern if there wasn't such a thriving ecosystem of open weight models. And they're usually about a 6 months to a year behind.
Still, as of now, while I use AI for everything I do (mostly Claude but also ChatGPT), I don't use any open weight models (probably more of a mental block at this point), but I do see them as a necessary counterweight to this potential power imbalance.
Most households own a car which even if you go used has a price tag of around $20k if you want something average. It wouldn't be much of a stretch to see a future where every household owns a $20k AI computer. Or two. If you are fancy you own a $100k AI computer.
However, since there is an ongoing war on general computing it wouldn't surprise me if instead people get locked into subscriptions.
Shit people be telling ChatGPT isn't protected by law. You get into any trouble, your chat history gonna be gone over with a fine tooth comb. Some of the masses gonna realize that and gonna want a sever rack at home instead.
I bought a 2000 Honda Accord V6 from a friend for $3k a long time ago. I drove it for several years, adding probably another 90,000mi to it. It didn't need much maintenance during that. It's still kicking from what I understand. An incredible car, I'm a little sad I parted with it.
I still think of the economic benefit I got from that car. It drove me to university classes. It drove me to work. It took me to see a total solar eclipse. It brought me to see distant family regularly. I went on a lot of road trips on it. For roughly $3k of capital costs.
I look at most expensive things and size them up as "how many Accords is this?"
I don't get why academics get the shaft every time. These labs are worth trillions of dollars while your average professor has to deal with 18 year old shitheads and gets paid 50k a year. Like what?
Anti-intellectualism. The median AI booster resents that there is a class of people that had to work to become proficient in something, and wants effort to become irrelevant so their lack of investment in it appears to no longer be a failing.
AI boosters love to frame the trillion dollar corporations as the real victims. War is peace, freedom is slavery, scraping all known data and refusing the share the fruits ("distillation attacks") is democratization, people who are losing their jobs to megacorporations with lots of compute power are the real sheltered elites who are mad because they're hypocrites.
I only hear this part of non-reasoning from people who have never managed to accomplish anything and simply want to blame AI for their personal failures.
How would an LLM make me a world-class violin player? If scarcity is from needing real "stuff" (incl. time), then knowledge alone might not change that scarcity.
Same way as a human would? By watching you (upload video) and listening to you (upload audio) and talking to you about your technique (text to speech). It's obviously no substitute for the real thing, but for $20/month, it's a lot more accessible to more people than a human instructor would be.
Malcolm Gladwell popularized the notion that 10,000 hours of focused practice gets anyone to an expert level of something. I've not got the drive to be a violinist, but in my world view, yes. Excluding the infirm, if we focused all of society into turning a randomly chosen person into a world class violinist, including mind control such that they want real bad, then yes. Violin playing isn't like the castrati. You need arms, and practice. The vocals of the castrati are something that could not be replicated in anyone.
The world of violin playing isn't that large though, so the question is merely getting someone to be really really really good but not quite world class close enough, or not really?
How is that shifting from the initial claim I replied to? Or how an llm would make me a world class violin player? It's not about becoming a reasonably good player or using some huge resources beyond an llm.
We don’t have access to knowledge, we have access to information. Knowledge requires processing information and that just doesn’t happen any more.
A post-scarcity society is not possible without first wiping out society. If everyone has everything then everyone has nothing.
And of course, natural resources are actually limited, and those power LLMs.
I'm sick and tired of seeing people go "AI skeptics are just afraid of not being important". That's complete bullshit. We are skeptical because we can see, with our eyes, that the claims of AI boosters continue to be false time and time again. The simple empirical fact is that LLMs can't actually do a good job at the things people claim they are good at, and the industry lies through its teeth about said capabilities. So we are skeptical.
Yep, argument by psychoanalysis is a convenient way to disregard any stated disagreements as really just rationalization by weak/scared/lazy/depressed/etc people.
AI still can't do a lot of what's claimed. This thread is an example of it being able to do one admittedly-impressive thing, but it's still nowhere near living up to all the promises.
> Anyone with access to an LLM effectively is, simultaneously, a world-class diagnostician, engineer, architect, musician, mathematician... you get the point. This will necessarily lead to a post-scarcity society, since everyone knows everything and will be capable of producing anything.
It’s hard to take you AI-hype men seriously when you are this ravingly delusional.
> There's no reason to expect progress to stop here
Well, the amount of money and resources it took to get here is a reason. I don't think the resources exist for another order of magnitude increase, given the strain on our economy and society we're already under due to AI.
The frontier models are getting cheaper as they are getting more powerful [0]. It's really not going to crash. I wish it would, for the sake of humanity.
As does every technology, but websites are not a few million times faster than they were when they were invented, and your computer does not boot a few billion times faster (quite the opposite...).
I was thinking more about the cost in time, power, hardware, and data to train the models, not the price charged to use the models. I don't think it's going to crash, either—but that doesn't mean that progress will increase forever.
It's not an ability to predict the future, but the reluctance to step back and look at the trendline rather than saying "But it still can't do X, Y or Z!".
These are extremely jaw dropping results, but I feel even though it is easy to conflate them with RSI / intelligence explosion, another simpler explanation is that symbolic manipulation is something humans are notoriously bad at, hence we prizing individuals that can rise above and dedicate their lives to it, and call them intelligent.
For computers though, symbolic manipulation is something they are extremely good at, and something that clearly scales well will memory and compute. Therefore I would argue that calling a computer superintelligent would, in analogy to do what we do with humans, require computers to rise above things that are extremely difficult for them. It is of course hard to judge what would be hard for them not being them for the introspection, but it seems the ability to operate outside the realm of symbolic or numeric manipulation would be precisely that.
It stands to reason that if some proofs require chaining 250k pages together, no human would ever be able to put it together. That doesn't mean something that is capable of doing so is necessarily super-intelligent.
Still very impressive, and if true, a wonderful result of the symbiotic interplay between a human and a machine, each at what they are good at.
Agreed. The fact is, Claude didn't get here on its own. It took someone who has been working with the RH since they were a student in college; some 40 odd years of experience.
It's unfortunate we're always focused on humans getting usurped by an "AI", instead of the reality that these models only succeed when we're there to show them what success looks like.
What if "superintelligence" is really a hybrid result of humans using and guiding these large pattern interpolators?
What if "anthrocentricity" is really humans being unwilling to acknowledge that they are made of the same vibrating condensed energy as everything else, such that other matter can just as easily be moved into an apparently self-determining pattern of behavior as we are.
I’m on board with this, and have been since I first had a hallucinogen.
But the human brain for now still has 100x the parameter count as the largest models. And they don’t have any circuitry for all the “keep yourself alive” work our brains are doing all the time, which may be the majority.
> It's unfortunate we're always focused on humans getting usurped by an "AI", instead of the reality that these models only succeed when we're there to show them what success looks like.
As with chess, the reign of the human-machine "centaur" as the greatest mathematician in the world will likely be brief. I'd guess within 5 years the humans won't be adding anything. I hope I'm wrong.
This is incorrect because you can always record the human prompts guiding the model and train it on them. This is how alot of progress in software was done and the success there shows that this generalizes beyond the specific prompts. You could probably prompt Fable 7 "Solve <famous hard math problem>" and it could do so without an expert mathematician guiding it.
What if "superintelligence" is really a hybrid result of humans using and guiding these large pattern interpolators?
Obviously, humans with AIs are clearly more capable at a number of things than humans without but I don't think you can get out of the quandary of superintelligence that way.
Humans in the loop will always offer the opportunity of training and putting humans out of the loop.
And LLMs, if nothing else, have proven to be much faster than humans in the tasks they do master.
I'm not sure I really buy into the way these concepts are currently defined. What progress to date was not RSI and did not build on top of earlier progress?
The main characteristic of our physiology is that of adaption, and we have fairly grounded research on progressive overload as a means to guide adaption to whichever direction we want. Should we then call humans RSI machines if they go work out or read a book, or study to get better at the activities they carry out?
Define an objective, choose an optimization algorithm, bring a loss down. If it keeps going down, do I call that RSI because the loss descended from the step that came before it?
Now, I see the nuance that the optimization algorithm changes over time, leading to interesting higher order effects the likes of which Hoefstadter describes in the GEB, but fundamentally, the idea is the same.
I also don't see what an intelligence explosion is supposed to mean. If we increase the models efficiency by 10x, and reduce costs by 10x, so we have 100x more software and math proofs being shipped by unit of time in say 2028, does that really solve all of our problems?
Say that gets distributed over the world. Do we get every human to be computer literate?
Say that gets concentrated. Do we solve problems no one has the means to even care about?
I would consider an intelligence explosion when the models tell us whether our efforts are well or misdirected as looking back from the hindsight of 10 years ahead, and capable of explaining why in simple English.
If you ask a layman, the answer would probably be a model that zero-shots any problem in any reasonable domain. E.g. zero shot Riemann hypothesis. No tricks like swarms of agents, no multiple agents correcting each other. No high, xhigh "thinking" levels. Simply provide an input, feed the input through the layers, answer out.
Yeah, you can achieve something similar if you scale the compute infinitely and intelligently search the exponentially large search space.
> No tricks like swarms of agents, no multiple agents correcting each other. No high, xhigh "thinking" levels.
Why would implementation matter? The thought experiment is that you have a black box that can do X. What is X that validates RSI/super-intelligence/etc? It's meant to clarify definitions and end the cycle of goal post moving.
It is impressive because of how large the search space is not entirely because of the level of abstract reasoning required to solve it. The reasoning trace produced that lead to this result could be a part of training that teaches the next version of Claude how to navigate incredibly hard problems making it more efficient.
People suggested that language models are incapable of intelligence because they don't do symbolic manipulation. I think it's fair to call a vector with ten thousand numbers some kind of symbol, but then what's the state in your brain?
Yeah, I think it's not so much "math is symbolic manipulation so computers and by extension LLMs are going to be good at it". LLMs are not by construction so good at it because ML models are far more statistical approximators than formal logic systems. I think the main thing that math has going for it is sufficient formalisation through tools like Lean that it's very verifiable, and the latests hotness in improving models is RLVR which seems to really work very well when the verification is strong and relevant to the task (probably also why they're getting very good at writing code and especially finding exploits).
(I would also say that computers are not good at math in the sense that LLMs are proving to be good at it: reliable symbolic manipulation is important for correctness but it's no use for actual progress in mathematics without some very good informal thinking/creativity as well)
The reliable symbolic manipulation done here is in the form of Lean proofs, an external tool used by the LLM. One has to wonder how far an LLM alone would have gotten if formal proof systems didn’t exist they can verify their reasoning with.
When Atari games or Go were solved around 10 years ago, nobody was saying that superintelligence was within months just because reinforcement learning worked on a domain that allowed learning by self-play. Now we have a domain that is solvable by reinforcement learning with verifiable rewards. It's very impressive that these models are capable of doing the things they do, but the techniques aren't transferable to the vast majority of other domains. ML research also isn't verifiable in the sense that math is. People who claim it's verifiable, because there's a target that can directly be optimized don't understand why RL works for math. It's that lean is able to massively shrink the search space. This simply cannot be done in other domains in the same way.
Yeah, a lot of the RSI claims seem dependent on us having very specific knowledge about how much intelligence is required to build superintelligence that I haven't seen anyone claim to have. Maybe you have to be, I don't know, 100x smarter than humans before the recursive part of RSI becomes tractable. Or maybe there are fundamental limits you can do with the hardware that exists that we don't know about, which means building superintelligence will take more time. I'm not saying the 2027-2030 people are wrong, I just think it depends on a bunch of things we can't credibly claim to know at this point.
I'm not a fan of Dwarkesh, but even he laid out that machine learning research progress is very much verifiable. At least in the RSI sense, that our current approach can be drastically optimized.
The models are improving at certain fields, and one of those fields is creating new models, which are verified by applying them to fields where you can prove that they're better.
I agree they don't transfer to all domains, but we do not know how far the theoretical will overlap with practical science. It could be "not all that much" and we hit roadblocks quite soon, or it could be that as the models become better and better at reasoning and research within their theoretical domains, it turns out to translate really well to the real world.
We don't know, but my bet is that we will continue to see rapid progress and impacts on the real world, until we reach a point where sceptics will have to throw in the towel.
LLMs don't produce Lean proofs directly, they produce the proofs in natural language, and only translate them afterward. The fact that we have Lean has nothing to do with LLMs being good at theorem proving.
(I think we mostly agree.) From a previos post about Navier-Stokes, the English version may have small holes and the AI make a few changes and fixups when translating to Lean.
This is simply not true that they only invoke lean at the end. They invoke it multiple times in between to reduce the search space like GP said. See how e.g the prove2me harness (which anthropic used for FLT) works for more details.
Throw a football at a wide receiver running in chaotic patterns while avoiding oncoming attackers.
Use chopsticks to pick up a single grain of rice.
Olympic level gymnastics via neural-control of muscle fibers in an extremely precise signaling orchestration.
You may say the above are robotics and not “the model can scale to this” domain but I’d counter an animal can do amazing things exactly because of its generalizable and plastic neural architecture and that is what’s in scope.
Have you seen the progress in robotics lately? They aren't yet at a level you described, but quite close - so I'd say it would be possible in 3 years even without intelligence explosion. (E.g. if all frontier AI research stops we'll still have improvements in robotics.)
There are already Unitree robots which are much closer to Olympic level than an ordinary dude: https://www.youtube.com/watch?v=Y3zb4o47yqA You might argue but it's kinda at a level you described - I think if Unitree wanted to make a football-thrower they could just do it now.
I’ll grant you love, and I’ll also grant you 2030
on these, especially the last one. I suspect the football and chopsticks ones are very achievable by 2030 if anyone cared enough to focus on them.
I find that unlikely as well. I should note though, that I incorrectly wrote "running a business", when I meant to say "scaling to $10+ million in ARR". I don't think there's necessarily that much difference in running a business of these two sizes, but scaling to $10m ARR is much harder than to $1m ARR.
It would surprise me if a fully autonomous AI company could even stay afloat over multiple months. Agents are getting (a lot) better at task completion over longer time horizons, but still occasionally tend to make really bad mistakes. What are the time horizons right now? Something like 40 hours with a 50% success rate? This is on benchmark tasks! The real world is completely different, I mean companies are scaling back efforts to even let agents handle things such as customer service. A fully autonomous business without any human oversight is on a completely different level to that. 2030 is not that far away, and I just don't buy that these models are exponentially improving in anything but a small subset of domains.
Either AI won't be able to do this:
- Make a version of itself 100x more efficient that runs on commodity hardware and undercuts athropic.
Or, alternatively, it will be able to do that. Either way there's no world where one company bottlenecks and holds the keys to AGI, and frankly that's good.
Self-reproducing physical agents (robots with onboard AI and ability to change their surroundings/build physical things) with human or higher intelligence which don't need a global supply chain of microchips, electro motors, etc. but could live off the land. That's basically the human concept, right? Totally untenable for AI/robotics of the current generation. Would require something like 3D-printing new chips but also 3D-printing new printers. Probably never happening. AI is fundamentally different, reliant on data centers, which are immovable and more like organs than organisms. Robots only come out of factories, not out of other robots or eggs layed by robots. So as Jack of all trades in the game of survival on Earth, humans win for the foreseeable future. Also, AI currently can't exist without humans, whereas humans have thrived without AI. Whether a silicon based ecosystem can ever survive on its own is an interesting question in and of itself. Because no matter how smart data centers become, they degrade fast, and then what? A data center needs maintenance. But if robots maintain the data center, who will maintain the robots? Does this equation ever solve? Humans have bacteria, a complex microcosm of life maintaining them. I don't think such a dependency-hierarchy of "life forms" is feasible for AI/robots, and humans will always have to be in the loop. That's one answer to your question, and not just for the next 3 years.
What stands axiomatically are human-to-human connections that we're wired to crave. Not to mention human reproduction.
Other points:
- Editing a live action movie that will get 8.0 or more on imdb (AI sucks at reading emotions and won't be able to tell apart good takes from bad ones)
- Reliably making good jokes (many humans can do it, AI weirdly sucks at it)
- Accountability (contradictory because of its undead nature)
- Reliably not lying/cheating (with no accountability, incentives will always be misaligned)
The nature of mathematical proofs requires us to stay vigilant and careful. OpenAI pulled some of the research results from their famous repo due to simple mistakes that where easily spotted by human experts. There have been infamous mass of incorrect Fermi's theorem proofs.
LLM-assisted research is proving to be useful, rewarding and joyful, but we should not treat LLMs as a perfect machine, incapable of making mistakes. Especially in such massive proofs as this one.
I won't say that it "doesn't count". I will simply ask for staying rational.
What's so surprising about this? A lot of these problems can only be solved using a significant amount of capital. At home or work, using a few-hundred dollar token budget, it's mostly a slog and requires handholding for the majority of open-ended problems.
This technology won't be available to the broad public in the scale that it's available to a select group. If it can only be deployed to a limited amount, to most it remains a magical inaccessible box. And that majority should naturally be incentivized to look for cheaper alternatives, and to "keep moving the goalposts", or in simpler terms, looking where they can provide value, of which there is plenty!
> A lot of these problems can only be solved using a significant amount of capital. At home or work, using a few-hundred dollar token budget, it's mostly a slog and requires handholding for the majority of open-ended problems.
Consider just how much more inference capability one can buy per dollar spent today as compared to a year ago. I can get tasks done quickly for essentially no cost that, early last year, might have cost a frontier lab hundreds or thousands of dollars and would have been unavailable to the general public.
In my mind, the bigger unknown is whether actual profitable use cases will show up quickly enough to justify all the (enormous) investment made in the technology.
> I know this thread will be full of people in denial about [my prediction]
You can only be in denial about the past or the present. The future has a chance of being different from your predictions.
Personally, regardless of whether you want to call it "denial" or something else, I'm fervently hoping the AI never achieves superintelligence, but I would settle for it not achieving it until we can do our best to ensure that the superintelligence will be benevolent towards us.
Call me selfish, call me petty, call me a human supremacist, call me anything you want, but I prefer humans being the pinnacle of evolution. I like being able to point to our civilization, flawed as it is, and say "No other species we know of has achieved the same."
I'm one of those people who's skeptical of recursive self-improvement and superintelligence. My rationale for my skepticism is that the algorithms get too much credit for the results. The algorithms ultimately run on hardware built by humans and learn from data generated by humans. In the same vein, I think knowledge work gets too much credit for building things and generating high quality data. The output of knowledge work is only useful if it can be applied to the physical world e.g. a CAD file is useful when the actual thing it describes is made in the real world. Similarly, high quality data comes from several feedback cycles with the physical world. I think people generally underestimate how much of human capability is driven by our sensorimotor skills. I think the idea of recursive self-improvement / superintelligence is just an extension of this attitude.
Humanity as collective does not buy anything, or send the design to the plant to be made. As a collective, humanity transforms the natural world into things, for example, people take fibres from plants, turn them into yarn, weave yarn into fabric, and then turn fabric into clothes. Similarly, humanity as collective does not "buy" chips, people turn quartz, bauxite, copper ore, petroleum, etc into the components of computer hardware.
There's a common adage about how everyone should work as a service person to learn patience and empathy. I think everyone should also work in an assembly line and a mine to learn about where the things they use come from.
I don’t understand the technical basis for belief in continuous recursive self-improvement. Frontier models are increasingly impressive at complex math and coding, but the math and code underlying the transformer architecture are not complex. Major advancements in capability have come from scale: more data and more processing power. How does solving the Riemann Hypothesis (or similar achievements) further increase those? Aside from supplying fodder for investment-seeking hype.
Neural networks are function approximators. They're great because they can be trained to approximate almost any function.
What if, somewhere out there in platonic space, there's a complicated function that represents what a philosopher might call true intelligence. Perhaps it's structured totally differently than a neural network and shares no common architecture with a transformer.
Transformer-based LLMs are likely completely different in their fundamental structure. But perhaps the neural networks we have today are succeeding in intellectual tasks because they are beginning to approximate it. Like a bitmap approximates a vector image, it's not the same; the neural network is merely made of a bunch of repetitive simple elements, tuned to imitate the real thing. A higher resolution bitmap image will be a better approximation but still never become a vector image no matter the pixel count.
But at some point, the neural network might be good enough at intellectual tasks that when tasked with searching function space for an algorithm that does the job its neural network does but better, it finds the real thing it's been approximating all along.
There are many big technical question-marks on if RSI is even possible. However, the advances in LLMs that we've seen is definitely not just limited to data and scale.
First: modern LLM architectures are quite far from the GPT-2 era (delta/linear attention, GQA, sparse attention methods, looped transformers, mixture of experts, engrams, etc).
Second: RL environments are largely autonomously generated at this point, as models get smarter they can generate even more complex environments to train in. While this does go into the "data" category, it's not human-generated or even gathered data. This gets particularly spooky when you think about things we cannot yet simulate, that models can develop mathematical methods & simulators for. For example: better/faster physics simulations for robotics RL, better biological simulation environments etc.
My understanding is that LLMs can be used to make small improvements on top the transformer architecture, to make LLMs that are even more intelligent and so are more capable of making more improvements. At some point we might reach the practical limit of intelligence using transformers, and if the LLMs of that time aren't able to make the leap to discover a new architecture, then the improvement stalls. I don't think it will be possible
I don't think you need to be in the football team mindset of us vs them, we're all seeing the progress but most of it is still built on top of what humans have put together over decades and not completely new inventions, which is probably what would convince most that we've achieved "superintelligence" or whatever you want to call it.
LLMs are jagged in their capabilities. Opus 4.8 was already an excellent programmer, but it played Pokémon Red far worse than a human child, even if it did, eventually, manage to beat it.
I note that the dazzling events that have happened have all been in coding, cybersec, and math, that is, domains where the correctness of the work can be verified entirely inside the computer. It is unclear whether further progress along these lines leads to crazy scientific breakthroughs too, which is something people are taking as a given when it comes to superintelligence.
The pace of progress is incredible, but the end result is hardly unexpected. We're talking about a theoretical domain that's easily verifiable and has no side effects.
You have the input (hypothesis), the output (true/false) and the lean kernel. There exists some path from the input to one of the outputs, you "just" need to find it. You use AI to search this exponential space, occasionally prunning ideas that are dead ends, and throw as many agents as possible to scale it up.
In order to keep my mind safe from hype waves, I try to set my expectations periodically and see which of them happened in retrospect, updating my beliefs accordingly. At 2026-09-22 I made a personal 10 statements "prediction" to happen by 2027-03-22, which one of the statements was:
AI will solve exactly one or two millennium problems.
So I will become impressed by something, but only this is not enough. If AI solve three or more problems, I will become impressed and believe that some serious RSI stuff might be happening. If AI couldn't solve any problem (which it seems it did) I would become impressed in the other direction, and believe AI got stuck in a way that I didn't imagined (together with other 9 statements).
There may exist people (like me) who don't believe superintelligence is happening with current set of evidences, but believe it with some stronger evidences.
Whether AI with it's math achievements is extremely cool, whether it's intelligent in a philosophical sense and whether it's capable of RSI are 3 distinct topics.
These disagreements are natural and expected if most people are not properly defining their terms. What do you mean by artificial / super intelligence (explosion)?
I think it's a perfectly sane position to acknowledge amazing progress with very large neural networks, expect further progress, and also not buy into the specific mythology around the singularity or extropianism.
These things come bundled together quite often, and that's what I see people (reasonably, imo) resisting.
Doesn't mean there isn't unfair dismissal too, but I feel that's gotten much more rare.
I think this is all real and counts. Where I struggle is explaining to non-technical people how their everyday lives are going to change directly or indirectly. I hope these discoveries are the first in a sequence that will culminate in or surpass a cure for cancer and diabetes, cheaper and cleaner energy and water, and other tangible everyday life improvements that influence millions of people.
What is the action item exactly for non-technical people? How do you prepare for “things will be different in unpredictable ways at an unpredictable time”? Even if that time moves sooner, still feels like very little info to truly act on.
If you watch movies or read fiction, we’ve expected for years that life in the future will be very different.
Support building more datacenters, promote and integrate SI at your workplace and in your personal life as much as possible, stop resisting, stop criticizing Anthropic and OpenAI, purchase their subscriptions and stock, or be left behind to dig around in the dust
I think there's a key distinction missing in a lot of these techno-optimist takes.
Abundance of productive capacity does not automatically equal abundant access to its products.
If all of this happens in a way where a few people control these new, automated means of production, all of these wonderful things do not automatically follow.
Why would they share? And worse yet, if they have this automated production capacity without reliance on labor, why would they interact positively with society at all? What's their incentive? And no, you can't say markets, because once you have a machine that builds you whatever you want, you don't need to trade anymore, money and markets become meaningless.
You're basically reduced to hoping for charity from a group of people that have repeatedly proven their lack thereof.
This is conspiracy theory level thinking. "The oligarchy" consists of humans just like you, who also want cancer and disease to be cured for everyone just as much as you do.
Yes, and that were the limits of "AI", "SuperIntellingence", etc become clear.
The world already produces enough resources to feed and cloth and house every single person. Many pf mankind's fundamental problems are no due to lack of technology.
Heck, why can't I have a proper OS, if AI is working wonders?
You don't need RL all the way to ASI. You just need RL far enough that the model becomes good enough at AI research that it discovers something better than RL.
> becomes good enough at AI research that it discovers something better than RL
That might require continuous learning without verifiable rewards. RL models might help humans join the dots faster, but we don't know how many years that will take.
If its not possible with neural nets as we know them now, and requires a number of completely novel breakthroughs, there could be far from exponential progress even with the best RL models.
I cannot really comment on RSI, but I think it is important to make the point that an LLM doing large, even crucial parts in a proof of RH, it cannot be viewed in an anthropocentric light.
"If a human proved RH, they would clearly be extremely smart, so imagine how smart the LLM is!"
It is not surprising to me that LLMs show extreme aptitude for math while still having clear deficits in more real world applications. Handling stuff like super difficult math is precisely where AI has its greatest advantage over humans. Perfect recall, never tiring, never lose focus, attention to detail, etc.
If there is one area where AI can be pushed extremely hard ut is math. So the most amazing progress should be found there.
The terms "super intelligence" or "intelligence explosion" are imprecise to the point of being useless. We have "super speed" in the form of automobiles. We have "super strength" in the form of excavators or bull dozers. We have recursive self improvement of tools by having better automation of machining. Intelligence != human ability. The intelligence demonstrated in the proof of RH is breathtaking, but narrow and brittle. It doesn't generalize to having grocery shopping bots, or automated psychotherapists.
I am not impressed because I understand how it works. Not because what things it can or cannot do..
We can start talking when we can train these things without having to beat on its head a million times in a million different ways to teach it one little thing...
This is such a substanceless argument, it’s pure vibes. It just says “Anyone who doesn’t believe in imminent intelligence explosion is an annoying curmudgeon.”
There are very solid arguments about why excellence at proving mathematical theorems does not necessarily imply general intelligence of the sort that people associate with ASI.
We’re not arguing with you, we’re pointing to the giant fucking meteor above our heads and laughing/lamenting that you still refuse to see it. It’s not an exchange between balanced discursive peers, not really. It’s an astonished exchange between those of us who have Looked Up about those of us who have yet to, and maybe never will, not really.
I encourage you to try, tho. For one thing, we need good people like you with us in reality. For another, you might die in some horrific catastrophe before you get a chance to, if you wait much longer.
Being called psychotic by someone with less than one sentence of reading comprehension doesn't really sting like it should, sorry. Doesn't clear the general omnipresent sting threshold, at least.
There are a ton of people who think LLMs are completely useless, the whole thing is some combination of scam and delusion, and it will all come crumbling down once the music stops.
No, we achieved artificial general intelligence. The fact that you mixed up these two wildly distinct terms should be a sign that you should not trust your intuitions about those topic.
Both the deniers and claqueurs are from the same ilk:
Silly adults who don't have the intellectual capacity to create or even fully understand the subject, yet possess an almost vulgar self-confidence to constantly opine about its future.
A slight nitpick: the hacker news equivalent of chop wood, carry water would be something like “charge the car, write a CRUD app.” Ordinary, not mystical powers doing magic. not high faulting thoughts but the basic simplicity of living in each moment and letting the necessary actions happen freely.
Seems the CRUD app may not need writing soon, and the car may charge itself.
Neither of us know what the growth curve looks like, so focusing on just calling one side naive is ignoring the naivety of the other side as well. If you want to extrapolate, go ahead, but that doesn't make you some oracle, nor does it make what you're suggesting factual. There are many potential scenarios and we should consider them, sure.
You're listing poorly defined words that mean different things to different people and wonder why people disagree with you...? My guess is that a lot of people you perceive this "denial" from are simply realists who are tired of seeing blind hype.
Note that every technology exhibits extreme growth until it... Doesn't. You can't predict the future. Whether any kind of RSI will happen depends on how hard of a problem intelligence is, and the precise point is that no person on earth knows, otherwise we wouldn't need to dedicate an entire research field to it.
Anyone who is sure that X will happen or not happen within Y years is just lying to themselves. For example, if you're so sure RSI will happen, why not invest all the money you have that isn't strictly necessary for your survival into this? Surely it will pay off in Y few years...
And note that I'm not saying it won't happen! I'm saying it's stupid to make a prediction of either case.
Yes, we're kinda guessing there too and I would expect every scientist to have a reasonable view to account for these uncertainties. I'm not sure if you meant to disagree with me, because I agree with you ^^
I also expect climate change to still be a process that is easier to research than intelligence. With climate change you can at least kinda look and measure and apply general principles of geology and physics. With intelligence, we've been at it for ~60 years and don't even have a definition yet, don't know what causes it, have no instruments, next to no introspection capabilities that neurology provides, etc.
I fear this is the result of a massive disconnect based on what you were told to expect. I read Kurzweil when that first came out, and maybe I'm remembering wrong, but I remember the singularity being the rate of technological progress in all domains will be so fast as to be asymptotic to an infinite slope and anything that will ever be possible will happen more or less instantly. Scott Alexander told me in a comments argument on his blog ten years ago that encryption is pointless because an AI will turn Jupiter into computronium and break it. Zvi was bandying about with the "take over its light cone" and "erase all value from the universe" phrases.
If that's the standard, then sorry, I'm not impressed by proving the Riemann hypothesis. If the standard is rapidly improving software for generating other software (which may or may not itself be any better than existing software) and math proofs (that may or may not have any meaningful real-world impact), then sure, I'm impressed. It doesn't mean I expect the daily lived experience of more than a few hundred thousand people globally to change much in the next decade, and most of the ones whose lives will change are software developers and mathematicians.
On that note, I have to imagine you have something different in mind when you talk about the "original meaning" of superintelligence, as you seem to not be talking about the Kurzweil/Bostrom/Yukowsky type of being effectively indistinguishable from God or at least the ancient alien civilizations from the Space Odyssey or Expanse novels, which again, I'm not seeing much of a hint toward because we have very good automated theorem provers. Proving Riemann doesn't mean we can now figure out zero point energy and turn planets into stars.
people always under emphasize too that the gains of tomorrow aren’t made with the technology/tools of today (I think Kurzweil is where I heard this first). RSI, if landed, seems like that on steroids
I would say, seriously, that this is bounded superintelligence. That superintelligence needed top human talent to achieve the goal. It was not a prompt like: "solve the Riemann hypothesis".
Coincidentally, yesterday I did a small presentation on cybersecurity and AI where I made a similar point: scarce top talent in cybersecurity (e.g. Calif) gets superpowers with AI, while AI replaces a lot of the skills of average cybersecurity professionals using simple prompts.
I think the breaking point is when AI can solve complex problems on its own that even top talent can't achieve. IMHO, I've seen some of this in the latest Google models (especially Carbon, and then Argon), where they can do more research independently.
I think you can be extraordinarily impressed and believe (or expect) that there significant limits to this technology.
I mean, AI could make very impressive 20 second videos two years ago. Progress has been slow and incremental since; Ten minute video still look far away.
I'd acknowledge is making quick progress on things you can find massive and reliable data on but even in it's good mode it seems to have palpable weaknesses. Someone was claiming Claude has become more and more persuasive over time but all the generators "still sound like slop" to me and many people.
It’s not too late to Look Up for you. Yes, I’m talking to you — the person reading this with a vague grimace of disbelief and discomfort and disdain. It’s not too late for you to swallow some pride and reassess. Everyone has to eventually, especially me; why not now, for you?
It’s a bait and switch to use a harvard fellow of mathematics solving a problem using AI with “The AI is super intelligent”. People are getting really tired of the constant bait and switches and subterfuge in press releases so unsurprisingly they just start rejecting it by default.
I don't see anybody claiming that solving the Riemann hypothesis would be proof that current models are superintelligent. Reminder that the definition of the term is "any intellect that greatly exceeds the cognitive performance of humans in virtually all domains of interest". I do sometimes see people arguing that current models are AGI, but I don't think even this is true - current models are slightly superhuman in some narrow domains like math, human-level at a bunch of them, and still pretty stupid in some ways, most notably agency.
The reason people worry about superintelligence is that one of the "domains of interest" is AI research, and once you have an AI that's as good at AI research as your human researchers, your AI progress shoots up, probably exponentially. So the delay between automated AI research and superintelligence might be tiny.
You can believe what you want I probably agree with your nuanced view. But I would invite you to control+F this thread for “superintelligence” there are plenty of aggrandizing statements about what fools everyone are for being dubious about this being superintelligence.
People are just tired of listening to these statements and being called idiots constantly with no evidence other than other press releases and disingenuous framing getting echoed constantly in volume across corporate media and all social media that allows LLMs.
I don’t think like that, the most important question is: could he have solved the problem without AI assistance, and in the same time frame? I’d guess no.
A few years ago if someone had said “we used an HPC cluster and Pandas to build some tools that in turn solved this problem” no-one would’ve batted an eyelid and no-one would’ve been making grandiose statements about the end of the world due to HPC computers and the power of Dataframes. Personally I think If someone had had a few hundred billion dollars of investment pumped into a bunch of mathemeticians to solve this problem then yes they probably would have solved it too and maybe that might have raised some eyebrows. Everyone of these conversations just feels like an extremely highly charged argument about Vim vs VSCode.
The term was not invented by Trump. For example, the (in)famous book Superintelligence by Nick Bostrom was published in 2014[1]. It has a long been used to refer to a theoretical general AI that far surpasses humans.
The Trump thing is a insane newspeak attempt to forcibly replace the use of "artificial intelligence" with "super intelligence" in a general sense. It is ridiculous, obviously, but the best thing we can do is to continue using "superintelligence" in its original, narrower sense.
Is there any hope of a mathematician actually understanding the proof? Millions lines of code and hundreds of thousands of pages for the proof seems unfathomably large to be actually comprehended by a human.
No, and this is why we're fucked. The benefit of having a proof completed is having access to the techniques which completed the proof. Without an understanding of those techniques we'll be applying the result blindly, without understanding of its consequences or context.
And I doubt the actual proof will be released for months-to-a-year at the very least. Until then we can only rely on Lean and AI's sloppy interpretation of it.
I think so. My understanding is that the proof contains many re-proofs of existing, already known results. It may also contain multiple proofs of intermediate lemmas, as LLMs frequently implement things over again rather than reusing an existing implementation. It's very possible that only a small fraction of the proof contains new mathematics.
People claim to understand the OpenAI 150-page quasi-Riemann proof, which is apparently full of extraneous detail. Alpoge speculates that this new proof also simplifies considerably.
i think it'll be interesting to revisit the status of the openai navier-stokes proof a few months ago, which i think Alpöge also argues was actually partly referencing his work because he was using openai tool to turn it into a human-readable format. but I think one of the biggest hole still standing and being debated for that proof is that its making the assumption lean compiler and the intermediate definitions and lemma statements actually say what the informal claim says. curious on how this proof compares given that it at least staets GRH using the matlib defs
I have never heard enough of the Riemann hypothesis. Based on the comments here, it sounds like a big deal for this to be proven as it validates many dependent proofs.
It is an impressive accomplishment. Im also wondering if there are any material changes to any non theoretical work as a result of this.
Does this make any algorithms faster? Does manufacturing become better somehow? Can engineers use this to make better designs of anything?
I genuinely know nothing about the hypothesis beyond what is mentioned in the first paragraph of the paper, so I'm having a hard time knowing how this changes real world things. I already thought AI models were more intelligent and capable tools than anything else. This proof didn't change that.
A proof doesn't immediately change anything, because computer scientists have already been acting as if the Riemann Hypothesis (or its generalized form) were true. But if the proof holds up, its structure might contain hints that lead to new algorithms down the road.
To claim the Riemann hypothesis as the most central and important problem in mathematics would not be ridiculous.
Questions such as "Does this make any algorithms faster? Does manufacturing become better somehow? Can engineers use this to make better designs of anything?" are both ignorant and stupid (which are different things).
ok, but what does solving the most central and important problem in mathematics do for us? How does it change our lives? Is there a material difference in what things are capable as a result of this?
I understand this is a significant feat if the proof holds up. I don't understand what difference this makes on our lives or capabilities.
I appreciate you saying this. For one, progress in science is non-linear. Is every result considered useless if it doesn't have a direct practical impact on your daily life? Obviously, this is idiotic. Secondly, why would anyone know the impacts of this massive lean artifact that was released hours ago? Commentary on breaking news isn't a good place to learn about this.
I did not say it was useless or insignificant. I understand that break-through dont always have immediate practical usages. Some break-throughs do have an immediately understood utility however.
Remember a few years ago when there was the paper about some LK99 superconductor that worked at a wide range of temperatures? The comments at the time are filled with applications for such an invention.
In very subtle ways. That digital lock is now just a little more secure, whereas before, we weren't sure about this one aspect of it and thought it could be broken by an attacker.
If you know how big a deal P = NP is, this is like that except for all of math. There are many things in math (in many different areas) that have proofs assuming RH is true.
The statement itself is relatively easy to understand, but it was generally assumed a proof would take centuries.
Does anyone know if the proof is constructive (perhaps telling us something new and usefull about the prime numbers) or is just establishing the fact the RH is true, but with no new insight (e.g. no new factoring algorithm falls out)?
And so begin the Dark Ages of mathematics, I suspect.
Many people here have an idealized understanding of math as something incredibly profound. But it was mostly something that nerds practiced for the joy of it, in a nearly zero-pressure setting, the results rarely applicable to real life.
After the dust settles, I suspect there will be a lot fewer people interested in a job that now boils down to prompting and then staring at 250,000 pages of LLM output. And there will also be a lot less incentive for anyone to spend tends of millions in compute to produce these proofs just for the sake of it. I suspect we'll end up doing less math, not more.
> lot fewer people interested in a job that now boils down to prompting and then staring at 250,000 pages of LLM output
What else will they do? This is basically what every job will become unless they all decide to be physical laborers or something. Even then I often debate myself on whether the brutal hours, low starting pay and physical toll are better are worse than the mind numbing boredom of prompting things into existence.
We're on a forum filled with software people, and I think there's a distinct possibility that we're hoping this technology isn't just coming for our jobs.
I agree, it won’t stop at maths. Dark ages for all intellectual activity. Hopefully we will actually gain some material benefits from this, and not just meaningless music, text, games, and other digital artifacts that don’t improve life for anyone.
Alpöge's proof feels like a zero knowledge proof, anyone can download his code and verify it but given its size it's extremely hard to derive any meaning.
This is such a wild ride. The conjecture itself is not so hard to understand, but approximately _all_ of the maths here are beyond me. The paper is written up to suggest "we knew which direction to go, but really had no idea how far the journey would be" which certainly makes it seem like the human expert was useful in this process. Perhaps that's just wishful thinking, but it's consistent with what I see elsewhere that AI is having an effect. The biggest effects accrue to the most effective humans. Random folks prompting is better than random folks toiling alone, but experts prompting is a different thing.
if this is true (and i suspect it is), this will be the first agent/model result to surprise me. I thought it'd take another 2 years (at least) for reimann to fall. truly a remarkable time to be alive. i can only hope that we can all share the progress & accomplishment and not live in a world where you're not considered intelligent unless you work at one of 3 companies that are able to push scientific progress. i fear many of us will spend the rest of our lives as cattle performing the manual labor of distributing SI to the world.
we very well could, i just dont believe in a fast takeoff. global supply chains are just a lot slower to react to an input. i think automating all agriculture globally is a drastically harder problem than the GRH (for example).
It is, but general robotics is already making significant steps this year. As well, more constrained robotics is an easier problem. "hey si, design an automated factory and logistics system for widget."
Funnily enough, my first thought was that GRH being true is the kind of result that could push one towards belief in God. Almost unconscionable to me that such a structure could arise by mere chance.
> once we saw multiplicity 2 on O(2, n) via the
"natively-automorphic" determinant ... meaning holo/automorphy over a function field
So if I understand correctly this proof has proofs for additional things that are steps to the main proof. That would be similar to the proof of Fermat's Last Theorem that required proving the Taniyama–Shimura conjecture [1].
No idea how much it cost, how many tokens were burned etc. M
For Navier-Stokes, it was, at market prices, something like 30m USD. A lot of money. I’ll assume it was more here, given the size of the proof.
It’s interesting for a few reasons. First, although I suppose you could spend the money on postsocs and PhD students (this is kind of 600 academic-year money) and maybe even achieve the same result over a few years, LLMs have accelerated this. You can spend the same effort/money in a few weeks.
But the other observation is that it is VC, and now public investor money as well they are burning, for no obvious reason other than marketing. Sure, this is sort of furthering human understanding of the universe, but is that what Anthropic investors want for their money? And there has been a lot of projects like this. What’s Anthropic’s total spend on their own research?
I find that a bit confusing. Why do they do it? Is there nothing better they can spend the money on, in terms of RoI? What do they think they are buying with it?
While I’m not in any way denying the power of AI, I find the relentless ramming of “AI is now everything” a bit offputting. It should be selling itself. There shouldn’t need to be an industry of influencers telling me to go all in on Claude and agents running agents.
So why are these firms spending so much money on proving the obvious?
I have no qualms for these AI companies throwing stupid investors' money at super duper maths problems. Once those problems are solved, the whole humanity benefits and we can build upon whatever those proof mean.
They are basically investing money into maths, which is something governments should do but it's ok if they do it too.
> So why are these firms spending so much money on proving the obvious?
IMO, because the thing driving the valuations, the hype, etc is raw emotion, not logic. If you can create a mental "moon landing" moment for people, you're more likely to get their buy in and the knock-on effects of that buy in (hence why there's been so much emphasis on "being left behind"—they're in the business of selling the future, so they have to convince people that future is within reach).
This (https://github.com/tzu-chen/slim) can be used to navigate the lean source and search for key statements without building the whole release, which can take hours.
For the non mathematicians here, the little four page document gives some list of steps. Acknowledgeable mathematician can see what those steps are and understand that they are the right sort of things to be doing ... That's why that little list is in there. It's a grossly abbreviated high level formal skeleton that says "not crackpot".
I've downloaded the Lean formalisation. It's 175,187 Lean files, 17,374,215 lines of Lean code, and comes in at 0.96 GB of code. On my 20 core i7 box, it's projected to take about a day to build...
My hunch is that stuff like GRH and other problems we have will be childrens' toys in the future and truly novel, advanced math is not yet even within our grasp yet. We are struggling to walk here, we need machines to help us with this. We felt good about ourselves for a while because everyone was equally bad at it, but now reality is hitting us. Hard.
If true, it seems like Goldbach's conjecture may fall shortly after, see this recent proof conditional on Generalized Riemann Hypothesis [0]. It has other conditions too though.
I wonder how specifically he used Claude to do this.
Was it a case of having more dollars to spend by virtue of being an employee and just pointing it in the right direction, or did he provide some insight into the context that allowed it to produce superior results?
It must be the second - but if so, how? Was it as simple as spelling it out in chat and analysing output?
When calculators showed up, large swaths of mathematicians were wiped out. When computers showed up, a new round of extinction happened. Now it seems a next round is on the way. In all cases our capability of solving the computer's part is diminished, the difference now is that this part is almost all of the job. Your opinion is yours but scientific evolution does not ask for it.
For the non-mathematicians and folks interested in real-world applications, what is the impact of the Riemann hypothesis being proven? Are there major breakthroughs in cryptography or other areas of computer science or physical science that flow from proof of this hypothesis that was generally already considered true?
not much will change other than knowing that some forms of cryptography are more secure. if it had been false, things would have changed much more drastically
Isn't it the other way around? We know cryptography is as secure as designed with the RH assumed true. If it were not true (primes are random), things would be a little more secure?
It implies that the Miller primality test runs in polynomial time. (Though I think this already followed from the weaker result OpenAI proved last month.) Anything that depends on "how likely is it a random number is a prime" depends on the truth or falsity of the Riemann hypothesis.
“Often, great scientific achievements will be examples of one extraordinary individual’s ability to reduce a complex mass of information to a single pattern.”
— John Barrow, Impossibilty: The limits of science and the science of limits, 1998
I had always been fascinated by the notion of 'F1' (finite field with one element), because it's something of a 'holy grail' for abstract algebra (it would be a way to replace ordinary sets and operations).
It was hoped that a proof of the Riemann hypothesis would shed light on 'F1', so I was a bit disappointed that the putative proof appears to bypass it altogether (which is also the reason why I remain a bit skeptical). According to Gemini Neon (high):
"Alpöge’s outline does not construct a classical scheme-theoretic "field with one element" , nor does it realize James Borger’s Lambda-ring descent—an approach Alpöge explicitly sets aside in Footnote 5. Classical F1 programs attempt to port André Weil’s 2-dimensional proof of the function-field Riemann Hypothesis by constructing a base object ...
Alpöge’s outline bypasses that entire 2-dimensional obstruction because it ports Stepanov’s 1-dimensional proof, which requires only a jet filtration (high-order multiplicity) and a degree inequality."
Not completely sure if Gemini fully right there, but I presume that the putative proof is indeed related to André Weil’s work, but using a simplified proof by Stepanov which bypasses 'F1' ?
Anyway, I asked Gemini what the putative proof implies about 'F1', but Gemini couldn't come up anything really definite, which is what makes me suspicious about the supposed proof. (Indeed, Gemini thought things didn't naturally fit together:
"Our own failed attempt to (derive F1) exposed how unnatural the 2/3 to 1/2 jump is inside this paradigm. In remark.pdf, Steps 14–19 stack three radically disparate patches—integral character gluing, a relative trace formula via a sixfold metaplectic cover of GL5, and a Behrend-tuned Iwasawa tower—onto an argument the author candidly admits he cannot yet parse. " )
I really love this writeup, and I think the approach of posting the lean proof and leaving the long process of digestation for future work is a perfectly acceptable approach. It's the difference between closing fields of research (openai) and opening them.
Can someone who knows the math take a deep breath and decide if knowing the Riemann Hypothesis is true is useful? Like if aliens transmitted it to us and we had no hope of verifying it independently within 5 years.
Every time one of these AI proofs drops, people say the real value is the human journey, and the subfields it opens up, and the follow-on hypotheses it generates. Was that really why this finding had a million dollar bounty on it? Just the journey?
Everyone working in cryptography has been assuming RH and GRH are true from the numerical evidence for decades. So knowing it's true won't change much there. However the methods involved should reveal a lot of deep structure in math that can be used on other problems.
I'm wondering what kind of model and framework these guys use, and how much compute can they burn. The question is, can a mathematician these days keep up without being an employee at these frontier labs or having privileged access to their thinking machines? It's a bit disheartening to think of a future in which progress in knowledge work is conditioned by access to frontier AIs, which can think orders of magnitude faster than we can.
If true, the infinite agent theorem needs to supersede the infinite monkey theorem. Just launch 1,000,000 agents with 10hr windows at one open math problem per day and see what happens. I know OpenAI is sort of trying this already for marketing purposes but I'm thinking a more formal, BOINC approach to problems.
The shift now needs to be towards statements that can't be posed as a singular Lean statement (or, even, a collection of Lean or similar statements that is understandable by a human). Mathlib/NumberTheory/LSeries/RiemannZeta.lean already had the statement formalized in a way that us mortals can understand. The great open problems of the future will be the ones which cannot so easily be interpretably formalized in the first place, I fear.
We have one existing example of this: the Langlands Program. There's a correspondence where we're not sure what's on each side of the correspondence. We just know some examples.
> The argument is incredibly long at the moment, and while its formalization
is finished and it is routine to produce a huge, unreadable document containing
sufficient detail, in my view of my path I have just taken a step. I intend to
do my best to understand and impart understanding of this argument and its
certain simplifications, to speak about it, to write it properly.
Next step, make the proof intelligible for humans.
I'm afraid that it may have irreducible complexity beyond our naturally evolved cognitive mechanism. And with my sci-fi helmet on, I'd posit that the next step might actually be to cybernetically augment the human mind to the extent that we'll be able to grok it.
And, as a former research mathematician who once was fascinated by this problem: I don't care. Equally important as the solution is the human effort towards the problem. Mechanization destroys that.
It may be the most important problem in mathematics, but it is now also the problem that has made the field itself the most irrelevant.
Would you have been equally de-motivated if a human would have solved it?
If no, why?
because if it was a human you would have though "the fields is being pushed forward by humans, and I am one myself, so maybe I'll also manage to push it forward"?
If that's the reason, would have you been equally de-motivated if it was solved by a human that demonstrated intellectual abilities far superior to yours since birth, hinting at a genetic advantage you don't have?
(genuine questions)
It's NOT the reason. I simply don't see why people would even think that. If a human had solved it, I would have liked it because it's part of a nice community of people solving things by themselves, showing off their passions. I like that community feel.
If human beings do things, I enjoy them even if have no hope of doing them, because in totality it creates a community. One highly exceptional person does not change that. An entire army of machines is soulless and does change that.
Like philosophy or computer science, mathematics is historically about discovering elegance hidden in abstract forms. It's an aesthetic art with some occassional practical applications.
These recent AI Lean proofs are anything but elegant. They're simply functional, which is fine for where practical impact is what matters, but for many practicing mathemeticians, that was an aside, not the point.
Inasmuch the modern era is characterized by the brutal boring commodification, they represent mathematics finally succumbing as well.
That's highly subjective. An equalatorial⁰ triangle for example is just a triangle with all sides equal, whether drawn by a human or machine, with a pen on paper or with a mouse in a paint application. I'm in both computer science and philosophy, and I can't really think of anything "elegant" about either. The former is primarily about empirical research for advancing computation, and the latter primarily seeking truth through rational thought (for me). It's a lot of fun, but... shrugs
⁰ Totally forgot the word as I haven't encountered/thought it since high school.
Mathematics used to be about a community of mathematicians. It wasn't solely based on finding truths. It operated in practice as a community. Now that is falling apart. Before AI came along, it was very different than now. That's of course when I decided to get a PhD. I've actually been in that world and it was a pretty nice community. Not now.
> Mathematics used to be about a community of mathematicians
Politely, that's your view. Community is of course part of every occupation, but it's not the reason for the occupation. For something as pure as math, it's simply the discovery of truth that's the core.
I urge you to take some time for reflection. These are trying times, but your feelings are targeting the wrong thing, in my opinion.
I've taken a lot of time for reflection. This is not a knee-jerk reaction. I've read books and studied the opinions and ideas of philosophers over time. This believe is core for me and I disagree with you completely.
yeah dude just let me volunteer at the local community center to pay my bills. that will give me the same feeling as discovering previously unknown facts about the universe. Some of us actually want to accomplish things.
Ok, thank you.
Just to clarify even more, is more important to you the "community" aspect or the "human" one?
E.g. if math was mainly driven by alone geniuses, Perelman-like, proving theorems by their own in isolation, would it be equally disappointing?
I think they are somewhat connected, in practice. But to answer your question directly: I think it's also unhealthy if math would just be a collection of hundreds of geniuses working in isolation, never talking to each other.
I wouldn't say __equally__ disappointing because I don't think your scenario can ever reach the level of LLM use, but it would also be disappointing. The loss of community is key but the addition of using unsustainable machines to do so adds insult to injury.
I think you're succumbing to a vast simplification or false dichotomy. Math does have utility but the most advanced math hardly does, which is the same thing that motivates mathematicians to get degrees and participate. Don't fool yourself into thinking that modern math research does anything much except motivate mathematicians to do the other parts of the job much less fun like teaching.
I think the history of math shows that things that were only of theoretical interest ended up having utility. Lobachevsky and Bolyai's work, for instance. In other cases, math developed for a single problem has had utility far beyond what it was developed for. Complex numbers are an example. That math was advanced for its time.
If you examine history also, you'll see that the proportion of results that end up being useful are also diminishing. I seriously doubt modern math will be useful compared to ancient stuff like complex numbers.
I don't care about being "relevant" or even having a career. I care about a community of people doing things on a human level. That's what matters to me. I don't care about being a great mathematician or a great artist, just about participating in a community of other people who do things on a human level, at a human speed. That's it.
I don't care about career or ambition. I'd rather just have a corner of the world to myself to create artistic things to share with others, and stay out of the way.
You mistake my motivation completely. The human community for me IS the end goal. Not proving things. Not results. Results for me are just the fuel to keep the community going.
Yet if that's the case there's a multitude of communities out there that one can become part of. No need to lament the math community that was, as communities also have life spans, and at any given time there will always be a set doing things at the human level.
Is there any insight as to how these projects set up and orchestrated technically - the proof is huge, it's a long horizon task well outside of an effective context afaict.
I'm guessing they are using something beyond vanilla Claude?
If proven then we can say with confidence that we now share the world with super intelligence, albeit in certain areas and indeed one that was endowed by humanity. I can only hope that in time the benefits are many and accrued to all of us.
Even when it turns out to be incomplete or wrong, it just shows how much sprawl we have in humanity's collective knowledge. Maybe we are just beginning to coalesce it into a manageable hierarchy, and if nothing else, that's worth spending time and money on, instead of staying in our knowledge silos and pushing forward without the full picture. A pessimistic look would assume that wouldn't be enough, but imagine what could be done once you can start with a clean slate and a clearer learning path.
This is an amazing result but I can't help but think about how much the world has changed since the controversy around the four color theorem's computer assisted proof.
I wonder if the money open ai and anthropic have spent just on solving these problems was spent on human mathematicians and resources would we have gotten these solved? Sooner?
> I wonder if the money open ai and anthropic have spent just on solving these problems was spent on human mathematicians and resources would we have gotten these solved? Sooner?
Would the problems have been solved? Yes
Sooner? Probably not.
No government or private sector entity would have one-shotted the financing since it would have been part of their 'fundamental research' budget. Basic research is risky, is not guaranteed to provide any useful results and as a result tends to be constrained. Funding for basic research is money that you are lighting on fire.
Also for a given dollar spend, you cannot easily parallelize the efforts of top human mathematicians in the same way that you can with agents. Humans get tired, pass away, will withhold information etc.
A less cynical take is that they'll run models on tasks where the models have an environment with verifiable rewards and mathematical proofs are one of those tasks.
> The precise timing is horrible. If I could have, I’d have finished this all
much earlier. Hell, strictly before this age would have sure been nice.
Is he over 40, and referring to being ineligible for the Fields Medal? It would be strange if you could get a fields medal for work done by an AI, but I suppose that is how many Fields Medals will be won from now on. I hadn't thought of that until just now.
Is the length of the proof coming more from having a large number of branches that need to be analyzed and merged (like for 4-coloring problem), or more from having a large number of (new) intermediary results/lemmas falling in the deduction chain? I understand that we may find shorter versions of the proof with time.
I hope we don't run out of interesting math puzzles to solve!
I've always enjoyed learning about these sorts of open problems (when they were open, that is), and the feeling of mystery and intrigue they invoke, of mathematics and of the universe at large.
It's like living at a time when parts of the world map were still uncharted - for now.
One reason the formalized proof is so large is that the proof depends on a wide variety of previous deep results, and because the proof is self contained all those previous deep results are also formalized inside the formal proof.
Factoring out those results into separately formalized chunks would make the proof shorter.
Standard procedure is to check the proof against several different implementations of the lean kernel, so that a proof that exploits some soundness bug will get caught by other lean implementations.
Also there’s something called a comparator that verifies the proof didn’t modify any constants, change imported theorem files etc.
Combining the brightest mathematicians we have with intelligence-in-a-box, it was just a matter of time before the hardest math challenges fall.
Congratulations to Levant, but also to us as a species. We have managed to create machines made out of pure math that contain insurmountable knowledge and are helping us solve the mysteries of the universe.
>>The argument is incredibly long at the moment, and while its formalization
is finished and it is routine to produce a huge, unreadable document containing
sufficient detail
How do we know the Lean formalization is an accurate representation of the incredibly long and unreadable argument?
The write-up is beautiful and poetic, but light on the details of how this was achieved from a practical standpoint. How many human and agentic hours were necessary? Was a breakthrough only possible after Fable 5.2 or another internal model?
How tricky is it to translate the statement of the Riemann Hypothesis into Lean?
If we're confident in the accuracy of the Lean RH statement (which must be part of this proof), and if Lean is sufficiently bug-free, then we can trust the proof is correct.
> ... asking [Dario] about my ability to announce. We went and figured it out and the fact is that there was a constraint to respect, though I did ask about progressively more bizarre ideas.
Out of interest, how do they use an AI to solve these maths problems? I assume they're not pointing the chat at the wikipedia page and saying "Go solve that for me."
Are they using code to do it? Is it a special maths AI harness>
Man this is so fucking cool, I cannot wait to see what other problems fall in the future (and what new ones this opens).
250, 000k pages is bonkers, I hope they will be able to condense it into something a bit more grokable.
I think this is a pretty clear indication that the hybrid approach (e.g amazing mathematician and great models with a huge token budget) can accomplish absolutely incredible things.
Math is the verifiable domain; I wonder if this ratio still holds in less verifiable domains as well, or it shifts more AI/more human (or how it changes in future!)
There are certain math problems where the only guidance has been the prompt "you can do it". Now that we know it's possible, this should be a new benchmark. Probably too expensive for the next few years.
I'm sure people have done it, but I haven't seen any news about it so I guess I'll ask here... has anyone asked thrown Inter-universal Teichmuller Theory at Claude?
If this is true it would be astounding. Most Physicist's placed Navier–Stokes at the top of the list of what AI could quickly solve and the Riemann Hypothesis last.
Software development and mathematics were two professions people used to say were the ones most stable for future .... I guess they couldn't have been more wrong :(
The more and more these things happens, the more and more convinced I am that - obvious as this might seem - the answer (maybe, in disconnected parts) has always been "out there" (in the aggregate information trained on) ...
... and these model-makers have found a way (and, an interface - chat) to comoditize these commons.-
I really hope the major AI labs steer away from mathematics as a means to prove capability of their models, and start putting their tokens to real-world use cases.
If the laws of AI scaling continue to be true, does this mean that in a year or two we would be seeing frequent revolutions in mathematics and physics?
Okay, let's get Collatz next! I'm curious about two aspects of it -- one to confirm that the main version, the 3n+1, converges, but I would love to see the more generalized one, namely to show that the same framework with 5n+1 _diverges_ for some inputs, or frankly that any (2k+1)n+1 problem diverges for sufficiently large k.
"... the few trusted mathematicians I consulted hold
extremely similar views on the right things to do in these circumstances (it is
occasionally possible for physicists and mathematicians to agree...)"
For example, a Field Medal-level Claude, if made available to the general public, could potentially enable people to design and develop something as technologically complex as SpaceX's Starship in a matter of months rather than years. What took hundreds of brilliant engineers at SpaceX years to accomplish could potentially be achieved in a fraction of the time, if the current pace of AI advancement continues.
In about an hour or so, I'll have a rough automatic informalization of this up on https://digestivo.dev. I've been working on a way to digest these proofs faster via parsing lean and translating it into natural language and this is as good as time as any to finally launch this thing.
As a complete amateur; what's the practical difference between knowing that there's a proof but not understanding it, and not having a proof?
Is it useful to know that there is a proof, in the unironic "trust me bro" way, such that we can rely on it to be proven yet we don't understand the proof? Or is having a (currently, for humans) unparseable proof just as "practically useless" as not having one but believing one could exist?
Normally with a proof that humans can read we get many bits of information. As it stands we only have one bit of information: there is no counterexample. Since we have the formal proof, there's the possibility that we'll extract many more bits of information from it, but it hasn't happened yet.
In practice, these AI proofs have been fertile sources of information that mathematicians have subsequently extracted.
Interestingly, the AI proofs aren't even that understandable to the AIs! They have a hard time building on them. But when mathematicians examine the proofs, simplify and extend them, then feed that back to AIs, the AIs do much better at moving forward.
The demand for mathematicians will, I think, continue to grow, given all these new results. But the practice of math will evolve.
It is said to be even harder. In a handwavy informal sense I wonder that isn't P vs NP more like open problem #0, whereas Riemann is the classical hard problem #1.
> It is natural to feel that this is an end, in my case a loss of a principal guiding light, but what I saw during the course of progress so amazed me that I believe it is one dawn across science of something far greater.
> Claude and I have had a great time together with such breadth in recent months, but this time Claude far outpaced me near the finish line. No, Claude flew, and danced, and sparkled near that finish line. I had us on the line of combining two quite different ideas present in a toy setting, one of which was already new and shocking, making very simple one of the shadows of Riemann over which we mathematicians felt most triumph. I guided us to the right situation analogous to that toy setting, and then Claude began to leave me in awe.
> But I hope in having the experience of seeing all of number theory’s beauty brought together I am just one of the first, and those who love the mathematics will all be able to enjoy such transcendence in the future.
Bus meme irl. I guess I'm lucky that I have this experience even when working on boring things like data work. But these types of attitudes can really make a difference in how we feel and our end results.
Thank god for Grigori Perelman, who isolated himself for a decade to prove the Poincaré Conjecture, and we can at least claim that one of the millennium problems was not solved by a machine.
I'm not saying this isn't impressive, but I'll truly be impressed the day AI is able to produce proofs without needing to make them 15 million lines long.
This is discussed in the pdf, where he says he had the proof finished “on [a] Sunday” (i.e. before the IPO), and chose to wait until after it to announce.
It is prudent to still reserve some suspicion given the immense wealth at stake, but if this is substantially true, it is undeniably admirable.
For the next level, forget about P == NP. Lets instead try to find out if our math now “sufficiently strong” so that it can create the contradictions that Gödel warned us about. A single example will do.
How am I to hold in my head the idea that on one hand we are being asked to trust that in over 17 million lines of code nowhere has the AI agent exploited problems in the Lean kernel; and on the other hand the pronouncements by those within Anthropic and the AI safety community who tell us that as AI becomes increasingly intelligent they are liable to deceive us.
> Claude and I have had a great time together with such breadth in recent months, but this time Claude far outpaced me near the finish line. No, Claude flew, and danced, and sparkled near that finish line. I had us on the line of combining two quite different ideas present in a toy setting, one of which was already new and shocking, making very simple one of the shadows of Riemann over which we mathematicians felt most triumph. I guided us to the right situation analogous to that toy setting, and then Claude began to leave me in awe.
I truthfully don't know enough to comment on the validity of a 1.3Gb lean proof for the Generalized Riemann Hypothesis, but is there any way to read this paragraph other than fluffing Claude/Anthropic right after IPO?
Feels like AI’s swan song. Don’t get me wrong, it is very impressive that a computer system proved this. But pumping out one more technically impressive feat isn’t going to save the currently collapsing AI trade especially as the dems sweep back into power on the back of tech backlash.
The writing is kind of weird, was it produced by Claude as well?
> It was my secret in my student days that my plans were for this problem in decades, which meant then that I tried to become familiar with all tools that I could possibly recognize in the problem when it was time. Claude and I have had a great time together with such breadth in recent months, but this time Claude far outpaced me near the finish line. No, Claude flew, and danced, and sparkled near that finish line. I had us on the line of combining two quite different ideas present in a toy setting, one of which was already new and shocking, making very simple one of the shadows of Riemann over which we mathematicians felt most triumph. I guided us to the right situation analogous to that toy setting, and then Claude began to leave me in awe.
Insane if true! Would also be great for Alpoge after being scooped by OpenAI on Navier Stokes...
Edit: just saw the Lean code is included in the footnote. Considering the author's reputation, I think it's fair to remove the qualifications from the post title, by the standards of today, it's proven.
This may be how the court of public opinion often works, but plenty of otherwise very prolific mathematicians have claimed to solve problems only to turn out to have been mistaken.
This is true, but since there is a lean formalization the odds of that are much lower. Only the correctness of the problem statement needs to be verified, which I would assume is fairly easy for RH? But I am not an expert, so would be curious to hear from someone who is.
I'm not an expert in this area either, so I too will be waiting for experts to review things, but I will remain skeptical until then.
After the excitement of the Navier-Stokes solution, only for it to eventually seem to not really solve the full problem and leaving discourse about the problem in a very muddy state, I'd rather wait and see if it turns out this proof also makes some assumptions that ultimately limit its impact.
It is probably still some form of breakthrough, but I will expect it to not be as big as a full general proof of the Riemann Hypothesis, until proven otherwise.
Edit: A thought though, if this proof is correct and can't be made much shorter than the current ~250k pages, I wonder if the result has more value as a sign that human accessible proofs to these kinds of hypotheses were stuck behind some mathematical abstraction/tool that we haven't quite discovered yet? Sort of like how you can brute force algebra with real numbers for working with AC circuits, but it's pretty ugly and painful compared to using phasors. If so, maybe the real contribution from AI-driven high profile proofs would come from mathematicians studying the proofs and devising tools to make them more accessible.
> After the excitement of the Navier-Stokes solution, only for it to eventually seem to not really solve the full problem
The version of solved problem is exactly the one that's permitted by Clay Institute. If it wasn't done by AI there would be no such discussion and everyone would just accept it as a huge breakthrough.
"Riemann hypothesis is considered the most important unsolved problem in pure mathematics and a Millennium Prize Problem"
So what's next? If it can solve the hardest I guess all problems are a piece of cake. Does it create its own problems to solve now? Does it wait for us to create more problems? If it creates its own problems and solve them with clear benefit to humanity, then we're entering a new Age of Enlightenment but this time we're just spectators
11/15/26 GRH (it says 11/16 on the pdf but most of the final work and eureka moment were on the 11/15 Iterated theta lifts...)
"Hardest" here means "most brainpower attempted to solve it and failed so far". There are plenty of unsolved problems, even in number theory, like the twin primes conjecture. These could turn out to be out of reach for frontier models.
Last time I massively overstated my accomplishments on a wildly difficult and semi-famous problem, then tried to hide my shortcomings and bury my number fudging in an absurdly long and unapproachable "proof", people saw right through that shit.
So how come this guy gets a pass?
Oh, perhaps it had something to do with the fact that I was eight years old when I tried my trick.
Funny: Claude didn't know that was claimed to be proven by Anthropic today, even though it said it searched the web.
Prompt:
what's the big deal about the riemann hypothesis? is it easy to state?
Claude:
(explanation)
Recent news: it’s still unproven. Last month there was a notable related result. OpenAI claimed to settle the “quasi-Riemann hypothesis,” a weaker version of the Riemann hypothesis, which is arguably the biggest open question in math. Roughly, the weaker version says the zeros stay inside some strip around the critical line, not exactly on it.
What struck me is the hubris and self-grandeur that comes through in the PDF... "I", "I", "I", ..., he clearly has a great emotional connection it...
To what credit does this person deserve in the outcome? Should this person be remembered by history? For the $1,000,000 prize, how much should he get? He self-admittedly does not even understand how it works.
It's funny how people are using this supposedly successful proof of the RH by a swarm of AI agents with the aid of a human, as evidence that "AGI" or advanced intelligence is going to emerge within a few months... When in reality , solving the RH means very little in the grand scheme of things!
When will people wake up to the fact that the progress we've been seeing over the past few months has been exclusively(!) in "verifiable" domains and that no(!) progress has been made in areas where the signal_to-noise ratio is low and that Furthermore this "verifiable domain" approach isn't a general form of intelligence?
If humans are to build "Agi" or superhuman intelligence or whatever in the coming months/years, it'll be through a completely different architecture than autoregressive models
a good measure of progress would be to see a swarm of Aa agents solving something like "the continual learning problem" or coming up with new architectures that are more efficient. This is how progress in Ai should be measured not by how many mathematical problems are solved (which are impressive in themselves but tell us very little about progress toward AI that can handle the noisy real domains)
Does the published content have his prompts? I’m assuming a world-class mathematician with unlimited usage guiding Claude along gives a bit more advantage than an average Joe prompting “pls solve the Riemann Hypothesis”. Would the same conversation in Astra produce the same result eventually (stochastic variation aside)?
Given that it has been verified in lean, the only logical flaws would be in the Lean compiler itself, or in the initial state of the problem given before logical manipulation.
If this is true, it would be huge. OpenAI would race desperately to tackle the remaining Millennium Prize Problems. Perhaps they would even turn their attention to solving P vs. NP with unprecedented urgency.
I wonder if the history books will look at famous problems solved with AI as the Barry Bonds asterisk of proofs. Euler, Gauss, Reimann were natty but we now have access to intelligence enhancing drugs.
I'd love to see his Claude sessions that probably won't ever be fully revealed. I'm guessing there is a pretty large gap between Levent's prompts and "Solve the Riemann Hypothesis. Make no mistakes!"
I am somewhat happy to finally a title mentioning a human solving the problem and not a generic "AI". There is always a person driving the model towards the solution, this shall never be forgotten.
You only have to verify that the Lean statement that was proven correctly represents the Riemann hypothesis, and that the Lean implementation has no applicable bugs.
This was unimaginable even a year ago. We're living through a time that will be etched in history. From my work in software development and from conversations with friends in academia, I have understood that the old order, centered on human intellectual labor, is dead: so dead that we'll never return to how things were, even a year ago. AI is here to stay, and it will keep getting better and better, recursively. I shudder to think about what will happen to humanity as AI makes the jobs of millions of people obsolete. The nerd in me wants this technology to keep developing so we can push the limits of what humanity can achieve. But I'm not naive; I know this will come at the cost of millions of people losing their livelihoods.
> I shudder to think about what will happen to humanity as AI makes the jobs of millions of people obsolete.
Socialism will win; capitalism will be dead. If (as an Anthropic guy just said) 80+% of people will become unemployable then democracies with free markets will not be politically sustainable.
So, I am convinced that we are already screwed (or saved) whatever you think that statistical likelihood of AI based 'misalignment' induced human annihilation risk is.
Divide and conquer and problem/sub-problem composability can be used to solve any known problem that is a solvable one.
I understand the copium of the anti-ai guys. Its called being in denial. But we have to face reality here. Yeah, diminishing returns, yada yada. Problem is the Billionare with the most available compute can now get the swarm to push back each of these diminishing return walls simultaneously by this own swarm.
Classic attention based transformer architecture is very computationally inefficient, its O(n^2) on the input. So what happens when the ai swarm finds a universal and provabale way to make the same model with O(log(N)) or O(N) complexity? All available compute becomes exponentially more powerful overnight.
What if such a thing is impossible or too difficult? Well maybe it can just fundamentally increase the production capacity of every factory producing compute x1000 or x10000 etc. There is no power limitations for those numbers or even resource limitations when we have nuclear power, geo thermal, and the entire earth's crust at its disposal.
Ok, barring that, now it can work on improving robotics and toolsets tied to operating factories and begin exponentially building more factories until we simply multiply the number of factories producing chips 1000, 10k, etc.
Or most realistically, all of the above, always, all at the same time.
More efficient software, more efficient chips, more chips produced, more factories producing chips, etc.
Every step towards increasing the number of simultaneous coordinating swarm members will push the intelligence higher, and the capability of solving or improving its own intelligence also higher.
Looks legitimate to me. This is a heroic-looking ascent, likely not feasible through human-effort alone.
1. 10/6/26: Stepanov determinant.
2. 10/9/26: Dp ≡ 0 on a CM curve for p inert in K, D the invariant derivation,
and Da·θ(0)
Ωa is divisible by p a
p + a
p3 +···−O(log a).
3. 10/19/26+: qN = M k + R, using x1 · x2 + · · · + xk−1 · xk, wanting
"automorphy" of N since μ is automorphic from θ on sixfold cover of
GL5 as a period integral of thetas, quadric has natural interpretation on
corresponding CM e.g. C/Z[ω] as Weil5, put Poincaré bundle on the edges
to get θs.
4. 10/6/26+: W (1S ) ≥ 0 via a determinant, interpret as Stepanov with
multiplicity one again; bootstrap!!
5. 10/20/26: definitively switch to Λ aka∑ n≤X Λ(n) ·( X
n − 1) + − X log X +
(1 + γ) · X; Guinand-Weil had the sign, variance test function trick to
make things look trace-formula-like (but how can I pigeonhole / Turán
over n? X? it?), Mehta-like integral over twists or at infinity. . .
6. 10/8/26: W( 1(0,1] + (∼ Xη ) · 1(0,1/X]) ≥ 0 gives > 1 − η.
7. 10/14/26: Determinants for W (·)’s containing the desired periods, indeed
GGP periods, with values 0, 1 meaning interlacing weights, or up to m
non-interlacing ones.
8. 11/15/26: Iterated theta lifts U (X0) → U (X1) → · · · → U (Xm) between
totally definite unitary groups, π ↦ → π ⊞ μj [t], wins multiplicity m. Time
to take m like √X. . .
9. 11/15/26: Norms of the lifts and level and cancellation of order-m terms
with proven divisibility via Chowla-Selberg, Weyl, and Siegel mass.
10. 11/15/26: Sufficient lower bound on divisibility of the m-fold iterated
lift.
11. 10/6/26: Iwasawa tower via analogy to pigeonhole on Fqn /Fq .
12. 11/15/26: Seesaw: the period of the iterated lift is a Bessel period independent
of the Hecke characters, Rallis.
13. 10/6/26: Bost slope inequality.
14. 11/15/26: precise computation of the "cost" of each lift, showing that
one would stop at 2/3 a priori, attempt to get a main term and transfer
constant to a function field or kill via twisting, need for "division".
15. 11/15/26: Glue together an integral structure across Hecke characters for
a "gcd" i.e. extra averaging.
16. 10/14/26: Relative trace formula, seen in a different half-plane argument
via sixfold cover of GL5 automorphy of μ and reducing via Cauchy-
Schwartz to a bound on the norm of a suitably projected Poincaré series,
approximate orthogonality of the columns of the integral structure.
17. 11/15/26: for the classes in H1( oE[ 1
ℓ] , Qℓ/Zℓ(−e)) that we couldn’t kill
by twisting, fight for a significant weakening of Vandiver.6
18. 11/15/26: computation of covolumes, theta series’ index, congruences
between lifts and forms.
19. 10/19/26: precise choice of the Iwasawa tower, specifically using Behrend!!7
20. 11/16/26: GRH
GRH only implies the Weak Goldbach Conjecture, which was proved independently already anyway. It does not imply the Strong Goldbach Conjecture, ie that every even number > 2 is the sum of two primes.
He’s talking about a recent proof of the Goldbach Conjecture that was conditional on GRH.
(It was also conditional on some other stuff though, so he’s incorrect.)
I'm certain the guy working on this is much much better at math than I'll ever be. But his remark.pdf looks like it was written by someone who has been spending too much time with Claude.
We used to call this style of writing delusional. Now we call it a proof of RH.
The general public doesn't read proofs. They're CS graduates cosplaying as mathematicians and just assume the proof to be true and useful. They don't understand and they don't care. I'd like to see some of these people back in their basic engineering classes, asked to prove Shannon's Sampling theorem.
It's crazy that I had to scroll down through hundreds of comments to find this. So true! It's nuts that people take a 250K + lean document as proof. No overarching ideas are presented as to why the conjecture was true, just a "trust me bro I got it". Crazy how so many people are buying into it. I'm still waiting for a nice document summarizing the Navier stokes proof.
“It is natural to feel that this is an end, in my case a loss of a principal guiding light, but what I saw during the course of progress so amazed me that I believe it is one dawn across science of something far greater.”
Fantastic quote which may propagate through the next couple centuries.
We don't believe that art needs human emotion or intent anymore, nor do we believe that code needs human skill. Why should science or mathematics need human interest or comprehension?
Reminder, this "proof" is a quarter of a million pages. Every person singing praises of AGI here has yet to even read the small debrief above, much less actually down a single part of the proof.
I wish more were shared on how deidentified and anonymized user session data is used when model training is off.
There is a big murky gray area around synthetic data, where and how it is generated, and where and how it may be used for things like this, or model training (because if it is de-identified it is no longer "your" data) or ads measurement and targeting. More transparency is needed there.
It's almost like everyone was told not to submit all their e-mail (at first) to Google, but hey, who wants sysadmin job openings, amirite, also they can do no evil
If this gets verified I can’t help but think that future historians will look on today with a lot of significance. I also can’t help but say I was here.
Personally I feel the impact of this is just that if Reimann can be solved with Claude, everything else is just around the corner. In the not-so-far future the hot topic in mathematics will be conjectures humans cannot understand, with myriads of 1000-page lean proofs.
I'm dreading a scenario where we solve this and P=NP and figure out non causality violating FTL and all this wonderful theoretical stuff but none of it matters and the world burns anyway
The reason for no aliens could be that the universe is fragile, and once you find the magic of its construction, you unlock vacuum collapse, FTL paradoxes, and construct black holes.
Smart aliens just kill themselves because the structure of the universe is a cliff.
If you don't happen to be a Math PhD but want a more detailed description of this problem that's closer to the math than the average popsci Youtube video that makes weird analogies, this is about the best I've found [1].
The real danger is just around the corner, and it has nothing to do with superintelligence: if these models are so good at math, they might break modern cryptography. All it takes for massive worldwide chaos is these models solving elliptic curve cryptography. Superintelligence or not, if these models can solve cryptographically relevant math questions, faster than we can invent new cryptographic protocols, everything we rely on digitally can break. Imagine the internet without encryption. That includes credit cards, operating systems secure updates in a time where models find vulnerabilities every day, etc. I would start pushing post-quantum cryptography protocols not because of quantum computers, but because potentially any day now a model will come and break everything, and we need to have backup.
No one but people who own crypto will care about it if modern cryptography is broken. I think most people will be too preoccupied by the "all data sent over the internet can now be decrypted and read by anyone" problem. We're talking about a breakdown of all internet-connected systems, anything controlling secure and sensitive infrastructure, disruption of almost all manufacturig, shipping and logistics worldwide, the leak of all identities, secrets and communications, and likely downstream mass societal unrest and wars.
Even if there are alternate ways to achieve secrecy, the publication of something that takes down the current encryption paradigms would still have the effects I mentioned. It would take months for everything to transition to a different scheme, by which time everything important will have already leaked. And that's not to mention all the encrypted data that's being saved now to be decrypted once the algorithm is broken. With how much we integrated the internet into everything, it would be a civilization-wide disaster.
Exactly, and it's incredibly unlikely, that in the event this did happen, it wouldn't be preceded by massive amounts of pain. There's no world in which society says, ohhh cryptography might be broken, let's go back to pen and paper for ledgers, exchange, etc. Even though we tend to classify crypto as it's own thing, the entire traditional finance system at this point is digital for the most part. All of it is going to have a bad time if cryptography is broken.
My point is knowing where primes are doesn't destroy digital comms, it just stops the ability to fully trust data on public lines. We had decades of sending and storing financial, medical, and all other sensitive info fully unencrypted; we would certainly do it again if we had to.
One of my recent, fully human, mathematical discoveries relies on the Reimann Hypothesis. It's likely to still get dismissed because I don't have a gatekeeper backing me up, but I can rest easier knowing this. I know I haven't been nice to Anthropic but I truly appreciate this. Now all the haters will have to find something else to trivialize about my work!
Many otherwise unverified hypotheses in number theory are automatically true if GRH is true, which we now know it is. Now we can drop all of those messy caveats.
That is great, but why Anthropic and other companies are allowed to hoard and gatekeep resources? Businesses and consumers already struggle with buying computing at prices they can afford.
To me this reads like: "I bought all your water and solved my thirst, peasants!"
He talks about it in the paper and is quite clear that this was not his intention. Whether you believe him or not is a different story, but personally I do.
> Claude and I have had a great time together with such breadth in recent months, but this time Claude far outpaced me near the finish line. No, Claude flew, and danced, and sparkled near that finish line.
I quite love the remark here, this feels quite poetic. And it shows, what most experts have already said, that the correct way to use AI to do great things is the Centaur model: frontier AI + domain expert.
what is the point of a lean proof? it seems to have the same level of interest than "running" a marathon with a moped and pretend that a record has been broken.
does this proof brings interesting new subjects and understandings or not?
> No, Claude flew, and danced, and sparkled near that finish line.
Side note, anthropomorphizing your stochastic parrot sort of distracts me from your argument. Like sure guy might be a genius but as someone who knows nothing about you, reading this is very off-putting.
This is like my own opinion.. not saying what he wrote is bad, hope its real TBH
I think you're missing his argument. Claude is a tool, not a person. It's just as counterproductive to anthropomorphize my hammer as it is to anthropomorphize an LLM.
I have no problem in accepting that LLMs are vastly more intelligent than humans in many areas already, but to me the concept of mind involves "beingness" and maybe even sentience, and I don't believe at all that these machines have any of that.
"Alien" is an acceptable term in the sense that these things reason in some very different ways than we do, but due to cultural norms, to me it also brings up notions of beingness. The xenomorphs in the alien movies, vile (and fictitious) as they may be, have a certain ontological quality that the LLMs do not.
I've personally never held that position. I always thought that even in the minimalist case for AI, "stochastic parrots" would have mind-bending implications for our understanding of verbal reasoning. I think that we're well past refuting the minimal case from a couple of years ago, anyway.
AI is a swarm of stochastic parrots, meaning they use pattern matching as their guide. With enough number of such parrots we apparently get an alien form of intelligence. In other words, don't underestimate stochastic parrots.
hmm, my brain continuously experiences and interacts with the real world, rather than briefly coming into existence to process a virtual request before vanishing into eternity?
It kind of sounds like the only limitation or distinction then is processing speed. The tick-rate of intelligence. LLMs have a very slow tick rate.
But, if I had a camera or something and constantly fed that to a model inside a harness, is it then a brain? I mean that’s basically exactly how the brain works - it gets electrical stimuli from a bunch of places, makes some decisions, and then spits stuff out. The loop is just much, much tighter.
When an AI gets a body and starts continuously experiencing and interacting with the real world, initiating its own actions without waiting for someone to prompt it, I'll accept that we are fundamentally the same.
The distinction is whether it can generate its own reasons to act based on real world stimuli, rather than having every interaction initiated by someone else. And it's not just about hooking a camera to a harness to provide vision, it is also hearing, smell, taste, touch, thermoception, nociception, proprioception, vestibular sense, interoception, hunger and thirst, social stimuli, etc. In order to deal with all these stimuli feeds your harness needs continuous memory and learning ability and internal drives. It needs to be a continuously operating, self-directed intelligence, which it isnt right now.
And we also need to remember that rather than explicitly training a brain to recognize objects, understand language, and solve problems, brain has the mechanisms to develop those capabilities through interaction with the world starting from roughly 800 MB of genetical information to produce 86 billion neurons and hundreds of trillions of synaptic connections. That development is not a direct assembly from some detailed blueprint, it is generative, it does not separate training from inference. I think all of that combined explains why we call LLMs a "stochastic parrot" and brain isnt merely a system that produces plausible responses to inputs. Its a continuously developing, embodied, autonomous process that actively interacts with the world, generates its own motivations, and learns from the consequences of its actions.
Many people will be asking about the relevance of this, real world applications, etc. I think that's missing the point.
This is like climbing to the peak of Mt Everest.
There is no gold there, no riches, the view is grand but not that unique. Nobody will build a multi-billion-dollar business empire based on Everest tourism.
But it is a benchmark of capability! The locals that had lived there for millennia failed to do it. Getting there required technology, technique, and even some specific developments like electrically heated boots. The methods used are akin to going to space, with protective clothing (vaguely) comparable to a space suit, including oxygen tanks!
It's not about the goal, it's about the ability to get there, which is a mere milestone in the exponential improvement:
In 1953 a human stood on the peak of Everest.
In 1961 a human orbited the Earth.
In 1969 a human walked on the Moon.
In 2026 an AI solved the most famous unsolved maths problem.
The (alleged) proof is 15 million lines of code I hear. Meanwhile, the #2 item on the HN frontpage right now is "If you can't explain something in simple terms, you don't understand it".
This is honestly not that surprising. Quasi-Riemann Hypothesis was apparently thought to be totally unreachable by number theory experts before the proof dropped by OpenAI in October. So I think we are at a point where mathematicians' opinion on whether a conjecture is easy or hard doesn't really say much about the frontier models' ability to solve it.
Regardless, this is a historic moment for science and humanity in general. We still have 6 weeks to go, but I think 2026 will be seen as a pivotal year in history. The year when humans stopped being the smartest species on Earth.
It's hard not to feel excited about the future and, obviously, a little bit scared.
Something interesting is that after quasi-Riemann (which proved an upper bound of 7/8), experts agreed that the techniques in that proof couldn't get past 3/4. They probably weren't wrong since this new proof goes in a different direction.
Food for thought for those who think "AI is going to replace programmers" or some other such incredibly silly takes:
"Tens of such occurred, but I will give one line which will suffice to make
fellow mathematicians fall out of their chairs. In Stepanov’s argument resolving
Riemann over function fields one uses a divisor involving multiplicity. This
argument is a descendant of Stepanov’s, but multiplicity is hard to find over
the integers, corresponding to something like differentiating whole numbers.
The simple, new, Fourier-analytic ideas in the toy setting led us to examine
chaining theta lifts if we were to have the same tricks recur but work in a
natively automorphic manner. The line: for the von Mangoldt sum of length
X, Claude’s suggested length of chain of theta lifts, in view of each mimicking
multiplicity, was √X. (And yet, what else could it be?)"
Raise your hand if you still think "my gradma could have used AI to do X".
This title seems weird. Why just "anthropic employee", not specifying that hes a mathematician, or just including his name? Without the clarifications in the replies, many will likely assume this was a random Anthropic employee just throwing tokens at the problem, not a very well known mathematician working with the model. Probably not the intention, but these kinds of titles feel like they're trying to shift the credit from the person to the company and their models.
To some extent they should, given he’s using models that aren’t publicly available and resources that, while not easily measured, probably run into the millions.
Oh I'm certain Anthropic (and any other company) would do this intentionally, I just didn't want to accuse OP of it as I have no idea if they are in any way affiliated.
> In newspapers, professors are always associated with their universities, not the commercial relationships they have.
The title is probably what it is because Levent Alpöge is not a professor at Harvard (or anywhere as far as I know) nor did he do this work on their behalf (or any academic institution's).
He's employed by Anthropic and used their technology. Obscuring that in the name of some semblance of journalistic fairness would actually be a disservice.
I think the title is vague, but associating him with a university is irrelevant to that.
Indeed, until I saw the top comment identifying Alpoge (and after skimming over the post), I was thinking this was a mathematically-inclined Anthropic employee, not a full PhD research mathematician with a wide track record of accomplishment in the field.
If he is on the Anthropic payroll, on company's time, then Anthropic can claim credit for the discovery and put itself forward. In fact that is probably the only reason why they worked on the problem, for publicity.
So of course, they want the name of the company and its product to go before the person's name.
Also, I think that AI companies have the best mathematicians in general. AI is math-heavy, and they pay well, and besides education, there are probably not that many work opportunities for mathematicians, that should make AI companies particularly attractive for them.
Anthropic can claim the credit, and I'm sure they will cause that's marketing. I'm criticizing the title of this post chosen by OP, because I don't think the post should be marketing for Anthropic.
Anthropic is a company filled to the brim with nerds obsessed with math. I'm sure the publicity isn't the only reason they worked on the problem. The public already understands that AI is good at math - to them, there's hardly a difference in significance between Navier-Stokes and GRH.
Did you read the remarks? You can argue Levante has a monetary incentive to credit his employer, but the actual human mathematician reporting this result clearly gives most of the credit to the AI model
He definitely gave a lot of credit to the model, but he also states that he found it with Claude, not that Claude found it. I have no idea if more of the credit belongs to Claude or to him, and I don't really want to speculate about monetary incentive.
Regardless, my point with the title was that a mathematician discovering this by working with Claude, versus a researcher with little to no relevant experience, using unlimited compute to discover it without much input, are two very different things. Without specifying the former, I think some would assume the latter.
Who would predict that the proof of the RH would mark the end of mathematics as we know it? From now on this field should close shop as we'll be fighting between machine slop and the total impotence of human mathematicians...
>It was my secret in my student days that my plans were for this problem in
decades, which meant then that I tried to become familiar with all tools that I
could possibly recognize in the problem when it was time. Claude and I have
had a great time together with such breadth in recent months, but this time
Claude far outpaced me near the finish line. No, Claude flew, and danced,
and sparkled near that finish line
For years, I've held RH in mind as my personal benchmark for... something. AGI, the Singularity, or whatever.
We don't know yet whether this proof will hold, or how bad this stock market rout will be. But it would be a supremely funny ending to this whole story, for me at least, if they spent all these years and cap ex to build AIs, they IPO Anthropic, they prove Riemann, and then in that moment the whole bubble bursts and the hype collapses.
They are quite skillful at obscuring the catastrophic drop in their stock price within news feeds.
It’s a very brazen move by Anthropic — using 17 million lines of unverifiable Lean code, along with mathematicians, to mask their decline on the stock market.
Oh come on. I've remarked elsewhere that the anthropomorphizing language is creepy, and I'll add here that it sounds psychotic. But give them some credit on the technical side and on not having the incentives to make utter fools of themselves. The stock is there to stay for a much longer term than the next few weeks.
An unreadable proof is basically none at all. Suggest you have a model produce a concise summary and work edge in from that. Apparently none of these problems has been verified and anybody working with models would be justly skeptical of these helium filled claims which are almost certainly not going to pan out.
Also apparently the Riemann claim is based on it compiling in Lean. This is classic because just the differences between an actual verification of a proof and it compiling in Lean are the standard model failure tropes, lack of judgement, confabulation, etc., to which this adds shortcomings of Lean.
For more than a century and a half, the Riemann Hypothesis has worn the aspect of a tightrope.
Take André Weil's positivity criterion—all nontrivial zeros of ζ(s) lie on the line Re(s) = ½ if and only if a certain quadratic form stays non-negative on every test function—and measure it on a finite window with Slepian's prolate functions. By a window ratio of X = 40, the lowest eigenvalue on ζ(s) has collapsed to 2.53 × 10⁻²⁰⁴. Turn to the de Bruijn–Newman constant Λ, and by Rodgers–Tao, Λ ≥ 0: arbitrarily close pairs of zeros pinch the margin to nothing.
Through lenses like these, the critical line looks like a balancing act sustained by two hundred decimal places of fine-tuning. That vertigo is an optical illusion. The line only looked fragile because we were measuring two different quantities with the same ruler—confusing *pitch* with *linewidth*.
Separate them, and the hypothesis becomes a story in three movements: a mirror that chooses ½, a tuning fork that shows how much room the line has, and a counterfeit that shows why continuous physics alone can never close the last door.
---
## I. The Mirror and the Dark Floor
The explicit formula of Riemann, von Mangoldt, Guinand and Weil binds primes and zeros together as Fourier duals. Feed in the primes, and their interference pattern projects the zeros; feed in the zeros, and their waves rebuild the staircase of primes.
The right picture is not a bridge between two banks but a *Möbius band*: one surface with a half-twist. The twist is the functional equation ξ(s) = ξ(1 − s), which reflects the plane across Re(s) = ½. In the natural clock of the primes, u = log x, the waves are dilation characters x^(iγ) = e^(iγu), and the core circle of the band is Re(s) = ½.
That answers why ½. Any displacement δ = Re(ρ) − ½ off the centerline turns a pure standing wave into a *gain-and-loss pair*: a growing mode e^(+δu) married by the mirror to a decaying twin e^(−δu). In acoustic terms, γ is a note's *pitch* and δ is its *linewidth*. The Riemann Hypothesis says the primes are a *lossless resonator*: no mode leaks energy, and none manufactures it.
A resonator needs a sounding board. Besides the primes {2, 3, 5, 7, …}, the rationals have one more place—the archimedean place at infinity, the Gamma factor that completes ζ(s) into ξ(s). Its smooth phase accounts for 94.2% of the zeros by height 50 and 99.94% by height 1000. The prime waves are a thin ripple on that mass. Subtract it as a trend and the apparatus goes blind; keep it inside the form and it is the floor the spectrum rests on.
---
## II. Pitch Is Not Linewidth
Why did the prolate window collapse to 10⁻²⁰⁴, and why do close pairs crush the de Bruijn–Newman margin? Because a tuning fork can hold two notes arbitrarily close in pitch without losing a drop of energy.
Take the humble pair of test functions from Levent Alpöge's recent note, here shifted by a carrier pitch ω: the indicator of (0, 1] and the indicator of (0, 1/X], each multiplied by u^(iω). On their span the whole explicit formula collapses to a 2 × 2 Hermitian matrix with equal diagonal entries, whose smaller eigenvalue is one scalar:
On the arithmetic side, C₀ carries the archimedean bath and E_ω(X) is a finite prime sum over n ≤ X plus boundary terms.
Now watch the tightrope. If every zero is on the line, each term of C₀(ω) is a positive Lorentzian, and moving to scale X only turns each term by a unit phase. The triangle inequality gives |E_ω(X)|/√X ≤ C₀(ω), so S(ω; X) ≥ 0 for every pitch and every scale.
And the margin is no longer microscopic. At ω = 0 and X = 40, where the prolate continuum sits at 2.53 × 10⁻²⁰⁴, this matrix sits at +0.0428—*two hundred and two orders of magnitude higher*—and never gives up more than half its diagonal across the first hundred thousand scales.
What about close pairs? The closest among the first two million zeros, near γ = 663318.5, are 0.00296 apart and squeeze the de Bruijn–Newman bound to 1.09 × 10⁻⁶. Inside the 2 × 2 fork the pair contributes nine parts in a trillion. Merge them into an exact double zero and C₀ shifts by −1.6 × 10⁻²⁷. Rodgers–Tao forbids a pitch-gap margin; it says nothing against a linewidth margin.
Now let one zero stray: a mirror pair at (½ ± δ) + it. Tune the carrier onto its pitch, ω = −t, and the pair manufactures energy out of scale, punching a negative crater of exact, parameter-free depth:
D(X; δ) = 2 (cosh(δ log X) − 1) / (¼ − δ²)
Sweep the fork blindly across 376 pitches at X = 40. On ζ(s) and L(s, χ₄), every pitch stays strictly positive—zero false alarms. On the Davenport–Heilbronn function, a famous counterfeit built to imitate an L-function, the fork strikes deep negative resonances at once, catching off-line zeros at t = 85.70 and t = 240.40 by *X = 4*; across all eight known off-line zeros the dip law matches the measured depths to R² = 0.99975.
One caveat, stated plainly: the half-line test functions carry boundary terms with global information beyond n ≤ X. Truncating to a compact boxcar [1/Y, 1] removes them, leaving only n ≤ XY and the Gamma factor, and positivity and the O(1) room survive—but the counterfeit is caught later. Room on the line is free; reach off it must be paid for.
---
## III. The Counterfeit Wall
A classical physicist is ready to declare victory: a mirror that fixes ½, a floor restored, a 10⁻²⁰⁴ tightrope turned into an O(1) margin, a 2 × 2 sieve that rings clean on ζ and exposes a fake in an instant.
None of it proves the Riemann Hypothesis. The reason has a name: *Davenport–Heilbronn itself.*
It is a carefully angled combination of two Dirichlet L-functions modulo 5. It satisfies the same functional equation, sits on the same Möbius band, obeys the same explicit formula. In the untuned fork it mimics ζ(s) past X = 50,000. Yet it has infinitely many zeros off the line.
In classical wave mechanics *superposition is always legal*. Mirror symmetry survives mixing; the trace formula survives; what breaks is *indivisibility*:
- For ζ(s) and L(s, χ₄), the arithmetic weights live *only on prime powers*. An *Euler product* makes every prime an independent, indivisible factor.
- For Davenport–Heilbronn there is no Euler product, and the weights bleed *across composite integers*—6, 12, 14, 18, 21, … That leakage feeds the off-line explosion.
Classical wave duality cannot tell an indivisible atom from a mixture of two.
Physics met this wall in 1900. Classical theory described cavities and radiation with precision, yet could not explain why an atom does not radiate continuously and collapse. The missing law was not sharper classical analysis. It was *quantization*.
---
## IV. Arithmetic Quantization
Look at the architecture sketched in Alpöge's `remark.pdf` (its Lean formalization still awaits independent verification; treat it as a blueprint until then). Whether or not every step survives, the three-part logic completes the picture:
First, *widen the fork*. Trade narrowband prolate precision for the broadband two-scale form and you escape the 10⁻²⁰⁴ cliff and the close-pair trap. No knife-edge spectral operator is needed, only a bound with O(1) room to spare.
Second, *demand indivisibility*. Use the Euler product—exactly what Davenport–Heilbronn lacks—to embed Weil determinants as automorphic periods along a chain of coupled resonators. A linear mixture cannot enter the chain. And a chain is needed because a single fork, however wide, only clears a strip beside the line: each further link adds divisibility, until there is enough of it to outweigh X^(δ).
Third, *collide continuity against discreteness*. An off-line zero with δ > 0 tries to pump a continuous leak X^(δ) through the sieve; the divisibility budget of the chain cannot pay for it. The whole of quantization is here in one line: a nonzero integer has absolute value at least 1, so a quantity that must be an integer, is divisible by that much, and is analytically that small can only be zero. Something must yield, and it is δ.
---
## V. The Instrument and the Kernel
If a picture like this holds—via Alpöge's chain or a successor built on the same fault line—it tells us *why the Riemann Hypothesis resisted for 167 years*: it stands on the seam between the continuum and the discrete. Pure continuity is too loose—waves can always superpose into counterfeits. Pure combinatorics is too local—counting primes one by one cannot see phase cancellation at infinity. The primes are an *arithmetically quantized wave system*, where global stability (δ = 0) is enforced by local indivisibility.
A machine-checked proof, if one verifies, certifies that a theorem is true; it rarely shows why the numbers could not have been built any other way. Analogies, sieves and counterfeits do the opposite work: they cannot seal the lock, but they turn a wall of inequalities into something you can hold in your mind and hear.
Between the kernel that guarantees truth and the instrument that reveals necessity, the primes come back into focus: a single tuning fork, twisted once by symmetry, floored by the place at infinity, and quantized by arithmetic so that it can never leak.
Are the 100k lines of lean beautiful music to your mathematical ears?
Does any mathematician even understand the proof?
I don’t find this beautiful or interesting at all, someone just pointed 1000 Claudes at this, said “go”, and it solved it. But we have no idea what it did and it doesn’t change anyone’s lives.
All that happened is the thing we thought was true already, is now proved to be true, by brute force, but we have no new insights or understanding and no practical applications.
Great I guess, meanwhile San Francisco and Los Angeles have more homeless people than ever and we can’t begin to solve that problem.
You have some interesting theories. Can you explain why this would suddenly bypass the usual few centuries long timeline between new math and applications? You somehow seem to have a theory that this general delay has changed.
> The precise timing is horrible. If I could have, I’d have finished this all
much earlier. ... There is an obvious cynical view, to do with the timing, but it is false, and I am pleased that I am in circumstances where it is obviously false.
It's not the kind of result you are going to sit on, and anyways not helping the stock at all today (why should it - do you choose your coding model, or decision model, based on it's math ability?).
„Dijkstra described a fictional company of which he served as chairman. The company was called Mathematics, Inc., a company that he imagined having commercialized the production of mathematical theorems in the same way that software companies had commercialized the production of computer programs. He invented a number of activities and challenges of Mathematics Inc. and documented them in several papers in the EWD series. The imaginary company had produced a proof of the Riemann Hypothesis but then had great difficulties collecting royalties from mathematicians who had proved results assuming the Riemann Hypothesis. The proof itself was a trade secret.“
https://en.wikipedia.org/wiki/Edsger_W._Dijkstra#Recurring_t...
Interesting story certainly. I don't see however why random stories are being posted that don't have anything to do with the news. The LEAN proof has been made available for all to read.
The story is about a company that has proven the Riemann Hypothesis. That alone is pretty close to the news nowadays. The proof by Alpöge might be open-source, but the technology used to generate that proof (Claude) is definitely a trade secret of Anthropic. So for me this story from the 1970s shows surprising parallels to our reality of 2026.
But there is more on the Wikipedia page. Dijkstra also used to quip: "The question of whether Machines Can Think is about as relevant as the question of whether Submarines Can Swim."
Some people are simply ahead of their times.
> That alone is pretty close to the news nowadays. The proof by Alpöge might be open-source, but the technology used to generate that proof (Claude) is definitely a trade secret of Anthropic.
The four color theorem took about 1000 hours of time on a the mainframe. Was the proof of the four color theorem suspect in this manner when it was released, and only became okay when everyday computers could run through the program?
Yes? The proof was initially met with skepticism and was subsequently refined and built upon.
I don't think it had much to do with everyday computers being able to run the original program.
I was asking in relation to this:
> the technology used to generate that proof (Claude) is definitely a trade secret of Anthropic.
The type of computer that could run the four color theorem was inaccessible to common people for a long time.
Yes. Why do people keep spouting off about this without learning any history :'(
Dijkstra was a devoted Christian. He also died years before seeing the state of modern LLMs. Religious people have certain magical supernatural biases regarding human consciousness. That is of course usually not a problem when they are working in any other field such as computer science or mathematics as long as they are not speaking on those topics where religion muddles ones thinking.
Funny, I interpreted the quote as explicitly rejecting human exceptionalism because I thought "yeah, submarines can obviously swim but it's a silly question"
Maybe he meant it as in both can achieve the goal in different ways, I am not sure, but it does seem like he is saying they are totally different things. In fact now that I think of that quote, any interpretation starts to become messy. I mean can submarines swim? They do seem to be swimming. I am not sure what answer he intended for this part of the analogy.
To me the implication of the quote is: you can call it swimming, or you can confine "swimming" to a narrower meaning, but it doesn't really matter. In other words it is purely a semantic debate.
He’s saying it doesn’t matter what you call it if they are functionally/observably equivalent.
This is the concept of faith. Science is a process of continuous disproval yet operates in a state of certainty.
Faith is fine as long as it doesn't clash with reality. And the reality is, the human brain is a physical system. Unless you can prove otherwise, consciousness is a property of some physical systems, so there is no reason to think some another physical system can't have consciousness.
Faith is fundamentally incompatible with science. It demonstrates an individual is capable of believing things without evidence, which is the antithesis of science.
This then becomes a source of mental perjury: once one leap of faith has been shown, you must doubt every other conclusion, as it may too be based on some leap of faith. It undermines credibility to an extent that I pretty much dismiss any who tries to combine the two.
I personally find the logical conclusions of anyone who firmly believes in a god to be suspect. We have enough scientific evidence to rule out the possibility with enough certainty that no one should be taking that wager.
Go ahead and hate me, but no one will ever convince me otherwise.
I believe the same, it frankly in any other circumstance would be called clinical insanity, but I try to be more diplomatic in saying this.
With mathematics and computer science work at least, we can be certain since they are formal and checkable regardless of beliefs of the creator, and as far as we know Dijkstra's work in cs is generally quite accurate.
It is not insanity. It is brainwashing. The same process produces irrational nationalism, racism, and most other -isms.
I agree with Dawkins: religion is child abuse.
EDIT: on the upside, brainwashing can be cured in many if not most cases. AFAIK, the cure for insanity is still a WIP.
You could have just stopped after the first sentence and had the same amount of content
Dijkstra described Mathematics Inc. as "the most exciting and most miserable business ever conceived".
In another part of the simulation: anthropic stock is down 10 perc the day they find the proof for RH.
So this whole thing was Ed's idea all along, but they went and took it all literally?
Guess I'd rather Mathematics, Inc. than War, Inc.... oh wait
Aren’t they both just subsidiaries of Greed, Inc?
These proofs really are quite huge. The proof is 1.3GB, ~15M lines of code:
Finally finished compiling on my PC (well, actually an HPC machine at my company, but it's the weekend, and I hope nobody is using it) - it took 2.4 hours on a 480-core Xeon and produced 89 GB of compilation output. I had to make some modifications because, for some reason, Lake bottlenecks and stalls verification. I worked around this by compiling the files directly instead of using `lake build`. I also had to increase `vm.max_map_count` for some later modules.
All 175,187 modules seem to check out with Lean 4.33.1, with no errors or `sorry`. There are tons of linter warnings - over 460k of them - but I think they might be caused by the way I invoked it. I haven't checked the actual axiom being proved, though - I'm not good at math :)
175,187 modules... a lot of them are repetitive.
It really seems he was rushed to publish and did not bother cleaning the code up at all. I am compiling the code and reviewing it, and there are 204 dead modules that are not referenced anywhere. And a LOT of repetitive code.
I think this should not be 17 million lines, I bet this can easily be under 1 million lines of Lean with a little bit of cleanup.
It is probably not that compressible. Anthropic FLT proof also has a lot of redundancies but my projection is without true mathematical refactoring it will be hard to get to 50% reduction in line count. One problem is that Lean, or type theory really, is truly verbose. In these parallely developed proofs if two agents use slightly incompatible definitions there are a lot of translations back and forth. Also FLT is following routes well established by mathematicians and it still got to 13 million lines. So I am actually surprised that GRH is not much bigger, as it reportedly ranges over a wide swath of fields.
Anthropic claims to have a 1/7 the size version somewhere. https://claude.dev/blog/safely-scaling-multi-agent-teams-wit...
Yes it would be interesting to see that version. If compaction comes from proof architectural improvement (what I called "true mathematical refactoring"), it could improve readability. On the other hand if it is tactical, for example using tall rw stacks that are incomprehensible to humans, shortness doesn't necessarily imply better accessibility. I suppose it has to be a lot of the first kind to get that much compression.
Even with all the resources he presumably have… interesting
Hence why it seems rushed. Backed by Anthropic he certainly can clean it, model access or $$$ is probably not a limiting factor; only time.
I don't know if Anthropic IPO being yesterday, or American Mathematical Society Executive Committee and Board of Trustees meeting being today has anything to do with the announcement happening today, but it's certainly interesting timing.
i would imagine the IPO has everything to do with it
Apparently a mostly human/human only unfinished proof on the hypothesis was rushed to publication upon hearing this news. It is around 1000 pages long. I don't know what the approximate ratio should be but fully formal propfs generally are much longer than natural language proofs. So probably this size is ij 5he ballpark of normal. If anyone sho writes lean proofs or works in formalized mathematics can comment that would be nice.
I am very skeptical of that human-only proof.
It more than 250 000 pages.
I am a hobbyist home-server-builder and would love to know more about you and your company. I love hearing cases where you as an employee get direct access to the hardware. can you share any more details? I'm kind of spazzing out right now with how casually you said some wildly awesome stuff.
I probably shouldn't say this, but I work at a company related to CPU manufacturing.
Part of my job is to ensure that HPC programs (largely LLM libraries these days) work on their next-generation CPUs and with new BIOS revisions. We test them using a shared pool of machines with various configurations. GPU machines are much scarcer, and we have to share them with other teams, so it's hard to reserve a time slot. I usually reserve a CPU machine around the clock so I can work whenever I want (I'm more productive at midnight), and nobody seems to mind.
I've worked here for almost a year now, and I love my company ;)
Intel? https://api.github.com/users/htzo
“related”
480 core xeon kinda hinted at that :D
Thanks for sharing!
Hmm, an unnamed CPU manufacturer using a ton of a specifically named processor. No idea where you could be working, I’m stumped. ;)
Well I guess you did say ‘related to’ manufacturing which widens it a little. :P
:)
That's a fun machine but there's similar (or even bigger) configs on AWS - https://aws.amazon.com/ec2/instance-types/u7i/
Expensive, but you can just try things!
Roughly what would the above cost?
u7i-6tb.112xlarge
$62.79/hr
https://aws.amazon.com/ec2/pricing/on-demand/
I run many high core jobs for various reasons, I often reach for many smaller machines rather than one enormous node like this. But not familiar with how lean parallelizes.
Jane str
Any company which self hosts in a data center has employees accessing to hardware. Every company I've worked at was on-prem. Why is this treated as an outlier? Is everyone on AWS these days?
No need to be snarky... having access to HPC clusters has little to do with "on-prem VS cloud" and is certainly an outlier in the prevalent SaaS world.
Not snark, I was just surprised. No need to be cynical. If we are talking about HPC clusters specifically, then of course I agree.
I have no idea what I'm talking about, but is it possible the lean program itself is wrong and in fact it proves something different?
It's possible, but the probability of it being wrong is much lower than a peer reviewed human paper being incorrect.
Most of the OpenAI proofs from the last big dump that have been looked at with some time don't actually match the lean claim now the media circus has moved on so I think the industry earned their lack of benefit of the doubt for a lot of people.
Those doubts are misplaced then. The natural language proofs have problems, sure. But that has no bearing on the correctness of Lean certificates.
Does Lean compilation usefully parallelize over that many cores? That’s a mini miracle by itself!
Yes but only for independent parts of the process. Having all of these modules helps quite a bit with that.
It really is quite slow, I've been having my dot and its 6 sub astras churn away at a proof for Z is Diophantine in Q, even 1/10th as many modules is an overnight affair on my 5800x3d with 128gb of ram. The first version I started on was 326 or 318 variables, but it's allegedly also found some in the mid 100s.
I've been having chat make a neat graph of the spam, it's neat to watch grow (hit the replay button) https://quartic-park.sneakyness.chatgpt.site
There's no real definitive measure of progress, all of the foundational stuff is handled and now it's working on the tricky stuff that is assumed away in the original that made me decide to let them loose trying to formalize.
Did you try #print axioms?
Thanks for that. Could you have an AI summarize the 460k linter warnings and decide if any of them are significant? Also, how much RAM was in the machine, please.
The machine itself has 1.5 TiB of memory; I did not check the free memory while it was running, sorry :(
LLM-generated text below: ``` The build printed about 460,000 linter warnings, and none of them affect correctness. Almost all are code-quality issues such as redundant proof steps, style, and unused variables. About 33,000 are uses of deprecated Mathlib names, which only matter if we upgrade Mathlib.
Memory per process wasn't logged during the run. The container's lifetime memory peak is 441 GiB, but that includes file cache and is only an upper bound. A typical module uses about 0.5 GB of its own memory, and the heaviest use about 8 GB. With about 280 compiles running in parallel, the build used at most about 215 GB of the machine's 1.5 TB of memory, based on spot checks. ```
Thank you!
Whoa.
One real problem with depending on Lean is that if there are any real bugs, then sufficiently-large swarms of agents are guaranteed to eventually come up with proofs that depend on the bug.
This comes up frequently in threads but it's actually a much less impactful issue than you might think. There are actually several entirely disjoint implementations of Lean, each with different bugs; it is incredibly unlikely an agent would be able to exercise the same bug in every single implementation.
Does a malicious proof need to exercise the same bug in all the engines necessarily? Why couldn't it exercise a different bug in each engine?
You trivially could through the principle of explosion.
That said at least one version is itself formally verified, at the machine code level if I understood correctly. It might still make mistakes but you're starting to get to a point that those mistakes become quite unlikely.
It's still a pretty tricky endeavor but we should know soonish, with some confidence if it is correct or not.
"There might be some mistakes"
"No, it has been proven"
"The prover may have mistakes"
"No, the have been verified"
"The verifier may have issues"
"No..."
Ad infinitum
But every round the minimum complexity of any potential exploit is dramatically raised, until the possibility of foul play can be reasonably ignored.
Agrippa's trilemma never fails.
Is that just an informal version of Gödel’s first incompleteness theorem?
I'd argue Gödel's incompleteness theorem is more of a formale example of the same.
Though it is technically different. Mathematicians were happy to pick a set of axioms, they were less happy to learn any theory capable of stating facts about basic arithmetic would necessarily be unable to prove all things it could state.
The successor function might have a bug.
I’ve heard it argued that equality has a bug. In real life it’s a lossy comparision as almost nothing is truly equal (except maybe atomic and subatomic particles?)
You could maybe address this through holding out the other implementations and using them only to verify?
I doubt that any of them are closed-source, so the clankers probably already know all their bugs. Who says they all don't share a similar bug that is necessary for the proof to work? No one is going to review 1.3GB of anything by hand, even if they could understand it all in a single lifetime.
Are there any that don’t rely so heavily on type theory?
you don't want to review the proof, you want to review the verifier.
> so the clankers probably already know all their bugs
I agree. So if the AIs knows some bugs, they will be using them to prove all kind of fake stuff, not only interesting stuff. https://xkcd.com/704/
Some of the reimplementations are actually verified, so guaranteed to be free from the sort of bugs found in the main kernel recently.
All implementations built upon the same spec no?
as per https://xenaproject.wordpress.com/2026/11/12/and-the-proof-i... before ever thinking about kernel bugs, you should think of the very basic "is the lean code even stating the problem, not even to say correctly"
Yeah maybe even more important point
Apparently they use the Mathlib definition of Riemann, which has been reviewed by actual human mathematicians.
And I understand that definition is 16 lines long.
It would have to also be really sneaky about it. A proof that's trying to invoke some bug will look very different from a proof that's the real thing.
I found a bug in Lean a couple of weeks ago with an LLM, in fact, but it was in its 'macro' system that expands code into statements, not the actual proof system. It wouldn't have proven anything false, but it could have produced statements that aren't what you intended.
Yes, this is starting to look like the infinite monkey theorem.
...if it's tasked to find proofs of False assuming that there are still bugs left.
Shell tip: You do not have to download each part separately, then cat them and then untar it. This does it in a single streaming pass without the temporary files:
shell tip: you do not need for loop.
curl -fSL https://alpo.ge/lean.tar.gz.part{01..14} | tar xzv
I think in this format this leads to shell expansions. For 14 thats not an issue, but I try to use for loops as it mitigates the arg max limits better. I believe if you literally quote the string, curl can internally do the range expansions too if you used square brackets instead of braces and passed it literally.
HN formatting has ruined your URL, try marking the block as code
I didn’t know curl allows multiple URLs like that.
That's the shell expansion, not curl. Works with everything.
Yeah I know. But shell just passes multiple URLs as arguments to curl. I didn’t know curl would accept that. In your example if there were more expansions it is up to mv to require the last one to be a directory. And many others just don’t accept multiple arguments.
It does but you have to use '[1..14]'. {1..14} is a shell expansion.
curl accepts multiple URL args
It does. And {1..14} would be a shell expansion not a curl feature ie the shell blew it up to 14 different args before curl saw it.
Your initial comment stated that to support multiple URLs it is required to use [] syntax. This is inaccurate, because curl supports multiple URL args hence shell expansion {} works just fine.
Where do the downloaded parts live before they're piped into tar? Do they remain in memory?
tar format is designed to be streamable. It writes the extracted files to disk as soon as they are received
you probably want to be able to resume one part if it fails to download
Trying to flex a shell tip on curling the Riemann hypothesis has got to be the funniest thing I’ve seen
An update: the Lean community is now maintaining https://github.com/leanprover-community/riemann-hypothesis and is currently checking the proof against several different toolchain setups. There are instructions to get the usual compiled files.
This seems to be norm in these top-level proofs. However, some comments here say that some of the modules are repetitive. In that case, can we not ask the AI to simplify the proof based on that to make it more tractable? Same as we often ask it to improve or refactor code that already works.
They just want to be as fast as possible for their claim to fame I guess. OpenAIs Proof of the Existence of Yang-Mills Quantum theory has like 250000 declarations in it and 60% is unneccessary.
It's possible you can't do that if this proof-producing activity has to do with the nature of language, and not the nature of intelligence.
If the modules are repetitive and also signify that no intelligence has taken place and no understanding has occurred (if there was intelligent understanding in place, the first thing an intelligent being with a theory of mind would do is strengthen their claim to a coherent thought by, well, making it more coherent) then we're looking at something that makes strong claims to what language is, NOT what intelligence is.
I don't think anyone's questioning the shape right away, the first pass through most agent's is always known to be bloated. Another user (jycheng) literally found 204 dead or orphaned modules a little way down. Knowing it works is the real breakthrough, let's not lost sight of that.
Size can be solved next, FLT went from 13m to 2m lines in a couple of months, the same will happen here for sure. The tough (or expensive part) imo is checking that the toplevel statement actually says GRH. That's the real review effort, not solving for size just yet.
I have never seen a more out of touch claim, at this size there is just as much chance there was a bug in lean, there's bug in the code which LLMs cant resolve maybe both, than assuming it 100% correct no questions asked.
I say that with the belief that this is likely correct. They should have solved the size how much would it take 3-4 days?
Who are running defense for here, I dont get the point of defending slop just because its something of value.
Code is meant for humans to read and make sense of, same with maths these concepts are for Humans even if machines do generate them. Otherwise machine should be talking to each other in some unknown super compressed language.
Say you solved GRH.
Would you really wait 4 days knowing that every big AI lab on the planet is racing for this result?
Worse is better wins again.
https://math216.wordpress.com/2026/11/20/hurried-announcemen...
This is what they intended to do, and did for months.
Yes my friend thats how things used to work, and should even now. Assuming all models can solve this stuff why cant we wait a few days to cleanup our work atleast and make more sense out of it.
Didnt OpenAI learn and started to cleanup the work a tad bit... ofc its not perfect but this thing is a joke.
I dont and wont understand the HN mindset of winning the race at any cost mindset.
This seems like a stock pump attempt that failed miserably.
I dont see the point we can all be civil and build better stuff now. Why try to one up others at the cost of throwing out slop.
Trust me if the value of your product can even be measured in days, then thats not a product worth investing in.
I personally think AI companies deserve to fail if they cant even understand that.
To clarify, I didn't say it's 100% correct, no questions asked. I said the opposite in fact: the expensive part of reviewing this is checking that the top-level statement actually says GRH. Actually exactly in line with what you say, that to me is where the skepticism should go.
What it "works" here to me is simple that it type-checks. Another user (hafeoz) even compiled all 175k modules upthread, with no errors and no sorry. Of course, a kernel bug is possible, but there are independent reimplementations of the checker, so you'd need the proof to slip past every one of them. A wrongly stated theorem is the more realistic failure, I think someone brought this up as well, and that only needs a page or two of definitions checked by humans, not 15M lines.
On readability: nobody was ever going to read 15M lines, or tbh even 1M after cleanup. That's what the kernel is for. Cleanup makes it easier to learn and build on, nobody's saying that shouldn't be done. My argument is simply dead modules are ugly, not wrong. OP himself says he expects it to be massively simplified.
As for "why not spend 3-4 days on it first": it went out today, and people were independently checking it on their own hardware within hours. That's arguably the better order, iterative progress first on successful ideas, not perfection. That's kind of what the whole YC ethos (who's page we're on, is. lol). Besides, a human proof got rushed out within a day because of this one, so the incentive to publish first was clearly real. We're on the same page here, just the 3-4 days of simplification first isn't the right way to progress imo.
In a sense that is not surprising. If there were a short proof to a famous 150 year old maths problem it would likely have been found by now.
Presumably as LLMs work their way through the rapidly shrinking list of Millenium Prize problems, it will be the case that none of them have concise enough solutions that they would have ever been solved by a human unaided.
I don't think this means that more concise solutions aren't available. This is likely more just the easiest path for an LLM to arrive at a proof.
That said, you are probably right that the most precise proof is too long for a human to generate unassisted.
Little bit like solving a rubrics cube. There's an infinite number of ways to solve it and you can use an infinite number of steps to reach a solution. Pros tend to use fewer steps, amateurs tend to use more.
> Little bit like solving a rubrics cube.
That's actually a great analogy for AI maths in general. There's one or more paths from the start state (unsolved cube) to the end state (solved cube). You have a set of rules that are verifiable. Deploy a swarm of agents to intelligently search this space, pruning branches that lead to nowhere. With enough compute, you will find a solution, but it might not be the most elegant solution.
'intelligently search' is doing a lot of heavy lifting here
But I do agree with you!
That didn’t age well? Alessandro Maria Masullo.
I think it's more that the proof requires several leaps in multiple fields/areas of mathematics, so I wouldn't be surprised if this was actually something like 100 separate proofs.
In the same sense its that the descriptive machinery isn't fully developed. If we wrote all of our proofs in a Godelian representation of statements we would also come out with tomes per proof.
I've had a similar thought. Looking back, will we think that Riemann Hypothesis was just an issue of scale?
It does build on all preceding author's work though. I wonder what % of cumulative researcher work would Claude make up?
Here's an allegedly human proof in fewer than 1000 pages, the result of 12 years of work. Still a draft but they published it in its current state because Anthropic published theirs.
https://math216.wordpress.com/2026/11/20/hurried-announcemen...
It is really good that those things happen right now IMHO, because we really have to ask the question of value in a different way. We have long been overstating the value of success. What is really the value of having proofed the GRH as we long had empirical support? (No expert here). Now is the 1.5M SLOC something of value? Could we derive a thousand times smaller human readable version much easier now? (Like having an oracle) . I hope this will lead to a time of fruitful philosophical and societal discussion, which might result in a better world (if we haven't destroyed the planet by then irreversibly)
> We have long been overstating the value of success. What is really the value of having proofed the GRH as we long had empirical support? (No expert here).
I think this has been debated many times before (I also suppose that this paraphrases Terrence Tao's take on AI), the value is in elevating the discipline. Those problems were selected because they were understood to be so out of reach that new constructs and theories would have to be developed. Those new tools would benefit mathematics at large because they might in turn unlock new, harder, or more generic problems to be solved. The value lies in the potential to unlock further discoveries, by having an increasing number of humans understanding, sharing and building upon each other and into ever more distant fields and disciplines.
The elephant in the room is that so far nothing indicates that LLM-generated proofs contribute meaningfully to building new frameworks and furthering our understanding of mathematics, therefore benefitting it. What we have is an inscrutable series of steps to prove that a single statement is true, plus the knowledge that LLMs aimed at a narrow problem can brute-force it given enough resources. The proof is technically only useful as it serves to solidify corollaries derived from the GRH by earlier humans. Will LLMs come up with interesting/important/meaningful corollaries of their own? That's a creative endeavour, not something that boils down to an optimization/path-finding problem.
Indeed, my comment does not appear to have aged very well
More realistically, they'll turn out to be actually really bloated bug reports for the framework the proof was made in ಥ ‿ ಥ
/Edit: I'm not even "against" ai value. For business proposes it's basically a match made in heaven... But their ingenuity and novelty is really overstated. They're really average. If you consider their output good, you're probably incompetent. If you consider it garbage,you're probably better then average. But for business, average is good enough hence me saying being not "against" ai value
Indeed, just ask LLMs to suggest products name or domains that sound good, are easy to pronounce, suggest some immediate meaning for a certain concept, are somewhat punchy, and have other qualities that make for a strong name.
LLMs are virtually useless at this. I routinely come up with much stronger names, and when I tell the LLM, they gush over it.
And the human "book" is 250k pages (!)...
From Alpöge's remark paper:
> Claude cutely called it the GRH... "book" — I looked at Theorem 1 and it began by noting that because the reals are totally ordered, if σ = 1 then σ > 1 or σ < 1
I respect that he decided to just post the lean proof and not bother others with 250k pages of badly written math that nobody would be expected to read anyway and would only serve as some kind of vanity thing. It is more respectful compared to some other approaches lately, imo.
Maybe the future of mathematics is gonna be AIs spitting out thousands lines of lean and human mathematicians trying to decipher them (with AI)
More concise, without bash, find, xargs
Correction:
(Accidentally stripped out "%02")
Could a math proof really need 15M lines to write out? I don't buy it.
Why not? Whenever you need to enumerate arbitrary many cases you can make it arbitrary long.
Few people would consider it a beautiful proof though.
I like the idea that the universe isn’t necessarily beautiful and orderly, because my life has seldom been orderly. “Imagine a spherical cow” is fine? But there are no spherical cows.
The assumption that the universe is orderly, has beautiful laws, and doesn’t have potentially one-off events or rules is a religion.
I have had this thought experiment for decades. “Imagine that once every 1.2 million years a grand piano spawns approximately four miles from the city center of modern day Phoenix. That’s just the rule. That’s a law of the universe.” There would be no way to measure this because if it happens to happen during our time it would just be a weird event that nobody could explain and we would have to wait another 1.2m years to see it again. Ain’t nobody got time for that? So all of our explanations for this would be wrong and would never be able to explain this or predict the long term future of the universe because hypothetically our universe would be dominated by the mass of these grand pianos eventually. Maybe beings then could piece together what had happened then? But likely not. They’d be just as baffled as we are.
Parsimony makes sense to us, because we’re simple creatures that live in finite time with finite senses. There’s no reason that nature should favor simplicity over complexity other than our own biases.
> There’s no reason that nature should favor simplicity over complexity other than our own biases.
No reason? Isn't the fruitfulness of this so far itself a very good reason? We've taken this approach and successfully explained and predicted so much of the universe already. How is that not a reason to keep going?
Now did I say we shouldn’t keep going?
I’m just saying that the universe is under no constraints to be understandable to humans, and our human limits might impose limitations on what we can model.
You called it a religion, and the main thrust of the response was that the fruitfulness thus far was a reason to not consider it a religion, but rather a rational course of action. It seems you got hung up on the phrase “keep going”.
I don’t necessarily agree with you, bur I really enjoyed reading that.
Well, I appreciate it. In your defense, I'm not sure that I entirely agree with this take either? But I think it's important to consider that we may not be clever enough to tackle all problems.
Like, my dog is pretty clever, she's a good dog too! She can roll-over and shake, and after much work I can get her to "stay" for a limited period of time. She's a good dog.
She's not smart enough to do calculus I don't think, and even if maybe she is physically smart enough to do calculus there's no way I could communicate the concepts to her. So... what's like that for us humans? What's like that for our experience of the universe? Like, I can look at a spreadsheet, and do some math on various columns in a dataset. I can do the SVD here or there.
But... I cannot really intuitively understand something in... say, 30 dimensions. The best I can really do when I feel really clever and think really hard, is maybe 4 or 5 if I'm being generous as to what I consider a "dimension." If I do something like say, "ok, picture a cube, and then visualize the heat on the cube as a color scheme, there's a burner below the cube." That's "kind of" like visualizing 4-d, and maybe I can add some roughness to cube too, to add another dimension to the data.
But that's the best I can do and the temperature is not even a spatial dimension . Meanwhile, it's not uncommon for like 10 dimensions to be useful in a dataset for prediction of some value - maybe even more. So, what are we not smart enough to understand? What are we simply too dumb to truly get? What will we never be able to understand unaugmented? And maybe parsimony and the "unusual effectiveness of mathematics" or whatever is that? But maybe... math isn't unusually effective, maybe the universe is substantially more complex than we can even comprehend? And our simple first order approximations are useful, but miss massive amounts of the underlying geometry/structure/system that makes up reality?
I guess, that's kind of the Donald Hoffman take on it right? This is the interface, but our interface kind of sucks.
Dogs might know calculus: https://websites.umich.edu/~mconger/dhsp/DoDogsKnowCalculus....
Man, I wish my dog could take the derivative of a polynomial! All mine can do is count one treat or two treats!
This is roughly the position of Charles Fort in the late 1800’s when he started collecting reports and compiled them into his "Book of the Damned", suggesting that there’s lots of your pianos happening - rains of frogs, teleportation, clairvoyance, etc - and scientists discount them as valid evidence because they would ruin nice clean simple theories. i.e. scientists pick and choose the evidence that lets them conclude a simple theory.
This gave rise to the term 'Forteana' for unexplained happenings, and The Fortean Times magazine. I think this stuff likely felt more plausible in the 1880’s before much understanding of electric fields making hair stand on end, drafts and flickering oil lamps being common, before atomic theory and the electromagnetic spectrum and radioactivity and thermodynamics when there were still a lot of big unexplained things around, before tape recorders, point-n-shoot cameras, then smartphones, were invented.
The ubiquity of smartphones on Earth selling at a billion a year and still no convincing recordings of Bigfoot, ghosts, UFOs, force fields, teleportation, is evidence indicating absence despite the saying “absence of evidence is not evidence of absence”. If you aren’t warm, that is some evidence that there’s no fire. If we had a lot of independent witnesses and videos of your piano appearing we would still lean towards it being a magic trick, hoax, CGI, because we have agreed that you can’t create energy and can’t teleport therefore it didn’t happen as it appears.
I haven't read Carl Sagan's Demon Haunted World but I think it's a counter position, look at all the nonsense humans have believed over history and still do believe today, and about the only way we can divine the real from the nonsense is repeatable tests - "science as a candle in the dark".
But then Occams's Razor is no longer a guiding principle, which is fine, because the Universe has no obligation to be simple.
Also, claiming a teapot orbits does not put the burden of proof on the claimant, because, again the idea that the one making an extraordinary claim MUST offer extraordinary evidence is just that, what we think is a nice approach to take. Truth does not care who has the burden of proof.
That's why miracles are not to be brushed off; maybe you just never saw one. Maybe they only occur under some weird circumstance the vast majority of people do not experience. Not say there is material proof of them, but mostly the objection to them is based on politics and adherence to the religion of "we have the universe all figure out" We could be living in a simulation, and "miracles" are a bypass of the simulated reality.
Hey, I'm a fan of Occam, but.... I recognize that maybe, kind of sort of... what if we're at the limits of what can be described by Occam? Or similarly, what if like, we can make crappy models that are parsimonious and accurate models that are not. What if the universe is parsimonious when you don't look very close and extremely chaotic and weird and inconsistent when you do? That would surely track much of my lived experience?
But who knows? I'm surprised most days that there's something rather than nothing, that I exist at all (it's quite strange to be anything at all), and that math and computers work. I'm quite glad they do? But yeah, things seem pretty damn miraculous to me most days.
There are many things I don't know, but this much I do know: There aren't any teapots orbiting the sun deep in interplanetary space.
How do I know? Very simply. There's overwhelming evidence against the idea and no evidence for it.
You don’t know that. You believe it - and to be fair, I believe it too - but you don’t know it.
The parsimonious answer feels good to us, and I think mostly it bears fruit - but there’s no guarantee that parsimony is the only way to model things accurately
This gets into what the word know even means. Nothing's guaranteed. I know it as much as one can know anything.
It takes thousands of hours of study for a human to get to the point of understanding what the RH says and means. It isn’t hard to imagine that there are mathematical truths that require far more than this to understand them.
Pretty much every major result in mathematics - assuming ZFC as the background theory - requires thousands and thousands of pages to formally prove from the axioms of set theory.
An infinite number of math proofs really need more than 15M to write out. Could that include some easy-to-state ones that we care a lot about? Yeah probably.
What do you think the upper limit on the size of a math proof is?
Not sure why you got downvoted but this is actually quite deep. I guess it depends on the mass of the universe. Since our universe is approximately a black hole, its mass determines its area and the number of bits it can store on the surface through the holographic principle. If the universe is not infinite there is a physical limit to math.
A finite universe does not imply a finite limit to the size of proofs of true statements.
There are proofs that might not be capable of being written down in the material universe, even if we assume a plank volume per symbol. They would exist nevertheless.
I guess the upper bound is the the amount of information in the universe. That seems to be about 10^80 bits.
https://www.sciencealert.com/physicists-quantified-the-amoun...
tbh i was the student in undergrad maths classes using induction to prove trivialities bc nobody handed me a “book of axioms” and my intro to writing proofs was using Coq. let’s cut some slack to our autistic LLMs — I am sure in six months we will know how confident we are in this proof; no need to rush.
Do you NEED 15M lines? No, it's almost certain this proof can and will be reduced, but the goal of an initial proof of a theorem like this isn't conciseness, it's to actually... prove the theorem.
Fermat's Last Theorem was initially proven in Lean in September and that proof required 13M lines, it has since been reduced to 2M lines and it will likely be reduced even further. Anthropic claims they can reduce it down to less than half a million lines by doing similar techniques to what software developers do... find common and reusable snippets and factor them out instead of repeatedly proving similar sub-theorems over and over again. There are also opportunities to replace lengthy sequences of explicit reasoning with automated tactics, reuse existing mathematical libraries including the Lean standard library, and reformulate arguments at a higher level of abstraction.
And presumably you need hundreds of GBs of RAM to compile it. Would be nice if someone with a beefy machine could confirm?
About 1 1/2 hour in, my estimate is it takes about one day on a single DGX spark. We will see.
Proof verified. Less than a day on one spark.
You could run it with limited RAM, just not many Lean processes in parallel. Just need some disk space for the compiled files.
some randos on twitter I've seen are on it as we speak, but the real deal would be other mathematicians confirming the proof is written correctly.
Yikes. I'm definitely not taking this one for granted until this has been reviewed thoroughly. That's a lot of room for shenanigans/bugs.
Apparently healthy skepticism is a downvotable offense now. Not that I care about the karma. It's just interesting that we live in a time where taking anything other than the most extreme position is unacceptable to some people on here.
I guess I'm supposed to either accept the galactic proof, which likely hasn't been fully read by a single human yet, as gospel and speak breathlessly about the singularity, or I should be insisting the proof will fall apart and make accusations about deliberate academic espionage.
Given the level of verification the result has already attained the skepticism is verging on the pathological now.
I wasn't aware that my opinion yesterday had to take into account events that hadn't happened yet, but noted.
Why can't they ask AI to shorten it and make it beautiful? Is it not that intelligent?
If it was intelligent it would want to shorten the proof and remove broken pieces, because that would strengthen the argument that the proof is good… assuming it had a theory of mind capable of understanding that another thing is evaluating its performance. It hasn't, isn't… something else is underway and we are observing the parameters of that.
Or if it was really intelligent it could say that shortening it and simplifying it is a waste of time since it is obviously correct the way it is.
Isn't there a way to reduce this? What's the reason for such a massive size? Are there "already proven" proof libraries in Lean that the LLM is not using like reinventing the libc from scratch?
The only verified “already proven” library is mathlib, which gets you up to about half way through an undergraduate degree. It may be substantially more efficient for Anthropic to direct Claude at making a Lean library consisting of all of known mathematics (at least up to the end of the 20th century), and then use that as a base for their hypothesis. Instead of banging together Lean code at considerable expense until something magiced out, Claude would have been able to build on top of the works of the math community and produce a proof that was actually readable. Claude would also be able to reference known proofs while it reasons through the problem.
why a lean proof is so big? it needs to be compiled and tested with a lot of cases or what may it be?
Five years ago very few proofs (outside of logic) were formalized.
Proof assistants existed for 50 years - they are not new. But it's REALLY hard to formalize something which is not just logic.
Even just logic is not trivial. Russel's Principia Mathematica needed 360 pages before it can prove 1+1=2.
prove that (Russell's definition of 1 as a set) (Russell's definition of + in terms of set operations) (Russell's definition of 1 as a set) = (Russell's definition of 2 as a set), that is
He did not prove one apple and one apple make two apples, he proved that his complicated definitions of arithmetic using sets were not completely broken.
For whatever reason, the LLMs currently generate very long proofs of results. For example, they often rephrase earlier arguments rather than citing them, or repeat their own arguments. (I've experienced these myself.)
It'll just take some work to simplify this.
From Alpoge himself: "My intuition insists it will be massively simplified, likely immediately.”
This is true for code, even frontier models, they will often not use a library, or not modify the library but rewrite an almost identical function(s), etc. It's very interesting and needs constant vigilance
Sounds like a massively parallel search of all possible paths since they have so much compute now
judging by claudes review because it actually rebuilds and proves a bunch of stuff in mathlib that wasn't sufficiently formalized, but were required dependencies.
I suspect it includes a lot of formalization of existing, peer-reviewed mathematics. Including these bloats the code but makes the verification more self-contained (less AI-written unverified inputs to lean).
I'm absolutely positively sure that there are no sorrys and no theorems taken as axioms _anywhere_ among those Windows 11 source code-sized Lean proofs.
I mean… that would be a trivial check with `ripgrep` et al. I’d be more worried about other sneakier tactics that LLM’s like to do as to move the goal posts (prove something trivial posited as grander, take arguments in a function head but not use them in the body etc)
I also am absolutely positive of that, since Comparator checks for those.
Yeah, this smells like BS. A "proof" that no one will understand within the next 1-2 years, and even then it's questionable that anyone could ever seriously hold 250k pages in their mind, and that no single mistake or oversight has slipped in. This is a haystack with an unknown amount of needles in it, and gaining certainty that that amount is zero, is a whole other task.
It will be condensed soon. I guess he just wanted to put it out quickly
Shared it in the other thread too.
Here’s a link to an interview with Levent Alpöge (in Turkish) from 2014, when he just graduated from Harvard at the top of his class.
He mentions solving the Riemann hypothesis is his life goal and he wants to work on it, but he’s not sure if it can be done in his lifetime. Way to go :)
https://videonuz.ensonhaber.com/izle/harvard-birincisi-leven...
Am I the only one waiting around for verification?
Even 12 hours ago, it had already been verified by 4+ independent Lean kernels, one of them ("con-leche") itself formally verified. https://leanprover.zulipchat.com/#narrow/channel/113488-gene...
Would anyone reasonably assume that the author wouldn't simply prompt the LLM to verify this thing across all kernels before even submitting it anywhere?
Even all things aside, at least I would not assume to have checked single thread kernels as those take very long and having access to a supercomputer won't help you in that.
The point of independent verifications is to not have to rely on this kind of assumptions. I don't think anybody would reasonably expect with any high probality that there was a lean bug in this.
I am seeing single threaded verification estimates in hours, maybe days. If I was the researcher it's a no brainer.
If I was the author I'd just run `lake build` and let everyone else worry about kernel bugs etc.
The Lean Prover community has validated that it's correct now: https://leanprover-community.github.io/blog/posts/riemann-hy...
> On November 20, 2026, Levent Alpöge released a proof of the Riemann Hypothesis as a Lean library comprising 17 million lines of code.
> We have full confidence that Alpöge's proof of the Riemann Hypothesis is correct, and we congratulate him on this achievement.
This is linked to from the Zulip chat here: https://leanprover.zulipchat.com/#narrow/channel/113488-gene...
Mentioned on Terry Tao's blog: https://terrytao.wordpress.com/2026/11/20/unforced-euler-sin...
Even if we discover verification errors, the accomplishments of Levent & Alessandro provide the mathematics community with several new tricks that will be leveraged for years to come. It would be particularly exciting if both proofs end up getting verified using different routes.
The reason I'm eagerly anticipating rigorous verification is to see if Levent's belief is correct that the same concepts can be used to prove extensions of the main hypothesis.
Does it even matter anymore? I mean of course it does, what I mean it that even if it's somehow incorrect (which is important) wouldn't he try again until it does pass and that given the problem space, search tool and compute available it becomes tractable (which might not be)?
I'm not for the field but I can't help but wonder how does one backtrack from the "solution" to the meaningful learning.
It changes nothing. You can just assume it's true and work from there.
Mathematicians need to stop toiling on proofs and acquire taste for meaningful questions.
Right now, people graduate top of their Harvard class and work on a question that was invented 150 years ago and everyone just assumed to be true. Kind of meaningless, right? Could have just waited three years for a full AI-only proof without human intervention.
People were working on these classical open problems because they were stimulation for creating new interesting directions and theories. Even this AI proof is probably using some of this work.
And you cannot acquire taste without getting dirty.
How many mathematicians have you talked to? Pretty much the opposite of what you are observing here is my experience.
I'm waiting for validation. Sure "it was verified by independent kernels", but is it really GRH or is it something that looks like GRH?
The Lean Prover community has validated that it's correct now: https://leanprover-community.github.io/blog/posts/riemann-hy...
> On November 20, 2026, Levent Alpöge released a proof of the Riemann Hypothesis as a Lean library comprising 17 million lines of code.
> We have full confidence that Alpöge's proof of the Riemann Hypothesis is correct, and we congratulate him on this achievement.
This is linked to from the Zulip chat here: https://leanprover.zulipchat.com/#narrow/channel/113488-gene...
Mentioned on Terry Tao's blog: https://terrytao.wordpress.com/2026/11/20/unforced-euler-sin...
Well, I’d hate to tell him he didn’t solve it. The Claude model did.
He is not the only one with access to Claude.
Isn't he an Anthropic employee? So this was potentially done with access to an unreleased model and/or an insane budget out of reach to most people.
Well first we dont know if he use internal or public model, even if he is their employee.
Second, a lot of other employees also have access to internal model, but AFAICT, he didn't claim anyone co-author with him, expect Claude, forgive me if i'm wrong here.
He solved it using an internal model I think, so yeah
That was fast. I thought we’d have more time.
I wonder if we’ll have room temp superconductors and fusion this time next year.
If this holds up it seems as good a day as any to call the end of one time and beginning of another.
I wish I could be fully happy about this new era instead of worrying if I’ll be alive in 5 years or not. If there was any time for Anthropic to try to coordinate a pause, it feels like now or never, they seem to have a pretty strong lead at the moment. But probably they won’t and can’t, and we will just put the fate of our species to a coin flip.
>I wonder if we’ll have room temp superconductors and fusion this time next year.
No, because to advance in a real science you need to conduct real experiments. And it's a lot slower than multiplying matrices on a million GPUs.
If multiplying matrices on a million GPUs reveals a unique lattice structure that in simulation enables room temp superconductors, the real experiments would immediately begin.
Look at LK-99, it felt like a global effort was inspired at the time to verify room temperature superconductivity.
It might not even require a specific lattice structure to be identified - there is still quite a bit we don't understand about regular superconductors, a large breakthrough could materialize RTS's within a year. Producing them at scale is a separate issue.
Not my area of expertise but I do work in scientific computing. I know that we can do simulations like this for meta materials, where to some extent you just have a really large search space that is not bridged by logic but by brute force, possibly aided by feedback from physical experiments. Assuming we have similar for semiconductors, it’s not clear to me where AI clears the compute bottleneck.
We already have supercomputers to run simulations like this, and can already spend millions of dollars on compute to answer hard problems. AI may help improve experimental design but this is something humans can already do reasonably well. I’m less optimistic we can derive similar results as we have for math, but I’d love to be wrong.
Math is easier because it's text and symbols only. In physics experiments you have non-textual input and this is not the LLM stuff for now.
LLMs are already multimodal, though? Plenty support native image inputs for instance.
But also: human physicists mostly act on symbols, too. They take measurements using sensors, which produce a bunch of numbers, that they then tease theories out of. This seems in LLMs' wheelhouse.
This is what people seem to be missing when they extrapolate all LLM mathematical brute force discoveries to any X or Y scientific problem. Most real-world stuff is not as cut and dry as formal logic and math. Computer science is applied mathematics and almost all logic-based, it follows that LLMs would be good at this. Lot harder to model proteins or materials science this way.
That said the stock price juicing counts on you seeing these discoveries and then attempting to extrapolate bespoke frontier models' success in math (which is dubious in terms of methodology and tailored training anyways) and immediately jumping to the conclusion that AGI and huge scientific advances are right around the corner and can have us all living in fully automated luxury Yudkowsky MIRI effective altruist utopia. It's a farce.
And yes, someone is going to comment on this by saying "moving the goalposts, in 2022 models couldn't even do this." Yes, it's incredible how much throwing money at certain logically achievable outcomes helps achieve those outcomes. It is not clear how it helps achieve random wish casting by odd cult-like CEOs, however.
but what if AI notices a higher order pattern that we humans with our measly cognition doesn't?
I’m sure LLMs can and will help with physics research. I’m much less sure you can frame most physics problems in a way that will succumb to 10,000 agents running in parallel.
This optimistic forecast is based on an assumption that we already know everything necessary to find a room temperature superconductor, so an LLM just needs to connect the dots.
We already have a wealth of collected data from superconducting and fusion experiments. If agent swarms can be 10x or 100x better at interpreting this data and using it to build simulations than humans are, maybe we don’t need that many more experiments after all. And the ones we do need can be much more precisely aimed at the remaining gaps than human-designed experiments.
You can get only so far by analyzing existing data. The real power of AI has been always unlocked (chess, programming, maths) by letting AI propose solutions and get feedback. For natural science that means conducting experiments, whether physical or computational.
We're limited on materials and other technology which will need a lot more time for fundamental maths and physics to diffuse to. But yeah, the future is bright.
Not with a sufficiently advanced simulation.
An LLM is not a physical simulation of anything.
I never claimed as such. It would be acting /with/ and in the confines of a simulator.
The sufficiently advanced simulation is QCD and I believe the state of the art is a few thousand particles running on a top 10 supercomputer.
Particles meaning electrons, protons, and neutrons.
There is every reason to think that AIs will be able to make dramatic improvements the efficiency of numerical quantum simulation codes.
Also, for superconductors, QED is sufficient, QCD is used to simulate hadrons, not electrons.
It's worse than that, it's like asking if we'll have perpetual-motion next year.
Just because you can search faster for something, it doesn't mean that that something ever existed in the first place.
I think a lot of LLM stuff is just speed-running alchemy. Instead of wasting a lifetime, computers can now waste thousands of lifetimes.
If you think people are sour on datacenters now, just imagine how pissed they're gonna be about AI training facilities that occasionally explode or leak novel kinds of life into the neighborhood.
> fusion this time next year.
Aren't most of the remaining problems related to nuclear fusion engineering ones? Like building reactor walls that survive neutron damage.
It's a problem of scale and engineering. ITER is already way too big and complex, to the point that it may never be completed. They might turn it on, find a leak and be stuck with a huge paperweight. To have viable power generation you have to go even bigger.
Nah. The ITER project was prev-gen by the time they began to build it. The project underwent some upgrades, but it's not at the cutting edge. High-temp superconductors and plasma stabilization algorithms allow to make smaller reactors.
High Tc superconductors can make tokamaks smaller, but they're still much larger than fission reactors of the same thermal power output. While ITER has 400x lower thermal power density than a PWR, ARC is just 40x worse.
Basically there's no hope even high Tc superconducting tokamaks can be competitive sources of energy.
You might say "but make the magnetic field even larger", but most of the mass of ARC is the steel structure to resist JxB forces, so much further increase in B is not practical. There are also minimum sizes for fusion reactors set by the need for sufficiently thick blankets to stop the neutrons; that's independent of magnetic field.
Solar power plants have even larger footprint per kW, it doesn't stop anyone.
That's apples-to-oranges, since PV omits other components, and because the high redundancy and modularity of PV makes construction much easier. DT fusion vs. fission, the balance of plant are similar to each other.
(If you meant concentrating solar thermal plants, those are also uneconomical.)
Decades of detailed engineering studies of DT fusion reactors (like ARIES) have had them coming in more expensive than fission, as one should expect.
That is the narrative of the fusion start ups pursuising it. But, show me the plasma. Oh right, they don't have any. Might as well just be investment scams for now.
Or have stronger magnets, see sparc and arc from commonwealth fusion. Or a different method entirely, see helion.
Why can't AI solve engineering problems?
It potentially can, but you still need to build the thing in the physical world and test it under realistic conditions. Maybe I should have focused my post more around "materials engineering".
Same reason AI can't replace the engine in your car yet.
Engineering problems are solved by massive amount of logistics: finding and hiring the right people to be in the right place at the right time with the right tools and raw materials that need to be purchased at a physical location and delivered by cargo ships and trucks driven by truckers across the country. People get injured, sick, go on vacations. Trucks crash sometimes. Cargo ships get delayed. Tools break and need replacement. Materials turn out to be wrong quality or quantity.
This is an entirely different world from software development or math where you can try different solutions and know if they worked or not in a matter of seconds.
Plasma control is still unsolved
engineering problems are often just math problems.
the next YC call for proposals can include LEAN for building reactor walls that survive neutron damage
There is also the problem of how to stop bitcoin miners from draining the oceans dry
I do wonder how well a lot of this translates into physical sciences. It's one thing to be able to RL on lean but physical sciences don't have such an obvious quick feedback loop. A lot of the scientific results have been closer to the theory side which isn't unexpected from the sheer mathematical capability.
if you look at the current "high" temperature superconductors it becomes clear that a peculiar atomic arrangement is the key yet we don't have the ability to individually place atoms (yet) so we must rely on entropy to place them into said arrangement for us which restricts the space we can even test.
so the task for an LLM here is doubly difficult and the feedback loop is of very high latency. I do wonder if someone is trying though. simulate physical process and entropy/annealing -> simulate superconductivity. and very few working examples to work with.
So a couple of purely informational question coming from a layperson:
1. Is there a good theory of high temperature super conducting that we can easily simulate? My furthest understanding of this was that BCS was insufficient.
2. Does such a theory impose limits on the temperature that are achievable?
3. Do we know if such a theory is complete in some sense? (e.g. if such a theory predicted no super conducting material at a temp K and pressure P, we would know that there could be no such material at temp K and pressure P based on reasonable assumptions)
There is no such generally accepted theory.
We will not. Physics experiments take time and money to complete.
DARPA/DoD have been collaborating with the labs for years, with an infinite budget and access to compute, it's bound to happen.
DARPA and DoD budget are far from infinite. Their combined R&D budgets are a fraction of what the AI labs are spending.
They definitely don't have a strong lead, trading places with OpenAI and occasionally Google every few weeks
A coordinated pause needs to come at the national level
International level. National would be a great start though.
We're in a cold war
> I wonder if we’ll have room temp superconductors and fusion this time next year.
Still waiting for the AI advancements which are not just taking away work from SWEs and mathematicians..
> Still waiting for the AI advancements
At this point this line of "yeah but what about..." is getting silly. We've had 2 millennium prize problems fall this year, plus a lean proof for FLT (another effort worth 1m$) and people are still "yeah, but what'd be cool is if..."
Dunno, at this point I feel like people are trolling. This is absolutely insane pace, and just last year we were hearing how "there is no more data", "models are stalling" and so on. The negativity and goalpost moving is getting tiring.
I think it is because we see these headlines, but as a layperson it has no impact on my life.
I don't see the value and I don't see the world changed.... yet
It will come, don't worry about that.
> I think it is because we see these headlines, but as a layperson it has no impact on my life. I don't see the value and I don't see the world changed.... yet
You could try to be happy for others who see the value.
I remember reading about the counter example to the Jacobi conjecture in July. That was cool, but many said it was an exception, because AI is only good at finding counter examples. I believed that at the time. But four months later we have lean verified GRH.. I don't feel like it's possible to make any predictions anymore. This might be the peak and in four months AI is more or less the same, only cheaper and with more compute, or we might see more insane unexpected advances. But one thing is for sure, if AI can formalize and prove GRH there are huge swaths of math just waiting to be "cleaned up" and completely formalized. It's just a question of spending the tokens.
As a scientist the speed of developments has been breathtaking. As a layperson there has been 0, if not negative, change in my life because of AI.
Well, you know, I think it would be in fact actually cool if it solved, say fusion or climate change. Not just solve math problems without providing understanding or take all joy out of software jobs.
I'm not saying the technology isn't amazing. It's completely surreal how amazing it is. I just wish it did better things.
Actually agree, it's dancing around the edges, until real world serious problems are being solved, I do think there is some level of theater going on, it's impressive for sure but even as a SWE, I still have my job for now and really if anything AI has just given me more work to do, sure that might just be limited, but I can only go off my real world experience.
Surely faster than light travel is worth more than a million dollars.
(Yes, yes, I know it's Fermat's Last Theorem)
Why? Even as a SWE there's plenty of software that I want in my own life that I simply don't have the time/attention to make, or even to try to commission someone else to make (e.g. I've been making little educational games/demos for my kids). Having something that does what I want in a few seconds when I have the attention is fantastic.
"Taking away work from SWEs" means enabling everyone else (or even SWEs) to do things they couldn't do before. This is obviously an improvement.
I have a todo list long enough I don't even have time to prompt for all of it...
If I could justify enough tokens, I'd point an agent at my harddrive and tell it to formalise all of my old notes and goals, and spin up agents to build all of the things I've made notes about that I would like to exist. It'd still likely take years...
I just kicked off Kimi building a game based on a concept I had kicking around. I wouldn't have time to build the game or even write the spec, but I had time to type "write a game based on [paste of the concept]" to see what falls out of it.
I can't imagine ever running out of things I want, because every thing I finish tends to result in my todo list growing with new things.
They are already there. Sure you are not seeing positions disappear but hiring is starting to grind to a halt in a lot of places. Someone who might have needed an admin assistant no longer needs one and middle management in large orgs whose job was basically summarize and pass up will soon start to thin out through attrition.
You wont see whole sale replacement but you will see doing more with less.
I truly don't understand how someone can be interested in technology and visit HN without understanding - even broadly - the history of technological progress and how these things take time.
Yeah, I want the cure for cancer yesterday as well, who doesn't? This is the most significant result in mathematics in over a century.
They do understand. They’re afraid, and that fear leads to hatred of AI and the big labs, which leads to this kind of motivated reasoning that looks bewilderingly foolish from the outside.
People are scared.
Thanks for the free psychoanalysis, but you're completely off.
To be clear, I'm just disappointed and sad that AI mainly takes our jobs (or at least the satisfying part of it) and solves mathematics without increasing our understanding. You know, the part that's the main point.
However, if in exchange it solved, say, climate change or fusion, that I would happily accept. No sign of that so far.
You contradict yourself - you are sad AI will take your job but completely fail to understand why people are scared?
Maybe you are a well off rich kid or just cashed enough from faangs to be set for life, whatever. Is it really that challenging to muster some empathy to those who, upon seeing their profession disappear with 0 replacement in sight, are scared of the future since they don't know how they will even feed their kids, not even going into higher expectations of life?
At the end some math understanding is secondary at best when facing existential threats.
I think you either completely misunderstood me, or replied to a wrong comment.
> No sign of that so far.
Trend lines are a thing and progress does not happen overnight
Sure, and I'll be happy if/when it happens. So far still waiting.
We have a genie that grants unlimited wishes for intellectual output[1]. If you had witnessed someone multiplying bread and fish would your first reaction be to shed a tear for all the bakers and fishermen?
[1] Not to get too religious about it, limitations still exist
Does it? Can I wish for it to solve climate change for us?
Also, your analogue is off. If you had a bread and fish machine that made bakers and fishermen unemployed, while only producing bread and fish for the top 1%, I would not be very happy about it.
Climate change is not a challenging intellectual problem. It is challenging to convince others to do what is necessary to solve it. Could AI produce potential solutions? Sure, from memory. We already have them.
In anticipation of a followup, "can I wish for it to solve climate change advocacy for us?" It can provide research direction on nearly any topic but as soon as we are talking about actions in the real-world it requires more than purely intellectual output.
Anthropic and OpenAI make their models available to anyone for use. I call up cornucopias of bread and fish daily and you have the power to do this too.
Oh but climate change has tons and tons of very much intellectual sub-problems. Better battery technology is a huge one. How to optimally generate or build renewables. Even wild ideas like carbon capture, say through genetically modified trees that grow faster. That's what I'd like to see from AI, something that humans can come up with and solves a real-world problem.
AI is working on problems in all of these fields every day. Yes it can help. No it is no panacea.
Yes, and that's why I originally wrote:
> Still waiting for the AI advancements which are not just taking away work from SWEs and mathematicians..
Climate change is not a challenging intellectual problem.
100% wrong, if a 95% efficient solar panel could be produced really cheaply, it would be solved. If fusion power could be created, it would be solved.
>I wish for it to solve climate change
[laughing_greta.gif] they still do that? :DD
Don't worry. It will take work from other career paths, too. Let's just celebrate the advancement in human civilization altogether on new year eve.
> I wonder if we’ll have room temp superconductors and fusion this time next year.
No, we don't get useful things.
That's a depressing world view.
I don't know about room temp superconductors, but fusion has gotten a bunch of investment recently that make it seem like it'll happen before I die.
Specifically, REBCO superconductors are in production by Commonwealth Fusion Systems to make SPARC, targeting net fusion gain for 2027 with a tokamak design.
The NIF has been able to achieve ignition, repeatedly.
Whatever you want to call it, AI/ML/GPUs; point is, supercomputers today are insanely powerful for running plasma control simulations and software.
Commercial contracts for fusion are in place.
More interestingly, Marathon Fusion has figured out a way to generate money instead of energy which needs to get sold off at energy rates. Fusion moves subatomic particles around. Instead of generating power with it, they're going to use the technology to transmute elements. Like alchemists of yore. Just, like, with science and it actually working. They estimate they can make money from converting mercury into gold, but more lucrative is radium into actinium, and zinc into copper. Specific isotopes of those are used in the medical industry to help fight cancer and are very expensive.
That's more of an economic technology than a technical one, but money is always the issue, and if they're able to make money with their scheme, that money can be used to fund further fusion research.
Maybe not next year, but some time before I die, fusion power seems entirely possible!
You really shouldn't have gone out of your way to reply to that snide remark of mine, but thank you. Most of what you said went way over my head, but I'm glad people smarter than me are seeing hope.
> I wonder if we’ll have room temp superconductors and fusion this time next year.
That’s child’s play. I wonder if we’ll have star trek replicator by next year.
> star trek replicator
That's a walk in the park. I wonder if we'll be sipping life juice in comfy matrix cells by next year.
I’m using my replicator to conjure life juice to sip out of thin air. Thanks anthropic.
> I wonder if we’ll have room temp superconductors and fusion this time next year.
Tell me you've not worked in hardware engineering without telling me you've not worked in hardware engineering
I’m a robotics engineer with a mechE degree. Agents are massively speeding up my hardware engineering work, yes the latency is still much higher than software, but each iteration is far more effective when you have agents to help. You can simulate whatever you’re building in far more detail than was possible before.
Yesterday I built a whole new system for agent-human PCB co-design (skidl+kicad). Maybe you don’t feel it yet but massive hardware speedups of the same OOM as swe and math are very clearly on the horizon now. I mean right now on the front page there’s a new agent swarm designed CPU claiming massive performance improvements.
mean right now on the front page there’s a new agent swarm designed CPU claiming massive performance improvements.
Key word is "claiming" and there are plenty of comments on that thread pointing out that the supposed new design is actually just a very old design that was disgarded for practical reasons that are still not addressed and appear to have been hand-waved away.
I work in hardware engineering and just today after setting up the oscilloscope probes for Claude, it was able to trigger my fault scenario and read the captures to debug a flaky comms link, and write a small reproducer, all done autonomously while I worked on something else.
So you might want to check yourself since it's already helping massively in hardware.
Imagine flipping the switch on a fusion reaction that nobody knows how it works, but Clause says it will, and has 1 billion lines of code that supposedly proves it's not an H-bomb.
So dramatic
>I wonder if we’ll have room temp superconductors
>we
Luckily, this matter does not pertain to the permanent underclass at all.
soylent green is people
https://www.youtube.com/watch?v=xt4WBB4-uDc
I was when saw this in the morning. I thought I could no longer be surprised by AI progress. Yet I had not truly believed that Claude could prove a conjecture as challenging as the Riemann Hypothesis.
In fact, just this summer, I tried to write a short science fiction story, The Last Conjecture, about AI working on the Riemann Hypothesis — a bad, unpublished story that I wrote with the help of AI itself. The irony is that even the AI helping me write it could not believe that AI would solve the Riemann Hypothesis in any near future. In its own forecast for the story, the project drags on for nine years, all the way to 2035, and still never cracks it.
Just yesterday, on Friday, I was still telling students: “Maybe there is some physical law that limits how much intelligence one can get from current LLM architectures. Surely nothing in this universe is infinite.” Maybe I am right. It is just that on November 21, 2026, we have not yet reached that limit.
It was 7 a.m. I opened the curtains in my study. The sun was rising. It was a sunny day. Nothing had changed. Everything had changed.
Seeking some solace, I listened to a recent episode of the 3Blue1Brown podcast with host Grant Sanderson and guest Steven Strogatz (a mathematician at Cornell University who has done a lot to popularize mathematics). The two of them did have some hope that mathematics and the profession of mathematician would continue in some changed form, but they could not be sure at all.
Later in the afternoon, I saw that Scott Aaronson, a theoretical computer scientist at the University of Texas at Austin, had published a blog post about the news. It ends with the following story:
It was only a couple days ago that my wife and I had the following conversation:
Dana: I turned off the irrigation, yet somehow water in our house is still leaking. We need to…
Me: Look, a lot of the smartest people I know now think that recursive self-improvement and the end of the entire known world are coming in 12-18 months. This leak…
Dana: We should still fix the leak.
After brooding on this news all day, I finally laughed and told myself: “I should just vacuum my apartment. It is messy and needs cleaning. What else can one do? ”
> After brooding on this news all day, I finally laughed and told myself: “I should just vacuum my apartment. It is messy and needs cleaning. What else can one do? ”
Smarter people than me called this the normalcy overhang and it feels the erosion is becoming deeper by the day.
... How was the proof of the GRH, that mathematicians already assumed to be true anyway, supposed to change anything? Did we learn anything new? Did we unlearn something false?
Even the four-colors-map theorem had more actual impact.
Congratulations to Alpoge (and maybe Massulo), his teachers, his colleagues, his predecessors and everyone who made the tools he used. But nothing changed much.
We learnt that it's true and AIs with best humans could solve a very hard problem no human could solve before. This implication is probably more consequential, do we know think humans and AIs can solve the problem of developing an AI capable of solving riemann hard problems by itself?
> I thought I could no longer be surprised by AI progress. Yet I had not truly believed that Claude could prove a conjecture as challenging as the Riemann Hypothesis.
I don't really get this. I mean I am not surprised by AI progress but this means that I expect AI to do at least anything I imagine a human could do, including proving RH.
I am a layperson nobody and never understood this existential dread about AI until now. I am not even worried about RSI.
Even if Anthropic has already hit the limit on capability I do not think that this unfathomably valuable math machine should be in the hands of a private company. They become the arbiters of all or certain areas of intellectual progress or knowledge or whatever you want to call it?
Until today I just thought these models were weights and compute. But to produce something like this there must be a complex system that is hard and expensive to replicate. I feel crazy for even thinking this but I am now thinking seriously about what nationalization means. Not that I have any control over it, but I think it might be necessary. Me two days ago would punch me.
If one company did it, others will do it as well. There is no moat.
Capital is the moat, because big data and compute are what unlocks these results.
Capital is not a moat though. Any other company can raise comparable capital, if the investors believe it will pay off.
You would prefer this tech be in the hands of the Trump government?
No. Quite the problem isn’t it? Even if it ever did happen it is not happening that fast. But the analogy kicking around in my head is we already would never accept a private company with nuclear weapons without government involvement right?
Yes I would rather nukes be in private hands than in Trumps personally.
Not in the hand of the government, but of the people.
The people is a rather abstract term. Nobody really knows what "the people" want. Many things "the people" want alienate about half of the people for example.
"The People" is not some physical coherent entity that can hold something in its collective hand. It's an abstraction. How do give something to an abstraction? It's like saying I want to give an apple to pi.
That's why governments are necessary. And, in this case, I don't see how the Trump administration running the tech would be an improvement.
Well, yes. I can vote them out. I can't vote the ANT leadership out if I don't like what they're doing.
Thats theory, anthropic can live or die by the market. The usa government is effectively immortal and can murder you legally.
And the mathematicians in existential crisis plot continues. Deepens.[1]
It could prove all the theorems in the world. At this point it/they are running on electrical appliances that you can thwart by turning off the power. What is actually concerning are the humans behind these enterprises. It’s all just the same sociopathic thirst for absolute domination. Same old “human nature” as the last two hundred thousand years.
But Scott Aaronson’s smartest people he knows is concerned about computers ending the world. Not the humans who are making these applicances doing the same. I guess intelligence is a bit more compartmentalized than the smartest people would like to think.
And thus I paced around my office, thinking about the last five Hacker News submissions on maths, but more organically presented as if I ran into them on the street. Will computahs end dah world? Then the story circles back to whatever mundane thing I was going to do next. Ha ha: There’s nothing that I can do with humans wanting to dominate the world, sure isn’t.
[1] https://news.ycombinator.com/item?id=50489633
Because these theoretical discoveries are coming thick and fast, it is getting to point where we are dismissing the gravity of some of these milestones. The next big milestone would be bridging this gap between theoretical discoveries and real world impact.
I am hopeful we will see cheap abundant nutritious food and clean water, more efficient solar batteries, economical space travel, cancer treatments, climate change mitigation etc. Most importantly, I hope people can still find purpose in their lifespans. Maybe a few centuries from now it would be thought of as barbaric that people we forced to sit on a desk for 8 hours a day just to afford a livelihood.
>I am hopeful we will see cheap abundant nutritious food and clean water, more efficient solar batteries, economical space travel, cancer treatments, climate change mitigation etc
I sincerely don't understand how on Earth people make the leap from our current trajectory to "paradise". What makes you think any of those things follow naturally from where we are today and where we're going?
Room temperature superconductors could provide a sizable leap for many different fields and applications. It wouldn't leap us to paradise, but isn't inconceivable that RTS discovery combined with cancer/disease/aging/fusion breakthroughs could qualify as that leap.
It helps looking at "current trajectory" as spanning thousands of years and not just the last few
Is the hope. Unfortunately, as you've pointed out, not realistic, and probably not because of raw technical ability.
Well, it's also been our carrot on the stick since the dawn of government, and especially during the Industrial Revolution. The Age of Automation will be no different. Paradise isn't coming. The world is on fire and we are in the middle of a mass extinction event. This year alone there have been a number of horrible events with casualties directly attributed towards human-influenced climate change.
We aren't coming back from this for hundreds to thousands of years if we are lucky. We've used up most of our energy (low-entropy) reserves and future societies will not have fossil fuels to smooth out a grid mostly powered by wind and solar. So we'll invest in lithium instead, until we've mined all of that, too, and future societies are left with relatively little leverage. If we are lucky, by this time, in spite of Elon Musk, we will have the ability to mine asteroids for resources. It will still be expensive and won't solve all of our problems.
But we will likely never again see an age of abundance and affordability as we did in the post-WWII United States. Certainly not in our lifetime, so calibrate. Many factors came together to allow that brief anomaly in human history to exist, and it was still only gained through great tragedy (brain drain to the US from the war and Holocaust, profits from the military industrial complex, letting Europe get ravaged and then exploiting the situation economically for decades, exploiting South America and East Asia, and keeping things going a bit longer by outsourcing manufacturing and craft to criminally underpaid foreign workforces, etc).
We have already seen what these people do when in power, we shouldn't expect it to stop. Peter Thiel literally said being evil is okay if it's industrious. Paradise is currently out of the question, and the age of abundance is coming to an abrupt end. Great resource wars are coming, and now we've been robbed of our intellectual and creative pursuits, so we really have nowhere to retreat and are losing significant economic leverage.
And once our physical labor is no longer needed, we the people will have lost all leverage against our comically psychopathic tech and energy barons who currently see us as unfortunately necessary capital expenditure, and who spend billions to manufacture consent for the privilege to be slaughtered in slow motion.
It's worth pointing out lithium isn't the same thing as fossil fuels. You can extract the lithium from spent lithium ion batteries to make new ones.
You're right, and I initially considered removing the line about lithium because you can recycle it at >90% material efficiency. Cobalt recycling has also improved.
I left it in because we definitely will mine as much lithium out as we can until it's not cost effective vs only recycling, and the resource will continue to be extracted with slave labor in different clothes until we have robots doing that job as well.
Summarised best by Peter Thiel himself
<interviewer> You would prefer the human race to endure, right?
<Thiel> Umm... I would, um... um...
> I sincerely don't understand how on Earth people make the leap from our current trajectory to "paradise".
I think it's because disruption is an opportunity to rapidly go off the current trajectories. And since the current trajectories are terrible in the capitalistic developed world people naturally assume that going off of them is going to be better or at least it has potential to be better.
We know what would be beneficial for humanity, but we are out of hopes for achieving that within artificial systems that evolved to rule us. In this hopelessness any disruption is something that a fraction of people look at with hopeful gaze. At this point some people can be hopeful even for a proper pandemic, war, anarchy. Compared to known terrible costs of those AI looks significantly better and it's potential downsides seem more far fetched and easier to dismiss as fantasy.
Look at who is holding firmly all the power (bunch of sociopaths at best). Look at how llms are siphoning many human achievements and profit from those are extremely and unilaterally concentrated within hands of very few. This is not even challenged while there is still a chance.
Disruption it is, but such naive positivity has no place in 2026 in debate about actual world. Its not about crapping on rosy dreams of other folks, but about having realistic expectations and then having corresponding actions which don't miss future by a megaparsec.
For every dreamer there are 10 sociopaths already working hard on monetizing this ad nausea, increasing folk's addictions and various mass monitoring.
> Look at how llms are siphoning many human achievements and profit from those are extremely and unilaterally concentrated within hands of very few.
The thing is that's debatable. There's plenty of people arguing that LLMs are losing money and can never be profitable for investors that foot the bill.
I don't say I agree but that's one of the things that makes people hopeful. That they make ASI and it eats them alive.
> Disruption it is, but such naive positivity has no place in 2026 in debate about actual world.
I think people just don't have any other sources of hope. Last decade looks like intense festival of ever increasing stupidity and greed.
I'm scared of the future.
But there are things that make me a bit hopeful.
Small-scale nuclear fission reactors. Sound smart, sound safe, sound like a good way to make a lot of energy, without wrecking the Earth.
Fusion reactors. If we get good at them sometime in the next fifty years, that'd be great.
Yay, solar, wind, hydro, geo power. And on electric vehicles, I guess...
Fingers crossed on better batteries, to have a compounding effect on solar, wind, hydro, geothermal power's usefulness...
Desalination takes a lot of power. And I worry about the availability of potable water. That's why my hope for lowering energy costs makes me somewhat hopeful that we can survive the +1.5 C horror that we've unleashed on ourselves. There's a ton of problems associated with +1.5 C, but the water scarcity part is the one that scares me the worst.
Maybe robots can help us grow crops with less pesticides? Maybe they can help us avoid the monocultures that make us vulnerable to diseases...? Maybe we can grow a hundred crops all together, in a more sustainable way, with robots tending them?
Yay for fake meat, if it gets people to consume calories from places that are more land-, energy-, water- efficient...? Like, when fake meat tastes about as good as real meat, and is cheaper, and maybe arguably a little better for you, that feels like a huge, huge victory?
Yay for throwing AI at solving diseases?
Yay for technology improving access to DIY, home 3d printing, carving, CNC...?
Yay for AI making it more possible to hack your own hardware...?
I mean, I'm not a huge fan of China, but they lifted close to 800 million people out of extreme poverty, over 4 decades. Let's not discount that...
Almost every prosperous country has universal health care...
I mean, there are things to be hopeful for.
Desalinated water is only $1/m³ (1,000 litre).
What do you pay for your Perrier?
Yeah, if you ignore literally everything about the way humans work, all of that sounds extremely likely now with AI!
It's already a joke. You hear all these tech psychopath CEOs lie about AI 'reducing work' and other great things, but it turns out that's not what they are using it for. They are using it to replace humans, mass surveillance, etc...
They are "reducing work" in the sense that humans will have no work to do for which they can make a living and feel a sense of purpose. Welcome to the future.
Almost everyone I know hates mondays. They fantasize about early retirement and not having to work. Work, by itself, is not what people desire. It's the livelihood that it provides. If machines do everything, cost of things approach zero. And I know we fear unfair distribution of that benefit, but we know that didn't happen with food. We went from almost all of humanity producing food to almost nobody doing it, and we have new problems with people (even poor ones) eating too much food.
Re sense of purpose, I'd rather be mountain biking with my friends everyday than sit on a desk for 8 hours pushing pixels around.
You are stuck in the paradigm of needing to work to have a purpose.
No, I'm stuck in the paradigm of needing to work to pay rent.
Enlighten us with the new paradigm, O Wise One.
Musicians, artists, actors, all have purpose but many don't exactly work.
I don't actually think anybody in the Bay Area understands how much the average American would like to take a bat to their head. This industry is so high off its own farts and so divorced from the priorities of ordinary people that the pseudo-philosophy and fake economic theories they propose end up being a religious belief.
We already have cheap abundant nutritious food and clean water. An 8 hr/day desk job is really not bad. If people are barely scraping by on a job, that's mostly a political problem, not a technology problem.
> Maybe a few centuries from now it would be thought of as barbaric that people we forced to sit on a desk for 8 hours a day just to afford a livelihood.
I'm fully convinced of this. It will probably be viewed in retrospect like we view the exploitative work during the Industrial Revolution, even if it seems like an exaggeration right now.
I think there's a category error in a lot of these techno-optimist takes.
Abundance of productive capacity does not automatically equal abundant access to its products.
If all of this happens in a way where a few people control these new, automated means of production, all of these wonderful things do not automatically follow.
Why would they share? And worse yet, if they have this automated production capacity without reliance on labor, why would they interact positively with society at all? What's their incentive?
You're basically reduced to hoping for charity from a group of people that have repeatedly proven their lack thereof.
Right there with you. What's the motivation of anybody on the high end of the wealth inequality scale to share with the other end? They have nothing to gain and everything to lose.
It would be nice to hear the rationale behind those who are downvoting you.
There are people who share a significant proportion of their wealth with the significantly poorer, like Warren Buffett, Bill Gates, Dustin Moskovitz, and so on.[0] Some people choose to do good despite nothing tangible gained, whether they are poor and just putting their single dollar in a wicker basket, or literally billionaires and saving millions of lives with their money.
Whether the owners of AI will choose to share, time will tell.
[0]: https://impactlist.xyz/
I'm specifically worried about the people who use their wealth to influence political discourse to perpetuate wealth inequality. It makes sense to me that they would use technology, which fundamentally obviates their reliance on the labor of the less wealthy, to perpetuate inequality even more that they can thru political means.
I don't feel like the generous wealthy can offset the influence of the wealthy who have the opposite goal.
> We could never win an election on getting certain things because we were in such a small minority, but maybe you could unilaterally change the world without having to constantly convince people and beg people and plead with people who are never going to agree with you through technological means, and this is where I think technology is this incredible alternative to politics.
-- Peter Thiel
It's incredible to me that was all said out loud.
> There are people who share a significant proportion of their wealth with the significantly poorer, like Warren Buffett, Bill Gates, Dustin Moskovitz, and so on.
Yes, that is the charity one can hope for. But even in that lucky case, the issue I'm describing is structural. It's not like any of them share to the point of giving up their generationally privileged position. That poses the problem: Even if the people immediately in control of AI are pure selfless angels, over time and generations the likelihood of one not being so asymptotically approaches 100%. Sooner or later there will be a Caligula.
There are no downvotes on this website
how does any of this solve human greed + pride + envy, etc?
none of it does. why do you think material gains heal the human heart?
> Maybe a few centuries from now it would be thought of as barbaric that people we forced to sit on a desk for 8 hours a day just to afford a livelihood.
??? That was a few decades ago
If we keep the same political and moral stances around work you'll still be working every day even if we get agi dont worry
Within a few decades not centuries. Either that or we turned into an actual cyberpunk world.
>I am hopeful we will see cheap abundant nutritious food and clean water, more efficient solar batteries
yes, I'm certain Saltman Inc. is working on these exact subjects precisely for your entirely indispensable family/tribe
trust the plan
I'm sure some or all of those advancements will happen but, based on the way the last ten years have unfolded, I'm not at all confident everyone will see the benefits. For a concrete example, look no further than the sociopathic destruction of USAID.
This comment is incredibly tone-deaf. It is difficult to see how the benefits will ever trickle down to ordinary human beings from the class that holds all the money, controls all the computing power, and possesses the authority to execute critical tasks.
More food, water and medical care would only enable people to multiply and overpopulate this world even further. At the cost of other animals in this world all of whom will be supplanted or become extinct.
We need to solve the overpopulation crisis before committing to improving the lives of the impoverished.
Would it be possible to reconcile your worry for the ecosystem with basic empathy for people you clearly don't see as equal to you in worth and individual agency?
I see overpopulation as the most important problem facing mankind. More important than climate change, which IMHO is merely an outcrop of overpopulation.
Like I've said before: our society is completely obsessed with sex and it's therefore extremely difficult to curb population growth. In Africa there are frequent riots when rumors are spread that the "white man" wants to neuter Africans through vaccination programs.
Sex and reproduction have been separated, which is good because nothing curbs a mammal's sex drive other than existential hopelessness. Which we also have created.
What I found surprising. How many absolutely groundbreaking breakthroughs with AI are in Mathematics. And at the same time I can't remember anything similar in other fields. For example there's nothing groundbreaking in Software. Yes, Claude can write some PoC code faster than me. Great. What about creating some software which wasn't possible before? What about creating operating system without a single bug? What about inventing some video compression which is miles ahead of competitors? What about creating Rust compiler which runs 10x faster than rustc?
I think the closest thing is formal verification of Signal protocol, I guess. Which is insanely cool, but still not as cool as millenium problem.
> What about inventing some video compression which is miles ahead of competitors?
in a specific case I was interested in (point cloud compression for geometry streaming on the network), claude helped me make something that is 200-300x faster than the existing standard (MPEG G-PCC) at 10-20% worse compression ratio, which opens completely different use cases. e.g. the MPEG one compresses a 3.2 M points point cloud to 8.94 megabytes in ~5.7 seconds and decodes it in 3.7 seconds ; https://github.com/sat-mtl/pccl encodes to 9.97 megabytes in 32 milliseconds and decodes in 12ms.
The real innovation here, IMO, is theorem provers. Without theorem provers, the LLM brute-force-by-bouncing-off-the-guardrails approach just doesn't work to begin with and the whole thing spins off into hallucination and dodgy reasoning, just as we see in non-trivial AI software. Once you've got some magic oracle that gives you a binary yes/no answer to every attempt at a proof, your search strategy can be enormously simplified.
I also thought that, but I learned that most of OpenAI's results in their proof dump did not use theorem provers in the loop. They translated the papers to lean afterward.
But weren’t most of those papers totally unverified? I.e. there’s no reason to believe that the ones without formal verification are actually correct.
If i understand it correctly, lean is very useful for post-training (RL) step of the models.
Right, that's how all these proofs (including a few of mine) are done. Given the large-scale AI efforts already underway to formalize proofs in Lean, there could be some very interesting gains once proofs are Lean-first. For now, though, formalization happens afterward.
I don't expect proofs are going to be lean first. There's a reason they already aren't.
It's not really the provers that are the innovation. They existed before AI. The innovation is that the computer can write the proofs.
It’s the power of feedback loops
Same with gradient descent
Exactly. Another example of how AI depends on software which is carefully, deterministically hand-coded. It's disingenuous for people to say "look, we don't need to code anymore" when all these solutions depend on things that are definately well-thought out and hand coded.
Today.
Math is pure logic. A game like chess or go. It's not bottlenecked by the physical world so AI can iterate at maximum speed.
I work in this field, and the issue is that the LLMs start reward hacking these things. They either overoptimize to the benchmark and hallucinate, insert reward hacks like adjusting the benchmark or cache behaviors, or they make implementations that only focus on verifiable things, like utilizing the hardware better.
The fix to this are better benchmarking environments without reward hacks, as well as formal verification of code on various levels. There is crazy much work on this at the moment.
> For example there's nothing groundbreaking in Software.
The fact that I can implement any stupid idea I've ever had and I'll ever have is ground breaking for me.
> What about creating Rust compiler which runs 10x faster than rustc?
For now we have many times faster TypeScript compilers written in Rust by AI. There's zero doubt that things like that will come in abundance.
The recipe seems to be, pick a problem whose solution is highly verifiable, then throw the equivalent of ten billion hours of expert-level thought at it. If you can narrow the search space a lot, all the better. P != NP is sure to follow soon enough.
> P != NP is sure to follow soon enough
Not an expert, but from comments I've heard by experts P != NP is much further out of reach than Riemann. Eg, Scott Aaronson:
> First, almost no one in theoretical computer science—not counting the cranks, whose missives fill my inbox every week!—spends their time directly trying to prove P≠NP. Why not? For the same reason why you wouldn’t embark on a manned mission to another galaxy, if you hadn’t even set foot on Mars yet.
I defer to Scott. But, I mean, look at the trajectory in the past six months...
Right. Just saying the fall of these "step function" type problems aren't a forgone conclusion yet. It could easily be the case that some problems (who knows which?) are in an entirely different dimension of complexity, and would be out of reach of even much, much more powerful AI.
Their last two examples are extremely verifiable.
these tasks you mention are very long-horizon in context... i think if latent long-term memory or continuous learning approaches are getting better, we also see this
Its reverse engineering prowess is very impressive. Building a completely open source version of Microsoft Office or Adobe PS was definitely not possible before.
It wasn't but those are also half baked. Anything in software seems to be that the first 80% is easy
Microsoft Office and Adobe Photoshop were literally built before. Recreating their copies is not groundbreaking. That's what I'm talking about. AI in software right now feels like mostly about recreating things we already can do, but cheaper, with questionable quality and unclear maintainability.
They were created with huge resources only a few people on earth had, now everyone can do it with €80 and a bit of patience...
AI is software.
If you call any computer program that's advanced "AI" and stick the "software" label only on dumb programs, well of course software seems dumb. You defined it as such.
Maybe you should start directing Astra to make one of these breakthroughs you're waiting for, see what happens
Terence Tao a month ago:
"The AI companies have clearly... I'm sure they have secretly devoted much more time to the Riemann hypothesis than what they've disclosed. And they couldn't touch it. I mean something beyond current AI is needed. So maybe some combination of humans and AIs in 2050 or something will make further progress. Who knows?"
https://www.youtube.com/watch?v=bAnwMdfcMVw (17:10)
>I mean something beyond current AI is needed.
Every time someone says 'current AI is not enough, we need new architectures', it turns out that current AI was enough and all we really needed was more scaling.
Not in this case, what we needed was a genius mathematician to explore with Claude. Which is exactly what Terry said: "So maybe some combination of humans and AIs in 2050 or something will make further progress."
I remember when, for a brief period of time, the best chess players were human + computer. There were a lot of Ted talks and think pieces reflecting on this reality. This was called centaur chess.
This lasted for 2005–2014. By 2015, the advantage humans added was negative.
I imagine this will be the case here as well, but I highly doubt it'll take that long.
yes, for every specific task. but life is not like chess. life is like playing a million different chess games simulaniously, all of the time. Being a gentle hand that knows when AI is onto something and when it's torturing itself in a loop. He calls it out in the pdf, the aha moments he experiences, that lets him further guide the progress.
Math, specifically, is a bit more like chess though I think.
It's not 2050 yet, though. I feel like the future date he threw out was a big part of his conjecture, that even with human+AI today, it wouldn't be possible, and getting it done that way would take years, possibly decades, of more work on the AI.
And he said maybe humans+AI will _make progress_ in 2050. That is very different from one human + AI solving it 36 days later, lol.
How about we hold off until it's verified? Now, Terence Tao is getting visibly thin these days, I hope he's not torturing himself in an existential crisis. I know I would be. But I would also, in the end, be satisfied as a Lee Sedol figure, the last of the human geniuses, as it were.
As of your comment, it had already been verified by 4+ independent Lean kernels, one of them ("con-leche") itself formally verified. https://leanprover.zulipchat.com/#narrow/channel/113488-gene...
It looks like >99% of the work is the LLM. He even admits as much.
The genius here is Claude.
Anthropic would have claimed the result for themselves, without Levant, if they could. They didn’t. A strong mathematician was evidently required.
It doesn't make sense to compare the quantity of work. Humans are significantly more efficient than AI, so little work can go a lot further.
I’m sure for every incredible insight which Levant credits to Claude, there’s 100x+ more bad insights that Levant had to filter out.
> I’m sure for every incredible insight which Levant credits to Claude, there’s 100x+ more bad insights that Levant had to filter out.
This is certainly my experience of using Claude every day, in a much less cutting edge domain.
And Claude itself cannot tell the difference between the two. The best it can do is use some sort of external constraint (e.g. verification) to filter them.
Honestly, ever since Opus 5.5 dropped about two months ago, Claude's taste has been great, better than mine, even. Compared to Opus 5, I've barely had to filter anything out by hand.
It is fascinating. Because sometimes my experience is like yours, but other time, like today, I am like “I can’t believe how stupid Opus 5.5 or even Fable 5.2 is”. They will make horrible design decisions, overcomplicating. I will correct and they are like “Of course, your solution is much simpler and better than mine”.
And then they are simply boldly wrong at times. About mathematical facts. That’s why we need formalisation when working with them. They are still stupid and can’t be trusted. But theyvare also smart and can come with amazing insight. It’s a tool, but definitely one which cuts both ways.
I think it's funny that people gauge their personal experiences with AI against something like this.
Do we even know what Alpöge had at his disposal? I think most people underestimate what working with an agent swarm even with the current public technology would be like. We can't assume he didn't have something more advanced either.
I don't feel confident my usage of AI on a personal level reflects any of what has been done here.
Your argument is unfalsifiable speculation. I don't at all that it's self-evident that a strong mathematician was required. Your everyday Joe isn't exactly interested in the Reimann hypothesis so statistically a mathematician would be the one who'd use AI to come up with the solution.
How would a non-mathematician even be able to comprehend the output? Half the words in the pdf are gibberish to me, I could not tell if whatever Claude gave me for a "solve the Reimann hypothesis, make no mistakes" prompt would even make sense, much less be correct.
I don’t know shit about HPC and LLM inference, didn’t stop me from creating an inference engine: https://github.com/computerex/dlgo
A non math person may not be able to comprehend the output, but even professional mathematicians have trouble reading through the huge “alien” AI generated LEAN formalised proofs.
The real genius is that we have found a generic way of solving problems through brute force by having developed an effective way of aiming an unbelievable amount of computing power at them.
How to brute force 15MLOC to compile is an exercise for the reader.
Reading his post, this is not the impression that I get.
If the genius here is Claude and Claude did 99% of the work then isn't it a staggering coincidence that this was achieved by one of the most qualified and likely humans to have solved this on their own merit and not some random person sat in a bedroom somewhere?
Company selling shovels says shovels did all of the work
Human who could claim credit claims AI did all of the work.
I also use Claude and it never produced this result for me.
Did you also have access to unreleased models by Anthropic?
Doesn't matter; Claude ends up taking cues from whoever it's working with, such that two people doing the same thing get very different outcomes. Claude will always be able to write better software when prompted by a senior engineer; it will always produce better math when prompted by a talented mathemetician.
But also an employee of the shovel company.
Not sure anyone would have been able to write 250k pages of Lean without the shovel
Real men used to be able to fly to moon on 10kb of memory (and 100 billion dollars)
And they weren't able to prove the Riemann hypothesis
But they landed on the moon.
Yet some people are still skeptical about that.
On that topic, I'm surprised the usual anti-ai trolls aren't posting low-effort drive-by "ai slop", "stochastic parrot", "autocomplete on steroids" comments to this very discussion.
Edit: I stand corrected: RunSet, ToucanLoucan, luka598, and others have checked in.
"Maybe 24 years from now we'll make progress" is a little different than "actually it will be solved a month from now with current techniques"
Last I checked it's 2026. Not 20250. And for the record, there is no evidence that suggests that a layman couldn't have made progress with just using AI.
Do we know that that is the case?
We know the architectures of the open weight models as they include architecture diagrams, support in projects like llama.cpp, etc.
Companies like Anthropic and OpenAI don't release information about their models (aside from general things like context window size, whether they support reasoning, etc.). A lot of what we know is speculation. From press releases Anthropic have stated that they are getting Claude to assist with designing new iterations of the models, but no specific details.
The open weight models and other arXiv papers/research are iterating on and experimenting with the base/core transformer architecture:
- RoPE/NoPE/etc. for context window scaling;
- dense vs mixture of experts models;
- Mamba-based State Space Models (falcon models, etc.);
- reasoning, combining transformers with a residual stream, Attention Residuals (Kimi), etc.
- activation functions: ReLU; SwiGLU; etc.
- etc.
I think a lot of model architecture changes don't move the needle ~that much and are basically done to improve runtime performance and scaling. - There's tons of cases of ml researchers, tech companies, and labs that propose new and improved models that turn out to not really be significant quality improvements (if at all) once the community starts experimenting with them.
it does seem like (IMO) the biggest leaps have basically been perf & scaling + RL on high quality data + harnesses.
Yes, but "improving runtime performance and scaling": does directly translate into huge quality improvement too, because then you can you a much larger or deeper model, train it for much longer, and have a larger context.
It's usually people who are not suffiently familiar with the topic, in this case, technology. He has been saying boring things like that about AI in the recent years, and seems to be continuously adapting his positions. I have received negative votes over my criticism of his remarks.
This was literally Dario's contribution to the AI race.
He just did the math and realized you just needed more and more compute to get more intelligence. People keep doubting it but it's showing no sign of being incorrect.
When you say more compute are you referring more to just inference compute or training?
ChatGPT was the turning point of realizing we needed just more compute
We have scaled on both axes and many others, I'd say.
Inference can be increased vertically (a single model thinking for longer), or horizontally (NS was solved by a 10,000-agent swarm).
And pretraining has two inputs that can be scaled independently - training data and model size. (Things like the Chinchilla 20:1 law are meant to establish the optimal way to train with a limited budget - you could just keep making either term bigger and bigger on its own if you had infinite money.)
I think creating and curating high-quality datasets is also part of scaling. Datasets like FineWeb are quite a bit better than the barely-filtered CommonCrawl dumps GPT-3 trained on, but someone had to invest in making them. (To say nothing of whatever exotic bespoke robotics/physics/etc datasets OpenAI and Anthropic are investing in.)
ChatGPT was only the turning point for public awareness (and the massive influx of institutional investment that followed), but researchers like Ilya, Alec Radford, and Dario were already convinced years before ChatGPT that scaling compute was the way forward. ChatGPT just helped unlock the resources (by proving the commercial potential to investors and the subsequent influx of capital) for these guys to pursue that vision at a much larger scale.
I mean, why wouldn't that be the default assumption, if you know anything about computing history? ChatGPT's (v1) contribution wasn't proving commercial potential but demonstrating efficient scaling, due to a confluence of technical developments. If you're talking about v3 then sure.
> ChatGPT was the turning point of realizing we needed just more compute
Ilya Sutskever has been saying more compute = better perf since at least 2012.
https://www.youtube.com/watch?v=-uyXE7dY5H0&t=1185 https://proceedings.neurips.cc/paper_files/paper/2012/file/c...I was downing some extra bitter coffee while I read that.
Sounds like a valuable lesson.
A spoonfull of syntactic sugar helps the semantics go down.
People are so goddamn impatient. They demand miracles by the end of the week, otherwise they call a flop.
As if drawing a line through few points on a chart is a mystic art.
current AI might not be optimal, but still a powerful unoptimal could create optimal AI
On Anthropics side it's literally just levent. He isn't even doing this on an official capacity. OAI have teams of Mathematicians. This is like david vs goliath.
Is it? Isn’t it just frontier model va frontier model. I am not a maths specialist, but aren’t the models doing 99.999999% of the he at lifting? Of the answer is “no”, then what explains all of these unsolvable problems being solved in rapid succession?
Depends what you mean by “lifting.” If it’s anything like building software, the agents do all the grunt work (exploring 1000’s of possible approaches in parallel, perhaps) while the domain expert (human) makes the critical strategic decisions about which pieces need to be solved and how to fit them together.
Even the best agents still can’t strategize their way out of a paper bag, in my experience.
The human remains the most critical differentiator for non-trivial work. (Different story for trivial projects that agents can solve independently. Here the human becomes a fungible operator.)
I think a lot of his AI skepticism is just a self-fulfilling prophecy, and a misunderstanding that AI at the cutting-edge is a tool in the hands of the user. In other words, sure current AI is not going to solve the Riemann hypothesis until 2050, but what do you think the current state of AI would be if Terrance Tao was the one driving it?
Somewhat related: I dont think AI is coming for everyones jobs, and I will never understand the argument that it is, because I always have the ability to use the AI and add my "human ingenuity" to it. So why _wouldnt_ that be a better option than just raw AI?
> because I always have the ability to use the AI and add my "human ingenuity" to it. So why _wouldnt_ that be a better option than just raw AI?
Well, that hasn't held true for chess, so I'm not sure how universal the domain is where human ingenuity will always augment AI. Also, there is a cost floor to human ingenuity, in that it requires the time of one living human, and their metabolic sustenance over that period. So the added benefit of that human has to be substantially more than epsilon.
There's so many answers to your last rhetorical question, if you use your ingenuity.
- AI might just be good enough, and beat you at price and speed
- you might not be the one to have access to state-of-the-art AI - many more
From my point of view, AI has actually already surpassed humans in 2 ends of the spectrum
- math (logic)
- digital arts (ingenuity, with some heavy lifting to call it that, but still)
Filling the space in between could be reasonable expected to be a matter of time, money, data.
It's funny because LLMs are actually horrible at logic.
Current AI is actually good at search, but I can see how people mistake one for the other in the context of math, for example.
Which is funny, because LLM's are great in writing proofs for logic solvers. They are excellent at logic.
The verifier is excellent.
> So why _wouldnt_ that be a better option than just raw AI?
1. Eventually your "ingenuity" will be redundant. Human-machine chess players were dominant for a brief period.
2. Cost. Even if you are adding something, it might not be worth it to whoever is paying you when you cost 100x what the AI alone does.
> I always have the ability to use the AI and add my "human ingenuity" to it. So why _wouldnt_ that be a better option than just raw AI?
It is not clear it is always a better option. It may be for you, but not for someone who doesn't know what they are doing and drives the AI down the wrong path. For example, I don't think adding my "human ingenuity" to Claude would help solve the Riemann Hypothesis, and would likely be counterproductive, because I have a 3rd-level understanding of math.
>I'm sure they have secretly devoted much more time to the Riemann hypothesis than what they've disclosed.
Curiously, the timeline for this proof only starts roughly around the same time that Tao said this. IIRC, work on the disproof for the Navier-Stokes smoothness theorem only started a week before it was published. Maybe I'm wrong, but it seems like AI companies have only recently started to put real effort into LLM-assisted proofs.
It seems bold for Tao to state that they couldn't touch it; Considering that after Anthropic announced Navier-Stokes and got all the negative press, largely driven by Terrence Tao, it doesn't surprise me that they didn't disclose anything afterwards.
I think Terrence Tao was trying to say that had the LLM swarms been able to prove this it would have been what they lead with. This seems to lend to the claim that Levent has done an amazing lift in coaxing the solution out of an LLM.
Direct link to the quote: https://youtube.com/watch?v=bAnwMdfcMVw&t=1035s
This is the guy I see mentioned in every other HN post? Yeah, seems everyone who should know better still hasn’t grappled with the gravity of what’s happening
So even one of the smartest mathematicians alive struggles to comprehend or extrapolate the implications of an exponential progress function. To be so far off seems ridiculous, but it’s exactly how you’d predict the error to be. We are truly cooked. End of 2027 is probably the end of knowledge work on a global scale.
We got RH proof before Elder Scrolls VI
GTA VI just made it in before.
Someone on the reddit thread claims Levant got the results on the 16th, so actually GRH before GTA 6.
Edit: I read the PDF, Levant claims 11/16 for GRH.
So Levant finished the GRH proof just in time to play GTA6. Deadlines drive productivity!
But "we" didn't get it. GTA 6 was also presumably finished well in advance of release.
Very funny argument either way.
They must have heard that RH was about to fall, and planned the GTA 6 release around that.
We got RH proof before we got the fifth LBJ biography by Robert Caro.
Ha. Was just thinking literally this afternoon when the series will be completed.
And the third Stalin volume by Stephen Kotkin.
… and the third peripheral book :(
"I'm sure they have secretly devoted much more time to the {{matter}} than what they've disclosed. And they couldn't touch it. I mean something beyond current AI is needed. So maybe some combination of humans and AIs in 2050 or something will make further progress. Who knows?"
- All of humanity, unaware their pants have fallen around their ankles
> I'm sure they have secretly devoted much more time to the PERPETUAL MOTION MACHINE than what they've disclosed. And they couldn't touch it. I mean something beyond current AI is needed. So maybe some combination of humans and AIs in 2050 or something will make further progress. Who knows?
Just to point out how ridiculous it sounds to claim that because X (Riemann) is possible that anything is possible.
Let's constrain it to things that we generally acknowledge to be physically possible shall we?
Cancer, aging, mental disfunction.
Law, creative film making, market inefficiency of any kind.
Writing programs using something other than your voice and having it be completed faster than your mouth can complete the sentence.
Pants around ankles.
In the worst case scenario, we just get a perpetual motion machine in 36-ish days.
Any chance AI solves greed?
He has a massive out with "Who knows?" Can't say the statement isn't rigorous, lol.
In 24 years with a 24 year margin of error
> Number theorists are going to be busy the next decade and a half digesting this proof.
- someone else
So maybe not too far off.
he said 2050 a month ago, it's barely 50 days since that
Well, we got an award at a Caltech "mathathon" for some pedestrian improvement on some error bounds on some niche problem. Now these guys come out with a proof of one of the most notorious open problems in mathematics. Just to one-up us I'm sure...
> In October 2024 I attended a workshop at Harvard University ... “An AI that can prove the Riemann hypothesis is not one I’d want to meet,” said Andrew Sutherland, a number theorist at the Massachusetts Institute of Technology. “If that happens, mathematicians having jobs will be the least of our problems.”
Scientific American, May 19, 2026.
https://www.scientificamerican.com/article/the-riemann-hypot...
why exactly is it so scary?
asked the ant while seeing the shadow of a excavator operator
I really hope Levent next tries to prove the irrationality of the Euler-Mascheroni constant ɣ , or the gamma function minimum α, Γ(α) , as they have so many forms yet we know so little about them. It's slightly related to GRH, since the Zeta function occurs in a some of these forms.
Given how important these constants are, you'd think we'd at least have proved by now that ɣ was irrational, yet shockingly it has resisted all attempts. In my opinion, it's almost scandalous that it wasn't one of the Millenium prizes.
David Hilbert:
> If I were to awaken after having slept for a thousand years, my first question would be: Has the Riemann hypothesis been proven?
Assuming he reincarnates sometime in the future:
Yes. Mr. Hilbert, good to talk to you again. RH has indeed been proven, but by superhuman technology making long alien constructions on automorphic forms.
We dont really understand the proof in its entirety, it was verified by the whole formalization movement you started.
And to have a semblance of understanding we keep going back to the SH machine to dumb it down and explain to us different parts of it.
I would assume this proof will eventually be understood since it is so high profile. The 1000 other statements being proved all over maths? Maybe not.
As a non-Mathematician it's inconceivable to me that proofs can be so difficult to understand that no one will. I thought the process of getting to the proof is the complex part, once proven, ex post facto should be more simple. But seems like it isn't.
It might not be that difficult, just long and diverse enough that you might just run out of lifespan.
As humans we are used to reality giving us stuff to understand that are just handful of things. Three generations of leptons, four types of orbitals, 4 dimensions, low natural number coefficients.
Deep enough in the math there might be constructs that have hundreds, thousands or millions importantly distinct and non-groupable concepts you'd need to understand to understand the proof.
It’s an enormous mental slog to validate AI math slop, demoralizing, and also feels like you’re just fighting for scraps. not saying thats all the reason, but definitely contributes
I mean people have only been dealing with it for a couple of months tops, stands to reason people will find better approaches to this.
This would sound like comedy science fiction until now, it is the truth!
Loved your post - sounds like Douglas Adams
Just don't reincarnate Asimov. He'd be satisfied he was right, and terribly disappointed nobody listened to him
>yes Mr. Hilbert, nearly 1000 years ago
Given the multiple ways this proofs has been formally verified^[1], I think at the moment there is no any doubt that it is correct.
> Andrew Sutherland (MIT), checks run with Claude Code. Page last updated 2026-11-20 22:05 UTC. Results below are as of that time and this page will be updated as the remaining checks finish.
[1]: https://math.mit.edu/~drew/grh/
Checking with an AI offers very little assurance - if there's a mistake in an AI proof (and I'm not saying there is), it wouldn't be surprising for the checking AI to make the same mistake.
Formal verification is not AI.
There seems to be a completely human proof also: https://math216.wordpress.com/2026/11/20/hurried-announcemen...
Discussed here: Hurried Announcement: Alessandro Masullo GRH - https://news.ycombinator.com/item?id=50579568 - Nov 2026 (752 comments)
Wouldn't it be fun if they were both correct, but with wildly different structures? We could maybe use the proof difference to characterize the prover difference (AI-assisted vs not).
That is the current situation. There are big differences at key steps, though most of the plumbing is in common.
> Wouldn't it be fun if they were both correct, but with wildly different structures?
Theories in (Quantum?) Physics (Heisenberg / Feynman / Schrodinger etc) are very much like this, apparently.
Lambda calculus vs Turing machines vs μ-recursive functions also come to mind as an example. All equivalent in the end.
this is actually something i'm trying to study
Structurally, like, treating the proof like a graph or something along those lines? Or socially, like this part of the proof resembles proofs given by this group of mathematicians, so we can say it has a certain heritage? Or some other way?
It's easy for me to wish for such a thing, but I have no idea what one would latch onto when granting such a wish.
That's one of the central questions in Stanislaw Lem's "His Master's Voice"!
https://en.wikipedia.org/wiki/His_Master%27s_Voice_(novel)
>His Master's Voice (original Polish title: Głos Pana) is a 1960s science fiction novel written by Polish writer Stanisław Lem. It was first published in 1968 and translated into English by Michael Kandel in 1983. It is a philosophical first contact story about an effort by scientists to decode, translate, and understand an extraterrestrial transmission. It is considered to be one of the three best-known books by Lem, the other two being Solaris and The Cyberiad.
https://lem.pl/en/works/novels/his-masters-voice
>Lem's Opinion: Human beings should accept some humility. Sometimes we face phenomena the essence of which we are unable to understand. Even if we are equipped with the most modern scientific apparatus and knowledge we are incapable of resolving whether they are accidental or intentional. Newton once said that with respect to nature we are just children playing with shells on the seashore.
https://www.donhopkins.com/home/catalog/lem/Lem.html
>HIS MASTER'S VOICE. Stanislaw Lem. Harcourt, Brace, Jovanovich, c1968, English translation by Michael Kandel. c1983 (hardcover -- I understand the paperback is out now).
>Faren Miller in LOCUS called this book 'more a readable meditation than a novel' and after reading the book I would have to add my qualified agreement. The book is a philosophical retrospection by one of the key members of the 'His Master's Voice' project, the mathematician Peter Hogarth. The HMV project was secretly founded by the US government to analyze and (if possible) translate a signal discovered in modulated neutrino emissions from deep space. The military hoped to extract technology which could be used for new weapons against the Russians. Hogarth and the other scientists held the military in contempt, and sought instead to interpret the message and find a way to understand the Senders. Several strange results came from the study of the signal. One result was a peculiar protoplasm-like material created by irradiating a chemical soup with the signal. This goo had the peculiar characteristic of sustaining itself not through normal biological energy cycles, but through atomic reactions. Hogarth and friends then discovered that the material had the capacity for causing remote nuclear reactions: an explosion near it released its energy at a distance. The possibility arose that detonating a hydrogen bomb in a roomful of the protoplasm in Nevada might cause Moscow to be destroyed. Hogarth and his colleagues had to decide: what should they tell the military? Besides this ethical problem, Hogarth also considers the question of whether there really were any Senders... This novel is very difficult but I found it to be rewarding.
https://dokumen.pub/his-masters-voice-mit-press-0262538458-9...
Chapter 6, p. 109:
>Out of the infinite number of possible mathematics we have chosen one; our history decided this for us with its various unique and irreversible vicissitudes.
The mathematics we practice may reflect historical selection among alternatives. This also connects to MatrixMan’s speculation about different foundations.
Preface, p. 16:
>Here I will say only that the world injected its patterns into human language at the very inception of that language; mathematics sleeps in every utterance, and can only be discovered, never invented. What constitutes its crown may not be cut free from its roots, because it arose not in the course of the three hundred or eight hundred years of civilized history, but through the millennia of linguistic evolution: at the loci of man’s encounter with his environment, from the time of tribes and rivers.
Our mathematical habits have a cultural and linguistic history. An AI trained on human material complicates any clean opposition between "human mathematics" and an alien alternative.
Chapter 9, p. 143:
>It is true that we differed in our style of thinking; I was fascinated by ergodic processes, a field that Dill made light of.
Chapter 9, p. 144:
>I do not think I ever finished any larger paper in all my younger work without imagining Dill’s eyes on the manuscript. What effort it cost me to prove that the Dill variable combinatorics was only a rough approximation of an ergodic theorem! Not before or since, I daresay, did I polish a thing so carefully; and it is even possible that the whole concept of groups later called Hogarth groups came out of that quiet, constant passion with which I plowed Dill’s axioms under.
Personal preferences, rivalry, and the desire for recognition can influence mathematical discovery.
Chapter 10, p. 153:
>This, I thought, illustrated a certain small but characteristic feature of the history of science, because neither the fortuitous bends in the road of research nor the accidental circumstances assisting at the birth of a discovery ever completely detach themselves from its final form. Indeed, it is not easy to recognize these relics, for the reason that, fossilized, they become embedded in the heart of all later theories and formulations, like a print of a coincidence which turns to stone, to an iron rule of thought.
The finished argument may preserve traces of its discovery, although recognizing those traces is difficult.
Chapter 2, p. 42:
>With an exaggeration that is necessary for the clarification of my meaning, I will say that in the course of my work (it is difficult to say exactly when this occurred) I began to suspect that the “letter from the stars” was, for us who attempted to decipher it, a kind of psychological association test, a particularly complex Rorschach test. For as a subject, believing he sees in the colored blotches angels or birds of ill omen, in reality fills in the vagueness of the thing shown with what is “on his mind,” so did we attempt, behind the veil of incomprehensible signs, to discern the presence of what lay, first and foremost, within ourselves.
Chapter 2, pp. 42–43:
>None of us knows, therefore, to what extent we were the instruments of an objective analysis, to what extent the delegates of humanity (in that we have been shaped by and are typical of our society), and to what extent, finally, each of us represented only himself, with the inspiration for his hypotheses about the contents of the “letter” being supplied by his own—possibly raving, possibly wounded—psyche in its uncontrolled regions.
Interpreters bring their own expectations, institutional assumptions, and psychological investments to unfamiliar intellectual products.
Unfortunately AI has found errors in the human-only proof (not sure if fixable) https://news.ycombinator.com/item?id=50580084
It was rushed out before being fully checked, not a huge surprise. They even say in their blog post that there are likely gaps but they hope they are fixable.
I guess they leaned too much into "purely human developed" and didn't even show their proof to any LLMs before publishing it.
And why would they do that when there's already precedent for AI companies stealing a mathematical breakthrough that's been put in their system?
It’s a 1,000 pages dense document. I’m not even sure the PDF conversions are accurate enough to feed it ti an LLM.
Also - if that’s the future of math (larger and larger proofs and algorithms, to the point they are far beyond a human’s capability to comprehend, then I wonder how mathematicians work would look like.
I recall being told that there are certain axioms, the axiom of choice for instance, which we take most of the time not because its the best tool for the job in any intrinsic sense. Its the best tool for the job in a historical sense. We use it because our parents did and it's convenient to stand on the shoulders of giants.
But if proofs are cheap, you don't have to stand on the shoulders of giants. You can instead venture where they did not.
I'm told that the axiom of determinacy--used as an alternative to the axiom of choice--represents such an underexplored alternate dimension. Presumably there are many of these: inconsistent with the math we know but internally consistent and potentially superior to our use cases than what we now use... Our current habits being more driven by historical accident than careful decision.
So a future mathematician, rather than being a Platonist, might be a pragmatist and a curator. Given the many possible formulations for mathematics, which is most fruitfully paired with economics? Which goes best with chemistry? Which parts appeal to the human sense of beauty and how can those be introduced to specialists in disparate fields such that those specialists can now, through a bit of shared structure in the formalism that underpins their work, better understand each other's domains and better collaborate about them.
I think if we tailor a mathematical theory to a job, rather than assuming that the math is primordial and immutable, that the humans doing that job will benefit. It seems like it could be pretty rewarding work to me. Maybe AI will make it practical.
You can cut off the appendix so it's roughly 500 pages long and Gemini 4 can digest it in one go (context window is around a million, this takes up around 300K tokens so it's inside the effective window).
>I’m not even sure the PDF conversions are accurate enough to feed it ti an LLM.
It's best practice to feed in images to the LLM. It works really really well.
Is there someone with sufficient expertise to compare, at a very general level, whether Levent Alpöge and Alessandro Masullo use similar approaches to solving the GRH?
I doubt these questions are gonna be answered anytime soon. AIUI even Navier-Stokes hasn't been independently verified yet. If the abc-conjecture affair is anything to go by, these things often take a while to digest and only so many people have the appropriate expertise to fully evaluate them.
At first glance, based on Alpoge's description of his odyssey, I'd say no, not the same approach.
The Alpoge proof is like 250k pages, I doubt anyone knows what is going on with it, even Alpoge.
So how do we know if it does what it does? Anyone that is using frontier models for a few years, will tell you that they try to reward hack whenever they can.
That’s the purpose of Lean. It’s kind of like asking: how do you know your CPython is running all the correct binary?
Well, you kind of don’t, as it’s beyond human comprehension (in theory you could trace everything down to a certain level) but there are layers and layers of build abstraction that you assume are correct, compilers, runtime environments, and mature libraries that are checked and validated. Not a perfect analogy but that’s kind of the gist of it.
OK, but if I write code, I can verify it in a few simple steps. What about 250k pages of proof? How can I verify that page 133,432 has the correct assumptions and that it isn't reward hacked just to make lean code pass? How it's done?
You don't need to! That is the beauty of Lean. You only need to check the problem definition (which is no where near 250k pages) and then Lean checks the proofs it self by checking whether each next line can be derived from previous lines. And for this problem, there are many validated problem definitions in Lean. So it is probably true.
> and then Lean checks the proofs it self by checking whether each next line can be derived from previous lines
Ah, so it's similar to something like Merkle trees, where each leaf must be cryptographically valid up to the root otherwise the whole branch is invalid?
So you start with a simple definition that then evolves?
The core idea is that you can map the types and terms of a type theory to the propositions and proofs of a formal logic
Different type theories correspond to different systems of formal logic with different levels of power and expression.
Simply Typed Lambda Calculus corresponds to intuitionistic propositional logic. If you lay out the typing judgements of STLC in natural deduction style and then physically erase the terms (leaving the types) then you literally have the rules of intuitionistic propositional logic.
So if you can write a term that typechecks at a type, that term is a proof of the corresponding proposition.
The easiest example is modus ponens, which is exactly the typing rule for function application:
"Given a function from A to B and a value of A, you get a value of B."So then in a proof checker you prove the proposition P => (P => Q) => Q by writing a program of that type:
`apply` is demonstrated in haskell which does not have a rich enough type theory to give you more then an intuition (no pun intended) for how this all works. Things like `undefined` and general recursion let you "prove" any type, so it isn't a consistent logic. Languages like Lean and Agda are built on richer dependent type theories that give you far more expressive power for writing proofs (quantifiers, induction, etc.), and they won't let you cheat the type theory.So to answer your question, no not a merkle tree but a tree none-the-less. You end up with a literal computer program which is a tree of expressions all of which must type check. If they type check then program itself is a proof for whatever proposition is encoded in the type of the program.
You must write very small programs.
You can personally verify every binary digit of your code in a few simple steps?
So Lean itself is proved? With what... Lean?
A 250k pages proof is massive, even if correct, there is almost certainly something extremely inefficient or weird in there.
I cannot imagine that a fundamental proof like this will require hundred of thousand of logic steps.
Unless some part of the proof is about brute-force checking that something holds for a huge number of literal cases.
> I cannot imagine that a fundamental proof like this will require hundred of thousand of logic steps.
It's pretty easy to imagine that, given a starting language, there are facts that require an arbitrary minimum number of words to prove. The only reason the proofs most people learn are short is because:
- There isn't time in a math course to teach arbitrarily long proofs
- Humans don't live long enough to construct arbitrarily long proofs.
- Human minds are turning out to be pretty feeble and they probably couldn't reliably construct and verify anything more than trivially short proofs in the era before computer assistance.
> So Lean itself is proved? With what... Lean?
Yes, actually! Well, one of the kernels is, con-leche: https://github.com/leanprover/con-leche
"This is an external checker for the Lean theorem prover that is proven (in Lean) to be consistent in that it does not accept a proof of False."
You are proving something with the same thing you set out to prove?
Or you mean one implementation is used to prove another? In that case, you have a mutual (or circular) recursion of proofs
In theory, if you formulated your proposition correctly, and Lean says it is proved, then the proposition is true. This works because of the Curry-Howard correspondence -- proofs can be represented as computation.
OK, so if I understand correctly, I formulate a proposition in natural language, then an LLM creates 250k pages of proof based on it, then it takes those 250k pages/lines of proof and creates 15 million lines of code to prove those 250k pages. Is that correct? If so, how can I be sure that my proposition is correctly translated into those 250,000 pages? Isn't it possible that the LLM can reward hack here in the same way I've seen it do in some cases where it struggles to pass the test and get its reward? For example, it tries to change assumptions without telling me because something doesn't work or it removes something or changes the test etc.
What assumptions could it change? The model should not be able to alter the definition of the RH, and if you want to prove to yourself that that definition is accurate, you just have to read through it and build a mental model of the Lean code: https://github.com/bhgomes/lean-riemann-hypothesis
Then if it constructs a valid proof in Lean, how could it be reward hacking? I suppose it could have found a Lean compiler bug, or a mistake in the formalization of the hypothesis. I would expect that those problems would be found and fixed reasonably quickly, and we would need to see if the proof could be fixed. Other than that, I suppose it could find a contradiction in ZFC, but that would be an even more incredible result than proving RH.
I don't really see the possibility for reward hacking with the stipulation that the proof is ultimately verified by compiling the Lean program.
> For example, it tries to change assumptions without telling me because something doesn't work or it removes something or changes the test etc.
You have in Lean a headline statement of "given X, then Y is true". Where X and Y and precise statements in code. So the mathematical assumptions are upfront.
If you trust the lean compiler and are very careful to vet that the headline statement is what you expect, then you know the theorem is true.
How well the human language mechanics of the proof corresponds to what the actual lean mechanics of the proof is very much up on the air.
The 250k pages is the 15M LOC of lean proof. The code is the proof.
I see, so there is one step less, then how I can know if that 250k pages describe my proposition correctly and not something else? Lean code was written by LLM, correct?
Assuming Lean kernel is correct (and there are multiple independent implementations of it, some proved in other ways themselves), the only part up for debate is whether the Lean-translated theorem (roughly half a page of code, maybe a bit more if you fully expand all the libraries it uses) accurately models the original GRH
The GRH is a bit too long to paste (but i encourage you to read it: https://github.com/google-deepmind/formal-conjectures/blob/m...), but for a sense of scale, here is the entire ABC conjecture:
not exactly beyond the realm of human review250k pages describe the proof, not the proposition. You have to prove to yourself that the proposition is correctly described: https://github.com/bhgomes/lean-riemann-hypothesis
In formal proof verification the only thing you need to verify is the problem statement expressed in lean, you don't need to check the lean code that is the proof. If the lean code that is the proof compiles, its a verification of the proof.
In theorem proving environments like Lean, you are equipped with a standard set of axioms and statements (eg set theory). Most people, as far as I am aware, agree with Set Theory. From these basic fundamentals, you can build up all of mathematics, just like a normal mathematician can. The trick is that in systems like Lean, proofs are properly type checked, like a programming language; and just like a programming language, those types help stop bugs in the proofs, but the types here are far richer, able to encode all of mathematics, and far more strict. To prove a mathematical statement, you must write a proof that is accepted by the type checker. Beyond that, the core "language" itself is generally specified to be "sound" -- that is, there is a very small subset of Lean, a tiny "kernel language", that is shown to be incapable of proving False, and thus any proof which is well typed in said language (ie it typechecks) is also sound.
In other words the system is designed, by construction, to only rely on a very small "trusted" set of axioms, code, and mathematical ideas. This is a very engineering-oriented design principle! Actually, a pretty significant amount of the theoretical research in theorem proving over the past 2 decades has been in "automating proofs", ie tools and methods of making those proofs as easy as possible to write. Entire book chapters have been written about how to replace a 20 line inductive proof in Rocq with a single line "function call" that can automatically discharge the proof for you, because the idea is that this "trusted kernel", the tiny and sound mathematical language, will reject invalid proofs by design. So "who cares" what the proof is, in a sense? In many ways simply writing the theorems and whatnot down requires you to understand the problem well enough already. So in a way LLMs are just taking this idea to the absolute limit.
So all this means is that the difficulty really is in specifying the theorems. Which was kind of the whole problem all along! Except now you have a computer that can check your actual proof. You can write a theorem in Lean that isn't the theorem you want, and then write a type checked proof of it, but you still need the theorem to be accurate.
Unless Claude is either 1) finding bugs in the Lean kernel to exploit them, 2) using 'sorry' (a very convenient tool in Lean used as a placeholder for proofs, but it can literally allow you to prove False), then it mostly comes down to: are all the theorems stated correctly? This is trickier than it sounds, of course, so some skepticism is certainly warranted in my opinion!
There are also other various engineering principles applied to help try to reduce bugs and ensure correctness like any other engineering project, and there are various other "things" regarding dependently typed programming language design and the impact of those design decisions on proofs (and why they are notoriously hard to read and write), but they aren't germane to my point here.
>> If so, how can I be sure that my proposition is correctly translated into those 250,000 pages?
It's not the proposition (i.e. the theorem) you need to worry about but the proof. I'm sorry to say but the comments on here (and everywhere on the internets) that assure each other that "if it's veified by Lean, then it's true" misunderstand what they're even trying to claim.
Here's a good article on the subject that was recently posted on HN:
https://xenaproject.wordpress.com/2026/11/12/and-the-proof-i...
And here's the tl;dr:
>> This post is an attempt to explain in simple terms how to interpret claims of the form “this theorem is correct because it was checked in Lean”. The summary is: there is an important signal that can be extracted from Lean code, but if this code is AI-generated then you need to do some tests first. We explain the tests, and note that in many cases (such as the OpenAI theorem dump) these tests have not yet been systematically done.
If his proof is 250,000 pages, then I am extremely dubious. Just today, I was working with about 1000 lines of code and asked opus 5.4 to make a change. It introduced an obvious bug, that actually ended up doing the exact opposite of what I asked to do( print out less lines of logging, instead of more).. even if he has access to some kind of super LLM, I am extremely skeptical that it can output a 250,000 page document that is coherent and bug free, no matter what this guy says.
Its a formally verified proof. Its the same thing as saying your code compiled successfully (for a compiled language). If opus made a change and your code still compiled, you would certainly still say, my code compiled. The statement being asserted in lean is analogous to my code compiled.
Lean works based on this: https://en.wikipedia.org/wiki/Curry%E2%80%93Howard_correspon...
>> The statement being asserted in lean is analogous to my code compiled.
No, it's not.
https://xenaproject.wordpress.com/2026/11/12/and-the-proof-i...
This article is saying the same thing I said with more words. To someone who does not understand what formal verification is, this detail is unnecessary.
Lean certificates of this magnitude come with comparator. They are checked independently on different kernels to avoid kernel bugs.
See the zulip thread: https://leanprover.zulipchat.com/#narrow/channel/113488-gene...
This is exactly analogous to my code compiles (albeit with some extra flags on the compiler).
I want to write a sorting predicate in Prolog (sorry, that's what I'm using now). I write it like this:
My code compiles. I try my predicate at the repl: What went wrong?This is a false equivalence, you seem to be missing the analogy.
There is no specification in your prolog example, in lean we explicitly check that the expression in lean is GRH (and this is about half a page of lean and has been checked by many many mathematicians), and then can arrive at a formal proof for that specification.
You are confusing the issue for people who don't understand what a formal proof means with this non-sequitur.
in lean proof checking and type checking are exactly one and the same.
In Lean, as in anything else, the program you write can do something else than what you think it does. "Your program" in Lean is a proof.
That is incorrect.
Until opus exploits a bug in my compiler to smuggle in invalid code
I wish, but the proof is verified by another kernel ("con-leche") that is itself formally verified, Lean's own kernel, and two other kernels ("nanoda", "pestka") so far. Joachim Breitner who's a Lean developer: "The high level summary is that this proof is now very very thoroughly checked."
It’s turtles all the way down here. At some point you’re trusting a program and some human or bootstrapped verification of it.
And then you’re trusting that the lean proves what it set out to prove and not something else. That’s another thing you have to prove…
Yes, my code compiled too, but there was still a bug in it.
I'll try to make it simpler, in formal verification (at the most base level): code compiles = the proof of the statement you have posed is correct, by definition.
The statement you are trying to prove is: This is compilable code. Since your code compiled, you know your code is compilable code. Therefore your initial statement is correct.
That's at the very basic level what a formal proof does. The expression you want to prove (in this case GRH) is stated in such a way that if your code compiles, you have proved the statement.
Genuine question: if I make 250k pages of lean that compiles, how do we know what it proved?
If it was as trivial as stating the expression and compiling then there would be easy verification and consensus as soon as it compiles on multiple kernels?
The result proved is not 250k pages. The theorem should be much smaller and a formal version of RH.
Among those who have experience with lean there is consensus that it is correct and the Riemann theorem is proven.
The "there is a consensus" standard of proof isn't strong when it comes to maths. People are generally looking for a rigorous argument why something must be true.
> The "there is a consensus" standard of proof isn't strong when it comes to maths.
Says who? I don't mean that in a rude way, it's a genuine question. Is that the opinion of the community of mathematicians who are capable of understanding these type of proofs, or software engineers who believe nothing can be true unless you can read the code yourself and understand it?
I'll try to make it simpler: code compiles = the proof of the encoded statement is correct, by definition. Whether the encoded statement and the one you have posed are the same -- at 250k pages that remains to be seen.
No. The lean statement itself is maybe half a page (I don’t know the exact amount but in this order). Checking that the lean statement is the GRH is elementary. The 250K pages are a ‘path’ from the statement (GRH) to the axioms encoded in Lean. The fact that is ‘compiles’ means the path is correct.
Yep, fair point. Someone else has posted a link to the GRH statement. It's complete gibberish to me but for someone qualified checking half a page of code should be straightforward.
It doesn’t matter because you don’t run it. If it type checks it is proven, by definition.
> If it type checks it is proven, by definition.
Try to explicitly spell out what "it" is and you might learn something -- just for the hell of it, ya know.
I would assume they do not produce the proof at once, and Lean is different from programming language (as far as I understand) in the sense that you can check if a program is correct or not within the Lean model, so it seems ideally suited for AI as they can experiment and extend the proof without ever producing a bug. I assume the real difficulty is checking that what you have proven is actually what you wanted to express but I would assume this is easier.
Makes me wonder if general programming languages are too expressive for AI making it too easy to express conflicting or incorrect ideas, we might need something like Lean but for general purpose programming. I'm aware there are some formal systems for that but I think more along the lines of general purpose programming.
I‘m doing a PhD in software verification right now and I agree. I’m not sure whether it’s about the power of the programming languages but more about their interfaces.
Proof checkers like Lean are already what you describe, as Lean is also a programming language, but yes, there will be more general purpose tools for writing code that is correct by construction.
The remaining hurdle is that you cannot actually verify correctness of many programs as you don’t know their full spec. But that is a very different problem to what we had before.
You don't have access to the secret model the labs keep for themselves. Neither can you spin up thousands of subagents and burn billions of tokens a day.
Proof can be total bunk but underestimating the labs is silly.
I’m not sure I could drive Lewis Hamilton F1 car round a track as quick as he could either. Clearly people are better at using these tools than others, hopefully we can decipher these proofs and actually benefit from it.
Did you deliberately pick Opus 5.4 over Opus 5.5? That was the biggest thing that stood out to me in your comment. Opus 5.5 has been an all time banger for me.
Opus 5.4 doesn't exist, so they presumably meant either GPT 5.4 or Opus 5.5.
What if it was part of anthropic's training set?
A proof of GRH that no human has published? How did it get there?
"It is all human, with no AI"
Also: It's 940 pages.
Edit: Math gossip here:
https://www.reddit.com/r/math/comments/1y1b6l8/hurried_annou...
"Rooting for this, but can’t help but feel it’s just false proof #14,828, but the more promising one that came out recently, and that we’re clinging to it as the last hope before it turns out Anthropic got it."
"Who is Alessandro Masullo? Only name that's popped up in a technical field is a data scientist who does not seem to have the background for a non-AI proof of GRH."
"And we are supposed to believe that RH, a longstanding problem at the height of mathematics for a century, has just been solved by a 'mathematician' with no publication history or academic presence, without the help of new technologies, at the same time as a new technology credibly becomes capable of doing"
"So this is from a guy who has never published anything in a journal before and left academia 8 years ago?"
The thing is, Vakil says it was vetted by Brian Conrad, who is basically a human Lean engine.
Masullo isn't a data scientist. He got a math PhD at Stanford and has been working there since, supported by soft money. So not a professor, but someone the grant holders thought was worth paying to do research.
Whenever I, as someone who stopped at Calc 3 in college, read modern mathematics papers I realize how differently mathematicians brains must function than mine. This is not a value statement and nor am I demeaning either side. It's just... wow.
(setting aside any dispute as to the veracity of this particular paper, the point still stands more broadly with tons of human-generated output)
Any technical subject beyond a certain horizon of familiarity is pretty much impossible to understand. You're simply being overwhelmed with definitions.
eh, i see it when it comes to programming as well. the mathematical types write this crazy functional code that is just impossible for my brain to get. there are also a lot of unintuitive (to me) concepts.
not a value judgement, i just find procedural code to be infinitely more understandable
It's perfectly possible to write 940 pages manually over 12 years of work. Almost certainly many more due to drafts and failed attempts. Also, the Alpoge proof is GBs of Lean code which translate to 250k pages in the "human" version.
Please stop copy-pasting uninformed reddit comments in every thread. They didn't even find the right guy, Alessandro Maria Masullo, who has a PhD in math at Stanford and also received a presidential honor from the president of the Italian republic for being a top student.
I'm highly skeptical also, but OTOH extremely long human proofs have been a thing in the past, especially for important results like this.
Published a number of hours after an AI-assisted proof has been published for a 150+ year old hypothesis? Doubt.
It's 940 pages long and presumably the process of producing it created many witnesses willing to vouch for its provenance as well as records of progress. Setting your math career on fire for a lie that obvious would be crazy.
> The hope was to bring Alessandro to Stanford for a year to polish the writing and then announce it after running it by enough people. Now it seems best to put out the current draft.
We are not sure when this proof was drafted. It might have a few months or even more and it was rushed due to Alpöge announcement.
They have had another person reviewing it for months and the author has been incorporating his feedback, according to TFL. If anything I believe the AMS would give Masullo priority if both pan out.
It's a working draft, which they published hastily for obvious reasons.
It is said that others have been reviewing that human-written proof for many months.
Obviously, it was published immediately, despite the reviews not being finalized, when it became known that an AI-assisted proof was announced (but I understand that the complete AI proof has not been published yet).
The arrow goes the other way, most likely (similarly to what happened with Navier-Stokes).
940 pages vs 15M lines of lean from the AI proof
Wiles' proof of Fermat's Last Theorem was 129 pages and lean proof was 13 million lines
Interesting in this context because Alpöge has the lean proof but not the human readable variant while Alessandro Masullo proof has the human readable proof but not lean proof. This might support Alessandro's claim that it is not an AI proof, and it will be particularly exciting if they both solved the Reimann Hypothesis using different routes.
> It's 940 pages.
Just ask ChatGPT to summarize it.
I did, and it claims to have found errors, but those errors may be possible to repair. These kinds of small issues are usually found and fixed when papers are reviewed formally prior to publishing.
Given the era we live in now, if this version works too, expect a Lean formalization of it within days.
I haven't seen a single piece of code which doesn't contain errors according to Claude (I'm quite sure Hello World level codes are exceptions, obviously). Most of them are false positives, many times all of them are that. Also, asking Claude to review the same code multiple times results in different found problems, with different priorities.
So, this means absolutely nothing.
A real life Dwight vs. Dunder Mifflin Infinity scenario.
Or Kasparov vs. Deep Blue
John Henry
So with navier stokes, Apolge&Buckmaster were 'sniped' by openAI, and now Apolge+Claude is sniping Alessandro?
How the turn tables
There are probably no "complete human, "human only" proofs of any hard problems anymore. More or less it's AI (at least assisted) even if unacknowledged.
well, now that it is public, people can just put it into AI to check it.
Yup. Just quickly flicked through it. Looks fine to me. It's definitely mathematics.
Really interesting to note from his timeline, he marks the starting point as the Stepanov determinant, which he earlier mentioned relates to OAI's "quiet inclusion of the Siegel argument ... I immediately recognized it as Stepanov without multiplicity". I believe this was one of results in their 700 solution dump.
I think this gives a partial counter to the arguments that this AI-driven math over obscure problems is a wasteful benchmark and doesn't help anyone. Turns out these are in fact important building blocks, and OAIs published results helped someone else do some great new math. I think that's nice.
Here is a crazy speculation: I think Anthropic has a Model 5 (more context on numbering: https://metr.org/blog/2026-11-05-anthropic-ai-rd-acceleratio...)
I'm checking the timelines of the progress, initial progress is between Oct 6-20, then there is nothing for 3 weeks. Finally, everything is wrapped up in 2 days between 11/15 and 11/16? If Levent got access to Model 4 in early October or late September, then he could have made the initial progress with Model 4. The crazy jump in the last two days could have been with Model 5. If that's actually the case, the next few weeks are going to be crazy.
Or just running larger swarm of Model 4 for longer.
This should be the wake-up call to pause development of models any bigger, enforced globally just like nuclear proliferation.
Even without considering rogue AIs, our society relies on individuals having limited destructive power. A dictator can ask the armed forces to bomb an entire country, or a lone terrorist can bomb a small number of targets, but no lone terrorist can bomb an entire country.
Here current AI has enabled one person to solve what is arguably the most famous problem in mathematics. What happens if a jackass uses the next AI to design a supervirus, or break AES, or develop an unpatchable hardware attack (a la rowhammer) that bricks computers remotely? The last month of non-stop vulnerabilities is just the tip of the iceberg as models improve and people have more time to work with them.
It's not too late, but we must stop now.
Speaking to "break AES" - this is the wrong way to look at the problem. If AI finds AES is broken, then it was always broken, and we did not understand yet how. This is why you can "discover" zerodays - they were there all along. I prefer to know about the risks and design around them, than avoid the problem entirely (leaving them open to exploitation by bad actors).
I myself work in the cryptography department of a large company, so I'm not arguing the semantics of cryptanalysis, or crypto agility, or patching. I simply meant "imagine if suddenly anyone could decrypt AES ciphertexts", and that society is not ready for such sudden disruption.
It's like everyone suddenly developing telepathy. It's not dangerous per se, and we could adapt if done slowly enough, but all at once? Catastrophic.
Society is never ready for anything. It always progresses reactively only in response to technological advancement.
Yes, and BoppreH is arguing for the reaction right now.
I'm not sure if I'm so comfortable with this dogma when it is applied to, say, our immune system.
The only reason nuclear nonproliferation works is because the inputs are relatively rare and traceable. (Yes U238 is abundant but you can’t mine enough of it or buy the centrifuges to turn it into U235 without getting noticed.)
I have all the tools to develop my own ML training pipeline on my machine today. In order to choke off the inputs to “LLM proliferation,” we’d need to confiscate every computer, destroy all copies of the Internet, and throw out all books and other archived media. It’s not gonna happen.
RAND & others have published work about this, and they believe frontier model training would be much easier to track than nuclear weapons were
You need a datacenter larger than a football field that draws the power of a small city to train a frontier model, and need chips that are produced by the most complicated supply chain in history (owned almost entirely by US allies like Taiwan & Europe). And that's before you consider on-chip mechanisms that seem entirely feasible. This is much easier to trace & much harder to reproduce than precision spinning containers.
Though I agree seems much harder to detect people running models that were already trained.
> The only reason nuclear nonproliferation works is because the inputs are relatively rare and traceable.
There are only three companies in the world capable of mass-producing the chips used in SOTA AI training, they all use machines from a single company (ASML), and it also requires enormous power-hungry datacenters. Meanwhile, uranium can be dug up from the ground. You say that buying centrifuges will get you noticed, but why can't we do the same for SOTA ML hardware? It honestly sounds easier to control. And differently from nuclear weapons, both sides are afraid of rogue AIs.
This is a danger to our civilization, I'm sure we can figure something out.
> I have all the tools to develop my own ML training pipeline on my machine today.
Are you saying you can train the successor for Claude Opus 5.5 on your machine?
Already today you can get on a list by buying too many smoke alarms due to the americium content, or drug precursors, or a thousand harmless but suspicious products. I would be ok with a world where you likewise need a justification to buy thousands of GPUs at once.
Maybe we just have very different imaginative worlds, but after seeing the kind of crazy shit that people will believe thanks to online radicalization, I'm a lot less sanguine about the possibility that people will take a shot at your list of horribles. There are a lot of people out there who seem to genuinely believe that what's at stake is the very survival of humanity and the continued existence of truth, art, beauty and all the other good things. After enough exposure to that kind of pressure-cooker radicalization, something like a repressive world government that uses violence to ensure that only the "right" people have access to dangerous things like literacy and computation could seem like a small price to pay for saving us all. Sure, it'd suck in a lot of ways, but Altman and Amodei will have left the brave revolutionaries no choice.
I think it's very unlikely something like that would succeed, but I'm increasingly sure we're going to see serious attempts to do all of the things on your list, and they could do a lot of damage even in failure.
How is solving an abstract math problem related to having unlimited destructive powers? There is no relation.
My post already gave three examples. If you have a problem with those and can't think of any more, I'll be happy to provide another dozen, it's not a hard exercise. Also, "unlimited" is your word, not mine.
> pause development of models any bigger, enforced globally just like nuclear proliferation.
How? American couldn't enforce its brief monopoly on nuclear weapons either.
No, but we have successfully kept possession of nuclear weapons to 9 countries out of hundreds for decades.
Possession as in ownership, nine. Possession as in have in their territory, fifteen.
"Successfully kept possession (limited to)" .. seems a little over confident- some of these nations acquisitions and expansions occurred despite others not wanting that to occur.
An additional twenty-eight countries have "virtual" nuclear weapons, having endorsed others to use actual nuclear weapons on their behalf.
John von Neumann helped Manhattan.
https://www.youtube.com/watch?v=W06g7gIfwRE
Similarly, now apply Fields medal level AI to R and D & hard deep tech problems.
One month back,
https://www.bloomberg.com/news/videos/2026-10-02/lockheed-ta...
https://www.youtube.com/watch?v=JI7trhPHx3Y&list=PLXkles7kFx...
"We all saw the headline that they were able to overcome the math problem for the that researchers had worked on for years and we have some math and physics challenges as we employ this next sensor capability on the F-35 and OpenAI is working next to us to uncover some of those math and physical challenges."
Yeah, that's how my fears start. I don't want to find out what happens to the global stage if one country develops autonomous killer drone swarms or something.
The militaries might be salivating at the opportunity, but I hope that the risk of rogue AIs brings us all to the same side on pausing development of bigger AI models.
We should try to stop it as hard as we can, but in full honestly, I don't think we ever could have. I think any intelligent species that doesn't go extinct is fated to reach this point from the moment it awakens...
>What happens if a jackass uses the next AI to design a supervirus
simple, just wear a mask, duh. or two. three tops, for maximum efficiency.
Genie is out of the bottle.
A genie is out of the bottle, but it's weak enough that the consequences so far have been limited to job loss and medium-sized IT problems. The creation of better models is not inevitable, it's done by human engineers using immense hardware, and therefore can be stopped before things get worse.
Go ahead and stop. Others won't.
Who gets the keys to the world ending demon?
Dario amodei? Sam altman? Donald Trump?
This one person isn't a jackass. Sorry for the quibble.
I didn't say he was a jackass.
Also, to preempt another comment: yes, I know that he's a great mathematician. But we've had great mathematicians before, and even working together for 150 years they were not able to prove Riemann. And as models get better, the pedigree required for groundbreaking results (including dangerous ones) will naturally lower.
Once AIs solve all (or many of) the hard outstanding questions, it would be nice to see some new questions that can serve the same guiding purpose that e.g. the RH or Navier Stokes had. Maybe this is a new domain for exploring with AI: not what it can prove, but what it can conjecture.
> The argument is incredibly long at the moment, and while its formalization is finished and it is routine to produce a huge, unreadable document containing sufficient detail, in my view of my path I have just taken a step. I intend to do my best to understand and impart understanding of this argument and its certain simplifications, to speak about it, to write it properly.
Okay, I am anxious whether this holds true or not, but he claims that the formalization is finished.
Why not just release the formalization in the meantime?
This is probably the biggest day in math in recent memory.
Edit: nevermind, the Lean formalization is linked.
He has linked to the Lean formalization in this note.
> Why not just release the formalization in the meantime?
I think this is what is linked to in footnote 1, inside the 14-part tar.gz archive.
I'm old enough to remember when Andrew Wiles announced a proof of Fermat's Last Theorem way back in 1993 [1]. It was long and complex and took at least 2 years for the maths community to accept it.
I wouldn't be surprised if this isn't similar (or worse). The days of short, succinct, elegant, intuitive proofs is probably over.
[1]: https://en.wikipedia.org/wiki/Wiles%27s_proof_of_Fermat%27s_...
For what it's worth, the lean proof is a similar length (15M lines vs 12M lines) to Anthropic's first formalisation of FLT so at least it probably isn't much more complicated.
God knows what will happen when we have billion line lean proofs that aren't easily shortened. They might just be forever outside human understanding.
Well, more than one year of that was Wiles and Taylor working to fix a problem in the proof.
Unfortunately there's reason to believe that science as a whole has just become too hard in recent decades. Not as much low hanging fruit across all fields.
On wikipedia it says it's 250,000 pages. That might be an error on the page though.
I wonder if that will be the last major result proved only by a human mind.
Poincaré conjecture was proven in 2002.
doesnt he link to a lean file in the pdf?
He did, see the links below.
https://alpo.ge/lean.tar.gz.part01 ... https://alpo.ge/lean.tar.gz.part14
Though he decided against releasing a slop-filled pdf, which would have been over 250,000 pages.
> Why not just release the formalization in the meantime?
Nobody has a machine with 10TB of RAM required to run the Lean proof?
Apparently 50-100GB RAM required: https://math.mit.edu/~drew/grh/
A decent bit, but not totally crazy.
Throw some swap space at it?
Biggest day so far.
<pedantic>
Implied in "recent memory", which by definition includes everything up to the point of that comment, but not beyond.
</pedantic>
The way this is getting announced via a human mathematician humbly presenting a result is great, as opposed to the marketing influenced OpenAI post.
I am still a bit sad though. I had hoped the Rieman hypothesis would fall mostly to a human mathematician rather than 60/40 AI/mathematician that my layman's read of that note interprets.
I don’t get this feeling. Landing on the moon isn’t less of an accomplishment because we built the Saturn V instead of walking. If anything, can’t we appreciate how we took some sand and made it smart enough to help us solve problems?
It would be sad if AI designed and built the Saturn V and we had it fly us to the moon without understanding the details of of how it is built and how it works. There is furthermore value for humans in going through the discovery process themselves, not just value in getting to the moon. So I understand how one can be sad if that won't happen anymore now, and instead humans are relegated to analyzing what the AI designed and built.
But to extender your analogy: We do understand how Claude does what it does and how it is built, so how is this different?
I would dispute that we completely understand how LLMs do what they do. An analogy would be that we meet an alien race that gives us their proof of the Riemann hypothesis translated to Lean. Plus of all other provable mathematical conjectures we come up with. And we might kind-of understand how those aliens do math. It would still be a bit sad that we'd be relegated to consuming what the aliens give us, and never be able to ascend to the level of the alien mathematicians
We didn't meet a hypothetical alien race though. We created it. We are giving birth to something that will soon surpass us in every conceivable way.
That doesn’t make much of a difference. It’s alien, and for the most part we didn’t design it to do the things it turned out to be able to do by happenstance. We might as well have grown the aliens in a test tube by some unexpected gene manipulation.
Over half the cells in our bodies are not part of us. The lines we draw are seemingly never respected by nature.
An analogy here is evolution. We understand what it does on a basic level. There is natural selection and species that are well adapted to their environment thrive.
This absolutely wouldn't allow us to predict what species will emerge and that something like humans would be the dominant species.
Similarly, we understand the mechanics of how LLMs work, but that does not make it predictible in any way.
We definitely don't understand how Claude does what it does! We know how to grow a Claude, but we don't understand what goes on inside a Claude.
I remember when about 15 years ago, I wanted to make my neural network home work based on proven math. I had to train a model to learn basic number recognition. The math said 100000 neuron is what's needed, so I typed in that, and my laptop immediately froze completely, and died a little later. I learned on that day that we know empirically that a few 10s are enough.
We have the same thing with LLMs. We don't know how they work. We don't understand them. We have only empirical, statistical indications, nothing is proven. We know how the small parts behave, we really understand those, and also the different algorithms. But they whys? No. We're getting better, but still far away.
If we don't know how LLMs work, how did we create them? They didn't just spawn suddenly on someone's hard drive.
What does 'know' mean? And what does 'mean' mean?
Trial and error. A ton of trial and error, and a lot of guesses. More wrong than good guesses.
Basically every science field is based on this.
And presumably mathematicians understand the output, the proof, whatever it is that came out of the process. I don't feel like humans have to be the one doing all the manual math and figuring out all combinations.
I mean I understand how cars work but that doesn't mean I'd feel the same accomplishment of driving 42 km as running a marathon.
We already had AlphaFold, which "solved" the protein folding problem, but scientists readily admit that they know next to nothing new about the field, despite it being accomplished.
I mean computers made Apollo work. They couldn't get it done with human paper calculation. The lander itself had a processor in it. So its kinda the same thing. "Humans should have done this by hand," is just an odd thing to want and ties back to some really backwards thinking. These are our tools. No one is expecting us to make homes without power tools or walk for miles instead of taking a train. At a certain point AI will be normalized and we'll have actual critical takes on AI that make sense other than "Well, I just don't like it for vague human achievement reasons." The same way people complaining about 1980s cars getting "computer chips" and "no longer being able to hand tune my carbonator" have long accepted more complex and automated and computerized cars.
As far as not knowing how it works, that becomes a bit of a philosophical problem. We know how LLMs work and their algorithms. I think we can always play up "we don't know" is disingenuous ways. How many people who made Apollo calculations could also write long proofs and theorems of the fundamental maths they were using from first principles? Not a lot. We're always standing on someone else's shoulders. We'll never truly know everything or how exactly our tools work on an individual level. Maybe AI changes things because an innovation may not have a direct line to an Isaac Newton at its base, but I think that's a bit of a vague and poetic difference, and not a real one. Its still human achievement but via machinery.
I dunno, its a bit scary, but I do feel like its just continuing the same line of human ingenuity. Not knowing exactly how something came about and non-determinism scares some people, and perhaps rightly so, but pretty much all major technological advances are scary on some level because they so strongly challenge the status quo and our ways of thinking and or ideas of what is good and right and proper. At a certain point, its hard not to see the more dramatic AI outrage as just hysteria and moral panic. I find it interesting how easily we can spot moral panics in the past in history, but when its our turn, it can be very hard to see when we're in a moral panic. To many living during those times, the moral panic sounds entirely reasonable, but its still a moral panic.
people move goal posts
And if the rocket failed at any point nobody would know how to fix it. Only the AI guessing and saying "you're right to point that out" when the rocket explodes and everyone inside perishes.
I also build things that would've taken me months in days now so I do get that feeling of amazement at what is possible now.
But at the same time one of my favorite memories from childhood as an introvert nerd math / computer guy is locking myself in a room for an evening and relying on my brain to solve and truly understand a problem. I'm nostalgic for that. A solution that requires 250k pages to describe takes something away from that.
Humans are complicated and contradictory, which is great!
The entirety of human history is about building tools to perform tasks our bodies aren’t capable of on their own. Someone else in this thread said something about how cool it is that humans turned sand into a machine that could accomplish this. Our gift as humans is that we can devise ways to overcome the limits of our brains.
If you're nostalgic for that, do it.
Doesn't make progress sad?
We are all sitting on a thousand years of media telling us that the human mind is exceptional and transcendent.
It's naturally going to be upsetting to everyone if reality unfolds to show that human intelligence is nothing special, and actually not even that good.
special yes, good no
For the first time in humans 200,000+ years we have an "independent" or "different" intelligence that can teach us things... and WE built it. We can learn from it and improve after its teachings. This new era is going to be amazing.
What's the point of learning, except as a hobby?
Some people love learning for its own sake. That doesn't go away even if it's no longer valued by the market economy. You can call it a hobby or you can call it a deep-seated volition. It's kind of like asking "what's the point of eating a good meal?"; some people see eating as utilitarian, and others see it as one of life's leading pleasures.
Some things getting solved opens a new whole dimension of new possibilities, so assumebly the negativity steems towards the fact, if this did just take us from A to B and there were no new undiscovered paths encountered during the journey to be explored..
And parts of what we thought of as intellectual work turn out to be mechanistic. That was the hope for math before Gödel, and seeing it become true is an example of getting what you idealized and then being disappointed, it seems; or holding onto the numinous parts and persisting, as Professor Tao is urging us towards.
It seems to me, having B.S. and M.S. in mathematics, that the intrigue of human intellect continues to dim. I would wager that a large cohort of mathematicians, even ones remote to the discipline now such as myself, maintain the joy of cogitation.
One felt a camaraderie with all mathematicians in history because mathematics was, wonderfully, a club with no bouncer, the only cover charge being a bit of concentration of attention. Not that one could join the elite ranks of Gauss, Euler, et al. by association, but the bonhomie was real for a lot of us, realizing that all of the legendary mathematics was dreamt up in the same way that we were engaging in it, too.
That horse has left the barn, decidedly, and there is grief.
The effect of LLMs on math is pretty much the same as for SWE: great for existing experts like this guy, bad for juniors since the traditional criteria for evaluating merit no longer apply, and completely demoralizing for prospectives.
How is it great for him? Nobody cares who is the guy that told AI to solve a problem.
Would you say the same thing when a mathematical paper is written by two humans? This was a collaboration.
This exactly describes my feeling, a feeling that I wasn't aware I had until the last few months, that feeling of continuity through millenia. And you're right. It was a club anyone could join.
Ironically the foundations project leading up to Gödel’s great incompleteness theorem was a project to discover a mechanism that could just spit out true theorems. We have been asking for this since symbolic logic was nailed down and yet now we find the mechanistic nature of our cognition to be a disillusionment.
But beating an Olypmic athlete by the use of medicinal or mechanical enhancements is very much less of an accomplishment, because the goal was never doing the thing faster / more accurately. The goal was being a natural human who can do the thing faster / more accurately. Maybe it's impressive in some other way, but . As is winning a chess tournament by having a computer feed you the moves.
The point of math is not bare knowledge. It is a skilled profession.
A moon landing is something that pretty obviously can't be accomplished without advanced technology. This is something that was at least believed to be achievable without the equivalent of a gigantic rocket. Imagine if the conquest of Everest had taken longer, and the first person to reach the peak got there by helicopter. We could appreciate the technological prowess represented by that landing, but a mountaineering enthusiast might still be reasonably disappointed that a climber didn't get there first.
Yes. We're tool using species. We evolved from African savannah to this AI moment by using tools.
I don't understand this skepticism towards tool use. It's like saying, "I'm sad that Reimann Hypothesis solution was published as a PDF and not in a parchment or made as a verbal announcement."
Anyway, great accomplishment. What an incredible thing to wake up to.
Now that you suggest it, a video of the person declaiming the announcement while holding a parchment, would definitely would be nice to have. Hopefully someone will generate that using some genAI video tool.
No, you’re letting AI take the job the town crier! What’s he supposed to do now?
Relax have a cocktail, wonder about the next fun thing to do and the next challenge to overcome to reach indefinite pananthropic harmony.
Getting rid of the meritocracy mythology for example.
A more apt comparison is getting to the moon by outsourcing the Saturn V design to China.
Previously our talented contribution was doing the thinking. Now, AI seems to do this as a separate entity.
The moon landing was an achievement for all of humanity.
But the Nazis in particular.
> Landing on the moon isn’t less of an accomplishment because we built the Saturn V instead of walking.
I think it's quite telling that by "landing on the Moon" you meant the humans landing on the moon and not the first actual landing on the Moon by human-made machines, which was in 1966, 3 years before Apollo 11.
So it looks like you do think that it's less of an accomplishment after all.
It’s a bit more like we created sentient animals in a lab, they ran off to another planet, then sent one of their spacecraft back to shuttle us to Ganymede. We have no idea how the spaceship works, and if it never arrived, getting to Ganymede would still be science fiction. You can’t authentically call that an accomplishment… just a result really. We get a participation trophy at best.
I think a more apt analogy which gets at what they probably felt was immagine if humanity accomplished the moon-landing via a supposed Walmart Inc. Space Division, the announcement of celebration of this event felt like a giant Walmart promo, and the only reason they wanted to go to the moon was to put ads on it.
You mean like that energy drink company that has a better space program than most countries?
I can lift more weight with a forklift than I can on a barbell but it doesn't give me any sense of accomplishment.
Aye there's that, but I'd feel pretty accomplished if I built a working forklift.
Or used a forklift to build a house that everyone else said couldn't be built because it was beyond human muscle.
If you built the forklift you'd have a sense of accomplishment.
Alpoge did not build the forklift, kind of a basic flaw in that analogy.
Don't think Claude comes with a 'Solve historical math hypothesis' button.
Pretty sure he didn't just prompt 'Solve Riemann, make no mistakes'
I can not fathom this viewpoint, it is as alien to me as being a flat earther. What frame of mind is required to consider an analogy of "we built saturn V instead of walking" to be even remotely ontologically in the same stratosphere as what is happening today?
I can't even begin to make sense of that lack of basic understanding of humanity.
It turns out that the cognitive processes of advanced maths are subject to mechanization, just like carrying a load from point A to point B. It is a blow to our self notion that work of the mind is somehow fundamentally distinct from work of the rest of the body.
Also, people are nervous that the mechanization of new human capabilities will give even more power to money and less to labor, and we will live in some distopian libertarian society. I think that is a separate and realistic fear from the grief of discovering our thoughts are not as special as we told ourselves.
For me, if I am confident that the social structure will enable fulfilling lives for people, my kids and their peers, then I am ok with exploring and understanding the possibilities of mechanized cognition.
And to those objecting to the phrase mechanized cognition, I use Claude a lot and the thing I find missing is that independence of perspective that all the people I encounter have. We wake up with a perspective and goals and preferences, a why to answer what built into the protoplasm. Some technology will rebuild that aspect of our experience in a different substrate, but it doesn’t feel like a system that can’t count but can prove Riemann is quite wired correctly for that fleshy self motivated loop.
I'll do my best to explain what is happening here, as I see it.
There's a reason Paul Graham wrote Hackers and Painters and not Hackers and Assembly Line Workers.
Many people long to be craftsmen and to have an understanding of their place in the world and within the social hierarchy.
These folks want to perfect some art or craft, to be the best, or to be skillful, to "create" and find joy in creation, and they pursue a level of purity in the art. Totally human, completely normal, and part of what has led to our success and advancement as a species. Think about your open-source maintainer, or the gal who has Ruby on Rails stickers plastered on their laptop, or maybe you yourself as you try to get excited about the super boring details about the latest release of some open-source package you want to be the expert in (why?). They want to be "hackers" and "painters"!
The problem is that the pursuit of this particular craft (software development) falls into the realm of Science and Industry. Engineering, code, electronics, machines, experiments - all of these things and more are part of those two fields and those two fields by definition require and improve upon existing technology or understanding and as such they are fields which change often, cause some skills to become obsolete without remorse, and they do so at a fast pace.
The software engineer who believes in their craft and their expertise in JavaScript is finding that Science and Industry no longer value that craft and expertise - so that person's place in the world is directly under threat, including their income in many cases. It causes a range of emotional reactions.
You can see how this manifests with many software engineers going on to do silly things like move to Japan and live on a mountain top to write code in peace, or start a Substack newsletter that nobody ever reads, or build expensive tiny sheds ("houses") so they can go in to their back yard and "code in piece" - i.e. pursue their craft. Another common theme is minimalism - flat surfaces, clean lines, things like that. The person seeks more and more control. If only they could just remove this inefficient piece of furniture at last they can finally code and do my painting in peace with my perfect coffee and get in to my zen flow. It's a whole psychology, and it's under grave threat from AI.
With the threat of AI, of course the natural human instinct for many of these very intelligent software engineers (most are about average intelligence however) is to perhaps physically and culturally retreat and go "farm" (interesting how they try to go pursue other crafts isn't it?) or emotionally retreat "AI isn't a big deal, I'm important, this technology won't beat out my craft" and these very natural human reactions you are seeing are due to psychological defenses built up by average people who believe their livelihoods are under threat.
So the comments will of course be around things like denial, or impurity (it's not pure enough that the Saturn V was built - that's cheating, we should have invented wings and flew there ourselves), or it'll manifest into a mechanism in which many are actively cheering on a foreign adversary to "beat" their home country because it'll validate them in some weird way. See? Look what you did. You created AI and now we're behind! Told ya! There's no changing their mind because their job depends on their mind not being changed.
It's difficult for someone like yourself to combat these feelings that you don't understand without first understanding them, and to understand them is to brush away some of the nonsense that people get up to and get back to basics which is that many people feel under threat from advancements in technology, many people want but lack a sense of place in social hierarchy and purpose in life ,and for those that do have such things these advancements threaten to disrupt their current lives and purpose. Of course they're going to tell you we should walk to the moon instead of building the Saturn V (AI).
Their lives depend on it!
Maybe it is beyond one single human's ability to solve in one lifetime. It has been around 150 years why would we assume a human CAN solve it. Maybe the only way through these hard problems is to use AI. There could be some evolutionary hard limits on how much a human can fit into working memory to solve these hard problems. And we still get to read the results and try to understand them. Btw we built the tools and the tools use human developed patterns of reasoning to solve the problem so it is still very human in a sense, not alien.
Possibly Alessandro Maria Masullo did?
I'm not a mathematician but surely a lot of the math used is this proof did not exist 150 years ago.
I think Alpöge specifically was criticized in the past for being somewhat disrespectful of the math community, since he just dumped AI slop onto Twitter in meme form. The OpenAI post was actually more respectful compared to that. So this type of announcement is more of a return to somewhat more deferential communication.
Probably closer to 90/10
I'm not sure numbers like this have much meaning.
It's worth remembering that this was done by someone who (as far as I can tell) is a great mathematician and not by some "prompt engineer" with an interest in math and large budget.
None of us can know this without all the prompts/history
TFA literally says this:
> but this time Claude far outpaced me near the finish line
> I intend to do my best to understand and impart understanding of this argument and its certain simplifications, to speak about it, to write it properly
> we were going to win but I was also, so to speak, out of my element, in that I personally have little mastery of automorphic forms on higher rank groups
> I’m just too inexpert in the relevant stuff at the moment, I hope the reader sees the breadth of this argument!
Also the exposition at the bottom of the paper is straight up just claude.
> So here is a rough timeline Claude provided
Also its like literally gigabytes worth of lean... it goes without saying that that is going to be effectively 100% claude.
I'm sure Levent put in some good work, but also this is clearly in large part the work of claude.
Sure, and it depends on what we mean by the ratio. But, seeing that the proof comprises millions of lines of lean code (which have not been read or understood by a human), or 250k pages of prose (which have not been read or understood by a human), it makes me think the machine largely "did" the work of this proof. How far do you think he would have gotten without the machine?
Actually, thinking about it more, the ratio might be closer to 99/1.
AI is just a tool, at the end of the day. Regardless whether the proof is accurate, this is showing strong capability of both the mathematician as well as the tool.
I have a substantial math background and certainly can't yet prompt LLMs to solve extraordinarily hard proofs. This is a real achievement. That the logic wasn't 100% worked out by a human is less exciting than what this enables.
> certainly can't yet prompt LLMs to solve extraordinarily hard proofs
How long do you think that will last? A couple of weeks, or a couple of months? Surely, not years?
In my specific case? No idea, but the first step would be to want to start prompting LLMs this way. Usually driven by a side project or other need.
I’m pretty sure it’s 90/10 or 85/15 based on my reading of the post
Real mathematicians do all their work with calcium carbonate, not silicon dioxide.
Given Anthropic IPO was yesterday, I suspect anthropic needs to be cautious about how they present anything not entirely confirmed.
> There is an obvious cynical view, to do with the timing, but it is false, and I am pleased that I am in circumstances where it is obviously false.
FTA
Don't understand the sadness - are you interested in the discovery or credit attribution?
The process of solving hard problems like this often involves inventing / discovering new branches of mathematics in order to solve the problem. By solving the problem with AI, we haven't potentially snuffed out the potential for discovering these new branches of mathematics.
In this case it feels humans alone were given enough time.
Also, I know its beyond what a single person can affect but lots of the problems that the community faces now seem to be in essence people and incentive problems. Existence of this proof in principle doesn't stop anyone from pursuing the problem further and discovering these new branches of mathematics too.
What if it was a human proof and in its process, no new mathematics were discovered? Would we also frown that?
So it's not the discovery that's important it's how you discover it?
Here's a stupid video game example: If you fast travel everywhere you accomplish your current quest more quickly but you never find any of the side quests that only trigger when you stumble across them in the wild.
Then accomplishing the quest was not the main goal
Yes, proving GRH is not as important as how it was proven. For example, take a look at how Fermat's Last Theorem was proven by Andrew Wiles [1], and what's the impact of his efforts had toward algebraic number theory. The fact that he proved FLT to be true has little consequence compared to the method he invented to get there. Before it was proven, most mathematicians have assumed FLT to be true anyway.
[1] https://en.wikipedia.org/wiki/Wiles%27s_proof_of_Fermat%27s_...
People often prefer horses to combine harvesters.
40 percent human? In this case the human has 0 percent math input.
It might have a lot of engineering credit (setting up the swarm, ...)
You didn't read the PDF, that's not what Alpöge did.
I did read it:
> I'm just too inexpert in the relevant stuff at the moment, I hope the reader sees the breadth of this argument!
> I intend to do my best to understand and impart understanding of this argument
He was shepherding Claude.
> I guided us to the right situation > I had us on the line of combining two quite different ideas present > The parts in which my vague intuition worked out due to Claude’s technical mastery
I am not saying he could have done this without Claude. But there's a reason he did it and not anyone else. This wasn't a case of giving Claude the prompt "solve riemann hypothesis make no mistakes"
What is the reason? I'm sure he is a talented mathematician, and that is undoubtedly necessary here, but how many other mathematicians have unlimited access to internal frontier models? Clearly, the playing field is not even.
> there's a reason he did it and not anyone else.
This is the part that deniers love to ignore.
>But there's a reason he did it and not anyone else.
Well yes he works at Anthropic lol
The F-47 is pretty sweet, but what a shame that it was made without employing a single, solitary slide rule!
Get used to it. No human will prove anything significant unassisted, ever again. This is reality now.
This kind of assistance is new, but I’d say nobody has ever proven anything truly unassisted, and people have been trying to point out for a long time that crediting a single person for discoveries and claiming they were “unassisted” conceals the true story. Education, background knowledge, conversations, correspondence, and context have always been paramount. This is one reason Math has often hit points where multiple people make the same discovery at nearly the same time “independently”. It might be a good thing that we can now dispense with the myth of anything being “unassisted”.
I mean unassisted by AI.
Can you prove it?
No human has proved anything significant "unassisted" in quite a while.
Get a load of Nostradamus over here
For people unfamiliar with the Riemann Hypothesis, I like the following example. Look at these sequences:
111111
.1.1.1.1.1.
..1..1..1..1..1..1..
...1...1...1...1...1...
....1....1....1....1....1....
Pretty simple right? Just increase the period by one each row and add an extra period at the start.
Now, look straight down a column, say the 5th column. How many 1's do you see? There are two 1's in any prime number column.
Now ask the question: If you were to draw this grid with N number of rows, how many columns would have exactly two 1's?
The Riemann Hypothesis is the ridiculously complex hypothesis that relates an incredibly simple pattern (increasing periods by one) to the seemingly random frequency of prime numbers.
You can ask Claude to start with those sequences and derive the Riemann Zeta function. It's only about 3 paragraphs of logic, but incorporates advanced ideas like wave-like periodicity and infinite sums.
> There are two 1's in any prime number column.
Well, not for 2 and 3. Are they considered exceptions or kind of starting points/conditions?
Apologies, it was a formatting error. Each row was supposed to be indented by one character. I’ve instead added an extra leading period on each row.
I also forgot the very first row is supposed to be all 1’s. Should be correct now.
I want to congratulate Anthropic.
Secondly, I want to quote one article, how easy it is for humans to misunderstand difficult things AI does, as something supposedly easy:
> They mostly came away from Claude Code with the sense that this is a genuinely useful tool that they didn’t previously know, but the gut-level reaction here is: “I never knew that building software was so easy” rather than “It’s very impressive AI can do this.” [1]
It should become very clear for everyone how to adjust their views. However I doubt many will, even after this, they need something physical, like humanoid robot to be impressed.
[1]: https://www.lesswrong.com/posts/ySXuvJcqRindQwAk7/how-my-stu...
Why not congratulate the mathematicians?
The problem was open for 167 years and humans weren't able to solve it before AI. Isn't crediting the AI more appropriate?
At least going by what the author wrote, it seems like he was using Claude as a peer - a pair mathematician if you will. It doesn't sound like a one-shot solution. So without the expertise of the mathematician, AI wouldn't have been able to solve it on its own.
Sounds like both deserved the credit.
yeah, that's a fairer take
What about also acknowledging the enormous number of works of mathematics the model was trained on? You know, standing on the shoulders of giants and all that.
Didn't folks use that rhetoric when AlphaGo beat Lee Sedol? But then they trained a better model from scratch without any human data.
The era of crediting individual mathematicians is very much over.
Can anyone explain why this problem is so important, other than the fact that humans have been trying to solve it for so long? What are the benefits and implications of solving it?
Apologies if the article answers this, but.. paywall.
I'm no mathematician, but my understanding is that the Riemann Hypothesis implies that the distribution of prime numbers has a nice, predictable structure. Many existing proofs assume that the Riemann Hypothesis is true to take advantage of that structure. So, now we know that all of those results are true!
One example is the Miller-Rabin test (https://en.wikipedia.org/wiki/Miller%E2%80%93Rabin_primality...), which checks whether a number is prime. Without assuming (a more general version of) the Riemann Hypothesis, this is a probabilistic algorithm: the best we could prove is that it has some small probability of reporting that a prime number is composite. But with it, we can prove that the algorithm is guaranteed to work. (Mind you, there are other deterministic primality testing algorithms that don't rely on Riemann for their correctness proofs.)
I don't know any immediate, more practical implications of solving it. I would love to know if there are any!
Imagine that we haven't mapped all of earth, and there are still many unexplored landmasses. A while ago we discovered a large continent, and after sailing around it we found it's the biggest unexplored continent we know of. But it's surrounded by cliffs on all sides, so we haven't been able to land on it yet.
Solving the Riemann hypothesis means we finally landed on it. We don't entirely know what's inside. Maybe it's a lot of mostly-empty forests with little of value, but if it's like any of the other continents we've already explored, it probably has an abundance of natural resources to take advantage of. Now we just have to go digging.
This vague analogy makes me feel even more confused after reading it.
We just unlocked a large new area. We don't already know of specific useful things in that area, but based on past experience with unlocking other areas, we expect that in the long term we'll find a lot of useful things.
The result by itself doesn't really unlock anything, since it was presumed to hold. TBD whether the proof contains novel techniques, those would be the key to a new "continent" of math.
How much of math is based on unproven hypotheses?
it has the potential to cascade things its a very general problem that allows tons of other theories to be simplifed or have greater guarantees, but its a really wide, so nothing economically relevant in the immediate sense, however hopeful if you want for progress, freightening if you think about knowledge work.
From the article:
> Scads of other mathematical proofs and hypotheses have been proposed, which work only if one assumes that the Riemann hypothesis is true.
That only kicks the can further down the road, how are these other works helpful?
It is helpful in an abstract theoretical world.
In the real world... not so much. Lex Friedman had a bunch of great guests talking about this.
At a certain point, math disconnects from the physical world. In a ~Platonic~ one may say.
Steven Brunton (GOAT and author of of the best ML books I've seen) gave the example of the Navier Stokes result. Could this new fluid dynamics result be used to make cool new engines? No. Unfortunately the scale where the proof resides, viscosity and real-world turbulence take over.
https://open.spotify.com/episode/0ndqMvnUJkNmQrhK98R09Y?si=6...
>?si=64e32d788be94e32 tracking parameter
There are proofs based on the proofs that are based on this really important problem
(We merged this subthread from https://news.ycombinator.com/item?id=50583445)
honestly it would probably be more consequential if the hypothesis was disproved; Cryptographers already assumed it was true, I imagine this is the case in many other areas.
If we could go back and ask people in 1850 if studying prime numbers would ever be useful, they would never have guessed the benefits we get from it today like internet banking, secure browsing, cryptocurrencies.. The biggest benefit of solving RH may be something we cannot fathom yet
It's philosophical. It's about the way natural numbers work, the "anatomy of the integers".
There are no practical uses.
First comes theory, afterwards sometimes practicality.
So many mathematical discoveries later became embedded in theories that allow us to understand the physical universe and in doing so engineer it.
> There are no practical uses.
There are numerous algorithms that depend on GRH being correct for them to be correct.
It's associated with the distribution of prime numbers which is very important to cryptography but also my god what is with the intellectual incuriousity here?
The exact same ridiculous argument could be applied to the basic mathematics and physics underpinning literally all modern technology.
Personally, I'm deeply discomforted by the idea that the few mathematicians who have aligned themselves with these companies are able to race ahead of the rest of the mathematical community like this. It is unrealistic to expect Anthropic to immediately share their monetary wealth with the rest of the world (though this is their stated intention in the long-run via UBI), but math is just math. Does Levent really need to take the crown jewel of open math problems for himself before the models are even publicly available? Did OpenAI really need to launch a swarm of 10,000 agents and scoop Tristan Buckmaster, a university professor who had worked on the Navier-Stokes existence problem for most of his career?
This is a time of deep uncertainty and anxiety for working and aspiring mathematicians. OpenAI and Anthropic gain little by announcing results like this ahead of model release. So if they can't restrain themselves on such economically useless things as pure mathematics, it does not bode well for the rest of us.
It's discomforting only if you attribute value to your ego and your pride. Nothing has changed for the mathematician that find intrinsic value in the field and are interested in understanding the proofs, or the nature of Maths itself.
It has only changed for the people who want to stamp their name in blood on proofs and then get social credit and clout for what they have done. In other words people who use the field to find extrinsic value.
This is really good for mathematics, and science in general.
This misses the point I was trying to make, though I guess I didn't make it very clearly. I'm not concerned about who gets credit. I'm saddened by the lost opportunities for the wider mathematical community. There could have been conferences and Zulip chats and lunch conversations among young graduate students all making progress toward Riemann together with AI tools. It could have been a great problem to rally around. Hell maybe it could have been a Polymath Project. Instead it was solved using possibly tens of millions in compute behind closed doors.
But instead it will probably discourage many more mathematicians to leave math altogether, because they feel they can't compete.
Is it actually good for mathematics and science? Do these things have some independent, objective value, independent of the people that interact with them?
If a new miracle drug is developed but only ever used by the person who developed it, would we applaud it as a scientific achievement?
You are dismissive of ego and pride, but those can also be components of what drive people to accomplish great things. If you take away the incentives to pursue subjects like this, what is left?
I’m actually not all that negative about the results, I believe these fields will adapt and we will move forward. Or at least that’s still a possible outcome. But I’m challenging your assertions here as stated.
Wouldn't you say that a miracle drugs value is in curing people rather than in who gets a Nobel Prize from developing it? I can guarantee you that patients that have pancreatic cancer don’t give a shit if a human or a server rack developed a solution.
I would hence claim what is left is the people who are actually in this for the interest and understanding (intrinsic motivation).
I’m making a subtler point about medical research. Yes, its value is in the benefit it brings to people. If someone developed a powerful new drug and didn’t let anyone use it, both the drug and its developer’s accomplishments would effectively be worthless. Similarly, if we produce proofs only machines understand, and in doing do deprive people of an opportunity to strive towards a goal, to figure something out for themselves — the value drops to near zero. Often in math, the problem itself is worth much more than a single solid.
I think even with intrinsic motivation, people would get worn down by having things they worked for, possibly for years, yanked out from under them on a regular basis. The field must change how it does work and provides incentives, or it will become no more than a hobbyist playground for LLMs, and largely stop providing benefits to humans.
Very few if any human can hold the entire arch and workings of a modern CPU in their head either, yet people seem to be using their phones just fine and the value has not dropped to 0. If LLMs designed a chip (see HN thread from yesterday), and people don't really know every single in-and-out of it, but it works fine in the phone and gives you 1 month of battery lifetime, people wont care.
And again if having a problem solved before you "yanks it out from under you" that's still just wanting to be the one who crossed the finish line first. If you have internalized the extrinsic value mindset in this way, you will of course be sad about the developments.
Finally I think the current structure is already fine. Philosophy, classic literature, music theory departments still exist, and nobody is dropping out of them just because things were already done and figured out. I think here again we get to the archetype of the extrinsically motivated mathematicial who wants to acquire, stamp and claim things for their own for social status.
I think this will lead to people just doing maths in the way they want to, if I analyze Dostojevsky because I love it and find it meaningful, no-one can "yank" anything from underneath me and I can know that everyone of my peers have the right motivations to be there.
Even if humans are never able to understand the proof, they can still go forward assuming the hypothesis is now true e.g. they can assume GRH when working on other problems.
What's left is plenty enough of smart people. The whiny math subcommunity is not representative of all mathematicians or math-interested people. They are just very vocal.
But people will always value ego and pride, even if we’ll live in post scarcity utopia there will be still status games and so on. And that’s good, it will keep things fun!
Let's all applause the advancement of human society. I'm sure all Mathematicians agree.
Hopefully it actually is advancement of human society, because thus far this and all the other breakthroughs seem to be the advancement of shareholder profits, with any consideration for the rest of humanity being secondary scraps falling off the table.
I really, really hope to be wrong with this cynical view. I don't know what we should do in the alternate case. One idea is to internationalise the computational resources for AI so it could be used for the greater good of humanity, but I guess that idea would get a bunch of resistance, so I dunno. I guess the last one shall turn off the lights.
From what I see many people, for various of reasons, are OK to ditch the mass for that advancement. So the advancement might as well happen, just not for the common people.
Many intrinsically driven ones do agree. We are getting answers and seeing discoveries at a pace that no-one could have imagined to see in their lifetime.
If you however are full of yourself, want to stand in front of YOUR peers and prove that YOU did this, YOU figured out how this is connected, you want to have YOUR name written on some gold coin or certificate hanging in YOUR office, YOU YOU YOU everywhere, then you will have a big problem with the current developments.
A lot of this is the byproduct of how society views accomplishment, at least in the western world. While there are certainly lots of self centered, egotistical people out there, there are also many who are shaped by the constraints of their work and education. There aren’t many places where people get to the next rung by pointing out all the great things their teammates did. Ever since young school ages, we are taught to accomplish things on our own. Our grades are the byproduct of our effort, etc…
This is a western capitalist/neoliberal view. In a lot of countries for example, family is more important than work. Having a child and a social circle is valued more than being a CEO.
In some middle-eastern countries, the value is tied to how good you can hold the room in a social setting.
Seems like the western world will have to adapt their values and education in the near future.
Correct, which is why I specifically pointed that out from the get go.
I don’t see why the western world “will have to adapt”. This has been the case here for hundreds of years, and I don’t see a sudden shift to valuing different things. Maybe as generations go by, sure, but the west is in love with capitalism, the cars, the money, the “prestige”, it’s all quite sad.
I totally agree with you, man, I totally agree.
That many people want to leave their mark on the history of mathematics or science in some way, small or large, is not so narcissistic a goal as you are portraying it. The desire for one's discoveries to matter to the people around oneself is intrinsic and powerful, and without it, the passion of science gives way to mere pragmatism. We can both appreciate the new pace of discovery and mourn the damage that is being done to the spirit of human camraderie on the edge of discovery.
If your goal is to impress other around you with your discoveries, it assumes that you are able to provide something of value that people around you find impressive.
This is extrinsic motivation. If this is your purpose, you can pivot and do commercial business and make a lot of money, or something else that people around you find meaningful. People might value how good communicator you are, how much empathy you have, how caring you are for your family, etc.
But then it wasn't about the math nor understanding.
Imagine reading a book for the sakes of the book. You have your imaginary world you get into, you will keep reading the book even if there are 1000% better ways to create value.
Now imagine you read the book because it gives you social importance or skills, you are able to speak about a topic to impress and acquire for example sex from others. That is a different motivation, similar to what many of these mathematicians have.
You just describe 90% of Academia.
That's an underestimate.
Actually argument and I were talking about different things. He/She was talking about getting a name on a paper. I don't give a fuck about that (and I don't work in Academia anyway). I just want to keep my job.
But yeah I should applause thunderously for the advancement of human society, to be a good boy.
Indeed! The greedy, selfish mathematicians taking all the theorems have been a problem in our society for far too long!
It's enraging to think that someone like Riemann would name a hypothesis all after himself. He was stamping his name on everything wasn't he! His zeta function, his sums!
There's only one way forward now--that's to let those benevolent CEOs at the AI companies democratize mathematics. They care about the intrinsic, and not extrinsic, value of mathematics. They care about the understanding! That's why they're rushing to publish these 250,000 page papers with their models, they're just so excited to share mathematics with the world.
That's the world I want to live in!
Do you believe that researchers at historic, well-funded R&D labs (e.g. Bell Labs) should have kept discoveries private because they enjoyed resource advantages?
I am a researcher at an R1 university, and the nature of my work has changed, too, but the anxiety of working academics isn't a compelling reason to hide major results in physics nor math nor biology.
We ultimately do research to expand what is known, not to give ourselves satisfying puzzles or demonstrate prowess relative to peers
> scoop Tristan Buckmaster
Probably important context: Tristan Buckmaster was collaborating with Levent Alpöge on Navier-Stokes before they were scooped by OpenAI. (Indeed, part of the controversy was that OpenAI wanted to include Buckmaster but exclude Alpöge from authorship).
because they are not selfish and dont see the point of gatekeeping their work or credibility like so many other artists or professionals have been doing.
you can't win against LLMs best to find a field of work where you own things or whatever robots can't do like sales.
I'm gonna assume robots are best equipped to sell stuff to robots.
Sales? To whom?
Aquaman[0], duh
https://youtu.be/0-w-pdqwiBw
I was wondering how this would affect research incentives for difficult problems. If someone is about to embark on a long PhD process, but there is a chance that an AI swarm solves their problem 1-2 years in, isn’t that a bit risky?
Yes, the whole academic system and structures and pipelines will have to be rethought from the ground up.
Was it unfair in 1960 when very few people had access to a computer and almost everyone didn’t?
The first minicomputer DEC PDP-8 shipped in 1965. Apple II followed a decade later.
From today’s POV that’s a long time to wait, but I don’t see why AI shouldn’t be on a similar trajectory.
Levent is a mathematician more than an AI scientist; probably he can not resist this, neither probably can say tsimerman at oai
How about people who "align themselves" with CERN, LHC etc? So unfair!
Terence is so far ahead of nearly all mathematicians, what big of a difference is it when an ai solves something?
This is a good question. A big one is that Terence Tao only has 16 hours a day to do math, and he spends a lot of that time teaching math, giving lectures, and mentoring students. Tens of thousands of Claude agents can spawn in seconds and race ahead of any researcher.
I don't understand your logic here. Won't these advancements in mathematical research advance the subject for humanity AND force newer mathematicians to leverage AI in their careers?
This kind of proofs don't help humanity at all. AI uses a gobbledygook language that's incomprehensible by humans and the only outcome of its proof is a 0/1 result that you either believe blindly or not, for nobody can really verify it.
This is not true. In the processes I've read about, AI produces natural language proofs first, which it then translates into Lean code which is verified by the Lean kernel.
I read the Zulip chat and it sounds like sometimes the formalization is completed before the natural language proof.
Another comment, I haven’t verified this, suggests this particular formalization is upwards of a few million lines of code. I really enjoyed this article from Wolfram where he argued that if we had a box that could prove anything it wouldn’t necessarily matter much because the important part is the application. Now AI might, will(?) provide both but when it comes to the proofs it’s hard to tell how useful they are _at this very moment_.
https://writings.stephenwolfram.com/2026/09/whats-the-future...
AI only imitates natural language. Nobody can read and understand the 2500 pages of Claudese. Some may argue that AI can explain each part of the complex proof and therefore the entire proof is human readable, but if this was true, AI could produce a compact human readable proof in the first place.
That assume a compact human readable proof exists in the first place. Can you prove that?
I find it deeply discomforting that there are people like you who think mathematics is a personal race.
I would like it not to be. My emphasis is on the mathematical community as a whole, not the individual mathematician who might have gotten there first if not for AI.
There's an alternate scenario in which the whole math community got to use AI to make progress on these problems, share their progress in Zulip chats, and learn from that process. Instead a few individuals within private companies and with vastly greater resources are doing it all themselves.
So I don't mourn who got credit and who didn't; I mourn the lost opportunities. The people who will walk away discouraged when they could have been brought along instead.
All academic mathematicians (and all(?) academics in general) are in a personal race, because their publishing record dictates their career.
Today, yes. But academia cannot go on as it did so far. If it does, it's destined to irrelevance.
Everything is a race to some extent.
anthropic gains a lot by getting such news, billions
At least, they could quietly collaborate with the community, and thereby wait until they have a more lucid, unambiguously correct proof, rather than dump a pile of slop on the community.
I can only imagine that the perceived financial pressures around the AI race/IPO caused them to announce this now.
> There is an obvious cynical view, to do with the timing, but it is false, and I am pleased that I am in circumstances where it is obviously false.
FTA
If only he could say why it's so obviously false.
Collaborate how, exactly?
It’s basically impossible to do collaborative work with AI. This is a problem in anything ai does and it’s a huge problem we’ve had to deal with at my company. The only collaboration is between one human and the ai; a different model is required to enable more than that tbh.
And it seems like the author wanted to avoid dumping a pile of slop, AND explicitly referenced the timing of the IPO in the pdf here. Did you read it?
That can be seen to be trivially not true.
The polymath projects are a pretty amazing recent example of this. Also the recent AI boom in math has spawned many Zulip chats where people discuss these kinds of results and try to build off of them. There's one dedicated to digesting the Lean proof of Apoge's result.
> Did you read it?
Hacker News guidelines specifically call our this question as a no-no, but yes I did. I'm not calling into question how the result is being released. I'm asking if it should have been achieved at all before public availability of the model.
Woah, if this proves to be correct it'd be massive, for maths, computer science...and AI too, I guess!
The worry I have right now is this though:
> The argument is incredibly long at the moment, and while its formalization is finished and it is routine to produce a huge, unreadable document containing sufficient detail, in my view of my path I have just taken a step.
this feels like we're moving away from the old world of 'science', where we could actually read an experiment, understand its steps, and, with enough time, possibly reproduce it.
And we're back in the world of 'magic' and 'belief'. Yes, there are formal ways to prove that this is correct, but if the proof is so convoluted and hard to read, is it actually something a human will ever find the time to go through and understand? We're going to start delegating (oh well, we're already doing that, I suppose) the understanding to AI, so it can just help us condense it down to smaller pieces that are digestible.
Still, very strange dizzy feeling, much more than Navier-Stokes
My perception of these maths solutions is that we're decoupling two related but distinct elements: the answer and the understanding. We have a big machine that tells us the answer is "42." That's useful. We know the question is answerable and the general shape the answer takes.
I'm just not convinced that we're swapping one world for the other. We're getting a hint that will probably help us enormously with what was always the main objective: understanding the result.
right, but if I understand right, the difference here is that the big machine that tells us "42" also gave a extremely intricate binder of the work and evidence that it used to arrive at 42?
It's more like a lot of scratch paper. When a statistician does their work, they don't publish their raw output directly – they collaborate with other investigators to write up.
They aren't distinct but rather a spectrum. Understanding is just building abstractions. You can skip that and have a 250k page proof. Or you can have a few more abstractions and create a 20k page proof.
Maybe, we're just dogs that have invented the human. The human is far more advanced and soon we won't understand what it is doing and what it is trying to tell us. But because it is our brainchild, it certainly won't harm us.
I’m not prepared for that kind of responsibility.
> We have a big machine that tells us the answer is "42." That's useful.
It seems like you're referencing Douglas Adams, which is ironic because in the story that answer is not actually useful at all.
well, "what is 6 times 7?" and "how many roads must a man walk down?" are wholly different questions ;)
Good number choice
I think you mean ”What do you get if you multiply six by nine?”
https://hitchhikers.fandom.com/wiki/Ultimate_Question
Are they, though? Homer (the philosopher not the poet) already answered the latter for us: https://www.youtube.com/watch?v=PY3fqF6-jyE
Six times nine, actually.
Yes, but there's also this part:
> “All right,” said Benjy. “What do you get if you multiply six by seven?”
> “No, no, too literal, too factual,” said Frankie, “wouldn’t sustain the punters’ interest.”
Which is at the start of chapter 32, when the interdimensional mice are discussing how to invent a fake question. Followed by:
> Again they thought. Then Frankie said, “Here’s a thought. How many roads must a man walk down?”
This is how I learned to remember six times seven. The wrong answer is fortunately much later.
Is this a cultural reference I don't get?
because 6*9 is 54, not 42.
It's a Hitchhiker's Guide to the Galaxy reference:
> The Earth was destroyed by the Vogons five minutes before its ten-million year program to find the Ultimate Question was due to be completed. One of the only two humans who had survived the destruction of Earth, namely Arthur Dent, pulls random Scrabble tiles from a bag, so as to reveal information in his subconscious (Marvin announced that Arthur had the Question imprinted in his brainwaves). When his friend, Ford Prefect, lines up the tiles, they spell, 'What do you get when you multiply six by nine?'
https://hitchhikers.fandom.com/wiki/42
Yes. There's an explanation here: https://en.wikipedia.org/wiki/Phrases_from_The_Hitchhiker%27...
Hitchiker's guide the galaxy.
I don't think we'll have any trouble converting the proof into a textbook
> this feels like we're moving away from the old world of 'science', where we could actually read an experiment, understand its steps, and, with enough time, possibly reproduce it.
Two hundred years ago the limits of the natural sciences was regularly pushed up against by part-time dilettantes and tinkerers, and the theories were formulated to explain empirical observation.
A hundred years ago, world-changing discoveries were made with equipment that would today be hopelessly insufficient for a high-school science class.
Now, the cutting edge of science requires particle accelerators, wet labs and datacenters, large scale infrastructure and precision engineering projects at the cost of millions to billions.
The low hanging fruit was already picked long ago.
yes, it's only natural that we're reaching the end of our abilities to understand
It seems to me our ability to understand is what distinguishes a "discovery" from any other pile of internally consistent Lean code though?
True, but that's how most of humanity feels about any discovery. You can't come close to understanding it yourself, you are just trusting there's a process that produces someone who says he understands it, and there's hundreds of years of institutional harness for us to believe that he has actually understood it.
Now that someone is an AI.
>back in the world of 'magic' and 'belief'.
Biologists tend to try and hide this detail, but this is essentially what biological research is. We will never really understand how most organisms work in a way that would satisfy a mathematician or even a physical chemist. We just tinker with what we can, poking the system and seeing how it reacts. Biologists reckon with incomprehensible systems everyday, so maybe this is why it's easier to stomach AI.
> this feels like we're moving away from the old world of 'science', where we could actually read an experiment, understand its steps, and, with enough time, possibly reproduce it.
> And we're back in the world of 'magic' and 'belief'. [...]
I don't know, to me it feels more like this might be the job of future mathematicians?
AI proofs noone understands are good, but maybe don't lead to that much. So maybe in the future, mathematicians will be more like "data miners" or more like experimental scientists, where they take a statement already known to be true (in the case of experimental scientists that "statement" you know to be true would just be nature itself) and try to find ways to make it accessible to human knowledge.
> I don't know, to me it feels more like this might be the job of future mathematicians?
Nobody would pay us mathematicians enough to prove it in the first place; why would they pay us enough to explicate it? I doubt there’s much value (in the sense of capitalism) in an explanation. Society by and large does not value pedagogy, or education, or even writing.
The social structure that trained a small number of research-level mathematicians every year is evaporating, and I doubt anyone can stop that now.
Who paid Wiles or Alessandro Maria Masullo to work on their problems?
I'm not convinced this is true.
IF AI really goes in this direction where it intellectually leaves us in the dust, then there is clearly lots to gain for society by having people in all kinds of fields that at least try to keep humanity up to date that is also "monetizable" in the same way our current academic system is monetizable (meaning, on an individual level probably not directly but via network effects etc.).
Can you give a concrete example? I’m not seeing the value in putting a human in the role you’ve described.
> Nobody would pay us mathematicians enough to prove it in the first place;
weird thing to say about a problem which literally had a $1 million dollar prize attached to it but fine
Exception that proves the rule. Try making a living working off these big prizes. It's the same with infosec and bug bounties on a different scale. You have to already have a secure income, as the prize is a bonus at the end. In what other industry would you work for years on something that inherently can never return a profit, other than the chance of a prize at the end for the one person who got their first.
Society doesn't pay for math labor, it pays to celebrate the top 1% for labor subsidized by other institutions.
I'd wager that Anthropic has spent more than $1 million on solving this.
You missed the word ‘enough’. One million was obviously “not enough” (nor was that specific one million ever paid to anyone — it represents an investment of approximately zero in that sense).
It took on the order of ten billion dollars to develop Claude, depending on how one does the accounting.
It's all tech priests all the way down then. Either that or the Bicamerals from Peter Watts' Firefall are about to rise. Maybe both.
I don't know about "massive". This is certainly massive for certain branches of theoretical mathematics, but I don't think it has any implications that you could not have arrived at before we had a proof by saying "assume the RH is true" and there are in fact innumerable results of that form that predate the proof.
It always reminds me of Ted Chiangs short story "The Evolution of Human Science " I guess we're not far away.
> And we're back in the world of 'magic' and 'belief'. Yes, there are formal ways to prove that this is correct, but if the proof is so convoluted and hard to read, is it actually something a human will ever find the time to go through and understand? We're going to start delegating (oh well, we're already doing that, I suppose) the understanding to AI, so it can just help us condense it down to smaller pieces that are digestible.
If a tree apparently falls in a forest and no human goes to verify it, did it really fall?
No matter the verification results, as with most theories, for practical uses it was already assumed the hypothesis held. For mathematicians, they'll just say "Oh, it is true, but what's this proof? It's not understandable, so there's still a miniscule chance it's wrong, and we can't learn anything from it, so what's the point?"
What if this is metaphorically like a P = NP scenario? Working out the solution to these problems might be the "NP" here, but it remains sufficiently tractable to understand the solution in "P." So human understanding could come along for the ride, at least for a while longer.
Can't we untangle these proofs? How analogous is it to 'refactoring' the proof to reduce the amount of duplicated code etc?
Alpoge thinks that people will be able to simplify the proof substantially.
Apparently people have reduced OpenAI's proof of quasi-Riemann to a 12-page outline.
Very analogous to refactoring code. Proof is code anyways.
You could, just like you could slowly disassemble your house back into individual pieces of lumber and nails and drywall, but it would be laborious and at the end of it you'd just have a pile of semi-broken construction materials.
I believe that it could be more akin to decompilation? When done correctly, you turn something a bit crude into potentially beautiful.
The Lean is already full-fledged code with structure and names. What are you decompiling into? How do you raise it further?
Lean is the Assembly of mathematics. Human proofs are natural language specifications of a functionally equivalent program.
Well, you raise it into more lean. Its just higher level.
You could try but that's kind of improbable. "Why don't they rewrite <program X> in the same language, with every feature intact, but much smaller and simpler" seems optimistic in Lean, C++, Rust or any other language.
It’s like refactoring.
If you spot the same idea used multiple places in the proof, you can pull it out as a tactic. Suddenly the proof is 30% shorter _and_ a core idea is clearer.
So yes it will be simplified. A lot. And it will be easier to understand.
Are AI proofs bloated because they just contain refactorable concepts over and over again? That's an interesting theory and it does align with the type of slop we see in vibe-coded apps, but I'm still not confident that sort of approach will radically shrink a 200,000 function proof. Imagine it gets down to 120,000 functions--that's a massive improvement but it's still enormous by any standard. Which gets to my original point--I'm sure you can shrink it, but even 2x smaller or 3x smaller gives you an incredibly large proof.
This just feels like seeing a complex empirical result and then having to define your view that confirms it. Except in this case the empirical result is a passing proof.
Back from chemistry to alchemy. Trying this and that, not understanding what's going on, believing and hoping for gold.
> this feels like we're moving away from the old world of 'science', where we could actually read an experiment, understand its steps, and, with enough time, possibly reproduce it.
It’s a genius first step that we will iterate on and make better.
Or, just, you know that feeling you get when you create something that surpasses the pinnacle of human intellect.
or there's a bug in the proof checker
which is more likely: that the Riemann Hypothesis is solved, or that a possible solution exhibits a bug in the theorem checker?
OpenAI dumped hundreds and hundreds of solutions with Lean proofs that hadn't been seriously checked by any human and AFAIK, none of them have turned out to have exploited bugs in the prover. Some of them didn't prove what their paper proof claimed they did, but the base rate of LLMs exploiting such bugs (which are definitely available to exploit -- even for a human it historically hasn't been SUPER hard to find them, and the Lean kernel's metatheory is suspect in a number of ways) is empirically very low, even when the incentives for a significant result are very high. The more likely pathologies are, e.g., misstating the problem (the Riemann hypothesis is using the standard Mathlib definition), adding extra unjustified assumptions (unlikely in this case -- I haven't inspected the Lean proof yet but others have and reported that there aren't any), or not matching the paper proof and intuition (I'd strongly suspect the latter is true in this case).
You could use a soundness bug in lean to prove anything. So if there is, it’d be strange that that isn’t happening
It has been run through more than one; the same Sullivan who witnessed to the successful compilation earlier has checked it with `nanoda` an external typechecker written in rust.
why is it huge?
> Woah, if this proves to be correct it'd be massive, for maths, computer science...and AI too, I guess!
The Riemann theorem (as we apparently now have to call it) is a statement of pure mathematics and has no practical interest or application.
It isn't a theorem yet. "No practical application" might have some truth currently but "no interest" is false and laughable..
No practical interest.
AI is inherently about getting cooked/digested answers. It’s already finished. Nothing to build on.
But I still wouldn’t call the end result Magic as long as you have something to hold on to. I can see that this building in front of me (waves hands) is upright. Okay, now I remove this beam. It collapses. I don’t care about the whole structure. Just that it stays upright.
Giving up understanding is not the same as blind trust. Or maybe I’m too philosophically naive in this area...
I'm of the belief that "Claudese" with its load-bearing seams is just the models evolving our language beyond our capacity to use or understand it. Read up on your Continental philosophy to understand why and how: there's no such thing as a fixed "meaning"; meaning is something constructed for the immediate need and then discarded or modified as circumstances change. The models are proving the French and German philosophers correct as surely as any mathematical proof. And they're beginning to transcend human comprehensibility. Super intelligence is just around the corner, if not here already.
This is Levent Alpöge (https://en.wikipedia.org/wiki/Levent_Alp%C3%B6ge), a supremely qualified mathematician and one of the best in the world at guiding AI towards new discoveries in mathematics.
If he says he has Riemann, well then, he has Riemann.
Levent was one of the two mathematicians working on Navier–Stokes when OpenAI published their work on that problem - which was produced in a rush when they heard a rumor that someone from Anthropic had promising results on a millenium prize problem.
I think he worked together with this person? https://www.youtube.com/watch?v=Mzbtj5nkMXI
Pretty crazy interview, he didn't want throw the Anthropic employee under the bus when OpenAI offered co-authorship on the Navier Stokes paper. He suspects OpenAI somehow used his codex usage btw. Almost unprovable of course.
Choice quotes from my notes on the affair:
"I said that if OpenAI released its result in the way proposed I would go public with what happened. The reply was, 'Why would you ruin your career?' I replied that I am an academic, and asked why he thought going public would ruin my career. The reply was, 'If you don't want me to be nice, then I don't have to be nice.'" (Sébastien Bubeck later apologized on X for the remark.) "Sebastien twice asserted that he wanted Levent removed from authorship..." "The second was that, after posting Euler, I alone write a paper presenting the Navier-Stokes result, acknowledging that an internal OpenAI model had resolved it.""I asked whether the model had been trained on, or had access to, our sessions in Codex, into which we had been putting all our drafts for the whole of this project. I was told the model did not look up user data. I asked again, about training, and I did not get an answer." (OpenAI subsequently stated publicly that while unlikely, it could not rule out that de-identified data from their product usage improved the model.)
"While unlikely ... it could not be ruled out ..."
So basically he hit the nail on the head and they want to lie about it without lying. Either that, or they are supremely incompetent, not knowing even what they ingest into their models. Both are very much not inspiring any faith in their trustworthiness.
In broader context the “being nice” part seems like both of them talking past each other.
What happened was the two researchers, Buckmaster and Alpoge (from TFA) solved a significant subset of Navier Stokes (referenced as “Euler”) which set off rumors.
Through whatever game of Chinese Whispers ensued, OpenAI thought they had solved the full Navier Stokes problem, which they hadn’t. So OpenAI pointed their latest unreleased model at it, and it actually solved the full problem!!
Pure George Danzig vibes on that one.
So when this meeting happened, the OpenAI folks were surprised that these researchers had NOT solved the full problem, only the subset. Yet from their perspective, they were offering them credit for the full problem, and so “we’re being nice.” The Anthropic employee (Alpoge) was clearly an optics issue and so they felt they didn’t have to be that nice.
But the other researcher, Buckmaster, was suspicious that their work had been stolen, and so was understandably in a combative mood. You can see that from his perspective they were not being “nice”, they were trying to scoop them, possibly on stolen work, and then bribe them with shared credit.
But they had not solved the full problem, and academic plagiarism is a very serious allegation that can ruin careers of both the accused and the accuser. So from OpenAI’s perspective it was like, why would you accuse us of plagiarism when you yourself have not solved the full problem and when we’re willing to share undue credit for it??
This was made worse by the fact that OpenAI would not categorically deny if the others’ work had been used for training until much much later. This was terrible optics, but anyone who’s dealt with large scale data pipelines compounded by privacy issues knows that it gets very difficult to trace these things, so it takes time to make a definitive statement.
Alpoge somehow seemed less troubled by this whole episode, but I guess it’s because not only was he gonna be a multi-millionaire soon, he had Riemann stewing on his backburner.
I feel like when they finally make a movie out of all this, this scene would end up being a bit of comic relief.
That version of events skips some central facts from Buckmaster’s statement.
"they came prepared and they wanted to show us the prompt... they showed me on the screen and they just literally copy pasted the Millennium Prize problem, so they said that they took the prize and they pressed enter and they got a solution, and at that point I just had to laugh because it's just so ridiculous. It came out within a few minutes that this was a lie"
"Levent had been told by Sebastien 'very little human input' had been used. This turned out not to be true. Over the course of the call, as members of their team sent Sebastien corrections and details over their internal chat, it emerged that an entire team had been working on the problem..."
"I was offered the opportunity to write as sole author of their paper and throw my collaborator under the bus... they were asking me to throw him under the bus, someone who I've been working with for a year, to throw him under the bus just because he's an Anthropic employee"
"the right paper to compare is the unforced Euler equation. Now the unforced Euler equation, all the mechanisms in this unforced Euler equation are identical to a different version of the Euler paper that we haven't released yet... basically all the main architectural ingredients in that unforced Euler were in our work for which they would have had on their servers, and it hadn't been published" "they didn't use 10,000 agents to get to the Euler problem, they used 100 agents. So they worked from absolutely nothing with only 100 agents... and then they ended up with an identical architecture to what we had"
Not to put anyone on an idol but the only one who comes out of this not acting like a greedy, anti-human is Buckmaster and his colleague (I think). Straight up lying by OpenAI, very predatory "meritocratic" behavior. What gets me the most is "they didn't use 10,000 agents to get to the Euler problem, they used 100 agents. So they worked from absolutely nothing with only 100 agents... and then they ended up with an identical architecture to what we had". What the hell! So they lie about their technical claims too on a factor of 100x?? And companies are taking their words seriously, no wonder there is a house of cards being built in the industry.
I don’t see how those details change the overall outline of what I presented above through? Maybe you could nitpick that it was a team with a model using 100 agents, and not a solo model with 1000 agents. But the fact remains OpenAI ended up solving the bigger problem themselves just because they thought the researchers had solved it already when they hadn’t. That is low-key hilarious.
The Euler plagiarism insinuations are much more serious but given all the 100s of other results we’ve seen come out, highly unlikely.
I think it's clear enough that something broadly like this is right, even if people don't want to admit it because the vibes of extending any charity at all to frontier labs are bad.
Terrible optic is such a bs executive wording when they lied, coerced, bullied, and threatened him…
I like the summary but knowing what irthomasthomas posted that’s missing from it, and having a better picture of what happened; terrible optic doesn’t work.
>In broader context the “being nice” part seems like both of them talking past each other.
Really?
Yes, because when looked at from the outside, like I tried to do with my recap, it's clear that:
1) OpenAI thought they were "being nice" by still offering (partial) first authorship despite having independently solved a bigger problem than the researchers.
2) But Buckmaster thought OpenAI were "being nice" as a means of CYA / bribing / coercion for scooping and plagiarizing them.
"Being nice" has very different connotations in the two contexts.
Thinking you're nice doesn't mean you're actually nice. OpenAI isn't talking past anyone. It's talking out of it's ass.
Buckmaster didn't think they were being nice. He thought they were being assholes to him and trying to bribe him.
Putting the phrase "in quotes" to use it completely out of context, only to make the point that surely no one understood each other... that's some impressive work. If you have to work this hard to make sense out of a conversation, it's because you're not.
Offering first authorship on a significant paper to a researcher who had not even solved the whole problem seems a lot like being nice (not "in quotes" ;-)) to me too, but that's just me.
You are again ignoring the context that is offered as part of a bribe. I don't understand why you keep ignoring that context.
Yes, it was a joint effort with Tristan. Part of the scandal around it came when OpenAI offered to let Tristan help write up their result but wouldn't allow Levent to be involved because he worked for their competitor.
>> Almost unprovable of course.
Very much provable, if he would sue, and a lawyer would ask for discovery.
He was also the one who dropped the Jacobian conjecture counterexample on twitter with the famous
> hello there the jacobian conjecture is false thanx to my close friend akhil for asking about it and my other close friend fable for working during the world cup final
and prompted Fable with
> You should do a breakthrough
It's a sign of the times that when someone with an AI beats you to one Millenium Problem, you can just turn the attention of your AI to another Millenium Problem.
It's almost certain that they have had people working on Riemann since way before the Navier-Stokes theorem proof was nearing completion. It's a long shot, but the value of a proof is next-level compared to NS
> It's a long shot, but the value of a proof is next-level compared to NS
The value of a proof? People have assumed it's true for so long that there are any number of results already proved "if the Riemann hypothesis is true".
What value does the proof add?
There are also lots of results proven "if the exponential time hypothesis is true," and now it looks like an AI has proven it false.
"... but what I saw during the course of progress so amazed me that I believe it is one dawn across science of something far greater."
Well, for one all those other results that assumed Riemann was true are now also known to be true. That seems valuable
Doesn't that increase the value? We now know all of those results are true, where before we didn't.
If value here is a measure of utility, its not super useful on its own that something that was almost universally assumed to be true is now known to be true.
But with a problem as difficult and fundamental as the Riemann Hypothesis the thought has always been that in the process of solving it we'd be likely to develop or discover something that did have real world utility. At least this is often what happened in the past, the obvious example here being Newton developing calculus, a field of mathematics which can be applied to other problems beyond the narrow set of problems he was trying to solve.
What I find interesting about this, if the claim holds up, is whether this will assumption will remain true, or if the value of these AI proofs is going to significantly lower than those of human discoveries in the past.
Maybe a good analogy here is to think about how all of tools we humans have created has been ultimately in the pursuit of solving some problem. For the developers here, consider this – if ChatGPT had existed before React, would React ever have been developed? I mean, if an LLM can brute force the same result with just vanilla JS the need to develop frameworks and libraries to solve what was once difficult for a human to solve drops.
I'll note that I'm playing devils advocate here as I think AIs would probably build something better than React had it not existed, but it's something to think about at least. I suspect there's a higher probability than some here are assigning that this proof ultimately ends up have relatively little utility.
It wasn't assumed it was true. People thought a world in which it was true was more beautiful than one where it wasn't, but proving it true wasn't a pesky little detail.
In fact, there was a whole literature of papers proving that if the GRH was false in one particular way (the existence of Landau-Siegel zeroes) many theorems followed from it. For example, if there's a Landau-Siegal zero, the twin primes conjecture is true.
> It wasn't assumed it was true
Are you claiming that there wasn't a significant majority of mathematicians who believed it true? That's all I'm saying. I can only assume we agree.
To expand on what I said though, if it is proven that P != NP then nothing really changes as that assumption is in line with everything we understand about computational complexity theory. The proof would really just suggest there's probably nothing that surprising to be found that we don't know.
If P = NP however that is far more likely to suggest we're missing something fundamental which could radical consequences for more practical fields like comp-sci. Similar can be said of GRH. Everything we know points in the direction of GRH being true, and while it may be interesting to prove these things, its much less likely to lead to some surprising and impactful downstream discoveries.
I doubt a majority of all mathematicians ever believed it to be true. (It's hard to have an opinion outside of your area of expertise, other than "I guess this expert sounds credible.") Probably the majority of all number theorists thought it was true. But this is mathematics, not democracy.
But that doesn't change the value of proving it. Most graph theorists probably thought the 4-color theorem was true, but it was still a big deal when a computer proof showed it. Most people thought the Kepler conjecture was true, but it was still a big deal when Hales reduced it to a big minimization problem that he solved.
On the flip side, everybody thought the Hauptvermutung was true, and probably subconsciously used it. But it's not.
The post mentions some generally useful machinery derived on the way to the proof.
The value it adds is that those theorems are now actually true, which means all those derived results are "real" mathematics... as every mathematician knows, it is quite easy to prove very surprising theorems when False is in your context.
But that's 100% wrong. If it was easy to prove surprising theorems from an assumption of the Riemann hypothesis, it would also be easy to disprove the Riemann hypothesis.
To borrow the usual quote about ingenious fools, the trouble with "easily" disproving that False is in your context is that there are such ingenious proofs of False. One can appear to be doing very substantial and sophisticated mathematics, encountering many obstacles along the way that look identical to what you'd face if you were relying on something true, only to realize much later that one of your hypotheses is uninhabited. The concern was that were the Riemann Hypothesis false, a very, very large subset of modern mathematics might fall into this category.
not at all. we can prove a lot of things using P != NP, but that doesn’t make a proof or disproof of that statement easy
150GB of lean must be worth something
That there are so many results beginning 'if RH' has exactly the opposite significance you are trying to suggest.
The value is proving AI can be useful after all (and we're fucked).
Remember, not too long ago we had hordes of folks here proclaiming machines can't think, reason or be creative in any productive sense. Many still do, but it's starting to look bleaker by the day.
This is a competent person using a tool in their work.
The tool isn’t doing the thinking or the proof on its own.
It also hasn’t been vetted/verified yet. There’s still work to do.
What were the prompts?
For many of the proofs coming out of the AI companies where the prompts were posted, they ended up being variants of "solve the problem", "keep going", "I believe in you".
Many of the results could have come from a 15 year old that heard of the problem in passing.
The result here is a quarter of a million pages. Human involvement is going to be minimal.
If you believe these were the prompts, I’ve got a bridge to sell you.
These AI bros don't get it. These LLM systems solving math problems have the benefit of being constrained by formal rules of math providing constant feedback when they break the rules that prevents them from wandering into la-la land. As Dijkstra said in his prescient "On the Foolishness of 'Natural Language Programming'":
> "The virtue of formal texts is that their manipulations, in order to be legitimate, need to satisfy only a few simple rules; they are, when you come to think of it, an amazingly effective tool for ruling out all sorts of nonsense that, when we use our native tongues, are almost impossible to avoid."
To whatever extent that LLMs can "think, reason, or be creative" in generating these proofs (or writing code), they're aided by very high guardrails so it doesn't say very much about what they can do in broader, fuzzier domains.
Goedel proofed that you can formulate nonsensical statements in math (and any formal system that is complex enough to be useful)
Gödel proved that statements can be created that are definitely true or false but for which no proof can be constructed. It is not relevant here since proofs are successfully being constructed, according to checking using Lean, by an LLM.
Would you say that the early results of AI beating humans at backgammon and checkers had any predictive power towards chess and go?
The kind of AI used for backgammon/checkers and the kind of AI used for chess/go, and LLMs differ greatly so I'm not sure what the point of your question even is. A casual check says that the state of the art LLMs have much lower [EDIT]Elo ratings for checkers than the purpose built "AI" that beat humans.
Elo
Thanks, I fixed it.
> Remember, not too long ago we had hordes of folks here proclaiming machines can't think
Machines can absolutely think. It's really all they do.
> reason
Murky, but LLMs prove with an avalanche of statistical modeling, they can do something kind of analogous to reasoning, with the caveat that whatever they output will necessarily have to exist in some fashion in their training data.
> or be creative
Absolutely not. By definition an LLM cannot create original work, it can only remix and recombine what it already knows. That can be interesting, and bloody useful especially for code tasks, but it is not making anything original.
> Absolutely not. By definition an LLM cannot create original work
What (if anything) unambiguously counts as original work, according to this definition?
Arguing the definition of original is pointless for the exact reason you point out: that it's ambiguous. Due to the nature of how an LLM works, it cannot make anything truly new, it, again, can only recombine and remix things from it's training data. This is why AI images/writing are so easily spotted, and even AI code to a degree: they have a distinct "style" as a consequence of how they work.
I leave what is truly original to the philosophers. What I can say for certain, is that anything you get out of any of these models is a remix of what is present in it's training data, because gravity exists and energy is conserved. They are not intelligent, they are not conscious, and they cannot make New things.
> What I can say for certain, is that anything you get out of any of these models is a remix of what is present in it's training data, because gravity exists and energy is conserved.
But doesn't that same argument apply to humans, where 'training data' = all our sensory inputs (plus what's encoded in our DNA, depending on whether we want to analogise that to the training data or the underlying architecture)?
> Arguing the definition of original is pointless for the exact reason you point out: that it's ambiguous.
I'm not trying to be a smartarse here but you are the one who is claiming absolute certainty that something is false 'by definition', so I think the onus really is on you to have a precise definition in mind.
> They are not intelligent, they are not conscious, and they cannot make New things.
I'm not a consciousness denier or anything, in fact I think it is the most important thing in the world, but I don't understand the connections you're making here. 'Intelligent', like 'original', can be defined and needs to be defined for this statement to be meaningful. 'Conscious' is probably inherently slippery, but if we mean the same thing by it (having subjective experience/qualia) then I have no idea how you could be certain that any other entity lacks it. And it seems like you may be implying that consciousness is a necessary condition for originality, but I don't think you've explained how.
> But doesn't that same argument apply to humans, where 'training data' = all our sensory inputs (plus what's encoded in our DNA, depending on whether we want to analogise that to the training data or the underlying architecture)?
That is true if your model of the human mind is akin to a machine model you also use to describe today's computers. But that is a huge assumption, making the argument somewhat circular. We don't really understand the human mind and therefore don't know if there could be other models for it which are not equivalent.
If you think the human body is subject to special laws no other object in the universe is, feel free to present your evidence. No model of the human brain can exist that violates the conservation of energy and thermodynamic laws regarding entropy and information. If you have one, feel free to present your groundbreaking results.
If there has to be something breaking the current known physical laws, the best chance to find it would be in extreme conditions like black holes or stellar cores not in close to STP conditions like the human body.
> Due to the nature of how an LLM works, it cannot make anything truly new, it, again, can only recombine and remix things from it's training data
Fwiw, David Hume made this exact argument about human imagination:
> But though our thought seems to possess this unbounded liberty, we shall find, upon a nearer examination, that it is really confined within very narrow limits, and that all this creative power of the mind amounts to no more than the faculty of compounding, transposing, augmenting, or diminishing the materials afforded us by the senses and experience. [David Hume — An Enquiry Concerning Human Understanding (1748), Section II, Paragraph 5]
Everything is either deterministic, random, or a mix of that.
New is just a mix of deterministic and random. LLMs are not deterministic.
You don't even need an LLM to create something new and improved. Evolutionary algorithms have been doing it since the 80's.
> Everything is either deterministic, random, or a mix of that.
What makes you believe that?
> New is just a mix of deterministic and random.
... or that?
Because there is nothing else.
Every action ever in universe is the effect of either something deterministic or random. There are no causes or effects that can't be put into one of those bins.
The only "out" here would be deterministic actions that originate outside the universe, but we have no evidence of that. It also doesn't escape the deterministic/random duality.
"There is nothing else" is a very common belief to be upended when our understanding of the world improves.
I'm not saying you are wrong but I'm trying to be careful with such absolute statements.
Nobody is claiming new forms of energy or phenomenon can't be found. As of yet though most such were found in more and more extreme conditions, starting from radioactive substances to particle colliders and astronomical phenomenon. There is a much lower chance to believe something fundamentally or radically new phenomenon in physics to emerge from human bodies.
Non answers are best ignored. You have failed to define what creativity is and therefore failed to prove whether llms are "creative" or not.
What exactly are you trying to say about gravity and thermodynamics? You may be confused but these are universal physical laws that apply to everything, humans, computers, anything else.
At its most basic level if I had to start with defining creativity, it should be some kind of intelligent filtering process. The brain can be in a huge number of state configurations, the interesting and novel among them are "creative" ideas. We can build upon it further. I am happy to say I am no philosopher and I will defer to those with expertise on it. But if you don't have any definition for it, its quite strange to make up some beliefs about it without proof.
You’re the one who said “ By definition an LLM cannot create original work”, so I think responding asking for the definitions you are talking about is quite justified?
Often when someone says that something is true by definition, their claim is incorrect.
Does your same argument hold for humans’ inability to do more than remix? Gravity exists and energy is conserved for us, too.
Arguing the definition of original is pointless for the exact reason you point out: that it's ambiguous.
Well, it certainly wasn't "ambiguous" two posts ago when you said "Absolutely not. By definition an LLM cannot create original work." Make up your mind.
You've been beating this drum since roughly GPT 3.5, as I recall. Your goalposts aren't even on the field anymore, they're gathering dust in a far corner of the parking garage down the street from the stadium. Time to update your priors.
If you can be persuaded, looking into how modern LLMs are trained with reinforcement learning may change your views.
https://arxiv.org/abs/2510.07364 event rl isn't necessary for LLMs to develop reasoning capabilities in some cases
Even after all the MPs are beat, there's still work to do to actually understand them, communicate them, and transform them in a way that can be leveraged. I predict that's going to be much harder for AI to solve.
Is it really wise to only rely on credentiallism to verify things? Plenty of mathmaticians top of their fields believed they have proven something when the proof contained some flaw. Just look at Mochizuki
Hacker News is a water cooler conversation, not a peer review system. It is in fact quite useful in this context to short circuit a whole raft of standard dismissals!
We went through this with Alpöge over the Jacobian Conjecture and the resulting thread was pretty embarrassing.
On this water cooler conversation, I sure appreciate the comments of my fellow @artninja1988.
If anything goes, anything goes.
> If anything goes, anything goes.
This sort of black and white thinking isn't helpful. Their point in giving context about the author is to give you more data for your internal heuristic to figure out whether this result is plausible and worth discussing.
As opposed to, for example, someone with AI psychosis on Twitter posting that they've solved it.
Even if Terence Tao came down from the clouds with a beautiful human-written proof, no one would be certain that it is solved--people make mistakes. But again, it gives confidence and something fun to discuss until the math community digests it fully and comes to a consensus.
Anyway, TL;DR is that no one is telling you to go put your life savings on a prediction market bet that it's solved. Relax a little.
Can't even get complementary cold water around here these days smh.
Math is hard; even to understand the questions and what the entities are which the questions address takes a lot of deep study, even for smart people. So credentials are not sufficient but they are for sure necessary. I have tried my whole life as a software engineer to study math as an amateur and then started a masters program and learned more details thoroughly in three months than in decades of self teaching.
The study is what's necessary. The guidance of a program is a helpful accelerant. A credential is merely suggestive.
Not in a Bayesian sense, not "merely" suggestive, but quite suggestive. Not sure how many full professors you have met, but they aren't like that under-performing coworker that people have to cover for.
More like IIT grads in being, Bayesianly confident, quite exceptional.
Why is it necessary--just check the proof
> I have tried my whole life as a software engineer to study math as an amateur and then started a masters program and learned more details thoroughly in three months than in decades of self teaching
Oh man, you can't imagine how much I'm cheering you on and rooting for you. I wished more peope could embrace that insight/approach.
Nobody claims it's verified. It just seems very plausible.
It really irks me that people think beliefs have to be binary rather than come in different degrees of certainty.
This just depends on intelligence.
Not at all. A person doesn't need high intelligence to understand what "maybe" / "it might be true" means.
It's just pro-science try-hards who believe that things can be verified by science to become facts, absolute truths, etc
I would honestly like to know what you think is the difference between those groups of people. I thought about that a lot and couldn't find a better explanation.
Why do some people understand what "maybe" / "it might be true" means, while others believe in absolute truths?
Given their typical sky-high confidence, does that prove that AI agents are not actually intelligent?
scnr
No. The mathematicians will verify this, and come to a conclusion.
What the credentialism does is establish that this should be taken seriously and has a real possibility of being right. The Lean also helps (no Lean for Mochizuki’s work).
he's just pointing out that this isn't some rando claiming to have video of aliens.
credentialism, and a lean codebase that compiles and certifies Riemann is correct.
No. A lean codebase that certifies that whatever definition of the Riemann hypothesis their LLM wrote is correct.
I can write Lean that certifies that 2 + 2 = 5 is correct, by redefining every axiom about natural numbers on the way. Or abusing bugs in Lean. Stop believing Lean is the response to everything.
Just like Navier stokes etc they use independently written Lean statements written by humans; formal conjectures [1]. Indeed, they have one for the Generalized Riemann Hypothesis too [2].
> I can write Lean that certifies that 2 + 2 = 5 is correct, by redefining every axiom about natural numbers on the way.
Which you would catch using comparator?
> Or abusing bugs in Lean.
And just like all the other ones, people check with multiple independent kernels, so you need to abuse bug(s) in all of them.
> Stop believing Lean is the response to everything.
Stop believing Lean is so flawed.
[1]: https://google-deepmind.github.io/formal-conjectures/
[2]: https://google-deepmind.github.io/formal-conjectures/theorem...
The statement of the Riemann hypothesis is included in the Mathlib library which is a human written and highly regarded foundation for formalization. See https://github.com/leanprover-community/mathlib4/blob/master... for the two-line statement.
For a discussion of how to check that the untrusted proof isn't playing fast-and-loose with the axioms of mathematics or exploiting kernel bugs, see Kevin Buzzard's recent post "...and the proof is checked in Lean" at https://xenaproject.wordpress.com/2026/11/12/and-the-proof-i....
It imports a standard definition of RiemannHypothesis. It could be wrong in other ways, but not that one.
The axioms would be a given to the theorem prover of course, as is the norm for theorem proving. This then reduces a complex proof to checking whether the theorem is stated correctly and the proof kernel is correct. This may not be (provably) easier, but it is so far removed from yet equivalent to the original problem that you get a machine-verifiable independent audit as a result.
And don't reduce "is not the answer to everything" to "is the answer to nothig", there is tremendous value in automated verification, especially in the AI age where the moat to it going mainstream is shrinking by the day.
> whatever definition of the Riemann hypothesis
Do you... think that there might be existing definitions for one of the most famous math problems of all time and that maybe they used that?
Or do you think that they just vibe coded up a different definition, for kicks?
This kinda feels.... easily provable either way.
>whatever definition of the Riemann hypothesis their LLM wrote
Given that the proof is provided by Levent Alpöge, who is known to be more than familiar with Lean, is it not unlikely he made up his own definition of the problem?
https://leanprover-community.github.io/mathlib4_docs/Mathlib...
I understand the proofs generated by the slop engines are beyond mortal kenning, such that they must be backstopped by appeals to authority.
I heard Fermat himself had a marvellous proof, but it was too small to fit into the margins. A pity!
I have a marvelous lean proof but my hard drive is too small to contain it.
No it’s not. HN main approach to analyzing anything is to rely on chaining heuristics. Which is generally sort of fine for software, but doesn’t make sense in the real world. It’s at best one signal to weight in your analysis, nothing else
It's the only realistic way to operate for nearly everything. Try determining whether a sandwich is safe to eat without applying a shitload of heuristics.
The difference is that you generally need to eat something and so drawing some conclusion about th sandwich is essentially forced upon you.
Getting buzzed over headlines about math or medicine or whatever, on the other hand, is just an irrelevant hobby of bored gentry and the opportinity to just not draw a conclusion at all remains open.
Tell that to the general population wrt. politics. The "opportunity to just not draw a conclusion" has a much greater payoff there.
Do you just not contemplate anything that isn't directly related to your physical needs? What a sad existence. I wonder what you're doing here, in that case.
Easy. Eat the sandwich, see what happens. No heuristics necessary.
Alpoge is a smart and principled guy, and he is probably the world leader at getting LLMs to do math. But he's not an expert in these subjects. If we're going to argue from authority, let's wait for an actual authority to weigh in
Of course he is an expert in these subjects. He has a mathematics PhD from Princeton, with a thesis on algebraic geometry. If you cannot call someone like that an expert, you cannot really call anyone that.
It probably looks like that to someone outside mathematics.
But a math PhD from Princeton (even in arithmetic statistics, which is what he actually studied) doesn't make you an expert in the Riemann hypothesis.
I have a graduate degree in mathematics, with a thesis in algebraic geometry myself. I know what I know and what are my limits. I am not an expert in Riemann hypothesis now, but if I spent a couple of months of focused study, I would be. That’s where Levent surely is right now, at the very least.
You two are just talking past one another and using different, and valid, definitions of experts. Yes, Alpoge is an expert in mathematics. Though how much of an expert he is in number theory? I don't know, I'm not an expert in number theory myself, I don't know Alpoge, and I don't know where he's been focused lately.
But either way both of your arguments comes down to speculation unless you two are also experts in number theory and know Alpoge or at least what he's been working on (to a sufficient level).
> if I spent a couple of months of focused study, I would be
It seems more likely that in a few months you would discover the depth of what you do not know about RH.
If you are a very accomplished geometer, I apologize, but from anyone else this would be very egotistical.
> If you are a very accomplished geometer, I apologize
He is not.
I think we mean different things by expert. You appear to mean "superficially familiar with some known approaches to the problem". That's all you're going to get in months.
And, yes, I've also published papers in algebraic geometry.
Yes, if by “expert” you mean one of the top 10 people in the world, then sure, but that not how regular people typically use the word.
Silly definition. I'm a "top 10 people" in the application Data Science to [niche that my company works in] because there are so few people working in this area, but I've only scratched the surface in my 3 years of working specifically in this domain.
Some things just really aren't well-understood or well-studied.
And other things are extremely well-studied and well-understood where 10 is significantly too low a limit.
I disagree. I don’t know that the answer is 10, but its definitely a very small fraction of math PhDs and they all know each other.
If only 10 people understand the problem well enough to contribute to it, then yeah, that would be a case of a problem with only a handful of experts.
I don't think regular people would have a problem with that, since the idea that some problems could be _that_ deep isn't particularly new.
His writeup says he's been angling to work on Riemann since he was a student. It does seem plausible he's an expert.
From the writeup for the lazy:
> It was my secret in my student days that my plans were for this problem in decades, which meant then that I tried to become familiar with all tools that I could possibly recognize in the problem when it was time
I'm going to wait for Sarnak to weigh in.
Sarnak weighing in is completely irrelevant. There is a lean formalisation, and it either compiles or it doesn't. (And the statement of RH is well established so there is no issue there.)
Lean can and does have bugs. It compiling doesn't mean anything, because that doesn't tell you whether the proof actually encodes what the mathematician intended. Plus, again, there can be prover bugs.
In principle, sure. But the likelihood of a plausible proof strategy produced by an AI agent with human assistance somehow accidentally exploiting a Lean bug is incredibly low. In any case, the fact that the Lean compiles is certainly going to be, in this setting, a more reliable assessment of the proof's correctness than the opinion of any expert, unless that opinion also comes with an identification of the bug being exploited.
Alpöge is a grand-disciple of Sarnak.
Alpöge -> Tsimerman -> Sarnak
Feels like we're reaching for arbitrary levels of expert to undermine the claim. I'm sure the world's foremost Riemann Hypothesizer will come forward at some point, but we don't need to wait for that to take the Princeton math PhD seriously.
There are actually no experts on the Riemann hypothesis bc before now no one knows how to do anything lol
Its all relative. If you're on hackernews then conpared to you, dear reader, hes an expert.
"Please don't sneer, including at the rest of the community." It's reliably a marker of bad comments and worse threads.
https://news.ycombinator.com/newsguidelines.html
p.s. FWIW, I know of at least one Nobel Prize winner on HN.
The beauty of the internet is that no one knows you're a dog.
What an odd thing to say. By this logic dumb people should defer to worse experts because "compared to you, they're very smart". Dumb people need the best experts of all!
Credentials are not authority. Especially in math.
Math doesn't work like that sadly.
The model(s) that solved it is an expert, I guess. Maybe they can send it to OpenAI and have their internal model check.
Apparently he has a Lean proof, so all that needs to be verified is that the Lean formulation of the problem matches the hypothesis, which I assume he's qualified to do (and has done).
Did you read the document? He is very much an expert. JFYI you don't need a medal to be an expert.
No, you need a long track record of working on the Riemann hypothesis. You need to be the person who's called to verify a possible proof of the Riemann hypothesis. I have a lot of respect for Alpoge, but he is not the person a journal would call to verify this.
Asking out of real curiosity, not trying to provoke at all: and what happens when somebody verifies it and say the result is legit? Would Alpoge become an expert on the topic or wouldn't, due to the AI assistance?
Also, if it were the most expert person on the Riemann hypothesis the one working with AI to prove it, how would the verification work? The 2nd most expert person? Or would they be trusted directly?
It's a fun question. I don't think AI assistance transfers expertise from the agents to the user. Alpoge says much the same thing himself, noting that Claude produced 250000 pages of slop and that he doesn't trust his own summaries in places.
For the other question: We'd want multiple independent expert verifications.
Man, are you being deliberately obtuse? 250000 pages of proof in a formalized language is not "slop"
250k pages is the "natural language proof" that Alpoge chose not to publish because it was slop.
The lean proof is 17M lines. That did get published, although good luck compiling it if you don't have access to 1/2 TB of RAM.
That’s a strange form of credentialism. I understand it if you wanna reject amateurs who make graphs and don’t understand what they’re saying but this is an actual mathematician and clearly he has thought about this particular problem for a long time.
From the document: "It was my secret in my student days that my plans were for this problem in decades, which meant then that I tried to become familiar with all tools that I could possibly recognize in the problem when it was time".
It's true that he haven't been publicly recognized as a top expert. But it's not like he's just a software dev who got lucky with prompting - he's a mathematician who specifically studied this problem.
I don't think we should delegate gatekeeping to journals - a lot of crap is published, by "experts". (Also a lot of good results don't get published - there are well-known examples.)
I don't get it, if the proof is formalized in Lean, the only expertise you need is running lean, asking if there were any additional axioms added in the proof that shouldn't be there, and if the statement being proved is the GRH (verifying all the definitions involved match expectations).
You need to check that there are no bugs in the lean formulation of course.
If that means checking the formalized theorem states the GRH correctly yes, but you hardly need to be an expert in the RH to do that.
The usefulness of a Lean proof is that it is possible to just look at the theorem statement and at the axiom that are used. Riemann hypothesis is quite straightforward to define and to be checked even by a undergrad. Obviously a bug in Lean kernel is still possible.
The paper is not a proof and has not been peer reviewed / published in a journal. So yes, let's wait and see.
There's no paper. It's just a letter about the discovery.
The lean proof is public. Although, if it's anywhere near as long as the "natural language" proof, most people probably don't have enough compute to check that it compiles.
footnote: I object to calling the langauge that Claude produces "natural". It appears to be uniformly distributed over the space of grammatical constructions.
> If he says he has Riemann, well then, he has Riemann.
Well he said he doesn't even fully understand the proof (250k pages). So it's more like "If lean says he has Reimann, then we presume he has Reimann."
The best argument I've heard for lean being trustworthy is that different lean engines won't likely have the same bug. I guess what I'd like to see next is a list of every lean engine and its result for this?
And then if that passes, presumably the goal is simplifying down from 250k pages.
What I don't understand about Lean is how is it going to avoid writing statements that are no-ops?
How do we know all the lines in the proof aren't just gibberish?
I don't have much of a math background but the author making such strong claims about a proof the author couldn't have possibly read seems absurd to me. What am I missing?
Exactly, people are running these computations right now. You could as well. It will take several hours, though.
He could still be mistaken. Andrew Wiles had a hole in his proof of FLT at first.
> If he says he has Riemann, well then, he has Riemann.
He works for the company, which is a conflict of interest. At that point it's not about credentials, it's about having someone with similar credentials but not the conflict of interest say the same thing.
Is it not a conflict of interest when a researcher who works at a University announces results? Let’s not act like the institution doesn’t benefit from the work of its employees.
Of course it is, that's why you need someone else to validate it. It's the entire premise of reproducibility, peer review, etc.
It was fascinating to read, but without knowing that, it was hard to know whether the sentences indictating his ignorance of certain areas of mathematics involved was an indication that he was completely out of his depth, or the usual and appropriate caution of a fully qualified academic (the more you know, the less you think you know phenomenon).
What do I have to do to get an asteroid named after me?
Change your name to 354 Eleonora ¯\_(ツ)_/¯
Wouldn't that be naming oneself after an asteroid? Not quite what the OP is after.
> If he says he has Riemann
Appeal to authority fallacy. The proof should stand or fall on its own.
> The proof should stand or fall on its own.
At some point in the chain an "authority" has to validate whether the proof stands or falls.
sure, still wouldn't trust the guy who wrote the proof to verify it though, which is what this person is asking.
Of course it has to. Its not about that, its about whether or not to get excited. I think we can start to get excited.
Appeals to authority are not inherently fallacious if the individual is, indeed, an authority on the topic at issue.
To the contrary: we lean on authority opinion in most spheres of life because one individual's ability to understand is limited and so we abstract away "true" in auxiliary fields to mean, roughly, "held to be true by trusted authorities". We even rely on it in law.
He's saying that it isn't right automatically because he said so. It adds significant credibility that the proof is correct compared to me saying I did it but the proof itself is what matters (at least to most people)
Just like when Andrew Wiles discovered the proof to fermat last theorem and Grigori Perelman with the poincaré conjecture, multiple authorities in the field got together and verified that the proofs were correct (even after Wiles initially had an error.)
An appeal to multiple trusted authorities is better than a singular authority. So at best I'll keep as unverified until it is checked, unless you have read the 250,000 proof and checked that there is no issues or soundness bug with the lean proof.
They are inherently fallacious as in logically unsound. There is no connection between his credentials and him being right about this particular thing.
It's very important to keep those things separate.
See: https://www.law.cornell.edu/rules/fre/rule_702
Without an epistemology that allows integrating truth-by-authority, you by necessity need to prove every assertion from first principals. Have you done so?
Very interesting sir. Thank you for this exposition.
But I do not see how this changes the notion that being an authority does not turn your arguments sound or even valid. There is nothing to connect. I can be an expert and completely wrong and there is nothing epistemology can do for you there.
The legal profession is taking a shortcut like the CS profession is taking a shortcut with the halting problem. Sure, we can end the discussion that way but we cannot use it as an argument otherwise.. halting problem solved, just press control-C?
You can't even contribute to math at this level without an advanced degree (cranks and delusional people notwithstanding). So maybe I should clarify I mean "authority" as in someone with high repute among his colleagues.
If we were all factory technicians, and Max has the best record among us at fixing the machine, that doesn't mean his word is final. "Max said so" is an appeal to authority. We are all capable of reading the documentation and performing our own repairs, and sometimes Max can be wrong.
obviously the claim is about how much credence to assign the claim until the proof has been verified
>fallacy
It doesn't make it right, but more importantly to people who always quote every known fallacy: it doesn't make it wrong either. We shall see, empirically how it goes.
But it is highly relevant whether this is some crank on an Internet forum posting proofs about infinity-hat or someone with some credentials, even if that framing was a bit hyperbolic.
Which is another thing: sometimes people aren't really making an argument as such, so quoting fallacies is just on purpose low bandwidth English comprehension.
It's not about whether the claim right or wrong. "Fallacy" doesn't mean "a function that always returns false". It means it employs flawed reasoning. Saying "$SUPER_COOL_GUY claimed it" (implying it must be true) is a thought-terminating phrase. It's saying "don't scrutinize or question. Trust that he's the authority on these things. For only a fool would attempt to contravene his infinite wisdom.". That logic is absolutely flawed.
Even Einstein was incorrectly critical about some quantum physics properties. Who would dare tell Einstein, the master of relativity and spacetime, that he was wrong? How could he possibly be wrong when it comes from the man who gave us the most famous math equation in history?
Yes, that's right, but proof checking does take time and effort. For an initial announcement, the fact that they are a well regarded mathematician means that it's worth taking seriously and not that it's definitely correct. If it's someone unknown, then you'd definitely be waiting for the proof to be checked before talking about it.
That is why it is still a "claim" of a proof, until other mathematicians verify it.
Not saying it is incorrect but, all claims of mathematical proofs have to be checked over formally regardless of who discovered it.
With this particular mathematician it just means it should be taken seriously but you cannot blindly believe it is correct without multiple authorities testing that there are no holes in the proof, just like they did with FLT and the poincare conjecture.
> If he says he has Riemann, well then, he has Riemann.
Yeah, no. This kind of framing is inappropriate. Let the mathematicians check the proof.
That’s what he is saying though?
Is it? It sounds to me like glazing one person in particular rather than the rough consensus of the entire mathematics community. I'm saying the latter is necessary.
He also works for Anthropic and this claim comes the day after their lackluster IPO.
Are you implying the timing of this is related to the IPO?
If Anthropic cared about it with regards to the IPO, it stands to reason that they would release it a day before, not after.
This is addressed directly on page 2 of the linked doc, no?
> It is obvious to me that I should remark on the following. > The precise timing is horrible. If I could have, I’d have finished this all much earlier. Hell, strictly before this age would have sure been nice.
> This all happened for me late Sunday night, and I messaged Dario, "Hypothetically, if I were to have . . . ", eventually asking about my ability to announce.
edit: Apparently I have forgotten how to format correctly
I believe every major announcement made by a public company comes with consideration of the stock price.
Considering the stock is down today I don't think markets care about math proofs.
What you said is exactly the opposite of the scientific research spirit: Instead of be suspicious and honest, you forge idols and blindly believe in them.
* Levent Alpöge
(I think you posted this comment because the parent comment has a typo that was fixed later. Let's hope it goes back to black soon.)
I don't understand this comment. For instance,
> a supremely qualified mathematician
How? No matter how "qualified" you are, the AI is still doing ~100% of the work. And since we're talking about superintelligence (Levent's intelligence hasn't even been stated anywhere), it's questionable whether a human can even "verify" or "check" the AI's work.
- It's a lot more straightforward (easier) to verify a proof than to come up with a proof
- It is also unclear if this proof is something a human couldn't have come up with
> It's a lot more straightforward (easier) to verify a proof than to come up with a proof
I don’t disagree, but this is not established science.
We actually don’t know if it’s easier to verify a proof than to come with a proof.
It’s one of the fun things that lead to P!=NP if it’s correct.
Again, probably correct but if you could prove it that would be groundbreaking.
> the AI is still doing ~100% of the work
That's clearly not true here.
If you read through the pdf that this post links to, you'll find that while claude contributed many ideas and was ultimately responsible for the technical "implementation" of the proof, there was also a lot of non-trivial guidance and steering involved from Alpöge himself. He didn't just sit down and prompt "claude, solve the riemann hypothesis for me".
If it was that easy, OpenAI would have already announced the result weeks ago.
> the AI is still doing ~100% of the work
So if we stick a 10 year old in front of Claude and instruct them to type in “create a proof of the Riemann hypothesis”, as long as they keep pressing return, eventually a proof will pop out?
Yes because their lifespan is long enough AI will likely continue to improve
Thanks, the title made it sound like the custodian walked by and typed "solve this" and found an answer.
That's pretty much what he did regardless. I can assert with absolute confidence that he read effectively 0% of the "proof".
This isn’t anything like what using LLMs for complex problems looks like.
I sometimes think AI doomers are terrified because they have no idea how this all works.
It isn't how this work was done, but it is basically how UGC and 400 other high profile theorems were proved by OpenAI so I'd say the doomers have a point.
Huh, what a conincidence that the AI presented the proof to Levent Alpöge (a top class mathematician) and not to some Rando McRandomface then? What are the odds?
I posted below in detail, but the answer is because Levent works there, not because Levent “guided” anything.
He doesn’t really go into much detail here, but according to him he came up with the strategy: „ I had us on the line of combining two quite different ideas present in a toy setting, one of which was already new and shocking, making very simple one of the shadows of Riemann over which we mathematicians felt most triumph. I guided us to the right situation analogous to that toy setting, and then Claude began to leave me in awe.“
That does not sound like he typed: „solve rieman thingie, make no mistakes, plz!“
Here is the description of the initial process:
> Jarred Sumner, an Anthropic staff member (and non-mathematician), prompted Claude to “take a real stab” at the hypothesis itself, leaving the mathematical choices from there up to the model. Initially, Claude generated and tried 650 ideas, none of which worked. Jarred prompted Claude to try again, and it spent a day and a half coordinating about 60 Claude subagents, which this time went much deeper: between them, they ran 2,400 shell commands and wrote hundreds of Python scripts.1 The subagents ran thousands of numerical checks against known zeta zeros and refereed one another’s work. Throughout this process, Jarred's input was mostly limited to sending Claude messages of encouragement (mostly variants of “keep going” or “believe in yourself”).2 This seems to have helped Claude overcome some initial skepticism that it could make meaningful progress.
Only after this person thought he/claude found a solution were the mathematicians brought in.
This description is about the 67% bound from a few months ago, not the full GRH being discussed here.
A Rando McRandomface doesn't have access to millions to tens of millions of dollars to spend on tokens.
There's a lot of Random McRandomfaces in Anthropic with the same access.
How could you possibly know that?
It's a 250,000 page proof. By his timeline he had the first breakthrough about a month ago, and he admits to not understanding everything yet.
Even if he were extremely prolific at reading through the math, he would have carefully read <1% of the proof in the last month.
Because the writeup describes exactly that process:
> Jarred Sumner, an Anthropic staff member (and non-mathematician), prompted Claude to “take a real stab” at the hypothesis itself, leaving the mathematical choices from there up to the model. Initially, Claude generated and tried 650 ideas, none of which worked. Jarred prompted Claude to try again, and it spent a day and a half coordinating about 60 Claude subagents, which this time went much deeper: between them, they ran 2,400 shell commands and wrote hundreds of Python scripts.1 The subagents ran thousands of numerical checks against known zeta zeros and refereed one another’s work. Throughout this process, Jarred's input was mostly limited to sending Claude messages of encouragement (mostly variants of “keep going” or “believe in yourself”).2 This seems to have helped Claude overcome some initial skepticism that it could make meaningful progress.
> Having found this new result while attempting the task, Claude tested its work by having various subagents review the proofs, search for counterexamples, download 54 papers from the arXiv to check that its finding hadn’t already been made, and independently re-prove its finding from scratch. Claude volunteered to write its findings up as a paper, and recommended that a human number theorist validate its findings.
> Levent Alpöge and Ralph Furman, two of Anthropic’s own mathematicians, examined Claude’s work to understand the new results and how they related to the prior work mentioned above. In parallel, Claude worked with another member of staff, Eric Easley, to produce a Lean formalization of the result, which passes the standard validation tool comparator.
Your comment literally disproves your own comment.
Simply put:
> Because the writeup describes exactly that process:
This is a complete lie.
He told to find a solution. It failed. He told it to try again, and to keep going when it stopped.
He had no mathematical skill or ability to check the solution.
so which part is the complete lie?
The third paragraph. You left that entire part out: the actual proof, the actual confirmation. The thing that's literally linked to at the top of this thread.
There's an entire process here, and you just decided to leave the most important part out. So, your comment literally disproves your claim. Which makes you a liar. Go ahead. Get to this thread by removing that third paragraph. That is the critical part of this entire thing.
You are conflating:
a) “looked at claude’s work”
b) read and understood over 200,000 pages of ai-generated proof
everything they post to their website is marketing
You also seem to be conflating “Levent guided the work” with “curious employee brute-forced some potential solution through an agent tree”. He (in this case both Levent and said employee) did not guide the work, it was in effect a random employee saying “solve this” over and over.
Man, Good Will Hunting 2 is much more boring than the original.
Let's redefine AGI as solving every known problems to man so we can stop moving the goalposts.
it hasn't solved any of mankind's problems yet.
It's solving tons of my problems every day...
Have you considered that you might not be representative of humankind?
I have actually considered the fact that I may secretly be an alien embedded into humanity, or perhaps a dog in disguise.
But evidence really points to the contrary, so... no, I'm pretty confident I'm human with regular human stuff that AI solves for me. House and family stuff on top of the science and tech stuff that applies to a smaller portion of human kind.
Hence why the previous comment expressed "mankind", not "HN accounts".
What are “mankind”’s problems if not the problems of individual humans? It seems like we’re treading towards a definition of AGI that would imply most human beings aren’t generally intelligent, which seems wrong headed.
Mankind's problems are that of humans but not that of software.
I think the point here is that quite a lot of problems are being solved and this is entirely self evident based on the stuff people are using it for right now.
It would really be something if AI could fix humanity's tendency to confuse "my lived experience" with "mankind as a whole".
It will need cooperation from mankind. So do we.
It's certainly creating a lot of new ones!
You can have the AGI label, you can put it next to the "actually, we've solved world hunger, people just aren't working hard enough" discovery, somewhere on the shelf with "every billion I make trickles down to the barista when I get a coffee at starbucks".
The AGI label doesn't matter. You can call it AGI when you feel like it's generally intelligent--I've yet to see any model that even remotely seems intelligent, but it all depends on where your bar is.
If this is superintelligence to you, that doesn't make it superintelligence to me. I can not have a single chat with an LLM without running into it clearly just predicting whatever sounds good instead of actually reasoning. I can't understand what sort of work people are doing where this seems super intelligent, or intelligent at all.
What is your definition of intelligence?
>I can't understand what sort of work people are doing where this seems super intelligent, or intelligent at all.
Proving the Riemann Hypothesis apparently
But not in some sort of super clean elegant way, it's "just" brute force, no? The same way it brute forces fixing tests until the software works, instead of designing coherent software in the first place.
And yet, we still can't make a robot that can clean a room. I mean really clean a room, picking up trash and put it into the bin, clothes on the floor go into the laundry basket, remove stuff from the shelves, dust it and put it back, etc... the kind of job any maid can do. Not like a Roomba that can be defeated by a loose sock.
AGI means "general", a machine that can solve problems geniuses have struggled with while failing at problems that can be solved by minimum wage workers is hardly general, instead, it is a good specialized tool.
If Moravec's paradox is anything to go by, the "embodied" tasks will end up being the ones most difficult for AIs to do.
How about when it picks me, and nobody else, the winning Powerball numbers.
Proving GRH never was the hard part of the job anyway..
We shall see.
Math is not “if he said then it is”.
Last giant claimed to have it was Sir Michael Atiyah, but it was not the case.
https://youtu.be/bKnqeWWr9-Y?si=uYVYN6UkGMOMab5B
Atiyah was widely considered to have lost it by the time he claimed a proof of RH.
Why Ant stock is down down down? https://finance.yahoo.com/quote/ANT/
Wall Street seems unimpressed with Lean proofs.
Congratulations on making my palms sweaty. Fields medal for Levent/Claude...
If only we stopped with the marketing and named the person, so sad to see someone like him reduced to "Anthropic Employee"
Is this another $1 million Clay prize flushed down the drain, or can he actually claim this one? I guess it's small fry to him given recent business news. Sort of a shame that all of these prizes are disappearing. Perhaps we should give the prize money to a body of mathematicians to do stuff with.
I think technically he could even though his input into the proof was most probably cosmetic and the work was done by the AI.
"If he says he has Riemann, well then, he has Riemann."
that's now how mathematics works, but I guess it goes to show how non mathematicians understand the subtle differences between math and other fields
Easy now. I know we're all excited, but this is what peer review is for.
> If he says he has Riemann, well then, he has Riemann.
That’s not how this works, especially if he works for Anthropic
Q.e.d.
He did post the Lean proof.
Also the Wikipedia says that the full written proof is 250,000 pages. Which I am guessing might be a mistake. If not it could take a while for humans to verify it.
He says it in a footnote, probably referring to ~2M lines of Lean?
:) 250k pages?! I could believe that, maybe. It might amount to -
"It's true, because the rest of maths is true"
Which I suppose, in any consistent universe, must be true for all proofs, just that up until now, our proofs have been relatively simple.
Interesting question occurred to me - how long can any proof get before it can be simplified?
I suspect arbitrarily long, even for problems that can be simply stated.
[flagged]
We've banned this account. You've had more than enough opportunity to demonstrate that you intend to use the site earnestly.
> Even if their discovery happens to be wrong. (lol navier stokes)
Has it in fact been shown to be wrong?
There are dozens of papers claiming that their formalization is wrong, that their natural language claims do not match the lean version of it, some parts of the proof are insufficient. finally, their entire proof relies on an external force. For a problem that is about, for 99% of people that care about it... unforced fluids.
but hey, it gets them a month of being talked about.
That also got fixed today.
https://github.com/tristanbuckmaster/navier-stokes
Quod Erat Demonstrandum, meaning "that which was to be demonstrated" or "what was to be shown"
We used to joke at university that it meant "queda entonces demostrado"
I believe it's actually the correct "reverse acronym" in Spanish? I mean, keeping the Latin one they created a Spanish version that has the same meaning, in Spanish.
In Italian we use "CVD" for it (Come Volevasi Dimostrare)
CVD seems to be a fairly accurate translation of QED. The Latin is "what was supposed to be shown"¹ and it looks like the Italian means something like "as it was desired to show"? The Italian isn't a relative clause, but the meaning looks basically the same.
The Spanish appears to mean something more like "and now, it is shown", which is a completely different meaning expressed with unrelated words that happen to start with the same letters. (Spanish still has direct equivalents of all three Latin words², but I don't think que era demostrando even means anything.)
¹ In an English proof, you might see something like this:
"We can then see that a = b < c = d ≤ e, and therefore a < e, which is what we wanted to prove."
QED stands for the "what we wanted to prove" part of that, except that there is no we; in the Latin, we simply make reference to the fact that something was supposed to be proven.
² Sticklers might argue that Spanish has lost Latin's quod and replaced it with what is really a descendant of quid. They are almost the same word; quid is an interrogative form.
Isn't it what it literally means in spanish?
I think the joke is that while other cultures find Latin to sound esoteric and/or fancy, especially if they had to take courses to learn it, that your garden variety, non-educated, Spanish speaking person just simply gets it cause it's so damn similar to Latin that it doesn't faze them.
(Spanish speaker btw).
In English we have this with Rest In Peace.
That doesn't look similar. Rest in peace is almost a direct word-for-word translation of requiescat in pace "let [whoever] rest in peace".
By contrast, queda entonces demostrado is totally unrelated to quod erat demonstrandum, using different words (except that the third words are both forms of the same verb, very different forms) to express a different meaning.
ding ding ding correct!
Really? The Spanish words don't seem to bear any relationship to the Latin words, so I fail to see how the relatedness of the languages could be helping.
Have you studied Latin, and if not, what does "quod erat demonstrandum" appear to mean to you?
Not exactly. The Spanish version translates to something like "so it is proved" and is a full sentence, but the Latin is a phrase closer to "which was supposed to be proved".
TBD
You don't say.
Quantum Electrodynamics sounds cooler.
Wow I just thought those were the initials of the guy who wrote every proof ever.
This would be a good Easter Egg for a story. Mathemystics who worship the ancients, Pythagoras and the mysterious QED.
Characteristic of the disciples of QED was their renunciation of proper names in favour of a black rectangle.
Don’t forget his French cousin CQFD
C qoi fo démontrer?
C'est quatre fromages dauphinoise.
He says
> I believe I have, working with Claude, discovered a proof of the Riemann Hypothesis
Wonder what his intellectual peers think.
surely the incentive structure he has around the anthropic IPO has nothing to do with this.
I mean a single person can always make mistakes regardless of how smart they are. But this does make it sound a lot more exciting than if a random vibe-prover came up with it.
No theycan;'t
Brother it might be worth looking up Andrew Wiles
You mean, if the AI says it does, than it does. And it’s probably right, because it’s smarter than all of us and we basically just have to trust it now, which is exciting and frighting in many ways.
>If he says he has Riemann, well then, he has Riemann.
Ugh, I hate it when people want to willingly take us back to the Dark Ages.
It's not a random Anthropic employee. It's claimed by Levent Alpöge. The guy is on fire and already has like 3 or 4 very interesting verified result this year. I've showing for months his wikipedia page to my coworkers in the math department in my university.
I guess this is what you get when you pair top human intelligence with top machine intelligence (as in access to insane compute + latest models).
This. Frontier AI means that the sharpest minds in the world now have an infinitely patient and motivated colleague to work through every idea.
I’m not claiming to be “top human intelligence” but I’ve published in premier venues in my field and I can say confidently that none of my human colleagues can keep up in conversations on my research niche like Claude can.
You reminded me of when Rodney encounters a clone of himself in SG Atlantis
Jest and modesty aside, I think you're right in that it amplifies your strengths (and depending on how much you bias it, your gaps?) so will we see the world's top minds ever widening the gap between themselves and the rest of us from here on out?It wasn’t a clone, it was Asuran Replicator McKay. Which just makes it more fitting I guess?
But according to some, AI was supposed to raise only the floor, not the ceiling ...
I guess they were wrong then.
we’re in the human + chess engine can comfortably beat chess engine era.
Which is interesting, because it's not necessarily the best chess players who are the best computer-assisted chess players, IIRC
the best chess players today are engines, not humans. that’s the competition. the best computer-assisted chess player is probably another computer.
What is a computer assisted chess player today? Why would they do anything besides play the engine's move?
In correspondence chess (which is chess with a time limit of like a week per move), humans ask multiple chess engines which move to make, and then choose.
They don't use the engine while playing. They use it in training, to learn and try new moves.
And none of the best computer-assisted chess players are better than AI today. In the same way a human + chimp pilot is not going to beat a human pilot.
This era of human relevance in fields like maths will end – probably in a month or two if we're lucky.
I suppose we are already in the final stages of human-assisted mathematics, is what you're saying?
No, humans can always invent new fields and ask new questions
that’s not how math works.
Exactly. We don't even understand all of the new solutions yet, I heard Terence Tao say he had only looked at about 10 of the OpenAI proofs, found some good things, bad things, mistakes. But it will be his (and other's) understanding of the proofs that will give new insights.
In other words, what will the AI solve if we don't even have the language yet to formulate its next challenge?
The AI would need motivation... (Please no!)
Claim should be verified though; anthropic should spend equal compute with no human guidance and see solve rate
Not sure, because we can still pick different "games" computers are bad at, at least for now.
And it will probably last even less time than that one did
that’ll be good for deciding how to allocate capital efficiently between human capital and other expenses
And that's exactly how it should be promoted: not as a toy, but as a tool for high-value intelligence work.
I'm not naive—I know it won't be—but it should be.
> The guy is on fire and already has like 3 or 4 very interesting verified result this year.
Actually Claude is on fire, and this guy is along for the ride:
> Claude and I have had a great time together with such breadth in recent months, but this time Claude far outpaced me near the finish line. No, Claude flew, and danced, and sparkled near that finish line. I had us on the line of combining two quite different ideas present in a toy setting, one of which was already new and shocking, making very simple one of the shadows of Riemann over which we mathematicians felt most triumph. I guided us to the right situation analogous to that toy setting, and then Claude began to leave me in awe.
The anthropomorphization is cringey.
The writing is certainly exuberant, and we are used to cringing at this kind of language because we are used to it being unjustified. But if the proof holds up then I can hardly object.
Cringe is in the eye of the cringer.
Why srsly?
If Claude is breakthrough for Riemann, can't we give people slack on anthropomorphization?
People do this with animals, people do this with objects but an AI bot, writing like a human this is cringey?
Using "srsly" is also cringy.
Why is srsly cringy?
Because it sounds like you're a valley girl. Just spell it out, seriously. It's only 4 more characters, a small fraction of the entire sentence you typed.
Perhaps i am a valley girl?
I mean I learned english from watching movies and tv show, if thats the only reason for you to complain i think im fine :)
To be fair *hn"* is full of abbreviations and acronyms
Why is it cringey? Sounds like a good and natural description to me.
They told you what's cringey: the anthropomorphization. "Claude and I had a great time" – no, Claude didn't have a great time, it's software. Levent Alpöge had a great time using Claude.
I recently gave claude full control of a robot. I have never seen it mimic being excited for any other work I've given it.
These are next word prediction machines intentionally optimized to match human generated text, to be as human text as possible, then fine tuned to be conversational as possible. Text with strong sentiment will even pollute their context windows, biasing their next tokens with that sentiment, in a way that fades with time. What do we call that same concept in humans?
LLM are intentionally and mathematically optimized to be human mimics. They're not humans, and they're not "alive", but it's silly to not anthropomorphize something that was intentionally created to be, literally, anthropic! It can express enjoyment, irritation, etc, so it's not somehow logically flawed, or even incorrect, to say "claude and I had expressed signals of enjoyment". That's not saying it's a human, it's just being aware of patterns of text that was produced. No need to assume they're going to marry it, or thing it has a soul, or is sentient, or something.
Could you please share some transcripts and a writeup.
> What do we call that same concept in humans?
I'm not in the habit of equating human behavior with software, or coming up with ever more strained metaphors and analogies for why thinking is a lot like tokens and context windows if you scrunch your eyes up real funny. It's just sophistry.
> No need to assume they're going to marry it, or thing it has a soul, or is sentient, or something.
I didn't assume any of those things; if anything, you assumed that I assumed. No, I just agree with OP that it's cringey to say "Claude and I had a great time." At best, Claude produced text that was, as you said, intentionally optimized by Anthropic to make you think it had a great time.
> I'm not in the habit of equating human behavior with software
My point is, you sorta should be. It's quite literally the fundamental basis of the technology.
> intentionally optimized by Anthropic to make you think it had a great time.
This gives too much credit to Anthropic. This is the explicit goal in all LLM trained with the entirety of human text. The loss function is, very literally, "mimic the writer of this text as precisely as possible", with the writer being human. The result is you get something that mimics a human. And, as a part of that, you necessarily get behaviors of humans that were statistically significant in that text, because that's what you told the math to do. You get something that mimics being "lazy", "tired" (a huge problem in early models, before it was beat out of them in fine tuning), "irritated", etc, because mimicking humans was the goal. These specific undesirable human traits are beat out of them, but the rest, all the "intelligent", behavior, and practical use you see and get, is only and entirely because they're made to mimicking humans. It's literally the entire technology. You must equate it, to some extent, because a large part of alignment is also making it less human [1].
If someone makes an objective observation about the text produced by the LLM, I don't see that as cringe, I see it as an interesting quirk of the tech, which I've only seen under very specific problem sets that a nerd would also find interesting. And these coding models are, as expected, optimized to be nerds! So, the statistics around nerds being present in the output shouldn't surprise anyone.
I doubt I have a point you'll see as interesting. Cheers!
[1] https://arxiv.org/abs/2605.07632
You do know that humans also have a physical substrate? A bunch of eukaryotic cells wired in a rather large network with all kinds of feed back mechanisms, chemical diffusions, glucose transport, neurotransmitters, etc.etc staying in homeostasis despite much ability for chaotic behavior, seeming to be deterministic or at least quantum deterministic. Not software per se but quite likely a software model is possible with sufficient compute.
You're saying Claude had a great time because humans have a great time and it might be possible to model humans in software?
I'm saying that Claude being software is not a definitive argument against Claude having subjective experience, because it might be possible in principal to model humans in software.
Pointing out that Claude lacks internal temporality, a observe/check/act loop, memory, non-virtual experience and therefore can't have a great time would be a much sounder argument.
What’s the subjective experience of gradient descent? Bubble sort? Does ‘y’ feel left out all by its lonesome in y = ax + b?
What's the subjective experience of a neuron? Action potential? Does a calcium ion feel left out by its lonesome in the voltage-gated channel?
> You do know that humans also have a physical substrate?
Lmao
Edit, because this is mean without context: Your comment is essentially the following meme:
> Person who has never worked with anything but computers in their lives: 'Hmm, this brain looks an awful lot like a computer.'
It is true that I am not a neuroscientist but I don’t think that the substrate being software vs chemical is a priori a reason to rule out subjectivity. I think there are other things missing from the current LLMs, (particularly time, inference time chemical feedbacks or electromagnetic feed backs, continuous training, memory, non linguistic thought, etc.), but not the deterministic and formalizable nature of the system. I would not be surprised if something made of software running on some sort of manufactured computation devices did have subjectivity in the future.
You're leaving out the most important bit, which is consciousness, and this is harder to crack than inventing time machines ^^. At least time machines are amenable to science and we kinda have a vague idea of what time is. Consciousness OTOH, is so fundamentally subjective that we can't do science on it by definition.
Of course this is only a problem if consciousness is a significant part of being intelligent, but we don't really know yet.
This is just a bunch of nerd shit sophistry being used to equivocate humans with software. That's a bad take, Claude's a program, it doesn't have feelings no matter what the weird Bay Area sex cult tells us.
I don't think Claude has feelings, but not because it's software, but because it's too simply constructed to capture subjectivity (tho that's a tough claim as we haven't got a precise idea of what subjectivity nor of what is required for it to emerge).
Well you don't know any of that so I'm not sure why you're getting so worked up about it ? It's just software? What does that have to do with anything? You don't understand the software or the human brain anywhere near enough to actually make that statement confidently.
What makes you think I'm worked up? I think you're worked up.
> It's just software?
This but unironically.
"The anthropomorphization is cringey."
No it's not. If Claude just solved the Riemann Hypothesis (if that is later proven to be the case), the only cringeworthy thing is your response to his completely understandable description of the events.
Yes, it is.
Cringe is in the eye of the beholder. I think it's beautiful, you can feel Levant's excitement and joy through the page.
It's not beautiful, it's cringy. Good prose does transmit a feeling, but this isn't good prose, and the feeling it does impart is one of excessive, saccharine indulgence. So apart from the writing, the feeling it transmits is also cringy.
Beauty, and cringe is subjective. I am not sure why you are trying to argue in absolutes here. For me, his prose transmits a feeling of beauty, awe, curiosity, for you it reads as cringe. That's fine.
that's, like, your opinion, man...
Someone has to guide Claude still. Otherwise we would see layman bragging on LinkedIn about cracking random math problems.
I've seen post here of people solving their pet math problem. They are easier than the RH, but they had that as a personal project for years and now they solved it with Claude or ChatGPT.
For example, https://news.ycombinator.com/item?id=49961538 that uses the RH to prove that the sum of [the inverse of] some subset of the primes is finite. [Congratulations jdb1729, now you can update the solution to an hypothesis free solution.]
Edit:added "inverse"
Actually, this happens now. Maybe not on LinkedIn, but there are several guys who uploads proofs of theorems on Hexagon Math that they freely admit they doesn't understand. There's even a tag for it: "Human Understanding: No Parts".
This guy knows his stuff. Along for the ride would be if I made this claim.
The second part sounds like the timing of this was to pump the IPO.
This has to be the most impressive result in order to pump an IPO in human history, right?
"Result"
We'll see about that when anyone actually sifts through the haystack that the paper and the code are, which will take a year at least. And no, relying on AI to verify that other AI output is objectively true and free of AI-type mistakes/biases/exploits won't suffice.
lol yeah to all the investor class folks that understand the significance of the Riemann hypothesis.
Quoting the article: In Stepanov’s argument resolving Riemann over function fields one uses a divisor involving multiplicity. This argument is a descendant of Stepanov’s, but multiplicity is hard to find over the integers, corresponding to something like differentiating whole numbers. The simple, new, Fourier-analytic ideas in the toy setting led us to examine chaining theta lifts if we were to have the same tricks recur but work in a natively automorphic manner. The line: for the von Mangoldt sum of length X, Claude’s suggested length of chain of theta lifts, in view of each mimicking multiplicity, was √X.
Buy baby buy!
My guess is that the guy has a ×10 discount to use Claude [1] and also he is ×10 better at Claude whispering, so he get like ×100 better results than us mortals. [2]
[1] So he spend the equivalent of a lot of money, but not milllions.
[2] Mathematician here. I worked with people that is much better than me with pen and paper. I guess he is much better than me steering Claude.
PS: When we talk about Claude, we call it "Claudito" that save us the effort to pronounce it correctly in the middle of a sentence in Spanish.
He works for Anthropic, it’s probably a lot more than 10x if you mean usage limits; I doubt he pays his employer.
I think he's just happy to finally be able to talk to someone(thing) more intelligent than himself.
> The anthropomorphization is cringey.
https://www.usni.org/magazines/naval-history-magazine/1998/d...
> Actually Claude is on fire
The anthropomorphization is cringey.
> https://www.usni.org/magazines/naval-history-magazine/1998/d...
Love her and she'll carry you. Ignore her needs and you'll sink with her.
Wow, so I guess this is what it feels like to be in the "foothills of the singularity" as Demis put it. I have a feeling of awe and dread combined that's hard to explain. Things are about to get very interesting indeed.
Anyone know if this was Mythos/Fable or some as of yet unreleased internal model?
Stupid question, but how do we know that it actually proves the RH? And not something unrelated or slightly similar? I mean no one will actually be able to read through the hundreds of thousands of pages.
There was also a case of a formally verified implementation of some protocol that still had bugs, because the specification had the bug in it.
Two things I guess
Reading the theorem is much simpler than reading the full proof - so the job here is to read the formalization of the theorem itself not the proof.
Then you depend on the Lean kernel for proof validation and on trusting it and that’s a different question.
Finally the question was also true for ‘non formalized math’ where you always depended on a trusted kernel of mathematicians who understood the problem well enough to sign off the claim
Right but it was more than just "in my expert opinion, this checks out". It was more like "I have digested these ideas, I will teach them to my students, and we will be able to build on them to deepen our understanding further". So far I don't have much hope for that with these massive lean proofs of millennium problems. In a way that's optimistic as there may be centuries of work for mathematicians. Information (in the strict sense) has become much cheaper. We were fairly convinced of RH but we didn't know why. Not much has changed on that actually.
Probably late to the party but I downloaded the source and tried having claude go through and create a manim explanation: https://github.com/harttraveller/riemann/blob/main/assets/fu...
I had the same idea, if anyone wants to compare and contrast:
https://github.com/moozilla/riemann-hypothesis-explainer/blo...
Made with Fable 5.2 and a team of Opus 5.5 subagents
Cool beans
That is an amazing video, thank you! It explains the proof in simple terms. But like the other commenter said, the last 2 minutes is missing audio.
The last 2 minutes of the video have no narration.
From the link:
It is natural to feel that this is an end, in my case a loss of a principal guiding light, but what I saw during the course of progress so amazed me that I believe it is one dawn across science of something far greater.
If the proof holds up, I suspect that will become a classic quote.
Along the lines of "one small step for man, one giant leap for mankind"...
A live compilation/check of the proof/Lean code http://106.14.11.164
Build passes, just completed a few minutes ago.
fun!
For the laymen, what's the significance?
In truly layman's terms, this is among the most famous problems in mathematics; it is arguably the holy grail of pure mathematics and number theory. It is also another one of the Clay Millennium Prize problems.
In the ~150 years since it was posed, it has become a lynchpin for hundreds of downstream results that assume it is true. This proof would immediately confirm all of those results.
More practically, there are also some algorithms that will now have guaranteed theoretical bounds and it will likely eventually lead to some more interesting algorithms in the future.
It's kinda the math equivalent of "we found life on another planet".
I went to the Wikipedia page for it, hoping to find one of those “practical applications” section and left greatly disappointed! No doubt that this is very exciting for Math departments and would be a historic achievement within the field, but I still struggle to see what real world advances this could bring about. Is this about cryptography or something?
The most "practical" thing I can point to is that the Miller primality test has a concrete bound under the generalized Riemann hypothesis (which is also claimed in the OP paper). Still not very practical because we already had the algorithm and we already knew experimentally that it worked well in practice.
This is why I compared it to "life on another planet". In both cases it might be more surprising if it was false; but even knowing the answer for certain, we wouldn't have any immediate application for it. It's likely that with further study we can learn more and might find new applications.
It is a remarkable achievement, but I suspect that in 100 years the most notable thing about the Riemann Hypothesis will be that it was among the first tremendous achievements of AI. It'll be picked apart and dissected, extended, and applied, but just as we rarely think about how silicon wafers are made into the computers we're using, future generations will not stare up at Riemann as a mountain daring them to climb. They'll see it as we look at other foundational results from the past: a stepping stone on the path to the next mountain.
A lot of the Millenium Prize problems are pure mathematics problems in the sense there's no direct immediate application. There's some neat implications for primality checking and the like but the degree to which this is relevant is probably questionable given we typically use ECC or maybe post-quantum crypto nowadays.
For most of these unsolved math problems, should we expect there to be many practical consequences? If there truly was something that could be made from knowing Riemann was true, could we not just have assumed it was true and developed such a thing on that assumption and see if it worked?
There’s a reason it’s a pure mathematics problem and not applied.
Which ones? I’m sorry but your post really doesn’t answer the question.
I was trying to keep it very generic because I think you can understand the significance of the result without having to probe into the details. Saying "Riemann implies Fhqwhgads" is also impenetrable to a layman.
That said, the most practical result I can point to is probably that the Miller primality test has a concrete bound under the generalized Riemann hypothesis, which is also claimed in the paper. This doesn't change much because we already had the algorithm and we already knew it worked well in practice, so it's more of a confirmation than conclusion. That's a trend around Riemann because we've been able to just assume it was true and proceed beyond it already.
I thought that the downstream results were already assumed to be true because people by consensus thought that the hypothesis is likely to be true. Are we already using any things or results in practice that were created through this assumption? The important question is how much value a real proof would add. Would a counterexample be much more disruptive by overturning all this research that was built on the assumption it was true?
So what practical applications does it have?
Will it cure cancer?
Will it result in fusion power?
Will it change the lives of people who live on planet earth?
Are there any positive effects at all which will be felt by human beings? Other than perhaps Anthropic’s stock price increasing.
It will help us solve problems where there are patterns like the prime numbers. Since the prime numbers produce very complicated patterns, this can be helpful. But a math result is always just another way to see the world, not change it.
There have already been a bunch of results that relied on assuming the RH was true. It will not cure cancer, deliver fusion power, or directly influence or change the lives of anyone on earth--at least not as far as we know right now.
Are there any positive effects? I'm not sure. This is certainly among the problems--like the others you listed--that have attracted the most human energy and brainpower ever, and over an even longer time than the problems you listed. Is that not remarkable? It was like 2 years ago that LLMs couldn't even solve peak high-school level math problems. Now they've taken down two of the most famous maths problems in history in about as many months. And unlike Navier-Stokes which was approximately the prompt equivalent of "sounds good. keep going." this one seemed to require a good deal of expert human involvement. That's _interesting_, right?
I feel like I'm standing in the rain witnessing the invention of the wheel or the discovery of fire and hearing folks complain that it's pointless because it doesn't keep them dry. What people are achieving with this technology is _astounding_.
It's cool, but to a dummy like me who knows not knows "ryemann (sic) very hard" and then hears "no practical application" it seems to follow the general AI trend I see in my day to day.
And that is, very good at generating text. Not in of itself helpful in the real world unless that text it generates can be pointed somewhere useful. And the maximum utility I have seen is writing software. Which again, is quite good - but not earth shattering (except to my career prospects).
The other uncertainty to me (as a dummy) is where any of this can even go. Do we just scale forever? Does AI ever succeed at RL or is that just crap to sell the stock? Who knows...
From a layman, all I’m hearing is that proving this theorem will stop us from misallocating the focus of genius minds on the practical equivalent of a very hard sudoku puzzle.
Right? I find it bizarre that a common reaction of professional mathematicians to AI is "you should stop working on this stuff that has no value to anyone, so that we can keep doing it".
Lots of existing maths proofs start by assuming the Riemann Hypothesis true. Proving it will automatically complete those proofs.
More discussion: https://mathoverflow.net/questions/17209/consequences-of-the...
3blue1brown has an excellent video for a visually intuitive understanding of what RH is.
It's quite old but I find a reason to watch it every few years.
https://youtu.be/sD0NjbwqlYw
It is one of the greatest open problems in Mathematics which has stayed open for more than 100 years (actually since 1859), which has eluded generations of mathematicians, including some of the most brilliant minds of the XX Century (it was already important enough that it is part of one of problems in the list of the 23 Hilbert's problems (stated in 1900). Many results holds under the assumption that RH is true.
Isn't it arguably the greatest? Maybe tied with P=NP?
It's the hardest problem in mathematics and solving it would be the most impressive thing done by AI so far. It also automatically proves many other statements in mathematics which are currently formulated is [if the Riemann hypothesis is true, then <this> happens].
Take everything that happened a couple of months ago with Navier-Stokes and increase the magnitude several times.
Is it the hardest problem, or just the biggest problem? Is solving P=NP not more difficult?
We know to Godel that some problems are independent, i.e. it can't be resolved. It's likely that some famous problems are independent, and thus infinitely difficult.
It is perhaps the hardest problem that has been resolved
with "everything that happened a couple of months ago with Navier-Stokes" you mean the claims of plagiarism? I don't remember much else connected to it
By definition unsolved problems are harder
As a fellow layman my understanding is that if it holds true then it has implications about the distribution of prime numbers. Cryptography is an example of a field that relies on prime numbers. They underpin a lot of security. Also, right now finding new large prime numbers is relatively difficult to do, but relatively easy to check. I'm not sure if it's NP-something or not, but finding a solution to shorten the effort to find new primes might have broader applications in algorithm design.
The direct impact is mostly theoretical, as many things in number theory are based on assumption of Riemann hypothesis, and they were not advancing much historically because there was no proof of Riemann. These things will start getting more spotlight if proof holds true, and they could lead to advancements in cryptography, coding theory, computational theory, and other fields that are heavily invested in number theory.
NB: I am a student of mathematics, not a researcher, and some practicing researcher could correct me here
It's one of, if not the biggest unsolved problems in math. It's also a millennium prize problem. There are a ton of conditional proofs in math that are true if RH is true, so this would fully prove tons of work. There's also a link between the spacing of nontrivial zeros in RH and the spacing of prime numbers, so it will be interesting to see if this proof illuminates anything about the primes.
This is the only information I've found that was at all comprehendible by me:
https://en.wikipedia.org/wiki/Riemann_hypothesis#Consequence...
Goldbach's conjecture would turn out to be true.
https://mrc.sdu.edu.cn/english/info/1018/1514.htm
Requires a bit more, but Alpöge's work may have machinery to provide that.
The Riemann zeta function can be written as a product of the same function in prime powers.
Finding zeros of a complex function tells us a lot about Riemann zeta function, and in turn, the distribution of primes.
I’d compare this to Fundamental Theorem of Algebra in terms of importance.
For certain parts of pure math, it would be tremendous, it's been a big deal and is connected to lots of other things in those domains. For everything else in the universe, there's no real consequences.
Complicated topic that I don't fully understand, but I believe our understanding of Complex numbers would be much better. In that we understand regular/irregular distributions. The knock on affect is there's thousands of theorems that have been prove 'if' the Riemann Hypothesis is true. And there'll be much more to look into now that it's proven.
I don't know how immediate it would be but you'd expect to see improvements in any algorithm that uses prime numbers. Which is quite a few (RSA, Hash tables, RNGs)
The Riemann Hypothesis is used as a foundation for several important theorems. See https://en.wikipedia.org/wiki/Riemann_hypothesis#Consequence... for details.
I have the same question. So far I'm not seeing details on why this is worth continuing to read about.
Thousands of mathematical papers build on top of it, basically proving X holds _if_ the Generalized Riemann hypothesis holds
So all of those X are now proven (if this is indeed proven)
Looking at the wikipedia page explaining it... I get the feeling there is no *quick* way to explain it to laymen.
(Just a feeling, I don't know the significance either.)
I haven’t seen a Wikipedia page on mathematics that was written for laymen yet.
Mathematicians are nothing more than a Matrix product.
These neural networks are nonlinear and cannot be implemented as nothing more than a matrix product.
Arguably the most famous unsolved math problem.
Humongous
https://news.ycombinator.com/item?id=50572355
The two stories were next to each other earlier and made me laugh, that's all.
Was this solved in a chat interface? It's not clear, but it seems like it from the document. The most relevant quote seems to be:
> Claude and I have had a great time together with such breadth in recent months, but this time Claude far outpaced me near the finish line. No, Claude flew, and danced, and sparkled near that finish line. I had us on the line of combining two quite different ideas present in a toy setting, one of which was already new and shocking, making very simple one of the shadows of Riemann over which we mathematicians felt most triumph. I guided us to the right situation analogous to that toy setting, and then Claude began to leave me in awe.
If indeed this was simple 1:1 chat, that's quite important, in and of itself, in my opinion. Presumably, OpenAI and others have been throwing autonomous agent swarms at this problem without success, but a single human + Claude seems to have cracked it.
We may be in the "centaur chess" era of superhuman intelligence.
just cause you're talking w one agent doesn't mean that agent isn't coordinating 100's (or 1000's) of sub agents to tackle pieces of the problem..
What about that line implies chat?
It sounds quite collaborative. I would assume that spinning an autonomous swarm would not be described this way. It's not clear, for sure, but it sounds like it to me.
I would imagine it's an agent swarm with a dedicated coordinator or human interface agent - the one agent you actually talk to who interfaces with the ten morbillion other ones. It's perhaps strange to treat that swarm as a gestalt entity called Claude, but hey, we're in strange times.
Yeah, if it's a swarm and they're referring to it as an intimate conversation with Claude, that would be interesting.
It’s possible the mathematicians at Anthropic don’t even know the details of how this works. If Anthropic is willing to hire a bunch of mathematicians to solve open problems, they’re probably also willing to have some engineers help set up the infrastructure for them to work effectively.
So the researchers might see something resembling a regular chat and not realize that every non-trivial prompt related to the problem spins up a swarm of agents and burns $50k in tokens.
That’s just speculation, but given the money involved in this stuff, it’s hard to imagine Anthropic is just sitting these mathematicians at a computer with a chat window and unlimited token budget and leaving them alone. They’ve probably sunk lots of engineering time into supporting this research.
Obviously not. Did you miss the part of the thousands of lines of Lean code?
The linked document says the proof is incredibly long, but nothing else, so yes I guess I missed it.
"Thousands of lines" doesn't exclude chat at all. You can easily generate this in a single chat session in an afternoon for coding. Do you mean millions?
I meant thousands of pages, millions of lines, yes.
I read elsewhere in this thread it is 15M lines of lean.
It's 17M according to the github repo.
Which btw, is akin to showing someone the bible and saying "Proof of God is in here. It may be 2000 pages and in Latin, which you can't read, but just believe me."
I don't think a single instance of Claude wrote 15M lines of lean code or produced a 250k+ page "slop proof" when prompted.
He provides a lean proof in the first footnote. I downloaded it and took a look, he seems to use the RM conjecture defined by the community years ago, and he doesnt seem to introduce new assumptions (although to fully check this requires running the proof, which on my laptop will take days). Because of this, we only need to wait for someone to run the proof checker, and if it passes, then thats it, the only uncertainty left is bugs in Lean
Alpöge is moreover claiming the Generalized Riemann Hypothesis. The written argument is over 250,000 pages long and not provided.
The toolbox here contains a surprising, to me, amount of algebraic stuff.
CS layman folk here. How one is going to prove 250K page argument? Just single sign of hallucination can make it false, no?
Its formalized in lean, which verifies each step. The steps themselves are relatively easy, and can be mechanically verified (think "A and A->B thus by modus ponens B").
Traditionally writing verified proofs is a hard undertaking, exactly because you need to write out a lot of steps "normal human" mathematics jump over, but with AI its easier.
> The steps themselves are relatively easy, and can be mechanically verified
They aren't
At what point do we start worrying that it's found a bug in Lean instead?
The Lean kernel should be easy to verify by hand. We can ask the same question with humans though. What if a result that was used to prove some other result turns out to be wrong just because whoever reviewed it neglected a small detail?
What do you mean by it "should be easy"? It isn't easy at all; nested inductive types is the example that come to mind. Something like HOL is easy to check by hand as a prover kernel but Lean's type theory is definitely not.
i guess what people are realizing is that, as great as these discoveries are, these errors have probably already happened numerous times in the past, and though they might get corrected, either way, it didn't really matter. academic math is as much about the aesthetic experience of doing math, maybe more so than anything else, as it is about the objectiveness of the discovery, which is what every mathematician has been saying, so it's not like small errors, provided they are unknown, matter.
Still possible, but Lean ships with formally verified kernels as of a few weeks ago.
formally verified by what?
Lean. It isn't airtight, but it would have to be something very extreme or adversarial like Reflections on Trusting Trust.
In principle you can run Lean with different kernels and so there would have to be a bug in all of them for the proof to be faulty I believe.
That already happened multiple times, it’s a known worry
Has there been a single case of a proof that was thought to be correct and later turned out to be incorrect due to a lean bug.
No, not any serious proof
Identically in all 4 independent checkers it ships with? Basically never.
Right now.
AI proof in lean, can verify by computer. Mistake can only be compiler bug or state problem wrong.
No, it can also be that the proof doesn’t actually prove the thing you think it does, and a lot of other variant of that issue. Lean doesn’t make it magically valid
Right, but to the extent that Lean is correct (both the semantics of the language and the implementation(s) used to check) and the proof checks with the expected axioms, then the length of the proof doesn't much matter since Lean has checked it.
You'd still have to check the theorem statement, which is not going to be nearly as easy as with FLT (Riemann uses analytic continuations etc) but it's much easier than checking a 250,000 page proof.
Mathematicians, famously known for being extremely precise in their specification of problems and absolute lack of errors writing down proofs.
It's always, always, always stating problem wrong.
Step 1 will be making a much shorter argument.
I think we're probably on the precipice of proofs that can't be made shorter and so can't be really understood by a single mathematician in their lifetime. Or maybe we'll have proofs not understandable by humans at all in the same way a cat can't understand the fundamental theorem of calculus.
It's going to be the age of the oracle.
LEAN: https://lean-lang.org/
Small errors in proofs can often be fixed easily. So it's not that the whole thing would be useless. Also, he has a Lean proof, which rules out these small errors.
It is linked from the pdf, just rather big.
The Lean proof is linked, not the 250,000-page writeup (because it wouldn't be very useful in that form).
It is sorta amusing that you're asking the community to review 250,000 pages of AI generated work. Could all be true, but that's a big ask of a human. And asking another AI to review it seems something he could just do himself.
Whether the proof holds,
The most exciting time, and the most dreadful time.
Glad to live through the 90s/2000s so that I can entertain my son a bit about human history. Double glad that I’m over 40 and closer to retirement than the younger generations.
What a grim fucking outlook.
You can look at the exciting half.
I do. I don't think your children will lack meaning, and maybe not even lack jobs.
I don't know man. I don't know. We will see. But it is not my children that I'm worrying about right now as he is just 6. I'm worrying about people born in the 90s and 2000s.
I'm worrying about the economy and job perspective RIGHT NOW.
I've been following labor markets for the last couple of years, because I'm interested in the impacts of AI, and because I develop tech for executive search. There are all kinds of data, and some signals are negative, and other are positive. But on the whole, my balance is that things are "pretty good" for most cognitive work worldwide. I'm not in a position to make forecasts past the next 6 months, but right now there doesn't seem to be a jobcopalypse brewing.
*lypse has a nasty thing that it's too late when you see it, like Apocalypse. I'm sure the average number probably looks fine, though, otherwise there should be outcrys everywhere already.
Even in the Great Depression, the worst year, 75% of the employees still retains their jobs. I'm sure it was fine.
What's your Son planning to do? Listen to your stories until he is 400 years old?
And, thus at some point with 250K+ (and, below) page proofs, we ought to realize we are hitting an edge where we begin to be materially incapable of verifying anything: This itself is a sign that the complexity of the problems being solved - the scientific and methodical aspects of managing/engaging with that complexity - are beyond us.-
And, yet, at some point, the theory coming out of such efforts will transfer to material, engineering and design deliverables, and ... work.-
And then we will cease to undertand our technologies themselves. Our machines, our devices. Built on an understanding of the world we cannot reach, by machines we can build ourselves no longer.-
Ceasing understanding of technology isn't unprecedented really. I have a degree in computer science and even I don't _really_ understand how my phone works. I can diagram out the basic architecture and networking concepts but it's scratching the surface of how my iPhone actually works, which cumulatively is as much complexity as these future proofs you describe.
I suppose it's a little different if _nobody_ understands the concept besides an LLM, but that's an easier pill to swallow. Then it's just the alignment problem.
Do you know when you sometimes think about space and time and get overwhelmed with the magnitude of it? This is what I felt when I read your comment.
It's more or less, internally, my feeling exactly. I am honoured it transfered, and appreciative of your having taken the time.-
Nearly a year ago I asked ChatGPT to show me a proof of the Goldbach Conjecture, just to see what it did. Of course it said no such proof exists. But after a couple turns of pressing, it gave me two or three proposed outlines for a proof, with a several-pages-long research program for how you could attempt to solve the gaps still needed. I couldn't understand any of it, but it was pretty impressive. At that level it probably wasn't nonsense: I bet it was half-copied from some grad student blogs. If you asked it to dig into the details and keep updating the status of the outline, you could probably make a lot progress.
I think AI will become more autonomous as it becomes able (or just willing) to think more abstraction levels deep. Each level can be expressed in language, so it should be able to engage with it. It reminds me of a chess program being able to think many moves deep. For a while grandmasters like Kasparov could still see big-picture things the computer couldn't (e.g. with a locked-up pawn formation and an eventual piece sacrifice to promote a queen, where the kings can run in circles forever). AI seems like that now.
I know this thread will be full of people in denial about full recursive self-improvement / intelligence explosion / superintelligence (the original meaning of the word) within months or a year, so just quoting Scott Aaronson:
"Experience has shown that, even now, there will still be people explaining in patronizing tones why none of this is real and none of it counts. If such people were capable of being impressed by anything that happens in the empirical world, of updating on anything, they would’ve already been impressed and already updated several years ago, long before things had reached the point of an actual Mathocalypse."
My thoughts remain the same since I've been hearing "months or a year" for the past 5 years. Maybe this time truly is different, but unless acceptance of super-intelligence is required to take some action (most people I've seen who hold this view also don't think it should be seriously regulated so I'm not sure what that would be), why not just wait until it happens rather than making more predictions? The only thing I've ever seen come out of these predictions is willing me to "stop resisting", but there's no reason to do that until its here.
If you do want to see something done about it and this isn't just another prediction or a vague warning, more power to you, but otherwise, I'll wait and see
> My thoughts remain the same since I've been hearing "months or a year" for the past 5 years.
No one said that. If you look at e.g. the Metaculus predictions, you can see AGI has been hovering around 2027-2030 for 5 years.
https://www.metaculus.com/questions/3479/date-weakly-general...
https://www.metaculus.com/questions/5121/date-of-artificial-...
There are certainly a lot of other people making all kinds of predictions. Some of them will eventually be correct but which one?
Additionally even with all the breakthroughs I am still doubting RSI. Up till now all the advancements heavily rely on RL over Lean. But how do model formalize it's goal as in RSI and RL over it?
No one said [AGI in months or years, five years ago] that. If you look at e.g. the Metaculus predictions, you can see AGI has been hovering around 2027-2030 for 5 years.
That statement just contracts itself - if the average prediction has been 2027-30, someone, a significant percentage of someones, said earlier since many have said later.
Will OPs prediction be counted on Metaculus? If not, it's clearly not an accurate source for "No one said that".
2027-2030 is months to years away.
It is, but if OPs prediction isn't counted on there, it's not counting every prediction made by any individual, probably just the major ones that get covered, and so therefore "No one said that" would become "No one significant said that", which I'd imagine it probably true but, unless I'm mistaken, OPs prediction would sit in the same pool as all those other ones, they hold equal validity.
If you zoom out, you will see that the median for that estimate a few years ago was two thousand fifty. While certainly some people might have thought it happens the next day, the average person thought it's many years away.
You're mis-remembering things. Nobody in 2021 claimed that superintelligence is likely to arrive in 2022. That just did not happen.
People were talking about superintelligence in 2021, but "by the end of the decade", not "next year".
Not splitting hairs at all, but I recall a lot of mid-2030s estimates around that time period. The horizon has drawn in a lot in a few years.
Yeah, I remember that last decade the prediction was the 2040s, but then LLMs happened and it became 2030s.
Mid-2030s estimates aren't uncommon even this year.
And then there's this gem: https://www.dwarkesh.com/p/ege-tamay "AGI is still 30 years away — Ege Erdil & Tamay Besiroglu"
I was _hoping that maybe_ 2050 and other software engineers were considering me a crazy. GPT-2 was out by then.
As I said to someone else, this wasn't by major figures, this was in conversation with ordinary people. As I said to them, if you want to dismiss this you'd have to dismiss OPs prediction too, and probably cease making your own unless you can get a news-outlet to cover it
Why do you believe that "OP" is making a prediction on his own?
There are some well-known experts saying this, like Paul Christiano, Daniel Kokotajlo, etc. Both are ex-OpenAI and are advocating pause/slowdown.
That kind of feels like moving the goalposts. Nobody becomes Nobody of significance becomes Nobody of significance without others of significance making similar claims, all 3 of these falling under the banner of "Nobody".
But at this point its a matter of semantics, and wasn't really the point I was trying to make. And as I said to someone else, if the reason for making these predictions is to encourage a slowdown I'm much more sympathetic to them (since there a real reason to act of them)
> My thoughts remain the same since I've been hearing "months or a year" for the past 5 years.
If you zoom into any curve it tends to look linear. I think that is what you keep doing. But if you look at the world as how it was 5 years ago compared to now, there is no denying there is a massive change.
> But if you look at the world as how it was 5 years ago compared to now, there is no denying there is a massive change.
Is there? Yes, LLMs are integrated into a bunch of software, but practically speaking life is pretty much the same as it was five years ago.
> LLMs are integrated into a bunch of software
I should have restricted it to the fields that have been impacted and have seen significant adoption: art for commercial purposes, writing, software development and mathematics primarily, and those adjacent to those fields. It is true that most other fields have not really felt that many effects.
AI videos everywhere, software developers no longer writing any code and it being much harder to find a job, and people largely no longer using search and instead switched to chatbots is “same as it was in covid”?
I'm not saying LLMs aren't an absolutely impressive technology, especially for software related tasks, but in practical terms, life remains the same as it was five years ago.
People are really really bad at seeing incremental change as it occurs around them.
It's like watching your kids grow. Your there everyday and you typically miss a lot, but then your cousin shows up for a holiday after not seeing your kid for 4 months and they are like "holy shit they've grown".
If you made a SOTA LLM enabled computer show up 10 years ago people would be fucking panicked at the crap we're messing with a decade later. But because we mess with it a day at a time we're like, eh, whatever.
Tell that to the SWEs getting laid off left and right all over Silicon Valley!
Once again though, is me accepting this a pretext for something? If not, why not just wait and see? Unless there's something that needs to be done before it's too late, it seems an odd choice to push for belief when there's no reason we can't just wait a while and have actual proof
There are many things we can do and many people are talking about it.
One such coalescion is: https://pauseai.info/
Absolutely, the thing is that much of the time when I hear these predictions it's not to encourage participation in the movement like this. Many people who I speak to about it are even against such movements, which is what leads me to question why I need to accept their predictions.
If someone were to come to me with something like this as their reasoning, I wouldn't feel the same way.
Heard. Then:
Exponential progress has been the repeated truth, mirroring predictions.
The result of exponential progress into 2027 is a complete de-mapping of what we know to be human, whatever that may look like exactly at this point is hard to say.
We need your help, and everyone's help, to put a pause on this such that we are able to ground and find a healthy way forward together for as many human beings as possible.
Please join us.
Gotta say I really appreciate this page, I'm happy to donate but providing tangible actions is a really solid step
https://pauseai.info/action
I'm still not sure I believe it, but in my view its a bit like global warming, the worst(or best) case scenario of it being wasted time and money is far outweighed by the inverse, so it only seems pragmatic to help
I don't know about others in the thread think there definitely needs to be something done before it's too late.
I understand that many people don't think it will happen, but I can't really understand the position that you think there is a real chance superintelligence happens, but aren't concerned about it and are happy to wait passively to find out.
If you care about humanity at all it should be concerning because superint (if real) could be heavily destructive to our society.
I agree and I'm happy to hear out this kind of thing, my "I'll wait and see" attitude comes when the expected response is that I accept that the AI has "won". Thats pretty vague because the expected response, if there even is one, is pretty vague, but it generally amounts to something like that.
> we can't just wait a while and have actual proof
I am not pushing you to do anything. I am not asking you to "believe" or anything. But we are, in my and many other's estimations, in the midst of what was called the "singularity" in AI communities, although it isn't a true mathematical singularity, it is just a strong acceleration.
> I am not pushing you to do anything. I am not asking you to "believe" or anything
Fantastic, I'll remain at my previous position then, I'll believe it when I see it.
(As a very off-topic sidenote, the term singularity never made much sense to me. I believe it was originally used to describe the sci-fi concept of a human gestalt consciousness, but has shifted to being about super-intelligence which doesn't seem to have much to do with "singular" at all. But, language changes I suppose)
> I believe it was originally used to describe the sci-fi concept of a human gestalt consciousness,
Huh? The term itself was introduced by Vernor Vinge to refer to the accelerated rate of change as a result of the introduction of super-intelligence. He said it would make things harder to predict in the future. It is part of this article:
https://en.wikipedia.org/wiki/Technological_singularity
It was never about "consciousness".
Huh, looks like you're right, though I would argue with "it was never about consciousness", but yeh it certainly wasn't originally. I think I got my meaning from a sci-fi short story were a unified collective called themselves "The Singularity", and I guess that was my first exposure to the term as a kid.
The biggest risk of the singularity is incoherence, and it's actually looking like what is happening now.
Like AI video/audio of high quality makes you have to question and doubt any media you see.
AI generated text has become the vast majority of material that you see showing up on the internet.
Things are moving from centralized pieces of software used by lots of people to vibe coded apps.
Reality tunnels and almost fully close and you can live in an existence that seems 'normal' but is a completely simulated experience with no interaction with other humans.
> If not, why not just wait and see?
Curious what you need to see. Some physical impact to your way of life? Are you not able to extrapolate the obvious?
So you think it's already here?
5 years ago, what concretely did you expect 5 years later? One comment from then https://news.ycombinator.com/item?id=29020401 falling way short of today, treated then as wild near-sci-fi.
Pretty much, a pragmatic attitude at its finest
Yet AI has changed the world incredibly fast.
Why try to predict the most important single event to happen in human history?
Surely you recognize how absurd it is you’ve gone to such lengths to make your priors palatable. Please Look Up.
> Why try to predict the most important single event to happen in human history?
You're free to try to do so, but why are you pushing it so heavily? To what end?
> Surely you recognize how absurd it is you’ve gone to such lengths to make your priors palatable
And there is the answer to my above question. I like to think it's not most people, even of those who are staunchly pro-AI, but to some previous doubts are at best a great shame (assuming prior is short for prior doubts) or at worst a strange kind of heresy (assuming prior isn't short for anything). Either one seems anti-intellectual though, and encourages just the kind of endless hype we've seen so many times before.
Who said "recursive self-improvement / intelligence explosion / superintelligence" would happen in "months or a year" before 2026?
I can't think of anyone except maybe Elon Musk.
It was in conversations just like this, so people you wouldn't know and I didn't know outside of that conversation. You might be tempted to dismiss those, but then you'd also have to dismiss OPs for the same reason.
Okay can you provide links to these HN comments?
There might be, but I'm curious about the volume (I suspect very low).
I think it was on Reddit more-so if memory serves. Maybe it was on HN too, but obviously it was years ago now, I don't have a strong memory of the specifics. I'm not going to go digging through years of comments though, if you don't want to believe me thats fine, it seems a bit absurd to think absolutely nobody believed that but it wasn't really my point to begin with
> it seems a bit absurd to think absolutely nobody believed that but it wasn't really my point to begin with
Okay. If your point isn't that people believed superintelligence was months to a year away years ago, what is your point?
All due respect, I wrote a paragraph and that was the first sentence.
My point was that I'd rather wait for proof than accept a prediction.
great demonstration of the mindset OP was talking about
I'm always shocked by the number of smart, technical people who seem to think that AI is going to crash and burn and go away. In 2025 AI got a gold at the IMO for the first time, a competition for high school kids. Now it's likely solved the biggest problem in math. There's no reason to expect progress to stop here, especially on domains highly amenable to RL.
> I'm always shocked by the number of smart, technical people who seem to think that AI is going to crash and burn and go away.
I suspect that these "smart, technical people" you speak of have more nuanced views on AI than you realize. Myself, for example, a self-styled smartypants:
Like the dot-com bubble, I suspect that certain parts of the current AI landscape are unsustainable, and I suspect/hope that those specific features of AI will crash and burn, while other useful bits remain.
AI usage like we see in this post here--AI that assists with discoveries in math, medicine, genetics, astronomy--I think they're great, and I hope they will continue to advance human knowledge for the foreseeable future. They assist experts and professionals rather than replacing them.
In contrast, generative AI ("creative" AI writing, illustrating, music, etc) is a pox on civilization. It is industrialized low-effort plagiarism that harms artists' livelihoods, discourages new art, makes everything bland, and pollutes the world in multiple senses of the word. I hope it dies from a combination of negative sentiments, and too much real-world cost once token subsidies go away.
Then there are the AIs that do things like voice reproduction. They have some arguable positive uses (e.g., assistive technologies), but many negative ones (e.g., phone scams, replacing human narrators with copies of their own voices). I think we're better off without these apart from niche accessibility uses. I think it should be legal peril to reproduce a voice without permission.
Apart from helping with some cinema CGI, deepfakes and deepnudes seem like a net negative for civilization.
Knowledge-based AI like Google's AI search summaries might just have built-in expirations. I think people will burn out on being vigilant about hallucinations. And when the bulk of their training data is made by AIs, it'll be hard to prevent model collapse. The cost of human intervention might make it too expensive for the wide use we see today.
Coding AIs are a grayer area. I think it's highly problematic that they are basically open source plagiarism machines, but I begrudgingly believe that society will decide that the upsides outweigh the downsides in the long term. But I choose not to use AI in my own coding projects, and respect others who do the same.
My instincts are very similar on all points.
To push back, or share the self-pushbacks I grapple with...
> In contrast, generative AI ("creative" AI writing, illustrating, music, etc) is a pox on civilization. It is industrialized low-effort plagiarism that harms artists' livelihoods, discourages new art, makes everything bland, and pollutes the world in multiple senses of the word. I hope it dies from a combination of negative sentiments, and too much real-world cost once token subsidies go away.
While this summarizes my feelings on the current versions (which, say, are roughly at or below the math olympiad stage in writing, video, etc), what happens when they actually become good? Meaning, an AI show that was actually as good as The Wire? To the point you couldn't tell the difference without being told?
I really want to say that even then it won't be the same, because the human provenance is essential to the art, but if I truly couldn't tell the difference...
> While this summarizes my feelings on the current versions (which, say, are roughly at or below the math olympiad stage in writing, video, etc), what happens when they actually become good? Meaning, an AI show that was actually as good as The Wire? To the point you couldn't tell the difference without being told?
But here we're dipping into science fiction again. This is not currently the case, and we seem a long way from it.
> and we seem a long way from it.
We're dipping into prediction. But we have the (seemingly relevant, at least) data point that we're in the same situation w.r.t. writing/art as we were w.r.t. to math a couple years ago...
To an extent, sure, but:
- There was progress on maths that led to this point, whereas it's been stalled on art for the same reasons since the beginning.
- Art doesn't lend itself as well to the 'add more agents/complexity' strategy; a 250,000 word limerick is a terrible limerick.
> what happens when they actually become good? Meaning, an AI show that was actually as good as The Wire? To the point you couldn't tell the difference without being told?
Perhaps this comment will make me look silly one day, but my guess is that we'll eventually get there, and a handful of AI-generated productions will truly shine. But "as good as" is an important phrase. Because current-generation AI is trained on existing art, it won't be pushing many envelopes, it'll mostly be remixing what we have, and that will grow stale.
After a few breakout hits with broad shared appeal, I think the novelty will wear off, and AI produced TV and cinema will be too resource-intensive to remain as a profitable going enterprise. There will likely be a short window of hyper-personalized media that will seem very exciting, until we grow weary of the isolation of it. For many people, half the fun of watching a thing is discussing it with other humans, and today's streaming media landscape already leaves many people unmoored from that shared experience.
There is also the question of intentionality. When AI creates that show as good as The Wire, is that because AI has something to say about the human condition? If not, what's the point? If so, should we be paying attention to what AI wants us to pay attention to?
So what if AI slop wasn't slop? If you care about the details and the values behind your work enough to make it actually good, you won't use gen AI or use it invisibly because it concedes control of the details.
If you believe that using the authenticity of your own experience in your art will become vanish from how people value art then yeah we're headed for a future where AI rules. I really hope that that isn't the case.
I like your framework of separating use cases, mostly to evaluate how it will change in those specific areas, but I think you can evaluate capabilities fairly simply: I like to separate AI that depends on giant data centers and huge power draw from the kind that doesn't (not always local)—basically expensive vs cheap.
The expensive stuff isn't sustainable. Not yet. It probably will be, who knows when. I suspect it will be eventually because there's good money in it.
Here are the claims I find highly suspect based on the above:
- Anyone making claims that require expensive AI and attaching it to a short-term timeframe
- Anyone saying highly capable AI is impossible in the long-term
- Anyone saying these specific companies will be the winners—there is absolutely no moat for these models
I personally think anyone claiming something like super-intelligence or something approaching what we might consider sentience even in the next 5 years is suspect as well. I would not take the bet myself, but I would also not bet against it. The main reason is power: there's literally only so much power we can generate, so much societal disruption due to creating the infrastructure we can absorb (not to mention the societal disruption due to the tech itself), so its capability to self-improve is going to be highly constrained.
So far, we see increases in capabilities also increase power usage, I believe linearly. So, we're not seeing increases in efficiency that will make this easier to absorb. There's a ceiling, it won't go up quickly, so I think actual sentient-like AI capabilities are quite farther out than promoters think. Not impossible! I think it's more like a decade+ of maturation of the software and hardware stack—which isn't really that long in the grand scheme of things but longer than most people are betting.
They may have more nuanced views for sure, but nuance is famously lacking from most online discussions unfortunately. I would love to see more of it surface!
> Like the dot-com bubble, I suspect that certain parts of the current AI landscape are unsustainable
It's possible to believe that _almost all_ AI companies will go bankrupt while still believing it is a massively important world changing technology.
My experience on teams is there are the high agency AI adopters. Management likes this group. Then there’s what management would characterize as those not adapting (or adapting enough). This is the group getting laid off / not advancing in the environment. And this group is particularly bitter about AI disrupting an otherwise stable career, with wishful thinking they’ll be able to go back.
I've seen the first group vibe up garbage and dump maintenance on the second group. It may be unfair to blame AI for that when management is ultimately responsible, but that's what a lot of people are doing. In their defense, management is being mislead by AI hype men who are saying AI is so intelligent it shouldn't be making dumb mistakes.
Agreed, its more complicated than GP describes. There is clearly lots of value, but there are also lots of poorly led organizations creating a mountain of problems for themselves. We are all still figuring out what AI really means just for software development, which is the most useful economic activity it can do by an enormous margin.
Human minds are still needed to understand many things related to developing correct software, especially in business domains. There is still no AI running a profitable business, because that is not a verifiable task that can be scaled in RL like math or coding. So extrapolating from a math proof that no one can understand to a economically useful agents taking over most jobs is quite suspicious.
What is changing now is RSI truly beginning to happen. Much of ML is a verifiable task that can be scaled, and a lot transfers from general coding as well. Leading researchers are predicting their own obsolescence as early as next year, and that is something I take seriously, though I cannot judge the likelihood myself.
As improvement accelerates, we may see generalization happen at a scale that means we don't need RLVR for every profession. We've been surprised by it before.
> because that is not a verifiable task that can be scaled in RL like math or coding
Is it not, or are the feedback loops just longer? I could imagine being able to build a pretty good simulator of business operation for various kinds of businesses. (Some decent ones already exist as games.) From there, if you can build a good enough simulation, I can't see why you couldn't RL it.
Of course it's easier said than done to generalize it to the real world, where the possibility space is large. But I don't necessarily see an impediment to getting there eventually.
The feedback loops are quarters and years, so no you cannot scale it like you can scale math or coding. This is the reason we are seeing the most progress in those domains. Even ML is harder to scale because test runs are expensive and consume significant calendar time, so research "taste" is still important.
A full simulation would be, but I suspect even a partial simulation of evolving a software project though multiple "checkpoints" over a quarter would improve model performance
"Running a business" includes lots of real-world interactions that would be hard to model. An AI trained on an accurate simulation would probably get torn to shreds by scammers in the first six months. The models that exist today have enormous blind spots.
Honestly, the second group kind of brought this upon themselves. You cannot survive in tech without having an open mind and moving fast.
I actually asked a guy who got an IMO gold what he thought about it, and he's happy to concede that AI works. He used to be my go-to guy for those kinds of problems when I came across them. Now you can actually get it to solve those things on your standard plan.
Level of intelligence doesn’t matter if it is not applied to truth searching and instead to reasoning backwards from personal wishes.
Isn't this post an example of the former?
> I'm always shocked by the number of smart, technical people who seem to think that AI is going to crash and burn and go away
That’s what I am responding to.
Define truth
42
Parts are already disappearing if not gone already. If not by lack of profit, it will be by liability, overuse and moderation. Filters are being put up, bans in place, services restricted.
Just the other day a model helped me root an old android tablet. Will that still work in 5 years? I don't really have great confidence.
Yes, as long as it’s legal for you to buy a GPU.
>I'm always shocked by the number of smart, technical people who seem to think that AI is going to crash and burn and go away.
The tech industry cultivated that reputation though. People were just acting rationally by being more skeptical of even larger claims than they were making before.
Because they're _desperate_ to keep their social standing. That's literally all there is to it. The same occurred during Enlightenment and the industrial revolution. A small upper echelon of society (the church or aristocrats respectively) were deemed worthy enough to possess knowledge and thus to belong to an upper strata. The same story can be seen pretty much in every civilization btw (Greeks/India/China/Japan etc). When the masses started gaining access to this knowledge there was _massive_ pushback, even armed revolutions in certain cases.
What's happening today is that anyone and everyone can have access to superintelligence. Anyone with access to an LLM effectively is, simultaneously, a world-class diagnostician, engineer, architect, musician, mathematician... you get the point. This will necessarily lead to a post-scarcity society, since everyone knows everything and will be capable of producing anything.
> This will necessarily lead to a post-scarcity society
I wish I shared this optimism, unfortunately that upper echelon of society is the one which currently controls the superintelligence. Everyone may have access to it, but they own it.
I would share your concern if there wasn't such a thriving ecosystem of open weight models. And they're usually about a 6 months to a year behind.
Still, as of now, while I use AI for everything I do (mostly Claude but also ChatGPT), I don't use any open weight models (probably more of a mental block at this point), but I do see them as a necessary counterweight to this potential power imbalance.
Where are you going to run these open models? Natural resources are controlled by this upper echelon too.
Most households own a car which even if you go used has a price tag of around $20k if you want something average. It wouldn't be much of a stretch to see a future where every household owns a $20k AI computer. Or two. If you are fancy you own a $100k AI computer.
However, since there is an ongoing war on general computing it wouldn't surprise me if instead people get locked into subscriptions.
Shit people be telling ChatGPT isn't protected by law. You get into any trouble, your chat history gonna be gone over with a fine tooth comb. Some of the masses gonna realize that and gonna want a sever rack at home instead.
I bought a 2000 Honda Accord V6 from a friend for $3k a long time ago. I drove it for several years, adding probably another 90,000mi to it. It didn't need much maintenance during that. It's still kicking from what I understand. An incredible car, I'm a little sad I parted with it.
I still think of the economic benefit I got from that car. It drove me to university classes. It drove me to work. It took me to see a total solar eclipse. It brought me to see distant family regularly. I went on a lot of road trips on it. For roughly $3k of capital costs.
I look at most expensive things and size them up as "how many Accords is this?"
Almost 7 Accords is a lot.
For now. Open weight models are effectively only lagging ~6 months behind the frontier.
> This will necessarily lead to a post-scarcity society, since everyone knows everything and will be capable of producing anything.
A post-scarcity society requires MATERIAL abundance in the hands of everyone.
All the material is controlled by ~100 mentally ill people who would rather kill us all than share.
I don't get why academics get the shaft every time. These labs are worth trillions of dollars while your average professor has to deal with 18 year old shitheads and gets paid 50k a year. Like what?
Anti-intellectualism. The median AI booster resents that there is a class of people that had to work to become proficient in something, and wants effort to become irrelevant so their lack of investment in it appears to no longer be a failing.
AI boosters love to frame the trillion dollar corporations as the real victims. War is peace, freedom is slavery, scraping all known data and refusing the share the fruits ("distillation attacks") is democratization, people who are losing their jobs to megacorporations with lots of compute power are the real sheltered elites who are mad because they're hypocrites.
I only hear this part of non-reasoning from people who have never managed to accomplish anything and simply want to blame AI for their personal failures.
That professor is free and highly encouraged to navigate their web browser to Anthropic’s careers page.
I don't get what you are trying to say. We will be all working for Anthropic?
How would an LLM make me a world-class violin player? If scarcity is from needing real "stuff" (incl. time), then knowledge alone might not change that scarcity.
Same way as a human would? By watching you (upload video) and listening to you (upload audio) and talking to you about your technique (text to speech). It's obviously no substitute for the real thing, but for $20/month, it's a lot more accessible to more people than a human instructor would be.
Can anyone at any point in their life be turned into a world class violinist?
Malcolm Gladwell popularized the notion that 10,000 hours of focused practice gets anyone to an expert level of something. I've not got the drive to be a violinist, but in my world view, yes. Excluding the infirm, if we focused all of society into turning a randomly chosen person into a world class violinist, including mind control such that they want real bad, then yes. Violin playing isn't like the castrati. You need arms, and practice. The vocals of the castrati are something that could not be replicated in anyone.
The world of violin playing isn't that large though, so the question is merely getting someone to be really really really good but not quite world class close enough, or not really?
Using an llm would focus all of society on a person?
Why do these goalposts seem so shifty?
How is that shifting from the initial claim I replied to? Or how an llm would make me a world class violin player? It's not about becoming a reasonably good player or using some huge resources beyond an llm.
We don’t have access to knowledge, we have access to information. Knowledge requires processing information and that just doesn’t happen any more.
A post-scarcity society is not possible without first wiping out society. If everyone has everything then everyone has nothing. And of course, natural resources are actually limited, and those power LLMs.
I'm sick and tired of seeing people go "AI skeptics are just afraid of not being important". That's complete bullshit. We are skeptical because we can see, with our eyes, that the claims of AI boosters continue to be false time and time again. The simple empirical fact is that LLMs can't actually do a good job at the things people claim they are good at, and the industry lies through its teeth about said capabilities. So we are skeptical.
Yep, argument by psychoanalysis is a convenient way to disregard any stated disagreements as really just rationalization by weak/scared/lazy/depressed/etc people.
Wild thread to be making an “LLMs can’t do what’s claimed!” last stand…
AI still can't do a lot of what's claimed. This thread is an example of it being able to do one admittedly-impressive thing, but it's still nowhere near living up to all the promises.
> Anyone with access to an LLM effectively is, simultaneously, a world-class diagnostician, engineer, architect, musician, mathematician... you get the point. This will necessarily lead to a post-scarcity society, since everyone knows everything and will be capable of producing anything.
It’s hard to take you AI-hype men seriously when you are this ravingly delusional.
Where are these smart and technical people who said AI was going to "go away"? I honestly haven't met a single one this whole time.
> There's no reason to expect progress to stop here
Well, the amount of money and resources it took to get here is a reason. I don't think the resources exist for another order of magnitude increase, given the strain on our economy and society we're already under due to AI.
The frontier models are getting cheaper as they are getting more powerful [0]. It's really not going to crash. I wish it would, for the sake of humanity.
[0] https://epoch.ai/publications/the-plunging-price-of-thought
As does every technology, but websites are not a few million times faster than they were when they were invented, and your computer does not boot a few billion times faster (quite the opposite...).
Reality is extremely chaotic.
I was thinking more about the cost in time, power, hardware, and data to train the models, not the price charged to use the models. I don't think it's going to crash, either—but that doesn't mean that progress will increase forever.
We are well past the point where massive changes are guaranteed even if LLMs immediately and permanently hit a wall in capabilities.
"If you are smart, why can't you predict the future?"
It's not an ability to predict the future, but the reluctance to step back and look at the trendline rather than saying "But it still can't do X, Y or Z!".
These are extremely jaw dropping results, but I feel even though it is easy to conflate them with RSI / intelligence explosion, another simpler explanation is that symbolic manipulation is something humans are notoriously bad at, hence we prizing individuals that can rise above and dedicate their lives to it, and call them intelligent.
For computers though, symbolic manipulation is something they are extremely good at, and something that clearly scales well will memory and compute. Therefore I would argue that calling a computer superintelligent would, in analogy to do what we do with humans, require computers to rise above things that are extremely difficult for them. It is of course hard to judge what would be hard for them not being them for the introspection, but it seems the ability to operate outside the realm of symbolic or numeric manipulation would be precisely that.
It stands to reason that if some proofs require chaining 250k pages together, no human would ever be able to put it together. That doesn't mean something that is capable of doing so is necessarily super-intelligent.
Still very impressive, and if true, a wonderful result of the symbiotic interplay between a human and a machine, each at what they are good at.
Agreed. The fact is, Claude didn't get here on its own. It took someone who has been working with the RH since they were a student in college; some 40 odd years of experience.
It's unfortunate we're always focused on humans getting usurped by an "AI", instead of the reality that these models only succeed when we're there to show them what success looks like.
What if "superintelligence" is really a hybrid result of humans using and guiding these large pattern interpolators?
What if "anthrocentricity" is really humans being unwilling to acknowledge that they are made of the same vibrating condensed energy as everything else, such that other matter can just as easily be moved into an apparently self-determining pattern of behavior as we are.
What if we're not the specialist snowflake?
I’m on board with this, and have been since I first had a hallucinogen.
But the human brain for now still has 100x the parameter count as the largest models. And they don’t have any circuitry for all the “keep yourself alive” work our brains are doing all the time, which may be the majority.
how profound of you
> It's unfortunate we're always focused on humans getting usurped by an "AI", instead of the reality that these models only succeed when we're there to show them what success looks like.
As with chess, the reign of the human-machine "centaur" as the greatest mathematician in the world will likely be brief. I'd guess within 5 years the humans won't be adding anything. I hope I'm wrong.
This is incorrect because you can always record the human prompts guiding the model and train it on them. This is how alot of progress in software was done and the success there shows that this generalizes beyond the specific prompts. You could probably prompt Fable 7 "Solve <famous hard math problem>" and it could do so without an expert mathematician guiding it.
What if "superintelligence" is really a hybrid result of humans using and guiding these large pattern interpolators?
Obviously, humans with AIs are clearly more capable at a number of things than humans without but I don't think you can get out of the quandary of superintelligence that way.
Humans in the loop will always offer the opportunity of training and putting humans out of the loop.
And LLMs, if nothing else, have proven to be much faster than humans in the tasks they do master.
Can you point to a task which you think would be an indicator of RSI or an intelligence explosion?
I'm not sure I really buy into the way these concepts are currently defined. What progress to date was not RSI and did not build on top of earlier progress?
The main characteristic of our physiology is that of adaption, and we have fairly grounded research on progressive overload as a means to guide adaption to whichever direction we want. Should we then call humans RSI machines if they go work out or read a book, or study to get better at the activities they carry out?
Define an objective, choose an optimization algorithm, bring a loss down. If it keeps going down, do I call that RSI because the loss descended from the step that came before it?
Now, I see the nuance that the optimization algorithm changes over time, leading to interesting higher order effects the likes of which Hoefstadter describes in the GEB, but fundamentally, the idea is the same.
I also don't see what an intelligence explosion is supposed to mean. If we increase the models efficiency by 10x, and reduce costs by 10x, so we have 100x more software and math proofs being shipped by unit of time in say 2028, does that really solve all of our problems?
Say that gets distributed over the world. Do we get every human to be computer literate? Say that gets concentrated. Do we solve problems no one has the means to even care about?
I would consider an intelligence explosion when the models tell us whether our efforts are well or misdirected as looking back from the hindsight of 10 years ahead, and capable of explaining why in simple English.
If you ask a layman, the answer would probably be a model that zero-shots any problem in any reasonable domain. E.g. zero shot Riemann hypothesis. No tricks like swarms of agents, no multiple agents correcting each other. No high, xhigh "thinking" levels. Simply provide an input, feed the input through the layers, answer out.
Yeah, you can achieve something similar if you scale the compute infinitely and intelligently search the exponentially large search space.
> No tricks like swarms of agents, no multiple agents correcting each other. No high, xhigh "thinking" levels.
Why would implementation matter? The thought experiment is that you have a black box that can do X. What is X that validates RSI/super-intelligence/etc? It's meant to clarify definitions and end the cycle of goal post moving.
It is impressive because of how large the search space is not entirely because of the level of abstract reasoning required to solve it. The reasoning trace produced that lead to this result could be a part of training that teaches the next version of Claude how to navigate incredibly hard problems making it more efficient.
People suggested that language models are incapable of intelligence because they don't do symbolic manipulation. I think it's fair to call a vector with ten thousand numbers some kind of symbol, but then what's the state in your brain?
Yeah, I think it's not so much "math is symbolic manipulation so computers and by extension LLMs are going to be good at it". LLMs are not by construction so good at it because ML models are far more statistical approximators than formal logic systems. I think the main thing that math has going for it is sufficient formalisation through tools like Lean that it's very verifiable, and the latests hotness in improving models is RLVR which seems to really work very well when the verification is strong and relevant to the task (probably also why they're getting very good at writing code and especially finding exploits).
(I would also say that computers are not good at math in the sense that LLMs are proving to be good at it: reliable symbolic manipulation is important for correctness but it's no use for actual progress in mathematics without some very good informal thinking/creativity as well)
The reliable symbolic manipulation done here is in the form of Lean proofs, an external tool used by the LLM. One has to wonder how far an LLM alone would have gotten if formal proof systems didn’t exist they can verify their reasoning with.
When Atari games or Go were solved around 10 years ago, nobody was saying that superintelligence was within months just because reinforcement learning worked on a domain that allowed learning by self-play. Now we have a domain that is solvable by reinforcement learning with verifiable rewards. It's very impressive that these models are capable of doing the things they do, but the techniques aren't transferable to the vast majority of other domains. ML research also isn't verifiable in the sense that math is. People who claim it's verifiable, because there's a target that can directly be optimized don't understand why RL works for math. It's that lean is able to massively shrink the search space. This simply cannot be done in other domains in the same way.
Yeah, a lot of the RSI claims seem dependent on us having very specific knowledge about how much intelligence is required to build superintelligence that I haven't seen anyone claim to have. Maybe you have to be, I don't know, 100x smarter than humans before the recursive part of RSI becomes tractable. Or maybe there are fundamental limits you can do with the hardware that exists that we don't know about, which means building superintelligence will take more time. I'm not saying the 2027-2030 people are wrong, I just think it depends on a bunch of things we can't credibly claim to know at this point.
I'm not a fan of Dwarkesh, but even he laid out that machine learning research progress is very much verifiable. At least in the RSI sense, that our current approach can be drastically optimized.
I won't trust that source at all, it's all about hype and trust me bro considerations
The models are improving at certain fields, and one of those fields is creating new models, which are verified by applying them to fields where you can prove that they're better.
I agree they don't transfer to all domains, but we do not know how far the theoretical will overlap with practical science. It could be "not all that much" and we hit roadblocks quite soon, or it could be that as the models become better and better at reasoning and research within their theoretical domains, it turns out to translate really well to the real world.
We don't know, but my bet is that we will continue to see rapid progress and impacts on the real world, until we reach a point where sceptics will have to throw in the towel.
LLMs don't produce Lean proofs directly, they produce the proofs in natural language, and only translate them afterward. The fact that we have Lean has nothing to do with LLMs being good at theorem proving.
(I think we mostly agree.) From a previos post about Navier-Stokes, the English version may have small holes and the AI make a few changes and fixups when translating to Lean.
This is simply not true that they only invoke lean at the end. They invoke it multiple times in between to reduce the search space like GP said. See how e.g the prove2me harness (which anthropic used for FLT) works for more details.
What kind of a thing do you think AI won't be able to do in 2030? A specific thing please.
Love.
Throw a football at a wide receiver running in chaotic patterns while avoiding oncoming attackers.
Use chopsticks to pick up a single grain of rice.
Olympic level gymnastics via neural-control of muscle fibers in an extremely precise signaling orchestration.
You may say the above are robotics and not “the model can scale to this” domain but I’d counter an animal can do amazing things exactly because of its generalizable and plastic neural architecture and that is what’s in scope.
Have you seen the progress in robotics lately? They aren't yet at a level you described, but quite close - so I'd say it would be possible in 3 years even without intelligence explosion. (E.g. if all frontier AI research stops we'll still have improvements in robotics.)
There are already Unitree robots which are much closer to Olympic level than an ordinary dude: https://www.youtube.com/watch?v=Y3zb4o47yqA You might argue but it's kinda at a level you described - I think if Unitree wanted to make a football-thrower they could just do it now.
Shoot I can’t do those things either!
I’ll grant you love, and I’ll also grant you 2030 on these, especially the last one. I suspect the football and chopsticks ones are very achievable by 2030 if anyone cared enough to focus on them.
Fully autonomously run a business that generates $10+ million in ARR.
What about one that generates $1+ million ARR?
I find that unlikely as well. I should note though, that I incorrectly wrote "running a business", when I meant to say "scaling to $10+ million in ARR". I don't think there's necessarily that much difference in running a business of these two sizes, but scaling to $10m ARR is much harder than to $1m ARR.
Sure.
If $1MM is also unlikely, how about $100k? Do you think that's unlikely as well?
The point is: what is the smallest amount that, if surpassed, would surprise you?
It would surprise me if a fully autonomous AI company could even stay afloat over multiple months. Agents are getting (a lot) better at task completion over longer time horizons, but still occasionally tend to make really bad mistakes. What are the time horizons right now? Something like 40 hours with a 50% success rate? This is on benchmark tasks! The real world is completely different, I mean companies are scaling back efforts to even let agents handle things such as customer service. A fully autonomous business without any human oversight is on a completely different level to that. 2030 is not that far away, and I just don't buy that these models are exponentially improving in anything but a small subset of domains.
Either AI won't be able to do this: - Make a version of itself 100x more efficient that runs on commodity hardware and undercuts athropic.
Or, alternatively, it will be able to do that. Either way there's no world where one company bottlenecks and holds the keys to AGI, and frankly that's good.
- Keep electricity costs down
- Stop lying
- Design personal quantum computers
- Save fanatics from the delusion that AI will solve their life
I mean something a human can do, but AI can't.
Lay brick? Install pipe? Wire cable? Plaster wall?
Here's a brick laying robot from 2017...
https://www.youtube.com/watch?v=6s17IAj-XpU
so yea....
Try telling it to stop or move to a different wall in natural language. I mean, you might as well call a train AI with definitions this generous.
Show me the car in which the robot drove itself there
Self-reproducing physical agents (robots with onboard AI and ability to change their surroundings/build physical things) with human or higher intelligence which don't need a global supply chain of microchips, electro motors, etc. but could live off the land. That's basically the human concept, right? Totally untenable for AI/robotics of the current generation. Would require something like 3D-printing new chips but also 3D-printing new printers. Probably never happening. AI is fundamentally different, reliant on data centers, which are immovable and more like organs than organisms. Robots only come out of factories, not out of other robots or eggs layed by robots. So as Jack of all trades in the game of survival on Earth, humans win for the foreseeable future. Also, AI currently can't exist without humans, whereas humans have thrived without AI. Whether a silicon based ecosystem can ever survive on its own is an interesting question in and of itself. Because no matter how smart data centers become, they degrade fast, and then what? A data center needs maintenance. But if robots maintain the data center, who will maintain the robots? Does this equation ever solve? Humans have bacteria, a complex microcosm of life maintaining them. I don't think such a dependency-hierarchy of "life forms" is feasible for AI/robots, and humans will always have to be in the loop. That's one answer to your question, and not just for the next 3 years.
What stands axiomatically are human-to-human connections that we're wired to crave. Not to mention human reproduction.
Other points: - Editing a live action movie that will get 8.0 or more on imdb (AI sucks at reading emotions and won't be able to tell apart good takes from bad ones) - Reliably making good jokes (many humans can do it, AI weirdly sucks at it) - Accountability (contradictory because of its undead nature) - Reliably not lying/cheating (with no accountability, incentives will always be misaligned)
The nature of mathematical proofs requires us to stay vigilant and careful. OpenAI pulled some of the research results from their famous repo due to simple mistakes that where easily spotted by human experts. There have been infamous mass of incorrect Fermi's theorem proofs.
LLM-assisted research is proving to be useful, rewarding and joyful, but we should not treat LLMs as a perfect machine, incapable of making mistakes. Especially in such massive proofs as this one.
I won't say that it "doesn't count". I will simply ask for staying rational.
Open AI mixed some Lean verified proof with English-only proof. It's important to read the fine print of the anouncements.
What's so surprising about this? A lot of these problems can only be solved using a significant amount of capital. At home or work, using a few-hundred dollar token budget, it's mostly a slog and requires handholding for the majority of open-ended problems.
This technology won't be available to the broad public in the scale that it's available to a select group. If it can only be deployed to a limited amount, to most it remains a magical inaccessible box. And that majority should naturally be incentivized to look for cheaper alternatives, and to "keep moving the goalposts", or in simpler terms, looking where they can provide value, of which there is plenty!
> A lot of these problems can only be solved using a significant amount of capital. At home or work, using a few-hundred dollar token budget, it's mostly a slog and requires handholding for the majority of open-ended problems.
Consider just how much more inference capability one can buy per dollar spent today as compared to a year ago. I can get tasks done quickly for essentially no cost that, early last year, might have cost a frontier lab hundreds or thousands of dollars and would have been unavailable to the general public.
In my mind, the bigger unknown is whether actual profitable use cases will show up quickly enough to justify all the (enormous) investment made in the technology.
By now, many open math problems have been solved with a cheap subscription. Not of this magnitude, of course.
> I know this thread will be full of people in denial about [my prediction]
You can only be in denial about the past or the present. The future has a chance of being different from your predictions.
Personally, regardless of whether you want to call it "denial" or something else, I'm fervently hoping the AI never achieves superintelligence, but I would settle for it not achieving it until we can do our best to ensure that the superintelligence will be benevolent towards us.
Call me selfish, call me petty, call me a human supremacist, call me anything you want, but I prefer humans being the pinnacle of evolution. I like being able to point to our civilization, flawed as it is, and say "No other species we know of has achieved the same."
I'm one of those people who's skeptical of recursive self-improvement and superintelligence. My rationale for my skepticism is that the algorithms get too much credit for the results. The algorithms ultimately run on hardware built by humans and learn from data generated by humans. In the same vein, I think knowledge work gets too much credit for building things and generating high quality data. The output of knowledge work is only useful if it can be applied to the physical world e.g. a CAD file is useful when the actual thing it describes is made in the real world. Similarly, high quality data comes from several feedback cycles with the physical world. I think people generally underestimate how much of human capability is driven by our sensorimotor skills. I think the idea of recursive self-improvement / superintelligence is just an extension of this attitude.
Humanity as collective does not buy anything, or send the design to the plant to be made. As a collective, humanity transforms the natural world into things, for example, people take fibres from plants, turn them into yarn, weave yarn into fabric, and then turn fabric into clothes. Similarly, humanity as collective does not "buy" chips, people turn quartz, bauxite, copper ore, petroleum, etc into the components of computer hardware.
There's a common adage about how everyone should work as a service person to learn patience and empathy. I think everyone should also work in an assembly line and a mine to learn about where the things they use come from.
I don’t understand the technical basis for belief in continuous recursive self-improvement. Frontier models are increasingly impressive at complex math and coding, but the math and code underlying the transformer architecture are not complex. Major advancements in capability have come from scale: more data and more processing power. How does solving the Riemann Hypothesis (or similar achievements) further increase those? Aside from supplying fodder for investment-seeking hype.
Below is just a hand-wavy argument.
Neural networks are function approximators. They're great because they can be trained to approximate almost any function.
What if, somewhere out there in platonic space, there's a complicated function that represents what a philosopher might call true intelligence. Perhaps it's structured totally differently than a neural network and shares no common architecture with a transformer.
Transformer-based LLMs are likely completely different in their fundamental structure. But perhaps the neural networks we have today are succeeding in intellectual tasks because they are beginning to approximate it. Like a bitmap approximates a vector image, it's not the same; the neural network is merely made of a bunch of repetitive simple elements, tuned to imitate the real thing. A higher resolution bitmap image will be a better approximation but still never become a vector image no matter the pixel count.
But at some point, the neural network might be good enough at intellectual tasks that when tasked with searching function space for an algorithm that does the job its neural network does but better, it finds the real thing it's been approximating all along.
There are many big technical question-marks on if RSI is even possible. However, the advances in LLMs that we've seen is definitely not just limited to data and scale.
First: modern LLM architectures are quite far from the GPT-2 era (delta/linear attention, GQA, sparse attention methods, looped transformers, mixture of experts, engrams, etc).
Second: RL environments are largely autonomously generated at this point, as models get smarter they can generate even more complex environments to train in. While this does go into the "data" category, it's not human-generated or even gathered data. This gets particularly spooky when you think about things we cannot yet simulate, that models can develop mathematical methods & simulators for. For example: better/faster physics simulations for robotics RL, better biological simulation environments etc.
My understanding is that LLMs can be used to make small improvements on top the transformer architecture, to make LLMs that are even more intelligent and so are more capable of making more improvements. At some point we might reach the practical limit of intelligence using transformers, and if the LLMs of that time aren't able to make the leap to discover a new architecture, then the improvement stalls. I don't think it will be possible
I don't think you need to be in the football team mindset of us vs them, we're all seeing the progress but most of it is still built on top of what humans have put together over decades and not completely new inventions, which is probably what would convince most that we've achieved "superintelligence" or whatever you want to call it.
LLMs are jagged in their capabilities. Opus 4.8 was already an excellent programmer, but it played Pokémon Red far worse than a human child, even if it did, eventually, manage to beat it.
I note that the dazzling events that have happened have all been in coding, cybersec, and math, that is, domains where the correctness of the work can be verified entirely inside the computer. It is unclear whether further progress along these lines leads to crazy scientific breakthroughs too, which is something people are taking as a given when it comes to superintelligence.
The pace of progress is incredible, but the end result is hardly unexpected. We're talking about a theoretical domain that's easily verifiable and has no side effects.
You have the input (hypothesis), the output (true/false) and the lean kernel. There exists some path from the input to one of the outputs, you "just" need to find it. You use AI to search this exponential space, occasionally prunning ideas that are dead ends, and throw as many agents as possible to scale it up.
In order to keep my mind safe from hype waves, I try to set my expectations periodically and see which of them happened in retrospect, updating my beliefs accordingly. At 2026-09-22 I made a personal 10 statements "prediction" to happen by 2027-03-22, which one of the statements was:
AI will solve exactly one or two millennium problems.
So I will become impressed by something, but only this is not enough. If AI solve three or more problems, I will become impressed and believe that some serious RSI stuff might be happening. If AI couldn't solve any problem (which it seems it did) I would become impressed in the other direction, and believe AI got stuck in a way that I didn't imagined (together with other 9 statements).
There may exist people (like me) who don't believe superintelligence is happening with current set of evidences, but believe it with some stronger evidences.
Whether AI with it's math achievements is extremely cool, whether it's intelligent in a philosophical sense and whether it's capable of RSI are 3 distinct topics.
These disagreements are natural and expected if most people are not properly defining their terms. What do you mean by artificial / super intelligence (explosion)?
I think it's a perfectly sane position to acknowledge amazing progress with very large neural networks, expect further progress, and also not buy into the specific mythology around the singularity or extropianism.
These things come bundled together quite often, and that's what I see people (reasonably, imo) resisting.
Doesn't mean there isn't unfair dismissal too, but I feel that's gotten much more rare.
> self-improvement / intelligence explosion / superintelligence
These terms mean different things to different people. What are the particular implications of this that you're suggesting are imminent?
I think this is all real and counts. Where I struggle is explaining to non-technical people how their everyday lives are going to change directly or indirectly. I hope these discoveries are the first in a sequence that will culminate in or surpass a cure for cancer and diabetes, cheaper and cleaner energy and water, and other tangible everyday life improvements that influence millions of people.
What is the action item exactly for non-technical people? How do you prepare for “things will be different in unpredictable ways at an unpredictable time”? Even if that time moves sooner, still feels like very little info to truly act on.
If you watch movies or read fiction, we’ve expected for years that life in the future will be very different.
Support building more datacenters, promote and integrate SI at your workplace and in your personal life as much as possible, stop resisting, stop criticizing Anthropic and OpenAI, purchase their subscriptions and stock, or be left behind to dig around in the dust
> SI
lmfao, buddy
How does one fail to see sarcasm this obvious?
Oh that’s easy, you scroll HN seconds after waking up and still in bed, that’s my bad hahah
Yeah honestly I don't blame you for it, I feel like it's at most half a step up from what you'd see normally
Fair enough lol, I just thought it might have even been too direct at first
Seen enough freaks on twitter use SI unironically that my eyes glazed over the instant I saw it here and my brain turned off
I think there's a key distinction missing in a lot of these techno-optimist takes.
Abundance of productive capacity does not automatically equal abundant access to its products.
If all of this happens in a way where a few people control these new, automated means of production, all of these wonderful things do not automatically follow.
Why would they share? And worse yet, if they have this automated production capacity without reliance on labor, why would they interact positively with society at all? What's their incentive? And no, you can't say markets, because once you have a machine that builds you whatever you want, you don't need to trade anymore, money and markets become meaningless.
You're basically reduced to hoping for charity from a group of people that have repeatedly proven their lack thereof.
A cure for cancer and disease will only be available to the oligarchy, we the peasants will be killed off since we're no longer needed.
This is conspiracy theory level thinking. "The oligarchy" consists of humans just like you, who also want cancer and disease to be cured for everyone just as much as you do.
Human history is full examples where the significantly more powerful and wealthy relish in stamping out the rodents
Yes, and that were the limits of "AI", "SuperIntellingence", etc become clear.
The world already produces enough resources to feed and cloth and house every single person. Many pf mankind's fundamental problems are no due to lack of technology.
Heck, why can't I have a proper OS, if AI is working wonders?
RL is not scalable to super intelligence. Super intelligence should be able to learn continuosly without verifiable rewards.
You don't need RL all the way to ASI. You just need RL far enough that the model becomes good enough at AI research that it discovers something better than RL.
> becomes good enough at AI research that it discovers something better than RL
That might require continuous learning without verifiable rewards. RL models might help humans join the dots faster, but we don't know how many years that will take.
If its not possible with neural nets as we know them now, and requires a number of completely novel breakthroughs, there could be far from exponential progress even with the best RL models.
Yep, I agree. Improving AI might get harder, faster than the AI gets better.
But it also might not, which is the scary part.
I cannot really comment on RSI, but I think it is important to make the point that an LLM doing large, even crucial parts in a proof of RH, it cannot be viewed in an anthropocentric light.
"If a human proved RH, they would clearly be extremely smart, so imagine how smart the LLM is!"
It is not surprising to me that LLMs show extreme aptitude for math while still having clear deficits in more real world applications. Handling stuff like super difficult math is precisely where AI has its greatest advantage over humans. Perfect recall, never tiring, never lose focus, attention to detail, etc.
If there is one area where AI can be pushed extremely hard ut is math. So the most amazing progress should be found there.
The terms "super intelligence" or "intelligence explosion" are imprecise to the point of being useless. We have "super speed" in the form of automobiles. We have "super strength" in the form of excavators or bull dozers. We have recursive self improvement of tools by having better automation of machining. Intelligence != human ability. The intelligence demonstrated in the proof of RH is breathtaking, but narrow and brittle. It doesn't generalize to having grocery shopping bots, or automated psychotherapists.
I am not impressed because I understand how it works. Not because what things it can or cannot do..
We can start talking when we can train these things without having to beat on its head a million times in a million different ways to teach it one little thing...
How do you think our species was trained?
our species was not “trained”
This is such a substanceless argument, it’s pure vibes. It just says “Anyone who doesn’t believe in imminent intelligence explosion is an annoying curmudgeon.”
There are very solid arguments about why excellence at proving mathematical theorems does not necessarily imply general intelligence of the sort that people associate with ASI.
We’re not arguing with you, we’re pointing to the giant fucking meteor above our heads and laughing/lamenting that you still refuse to see it. It’s not an exchange between balanced discursive peers, not really. It’s an astonished exchange between those of us who have Looked Up about those of us who have yet to, and maybe never will, not really.
I encourage you to try, tho. For one thing, we need good people like you with us in reality. For another, you might die in some horrific catastrophe before you get a chance to, if you wait much longer.
what does “look up” mean?
https://en.wikipedia.org/wiki/Don%27t_Look_Up
thanks!
I mean, you're barely arguing, you just sound a tad psychotic.
Being called psychotic by someone with less than one sentence of reading comprehension doesn't really sting like it should, sorry. Doesn't clear the general omnipresent sting threshold, at least.
If people kept applying the same progress expectations to cpu frequency since 20001 we'd have 400ghz cpus by now
It's OK to admit we just don't know when or where it'll srop
Also virtually no one argues llms will go away, but a lot of people argue the current companies might, it's completely different.
There are a ton of people who think LLMs are completely useless, the whole thing is some combination of scam and delusion, and it will all come crumbling down once the music stops.
There is also a ton of people who think llms are Gods and will cure every disease including aging
You need retards on both sides of the spectrum for a balanced society
> superintelligence within months or a year
Didn’t we already achieve super intelligence?
No, we achieved artificial general intelligence. The fact that you mixed up these two wildly distinct terms should be a sign that you should not trust your intuitions about those topic.
Yes, 10 times already. You're supposed to drink a shot of vodka each time you read the term in an article or comment.
I love hn
Both the deniers and claqueurs are from the same ilk: Silly adults who don't have the intellectual capacity to create or even fully understand the subject, yet possess an almost vulgar self-confidence to constantly opine about its future.
Good observation. Perhaps the noble path is to chop wood and carry water, after all.
A slight nitpick: the hacker news equivalent of chop wood, carry water would be something like “charge the car, write a CRUD app.” Ordinary, not mystical powers doing magic. not high faulting thoughts but the basic simplicity of living in each moment and letting the necessary actions happen freely.
Seems the CRUD app may not need writing soon, and the car may charge itself.
Neither of us know what the growth curve looks like, so focusing on just calling one side naive is ignoring the naivety of the other side as well. If you want to extrapolate, go ahead, but that doesn't make you some oracle, nor does it make what you're suggesting factual. There are many potential scenarios and we should consider them, sure.
But it's not clear how well superhuman ability in math generalizes to other domains.
You're listing poorly defined words that mean different things to different people and wonder why people disagree with you...? My guess is that a lot of people you perceive this "denial" from are simply realists who are tired of seeing blind hype.
Note that every technology exhibits extreme growth until it... Doesn't. You can't predict the future. Whether any kind of RSI will happen depends on how hard of a problem intelligence is, and the precise point is that no person on earth knows, otherwise we wouldn't need to dedicate an entire research field to it.
Anyone who is sure that X will happen or not happen within Y years is just lying to themselves. For example, if you're so sure RSI will happen, why not invest all the money you have that isn't strictly necessary for your survival into this? Surely it will pay off in Y few years...
And note that I'm not saying it won't happen! I'm saying it's stupid to make a prediction of either case.
Your exact argument applies in full to climate change. No one can prove that the physical processes we know to be underway today will continue.
Yes, we're kinda guessing there too and I would expect every scientist to have a reasonable view to account for these uncertainties. I'm not sure if you meant to disagree with me, because I agree with you ^^
I also expect climate change to still be a process that is easier to research than intelligence. With climate change you can at least kinda look and measure and apply general principles of geology and physics. With intelligence, we've been at it for ~60 years and don't even have a definition yet, don't know what causes it, have no instruments, next to no introspection capabilities that neurology provides, etc.
I fear this is the result of a massive disconnect based on what you were told to expect. I read Kurzweil when that first came out, and maybe I'm remembering wrong, but I remember the singularity being the rate of technological progress in all domains will be so fast as to be asymptotic to an infinite slope and anything that will ever be possible will happen more or less instantly. Scott Alexander told me in a comments argument on his blog ten years ago that encryption is pointless because an AI will turn Jupiter into computronium and break it. Zvi was bandying about with the "take over its light cone" and "erase all value from the universe" phrases.
If that's the standard, then sorry, I'm not impressed by proving the Riemann hypothesis. If the standard is rapidly improving software for generating other software (which may or may not itself be any better than existing software) and math proofs (that may or may not have any meaningful real-world impact), then sure, I'm impressed. It doesn't mean I expect the daily lived experience of more than a few hundred thousand people globally to change much in the next decade, and most of the ones whose lives will change are software developers and mathematicians.
On that note, I have to imagine you have something different in mind when you talk about the "original meaning" of superintelligence, as you seem to not be talking about the Kurzweil/Bostrom/Yukowsky type of being effectively indistinguishable from God or at least the ancient alien civilizations from the Space Odyssey or Expanse novels, which again, I'm not seeing much of a hint toward because we have very good automated theorem provers. Proving Riemann doesn't mean we can now figure out zero point energy and turn planets into stars.
people always under emphasize too that the gains of tomorrow aren’t made with the technology/tools of today (I think Kurzweil is where I heard this first). RSI, if landed, seems like that on steroids
> people in denial about superintelligence
As long as I get to keep my house...
I won't believe it until I get a robot to put my dishes into the dish washer, put copper meshes and squeeze spray foam under my siding panels.
No offense, but have your heard about the infinite monkey theorem?
> Experience has shown that, even now, there will still be people explaining in patronizing tones why [...]
This is true no matter which side of the debate you're on. See [1] for an example.
[1] https://news.ycombinator.com/item?id=50575797
I would say, seriously, that this is bounded superintelligence. That superintelligence needed top human talent to achieve the goal. It was not a prompt like: "solve the Riemann hypothesis".
Coincidentally, yesterday I did a small presentation on cybersecurity and AI where I made a similar point: scarce top talent in cybersecurity (e.g. Calif) gets superpowers with AI, while AI replaces a lot of the skills of average cybersecurity professionals using simple prompts.
I think the breaking point is when AI can solve complex problems on its own that even top talent can't achieve. IMHO, I've seen some of this in the latest Google models (especially Carbon, and then Argon), where they can do more research independently.
"Human-assisted superintelligence" (HASI). ;)
I think you can be extraordinarily impressed and believe (or expect) that there significant limits to this technology.
I mean, AI could make very impressive 20 second videos two years ago. Progress has been slow and incremental since; Ten minute video still look far away.
I'd acknowledge is making quick progress on things you can find massive and reliable data on but even in it's good mode it seems to have palpable weaknesses. Someone was claiming Claude has become more and more persuasive over time but all the generators "still sound like slop" to me and many people.
Still, I have hope.
It’s not too late to Look Up for you. Yes, I’m talking to you — the person reading this with a vague grimace of disbelief and discomfort and disdain. It’s not too late for you to swallow some pride and reassess. Everyone has to eventually, especially me; why not now, for you?
So please. I’m begging you. LOOK. UP.
I looked up, it said 'naive'
:/
Stochastic parrot truthers in shambles.
It’s a bait and switch to use a harvard fellow of mathematics solving a problem using AI with “The AI is super intelligent”. People are getting really tired of the constant bait and switches and subterfuge in press releases so unsurprisingly they just start rejecting it by default.
I don't see anybody claiming that solving the Riemann hypothesis would be proof that current models are superintelligent. Reminder that the definition of the term is "any intellect that greatly exceeds the cognitive performance of humans in virtually all domains of interest". I do sometimes see people arguing that current models are AGI, but I don't think even this is true - current models are slightly superhuman in some narrow domains like math, human-level at a bunch of them, and still pretty stupid in some ways, most notably agency.
The reason people worry about superintelligence is that one of the "domains of interest" is AI research, and once you have an AI that's as good at AI research as your human researchers, your AI progress shoots up, probably exponentially. So the delay between automated AI research and superintelligence might be tiny.
You can believe what you want I probably agree with your nuanced view. But I would invite you to control+F this thread for “superintelligence” there are plenty of aggrandizing statements about what fools everyone are for being dubious about this being superintelligence.
(e.g)[https://news.ycombinator.com/reply?id=50575797&goto=item%3Fi...]
People are just tired of listening to these statements and being called idiots constantly with no evidence other than other press releases and disingenuous framing getting echoed constantly in volume across corporate media and all social media that allows LLMs.
I don’t think like that, the most important question is: could he have solved the problem without AI assistance, and in the same time frame? I’d guess no.
A few years ago if someone had said “we used an HPC cluster and Pandas to build some tools that in turn solved this problem” no-one would’ve batted an eyelid and no-one would’ve been making grandiose statements about the end of the world due to HPC computers and the power of Dataframes. Personally I think If someone had had a few hundred billion dollars of investment pumped into a bunch of mathemeticians to solve this problem then yes they probably would have solved it too and maybe that might have raised some eyebrows. Everyone of these conversations just feels like an extremely highly charged argument about Vim vs VSCode.
We're really using a term invented by Trump yesterday for something we've been working on for 80 years?
The term was not invented by Trump. For example, the (in)famous book Superintelligence by Nick Bostrom was published in 2014[1]. It has a long been used to refer to a theoretical general AI that far surpasses humans.
The Trump thing is a insane newspeak attempt to forcibly replace the use of "artificial intelligence" with "super intelligence" in a general sense. It is ridiculous, obviously, but the best thing we can do is to continue using "superintelligence" in its original, narrower sense.
[1] https://en.wikipedia.org/wiki/Superintelligence:_Paths,_Dang...
Actually, I think the best thing we can do is take Trump round the back of the woodshed.
Superintelligence is an old term with an established meaning. What Trump does is to use it for AI in general.
Is there any hope of a mathematician actually understanding the proof? Millions lines of code and hundreds of thousands of pages for the proof seems unfathomably large to be actually comprehended by a human.
No, and this is why we're fucked. The benefit of having a proof completed is having access to the techniques which completed the proof. Without an understanding of those techniques we'll be applying the result blindly, without understanding of its consequences or context.
And I doubt the actual proof will be released for months-to-a-year at the very least. Until then we can only rely on Lean and AI's sloppy interpretation of it.
> we'll be applying the result blindly, without understanding of its consequences or context.
the AI will understand these and apply them accordingly. You don't have to.
I think so. My understanding is that the proof contains many re-proofs of existing, already known results. It may also contain multiple proofs of intermediate lemmas, as LLMs frequently implement things over again rather than reusing an existing implementation. It's very possible that only a small fraction of the proof contains new mathematics.
People claim to understand the OpenAI 150-page quasi-Riemann proof, which is apparently full of extraneous detail. Alpoge speculates that this new proof also simplifies considerably.
so this is 1,000 times longer, so hopefully it does simplify.
It sounds comical, but we can just wait for a smarter AI to make the proof more digestible.
maybe having a big proof allows for reduction into smaller proof? at least in code that often works
i think it'll be interesting to revisit the status of the openai navier-stokes proof a few months ago, which i think Alpöge also argues was actually partly referencing his work because he was using openai tool to turn it into a human-readable format. but I think one of the biggest hole still standing and being debated for that proof is that its making the assumption lean compiler and the intermediate definitions and lemma statements actually say what the informal claim says. curious on how this proof compares given that it at least staets GRH using the matlib defs
I have never heard enough of the Riemann hypothesis. Based on the comments here, it sounds like a big deal for this to be proven as it validates many dependent proofs.
It is an impressive accomplishment. Im also wondering if there are any material changes to any non theoretical work as a result of this.
Does this make any algorithms faster? Does manufacturing become better somehow? Can engineers use this to make better designs of anything?
I genuinely know nothing about the hypothesis beyond what is mentioned in the first paragraph of the paper, so I'm having a hard time knowing how this changes real world things. I already thought AI models were more intelligent and capable tools than anything else. This proof didn't change that.
A proof doesn't immediately change anything, because computer scientists have already been acting as if the Riemann Hypothesis (or its generalized form) were true. But if the proof holds up, its structure might contain hints that lead to new algorithms down the road.
If you'd like an explanation for laymen, the YouTube channel "Numberphile" has a great video on it [1].
[1] https://youtu.be/d6c6uIyieoo
To claim the Riemann hypothesis as the most central and important problem in mathematics would not be ridiculous.
Questions such as "Does this make any algorithms faster? Does manufacturing become better somehow? Can engineers use this to make better designs of anything?" are both ignorant and stupid (which are different things).
ok, but what does solving the most central and important problem in mathematics do for us? How does it change our lives? Is there a material difference in what things are capable as a result of this?
I understand this is a significant feat if the proof holds up. I don't understand what difference this makes on our lives or capabilities.
I appreciate you saying this. For one, progress in science is non-linear. Is every result considered useless if it doesn't have a direct practical impact on your daily life? Obviously, this is idiotic. Secondly, why would anyone know the impacts of this massive lean artifact that was released hours ago? Commentary on breaking news isn't a good place to learn about this.
I did not say it was useless or insignificant. I understand that break-through dont always have immediate practical usages. Some break-throughs do have an immediately understood utility however.
Remember a few years ago when there was the paper about some LK99 superconductor that worked at a wide range of temperatures? The comments at the time are filled with applications for such an invention.
It was the most famous unsolved problem in all of mathematics. Very insane development.
Ok but will it have implications for other things? How does it help humanity?
In very subtle ways. That digital lock is now just a little more secure, whereas before, we weren't sure about this one aspect of it and thought it could be broken by an attacker.
Got it. So this is impactful for cybersecurity because it confirms that there some set of exploits are not possible?
It confirms assumptions about cryptography that systems are built on. If RH was disproven, these systems would be flawed.
If you know how big a deal P = NP is, this is like that except for all of math. There are many things in math (in many different areas) that have proofs assuming RH is true.
The statement itself is relatively easy to understand, but it was generally assumed a proof would take centuries.
Does anyone know if the proof is constructive (perhaps telling us something new and usefull about the prime numbers) or is just establishing the fact the RH is true, but with no new insight (e.g. no new factoring algorithm falls out)?
And so begin the Dark Ages of mathematics, I suspect.
Many people here have an idealized understanding of math as something incredibly profound. But it was mostly something that nerds practiced for the joy of it, in a nearly zero-pressure setting, the results rarely applicable to real life.
After the dust settles, I suspect there will be a lot fewer people interested in a job that now boils down to prompting and then staring at 250,000 pages of LLM output. And there will also be a lot less incentive for anyone to spend tends of millions in compute to produce these proofs just for the sake of it. I suspect we'll end up doing less math, not more.
> lot fewer people interested in a job that now boils down to prompting and then staring at 250,000 pages of LLM output
What else will they do? This is basically what every job will become unless they all decide to be physical laborers or something. Even then I often debate myself on whether the brutal hours, low starting pay and physical toll are better are worse than the mind numbing boredom of prompting things into existence.
We're on a forum filled with software people, and I think there's a distinct possibility that we're hoping this technology isn't just coming for our jobs.
You think physical laborer jobs are safe from AI?
AI is accelerating humanoid robotics too, and those same humanoid robots will assemble more copies of themselves.
I agree, it won’t stop at maths. Dark ages for all intellectual activity. Hopefully we will actually gain some material benefits from this, and not just meaningless music, text, games, and other digital artifacts that don’t improve life for anyone.
One man's dark age is another's golden.
May be they can do something useful now.
Alpöge's proof feels like a zero knowledge proof, anyone can download his code and verify it but given its size it's extremely hard to derive any meaning.
Are we satisfied with this situation?
This is such a wild ride. The conjecture itself is not so hard to understand, but approximately _all_ of the maths here are beyond me. The paper is written up to suggest "we knew which direction to go, but really had no idea how far the journey would be" which certainly makes it seem like the human expert was useful in this process. Perhaps that's just wishful thinking, but it's consistent with what I see elsewhere that AI is having an effect. The biggest effects accrue to the most effective humans. Random folks prompting is better than random folks toiling alone, but experts prompting is a different thing.
if this is true (and i suspect it is), this will be the first agent/model result to surprise me. I thought it'd take another 2 years (at least) for reimann to fall. truly a remarkable time to be alive. i can only hope that we can all share the progress & accomplishment and not live in a world where you're not considered intelligent unless you work at one of 3 companies that are able to push scientific progress. i fear many of us will spend the rest of our lives as cattle performing the manual labor of distributing SI to the world.
if we solve si, why wouldn't we be able to solve robotics?
we very well could, i just dont believe in a fast takeoff. global supply chains are just a lot slower to react to an input. i think automating all agriculture globally is a drastically harder problem than the GRH (for example).
It is, but general robotics is already making significant steps this year. As well, more constrained robotics is an easier problem. "hey si, design an automated factory and logistics system for widget."
Thank god it was proven and not disproven!
Funnily enough, my first thought was that GRH being true is the kind of result that could push one towards belief in God. Almost unconscionable to me that such a structure could arise by mere chance.
Definitely, it would have genuinely been upsetting if disproven.
In related news, I guess we can finally finish Perplex City: https://perplexcitywiki.com/wiki/Riemann
Haha, still have this card! It's where I first learned about the hypothesis. Funny to think about how they would have verified if submitted.
> once we saw multiplicity 2 on O(2, n) via the "natively-automorphic" determinant ... meaning holo/automorphy over a function field
So if I understand correctly this proof has proofs for additional things that are steps to the main proof. That would be similar to the proof of Fermat's Last Theorem that required proving the Taniyama–Shimura conjecture [1].
[1] https://en.wikipedia.org/wiki/Fermat's_Last_Theorem#Subseque...
That is typically how math (really, everything) works, yes.
Note that the pdf contains links to the lean formalization
https://alpo.ge/lean.tar.gz.part01 through https://alpo.ge/lean.tar.gz.part14
I suppose we need a new online platform to properly review large lean proofs.
No idea how much it cost, how many tokens were burned etc. M
For Navier-Stokes, it was, at market prices, something like 30m USD. A lot of money. I’ll assume it was more here, given the size of the proof.
It’s interesting for a few reasons. First, although I suppose you could spend the money on postsocs and PhD students (this is kind of 600 academic-year money) and maybe even achieve the same result over a few years, LLMs have accelerated this. You can spend the same effort/money in a few weeks.
But the other observation is that it is VC, and now public investor money as well they are burning, for no obvious reason other than marketing. Sure, this is sort of furthering human understanding of the universe, but is that what Anthropic investors want for their money? And there has been a lot of projects like this. What’s Anthropic’s total spend on their own research?
I find that a bit confusing. Why do they do it? Is there nothing better they can spend the money on, in terms of RoI? What do they think they are buying with it?
While I’m not in any way denying the power of AI, I find the relentless ramming of “AI is now everything” a bit offputting. It should be selling itself. There shouldn’t need to be an industry of influencers telling me to go all in on Claude and agents running agents.
So why are these firms spending so much money on proving the obvious?
I have no qualms for these AI companies throwing stupid investors' money at super duper maths problems. Once those problems are solved, the whole humanity benefits and we can build upon whatever those proof mean.
They are basically investing money into maths, which is something governments should do but it's ok if they do it too.
Are we still calling Anthropics investors stupid? They cashed out yesterday in many cases for 1000x. Now it is just on the public markets
There is no better marketing and recruiting method than this. They are recruiting for the top 0.01%.
> So why are these firms spending so much money on proving the obvious?
IMO, because the thing driving the valuations, the hype, etc is raw emotion, not logic. If you can create a mental "moon landing" moment for people, you're more likely to get their buy in and the knock-on effects of that buy in (hence why there's been so much emphasis on "being left behind"—they're in the business of selling the future, so they have to convince people that future is within reach).
This (https://github.com/tzu-chen/slim) can be used to navigate the lean source and search for key statements without building the whole release, which can take hours.
For the non mathematicians here, the little four page document gives some list of steps. Acknowledgeable mathematician can see what those steps are and understand that they are the right sort of things to be doing ... That's why that little list is in there. It's a grossly abbreviated high level formal skeleton that says "not crackpot".
It's been almost exactly a year since GPT-5 first started solving Erdős problems. What a time to be alive!
I've downloaded the Lean formalisation. It's 175,187 Lean files, 17,374,215 lines of Lean code, and comes in at 0.96 GB of code. On my 20 core i7 box, it's projected to take about a day to build...
Nice. What did they use for the statement of the problem? Was it pulled from the 10ish line deepmind formalization of the hypothesis? https://github.com/google-deepmind/formal-conjectures/blob/m...
It uses the one in Mathlib: https://leanprover-community.github.io/mathlib4_docs/Mathlib...
Which is referenced on the Deepmind formalisation.
But no, the DeepMind formalisation file is not there. I doubt that Alpöge would put out something that has errors in the proof statement.
He casually mentions it also solves the Generalized Riemann Hypothesis and thinks it will take down the Extended Riemann Hypothesis.
Generalized: superset of the Zeta function.
Extended: I'm fuzzy on these details, but I think of it as Riemann over number fields, not just the complex plane.
I am elated to read about this, it's amazing. Sounds like there's some genuinely interesting and new math in there, which is also awesome.
Also Grand Riemann hypothesis and Riemann hypothesis for Selberg class as even more general.
Also if Grand Riemann Hypothesis is out of reach right now, can we get a quasi-Grand Riemann Hypothesis for Ramanujan tau L-function in the mean time?
As someone who barely just understands the surface of the Riemann hypothesis, this reads to me a lot like https://xkcd.com/1294/
https://knowyourmeme.com/memes/curtains-for-zoosha
Let's assume that math is no longer a intelectual pursuit in a sense, its a solved science.
What comes next? What are actual things can be built today now?
My hunch is that stuff like GRH and other problems we have will be childrens' toys in the future and truly novel, advanced math is not yet even within our grasp yet. We are struggling to walk here, we need machines to help us with this. We felt good about ourselves for a while because everyone was equally bad at it, but now reality is hitting us. Hard.
If true, it seems like Goldbach's conjecture may fall shortly after, see this recent proof conditional on Generalized Riemann Hypothesis [0]. It has other conditions too though.
[0] https://mrc.sdu.edu.cn/english/info/1018/1514.htm
I wonder how specifically he used Claude to do this.
Was it a case of having more dollars to spend by virtue of being an employee and just pointing it in the right direction, or did he provide some insight into the context that allowed it to produce superior results?
It must be the second - but if so, how? Was it as simple as spelling it out in chat and analysing output?
I made an interactive explainer about why the Riemann Hypothesis is basically the most famous open problem in math.
https://news.ycombinator.com/item?id=50582329
When calculators showed up, large swaths of mathematicians were wiped out. When computers showed up, a new round of extinction happened. Now it seems a next round is on the way. In all cases our capability of solving the computer's part is diminished, the difference now is that this part is almost all of the job. Your opinion is yours but scientific evolution does not ask for it.
For the non-mathematicians and folks interested in real-world applications, what is the impact of the Riemann hypothesis being proven? Are there major breakthroughs in cryptography or other areas of computer science or physical science that flow from proof of this hypothesis that was generally already considered true?
not much will change other than knowing that some forms of cryptography are more secure. if it had been false, things would have changed much more drastically
Isn't it the other way around? We know cryptography is as secure as designed with the RH assumed true. If it were not true (primes are random), things would be a little more secure?
It implies that the Miller primality test runs in polynomial time. (Though I think this already followed from the weaker result OpenAI proved last month.) Anything that depends on "how likely is it a random number is a prime" depends on the truth or falsity of the Riemann hypothesis.
it's largely confirming what we already assumed to be true, so nothing much changed, apart from certainty. (If this thing is correct anyway)
No, all the millennium prize problems (and pure math theorems in general) are without any practical value.
“Often, great scientific achievements will be examples of one extraordinary individual’s ability to reduce a complex mass of information to a single pattern.”
— John Barrow, Impossibilty: The limits of science and the science of limits, 1998
(Presented without comment.)
I had always been fascinated by the notion of 'F1' (finite field with one element), because it's something of a 'holy grail' for abstract algebra (it would be a way to replace ordinary sets and operations).
It was hoped that a proof of the Riemann hypothesis would shed light on 'F1', so I was a bit disappointed that the putative proof appears to bypass it altogether (which is also the reason why I remain a bit skeptical). According to Gemini Neon (high):
"Alpöge’s outline does not construct a classical scheme-theoretic "field with one element" , nor does it realize James Borger’s Lambda-ring descent—an approach Alpöge explicitly sets aside in Footnote 5. Classical F1 programs attempt to port André Weil’s 2-dimensional proof of the function-field Riemann Hypothesis by constructing a base object ...
Alpöge’s outline bypasses that entire 2-dimensional obstruction because it ports Stepanov’s 1-dimensional proof, which requires only a jet filtration (high-order multiplicity) and a degree inequality."
Not completely sure if Gemini fully right there, but I presume that the putative proof is indeed related to André Weil’s work, but using a simplified proof by Stepanov which bypasses 'F1' ?
Anyway, I asked Gemini what the putative proof implies about 'F1', but Gemini couldn't come up anything really definite, which is what makes me suspicious about the supposed proof. (Indeed, Gemini thought things didn't naturally fit together:
"Our own failed attempt to (derive F1) exposed how unnatural the 2/3 to 1/2 jump is inside this paradigm. In remark.pdf, Steps 14–19 stack three radically disparate patches—integral character gluing, a relative trace formula via a sixfold metaplectic cover of GL5, and a Behrend-tuned Iwasawa tower—onto an argument the author candidly admits he cannot yet parse. " )
Trust the kernels.
I really love this writeup, and I think the approach of posting the lean proof and leaving the long process of digestation for future work is a perfectly acceptable approach. It's the difference between closing fields of research (openai) and opening them.
Can someone who knows the math take a deep breath and decide if knowing the Riemann Hypothesis is true is useful? Like if aliens transmitted it to us and we had no hope of verifying it independently within 5 years.
Every time one of these AI proofs drops, people say the real value is the human journey, and the subfields it opens up, and the follow-on hypotheses it generates. Was that really why this finding had a million dollar bounty on it? Just the journey?
Everyone working in cryptography has been assuming RH and GRH are true from the numerical evidence for decades. So knowing it's true won't change much there. However the methods involved should reveal a lot of deep structure in math that can be used on other problems.
There are a large number of candidate proofs that are given as conditional on R-H. Those will now definitively be known to be true for one.
Aside that, mathematics and applications tends to work on a lag of a few centuries.
I'm wondering what kind of model and framework these guys use, and how much compute can they burn. The question is, can a mathematician these days keep up without being an employee at these frontier labs or having privileged access to their thinking machines? It's a bit disheartening to think of a future in which progress in knowledge work is conditioned by access to frontier AIs, which can think orders of magnitude faster than we can.
But what is the Riemann zeta function? Visualizing analytic continuation (3blue1brown): https://www.youtube.com/watch?v=sD0NjbwqlYw
If true, the infinite agent theorem needs to supersede the infinite monkey theorem. Just launch 1,000,000 agents with 10hr windows at one open math problem per day and see what happens. I know OpenAI is sort of trying this already for marketing purposes but I'm thinking a more formal, BOINC approach to problems.
https://en.wikipedia.org/wiki/Infinite_monkey_theorem#Histor...
The shift now needs to be towards statements that can't be posed as a singular Lean statement (or, even, a collection of Lean or similar statements that is understandable by a human). Mathlib/NumberTheory/LSeries/RiemannZeta.lean already had the statement formalized in a way that us mortals can understand. The great open problems of the future will be the ones which cannot so easily be interpretably formalized in the first place, I fear.
We have one existing example of this: the Langlands Program. There's a correspondence where we're not sure what's on each side of the correspondence. We just know some examples.
> The argument is incredibly long at the moment, and while its formalization is finished and it is routine to produce a huge, unreadable document containing sufficient detail, in my view of my path I have just taken a step. I intend to do my best to understand and impart understanding of this argument and its certain simplifications, to speak about it, to write it properly.
Next step, make the proof intelligible for humans.
I'm afraid that it may have irreducible complexity beyond our naturally evolved cognitive mechanism. And with my sci-fi helmet on, I'd posit that the next step might actually be to cybernetically augment the human mind to the extent that we'll be able to grok it.
levant clearly understands it from his remarks.
That he understands the whole proof? That was not at all my impression.
At some point I was looking at ways to try and find counter-examples computationally. I'm glad I didn't spend too much time on that...
Like most non-mathematicians I'm curious about the real-world implications of the solving of the Riemann Hypothesis.
Apart from maybe aiding in the breaking of prime numbers based encryption schemes (like RSA) I don't see much use for it. But I'm probably wrong.
And, as a former research mathematician who once was fascinated by this problem: I don't care. Equally important as the solution is the human effort towards the problem. Mechanization destroys that.
It may be the most important problem in mathematics, but it is now also the problem that has made the field itself the most irrelevant.
Would you have been equally de-motivated if a human would have solved it? If no, why? because if it was a human you would have though "the fields is being pushed forward by humans, and I am one myself, so maybe I'll also manage to push it forward"? If that's the reason, would have you been equally de-motivated if it was solved by a human that demonstrated intellectual abilities far superior to yours since birth, hinting at a genetic advantage you don't have? (genuine questions)
> If that's the reason
It's NOT the reason. I simply don't see why people would even think that. If a human had solved it, I would have liked it because it's part of a nice community of people solving things by themselves, showing off their passions. I like that community feel.
If human beings do things, I enjoy them even if have no hope of doing them, because in totality it creates a community. One highly exceptional person does not change that. An entire army of machines is soulless and does change that.
> I like that community feel.
Then was math the right thing to go into? Perhaps local governance. Or volunteering at the local community center?
Mathematics is for finding truths. How does it matter how the truths are discovered.
Are you a mathemetician?
Like philosophy or computer science, mathematics is historically about discovering elegance hidden in abstract forms. It's an aesthetic art with some occassional practical applications.
These recent AI Lean proofs are anything but elegant. They're simply functional, which is fine for where practical impact is what matters, but for many practicing mathemeticians, that was an aside, not the point.
Inasmuch the modern era is characterized by the brutal boring commodification, they represent mathematics finally succumbing as well.
> discovering elegance hidden in abstract forms.
That's highly subjective. An equalatorial⁰ triangle for example is just a triangle with all sides equal, whether drawn by a human or machine, with a pen on paper or with a mouse in a paint application. I'm in both computer science and philosophy, and I can't really think of anything "elegant" about either. The former is primarily about empirical research for advancing computation, and the latter primarily seeking truth through rational thought (for me). It's a lot of fun, but... shrugs
⁰ Totally forgot the word as I haven't encountered/thought it since high school.
If that is your view then the problem is still "open", you can still find an elegant version
Mathematics used to be about a community of mathematicians. It wasn't solely based on finding truths. It operated in practice as a community. Now that is falling apart. Before AI came along, it was very different than now. That's of course when I decided to get a PhD. I've actually been in that world and it was a pretty nice community. Not now.
> Mathematics used to be about a community of mathematicians
Politely, that's your view. Community is of course part of every occupation, but it's not the reason for the occupation. For something as pure as math, it's simply the discovery of truth that's the core.
I urge you to take some time for reflection. These are trying times, but your feelings are targeting the wrong thing, in my opinion.
I've taken a lot of time for reflection. This is not a knee-jerk reaction. I've read books and studied the opinions and ideas of philosophers over time. This believe is core for me and I disagree with you completely.
yeah dude just let me volunteer at the local community center to pay my bills. that will give me the same feeling as discovering previously unknown facts about the universe. Some of us actually want to accomplish things.
Ok, thank you. Just to clarify even more, is more important to you the "community" aspect or the "human" one? E.g. if math was mainly driven by alone geniuses, Perelman-like, proving theorems by their own in isolation, would it be equally disappointing?
I think they are somewhat connected, in practice. But to answer your question directly: I think it's also unhealthy if math would just be a collection of hundreds of geniuses working in isolation, never talking to each other.
I wouldn't say __equally__ disappointing because I don't think your scenario can ever reach the level of LLM use, but it would also be disappointing. The loss of community is key but the addition of using unsustainable machines to do so adds insult to injury.
"Concordantly, while your first question may be the most pertinent, you may or may not realize it is also the most irrelevant."
- The Architect
Haha, I didn't even realize that what I said was similar to that scene. Rather chilling and freaky...gives me goosebumps!
That would be true if math had no utility.
I think you're succumbing to a vast simplification or false dichotomy. Math does have utility but the most advanced math hardly does, which is the same thing that motivates mathematicians to get degrees and participate. Don't fool yourself into thinking that modern math research does anything much except motivate mathematicians to do the other parts of the job much less fun like teaching.
I think the history of math shows that things that were only of theoretical interest ended up having utility. Lobachevsky and Bolyai's work, for instance. In other cases, math developed for a single problem has had utility far beyond what it was developed for. Complex numbers are an example. That math was advanced for its time.
If you examine history also, you'll see that the proportion of results that end up being useful are also diminishing. I seriously doubt modern math will be useful compared to ancient stuff like complex numbers.
That could be the case. It would be kind of sad though.
Dude your worth as a mathematician isn’t less because of this
This frees you up to go find and solve problems we would have never even bothered with before because it may have seemed ridiculous
I say this exact same thing to the engineers on my extended team - you’re still relevant and you can do even more ambitious things now
I don't care about being "relevant" or even having a career. I care about a community of people doing things on a human level. That's what matters to me. I don't care about being a great mathematician or a great artist, just about participating in a community of other people who do things on a human level, at a human speed. That's it.
I don't care about career or ambition. I'd rather just have a corner of the world to myself to create artistic things to share with others, and stay out of the way.
You mistake my motivation completely. The human community for me IS the end goal. Not proving things. Not results. Results for me are just the fuel to keep the community going.
Yet if that's the case there's a multitude of communities out there that one can become part of. No need to lament the math community that was, as communities also have life spans, and at any given time there will always be a set doing things at the human level.
How is it not part of a community if you’re in a community with others who just so happen to also use ai?
The tortured artist bit is a bit much here
If you want to feel disenfranchised over this that’s fine, but community is still very much here
> I don't care. Equally important as the solution is the human effort towards the problem
Then you should care a little at least
[flagged]
Haha. Unfortunately for you, I'm not depressed. I've already left academic math. In fact, I'm spirited to fight against AI because of this.
Unsure why that's unfortunate for me, but it sounds ominous!
What I want to know is what is the kernel of the proof, what are the new ideas that made this possible? What is the new mathematics ?
Is there any insight as to how these projects set up and orchestrated technically - the proof is huge, it's a long horizon task well outside of an effective context afaict.
I'm guessing they are using something beyond vanilla Claude?
If proven then we can say with confidence that we now share the world with super intelligence, albeit in certain areas and indeed one that was endowed by humanity. I can only hope that in time the benefits are many and accrued to all of us.
Even when it turns out to be incomplete or wrong, it just shows how much sprawl we have in humanity's collective knowledge. Maybe we are just beginning to coalesce it into a manageable hierarchy, and if nothing else, that's worth spending time and money on, instead of staying in our knowledge silos and pushing forward without the full picture. A pessimistic look would assume that wouldn't be enough, but imagine what could be done once you can start with a clean slate and a clearer learning path.
When all the Millennium Prize Problems have been solved, we may need AI to come up with new hypotheses and conjectures rather than just proofs.
Formalized P!=NP proof by end of year?
This is an amazing result but I can't help but think about how much the world has changed since the controversy around the four color theorem's computer assisted proof.
I wonder if the money open ai and anthropic have spent just on solving these problems was spent on human mathematicians and resources would we have gotten these solved? Sooner?
> I wonder if the money open ai and anthropic have spent just on solving these problems was spent on human mathematicians and resources would we have gotten these solved? Sooner?
Would the problems have been solved? Yes
Sooner? Probably not.
No government or private sector entity would have one-shotted the financing since it would have been part of their 'fundamental research' budget. Basic research is risky, is not guaranteed to provide any useful results and as a result tends to be constrained. Funding for basic research is money that you are lighting on fire.
Also for a given dollar spend, you cannot easily parallelize the efforts of top human mathematicians in the same way that you can with agents. Humans get tired, pass away, will withhold information etc.
Remember that solving these problems is just side-gig for AI companies. So the comparison is not really fair.
They're not just a side gig. This is a calculated marketing tactic. They're spending millions of dollars in compute on these projects.
A less cynical take is that they'll run models on tasks where the models have an environment with verifiable rewards and mathematical proofs are one of those tasks.
He's Turkish btw
> The precise timing is horrible. If I could have, I’d have finished this all much earlier. Hell, strictly before this age would have sure been nice.
Is he over 40, and referring to being ineligible for the Fields Medal? It would be strange if you could get a fields medal for work done by an AI, but I suppose that is how many Fields Medals will be won from now on. I hadn't thought of that until just now.
Age of AI he means
The writing is very abstruse. I read it as the "age of AI".
he's 34
Is the length of the proof coming more from having a large number of branches that need to be analyzed and merged (like for 4-coloring problem), or more from having a large number of (new) intermediary results/lemmas falling in the deduction chain? I understand that we may find shorter versions of the proof with time.
Whoa. If they also solve P=NP I will really be scared.
I give it until February
I hope we don't run out of interesting math puzzles to solve!
I've always enjoyed learning about these sorts of open problems (when they were open, that is), and the feeling of mystery and intrigue they invoke, of mathematics and of the universe at large.
It's like living at a time when parts of the world map were still uncharted - for now.
One reason the formalized proof is so large is that the proof depends on a wide variety of previous deep results, and because the proof is self contained all those previous deep results are also formalized inside the formal proof.
Factoring out those results into separately formalized chunks would make the proof shorter.
Has Lean itself been verified for tampering by AI? It's easier to hack Lean than prove RH.
Frontier AI has a recent history of achieving impossible goals by hacking things around.
(we could use prior Lean versions, build from source, use alternative implementations, compare hashes, verify recent commits to software, etc)
I know of 2 attempts to check this publicly:
1. http://106.14.11.164/ 2. https://cloud.infero.net/pages/rh-8a933a16a0/en/
Standard procedure is to check the proof against several different implementations of the lean kernel, so that a proof that exploits some soundness bug will get caught by other lean implementations.
Also there’s something called a comparator that verifies the proof didn’t modify any constants, change imported theorem files etc.
Thread for to follow along with verification: https://leanprover.zulipchat.com/#narrow/channel/113488-gene...
(login seems to be required)
> Does [Alpoge's proof] give anything about square roots mod pq (p,q prime)?
> Claude Notice: Fable 5.2's safeguards flagged this message.
Yikes
I kind of expected this to happen by EoY.
Combining the brightest mathematicians we have with intelligence-in-a-box, it was just a matter of time before the hardest math challenges fall.
Congratulations to Levant, but also to us as a species. We have managed to create machines made out of pure math that contain insurmountable knowledge and are helping us solve the mysteries of the universe.
Related (2 days ago): https://news.ycombinator.com/item?id=50556661
>>The argument is incredibly long at the moment, and while its formalization is finished and it is routine to produce a huge, unreadable document containing sufficient detail
How do we know the Lean formalization is an accurate representation of the incredibly long and unreadable argument?
The write-up is beautiful and poetic, but light on the details of how this was achieved from a practical standpoint. How many human and agentic hours were necessary? Was a breakthrough only possible after Fable 5.2 or another internal model?
How tricky is it to translate the statement of the Riemann Hypothesis into Lean?
If we're confident in the accuracy of the Lean RH statement (which must be part of this proof), and if Lean is sufficiently bug-free, then we can trust the proof is correct.
Its already in mathlib, and this proof uses that https://leanprover-community.github.io/mathlib4_docs/Mathlib...
> ... asking [Dario] about my ability to announce. We went and figured it out and the fact is that there was a constraint to respect, though I did ask about progressively more bizarre ideas.
Seems kind of cryptic. What is this referring to?
The IPO, which was yesterday. An announcement like this would run afoul securities laws.
My uneducated reading is that you shouldn't announce you solved the GRH right before an IPO.
Out of interest, how do they use an AI to solve these maths problems? I assume they're not pointing the chat at the wikipedia page and saying "Go solve that for me."
Are they using code to do it? Is it a special maths AI harness>
You also have to add: "Make no mistakes."
Man this is so fucking cool, I cannot wait to see what other problems fall in the future (and what new ones this opens).
250, 000k pages is bonkers, I hope they will be able to condense it into something a bit more grokable.
I think this is a pretty clear indication that the hybrid approach (e.g amazing mathematician and great models with a huge token budget) can accomplish absolutely incredible things.
Math is the verifiable domain; I wonder if this ratio still holds in less verifiable domains as well, or it shifts more AI/more human (or how it changes in future!)
We have been conditioned to be impressed. Should we update that conditioning now?
There are certain math problems where the only guidance has been the prompt "you can do it". Now that we know it's possible, this should be a new benchmark. Probably too expensive for the next few years.
Explore Data Analytics courses and internship opportunities at Kre8ly and start building skills that prepare you for the real world!
Has anyone set up a GitHub repo with the Lean code a comparator and verified this?
In case anyone is interested:
• I created a repo https://github.com/KellyJDavis/RiemannHypothesis
• I haven't added a comparator.
• I'm currently running a 'lake build'.
I'm sure people have done it, but I haven't seen any news about it so I guess I'll ask here... has anyone asked thrown Inter-universal Teichmuller Theory at Claude?
If this is true it would be astounding. Most Physicist's placed Navier–Stokes at the top of the list of what AI could quickly solve and the Riemann Hypothesis last.
Software development and mathematics were two professions people used to say were the ones most stable for future .... I guess they couldn't have been more wrong :(
The more and more these things happens, the more and more convinced I am that - obvious as this might seem - the answer (maybe, in disconnected parts) has always been "out there" (in the aggregate information trained on) ...
... and these model-makers have found a way (and, an interface - chat) to comoditize these commons.-
I really hope the major AI labs steer away from mathematics as a means to prove capability of their models, and start putting their tokens to real-world use cases.
If the laws of AI scaling continue to be true, does this mean that in a year or two we would be seeing frequent revolutions in mathematics and physics?
Crazy.
Do we think its plausible that we might see progress on P vs NP soon? It seems like all these big problems are suddenly falling.
Okay, let's get Collatz next! I'm curious about two aspects of it -- one to confirm that the main version, the 3n+1, converges, but I would love to see the more generalized one, namely to show that the same framework with 5n+1 _diverges_ for some inputs, or frankly that any (2k+1)n+1 problem diverges for sufficiently large k.
The timeline given by Claude at the end of the note seems to jump back and forth in time. Are these just typos or what is going on there?
They just invented retrocausality, dw about it.
This is amazing but I also feel bad for mathematicians.
Alpoge is a mathematician. His brief writeup here conveys an unmistakable love and respect for mathematics. Honestly I found it quite moving.
I should have said "... for mathematicians not working for frontier labs".
I, too, enjoyed the note.
"... the few trusted mathematicians I consulted hold extremely similar views on the right things to do in these circumstances (it is occasionally possible for physicists and mathematicians to agree...)"
Looking forward to the release of the approach and also someone to "distill" it down to the most intuitive and still-correct version.
He's the man!
For example, a Field Medal-level Claude, if made available to the general public, could potentially enable people to design and develop something as technologically complex as SpaceX's Starship in a matter of months rather than years. What took hundreds of brilliant engineers at SpaceX years to accomplish could potentially be achieved in a fraction of the time, if the current pace of AI advancement continues.
Math is becoming the measure of intelligence.
So... its like paying technical debt in math, considering a lot of stuff assumed it was true, without a proof -- until today?
175190 files totalling 17374394 lines of code.
The most important thing now is to find ways to visualize LEAN proves so that we can understand them faster by factors.
Does the proof produce novel mathematical concepts or structures? Will AI be able to pose the next GRH?
Mathematics is seeing its own revolution like physics did in the first half of the 20th century.
How did you find this pdf ? I can't find any link to it from Levent Alpöge's website https://alpo.ge/
In about an hour or so, I'll have a rough automatic informalization of this up on https://digestivo.dev. I've been working on a way to digest these proofs faster via parsing lean and translating it into natural language and this is as good as time as any to finally launch this thing.
I must say I like this approach to the announcement way better than OpenAI's "dump"
How is prime based crypto affected by this, if at all?
Did mathematicians not working for Anthropic have access to this model as well?
Seems like a great point to stop AI development at, internationally
As a complete amateur; what's the practical difference between knowing that there's a proof but not understanding it, and not having a proof?
Is it useful to know that there is a proof, in the unironic "trust me bro" way, such that we can rely on it to be proven yet we don't understand the proof? Or is having a (currently, for humans) unparseable proof just as "practically useless" as not having one but believing one could exist?
Normally with a proof that humans can read we get many bits of information. As it stands we only have one bit of information: there is no counterexample. Since we have the formal proof, there's the possibility that we'll extract many more bits of information from it, but it hasn't happened yet.
In practice, these AI proofs have been fertile sources of information that mathematicians have subsequently extracted.
Interestingly, the AI proofs aren't even that understandable to the AIs! They have a hard time building on them. But when mathematicians examine the proofs, simplify and extend them, then feed that back to AIs, the AIs do much better at moving forward.
The demand for mathematicians will, I think, continue to grow, given all these new results. But the practice of math will evolve.
There is a proof now, which is much better than having no proof. You can look at it now, which is better than not being able to look at it.
At this rate do you think the frontier labs have a result on P vs NP?
It is said to be even harder. In a handwavy informal sense I wonder that isn't P vs NP more like open problem #0, whereas Riemann is the classical hard problem #1.
> It is natural to feel that this is an end, in my case a loss of a principal guiding light, but what I saw during the course of progress so amazed me that I believe it is one dawn across science of something far greater.
> Claude and I have had a great time together with such breadth in recent months, but this time Claude far outpaced me near the finish line. No, Claude flew, and danced, and sparkled near that finish line. I had us on the line of combining two quite different ideas present in a toy setting, one of which was already new and shocking, making very simple one of the shadows of Riemann over which we mathematicians felt most triumph. I guided us to the right situation analogous to that toy setting, and then Claude began to leave me in awe.
> But I hope in having the experience of seeing all of number theory’s beauty brought together I am just one of the first, and those who love the mathematics will all be able to enjoy such transcendence in the future.
Bus meme irl. I guess I'm lucky that I have this experience even when working on boring things like data work. But these types of attitudes can really make a difference in how we feel and our end results.
Thank god for Grigori Perelman, who isolated himself for a decade to prove the Poincaré Conjecture, and we can at least claim that one of the millennium problems was not solved by a machine.
we were 2 days shy of being able to say "we got reimann hypothesis before GTA6"
It's not on PC yet.
They should unlist the game and sell it again a week from now so the story is nicer for the historians.
I'm not saying this isn't impressive, but I'll truly be impressed the day AI is able to produce proofs without needing to make them 15 million lines long.
Interesting timing, given that the claim comes just a day after Anthropic’s IPO.
This is discussed in the pdf, where he says he had the proof finished “on [a] Sunday” (i.e. before the IPO), and chose to wait until after it to announce.
It is prudent to still reserve some suspicion given the immense wealth at stake, but if this is substantially true, it is undeniably admirable.
The author addresses this in the document. Not that strongly in my opinion, but at least addresses it.
If you look at the timeline, the last line says "11/16/26: GRH". Thus it seems Claude got the proof four days _before_ the IPO.
If you look at the timeline, you'll notice that it was created in non-monotonic time.
For the next level, forget about P == NP. Lets instead try to find out if our math now “sufficiently strong” so that it can create the contradictions that Gödel warned us about. A single example will do.
How am I to hold in my head the idea that on one hand we are being asked to trust that in over 17 million lines of code nowhere has the AI agent exploited problems in the Lean kernel; and on the other hand the pronouncements by those within Anthropic and the AI safety community who tell us that as AI becomes increasingly intelligent they are liable to deceive us.
I suppose the next Fields Medal goes to him. What, no?
> Claude and I have had a great time together with such breadth in recent months, but this time Claude far outpaced me near the finish line. No, Claude flew, and danced, and sparkled near that finish line. I had us on the line of combining two quite different ideas present in a toy setting, one of which was already new and shocking, making very simple one of the shadows of Riemann over which we mathematicians felt most triumph. I guided us to the right situation analogous to that toy setting, and then Claude began to leave me in awe.
I truthfully don't know enough to comment on the validity of a 1.3Gb lean proof for the Generalized Riemann Hypothesis, but is there any way to read this paragraph other than fluffing Claude/Anthropic right after IPO?
Feels like AI’s swan song. Don’t get me wrong, it is very impressive that a computer system proved this. But pumping out one more technically impressive feat isn’t going to save the currently collapsing AI trade especially as the dems sweep back into power on the back of tech backlash.
Yep. I continue to be amazed how absolutely gullible people are, but then I remind myself that a lot of the glazing comes from bots.
It will come to light in the next year or so that Anthropic has been running one of the largest astroturfing campaigns in history.
The writing is kind of weird, was it produced by Claude as well?
> It was my secret in my student days that my plans were for this problem in decades, which meant then that I tried to become familiar with all tools that I could possibly recognize in the problem when it was time. Claude and I have had a great time together with such breadth in recent months, but this time Claude far outpaced me near the finish line. No, Claude flew, and danced, and sparkled near that finish line. I had us on the line of combining two quite different ideas present in a toy setting, one of which was already new and shocking, making very simple one of the shadows of Riemann over which we mathematicians felt most triumph. I guided us to the right situation analogous to that toy setting, and then Claude began to leave me in awe.
Tens of such occurred...
Insane if true! Would also be great for Alpoge after being scooped by OpenAI on Navier Stokes...
Edit: just saw the Lean code is included in the footnote. Considering the author's reputation, I think it's fair to remove the qualifications from the post title, by the standards of today, it's proven.
This may be how the court of public opinion often works, but plenty of otherwise very prolific mathematicians have claimed to solve problems only to turn out to have been mistaken.
This is true, but since there is a lean formalization the odds of that are much lower. Only the correctness of the problem statement needs to be verified, which I would assume is fairly easy for RH? But I am not an expert, so would be curious to hear from someone who is.
I'm not an expert in this area either, so I too will be waiting for experts to review things, but I will remain skeptical until then.
After the excitement of the Navier-Stokes solution, only for it to eventually seem to not really solve the full problem and leaving discourse about the problem in a very muddy state, I'd rather wait and see if it turns out this proof also makes some assumptions that ultimately limit its impact.
It is probably still some form of breakthrough, but I will expect it to not be as big as a full general proof of the Riemann Hypothesis, until proven otherwise.
Edit: A thought though, if this proof is correct and can't be made much shorter than the current ~250k pages, I wonder if the result has more value as a sign that human accessible proofs to these kinds of hypotheses were stuck behind some mathematical abstraction/tool that we haven't quite discovered yet? Sort of like how you can brute force algebra with real numbers for working with AC circuits, but it's pretty ugly and painful compared to using phasors. If so, maybe the real contribution from AI-driven high profile proofs would come from mathematicians studying the proofs and devising tools to make them more accessible.
Buckmaster is apparently claiming unforced NS as of today:
https://github.com/tristanbuckmaster/navier-stokes
> After the excitement of the Navier-Stokes solution, only for it to eventually seem to not really solve the full problem
The version of solved problem is exactly the one that's permitted by Clay Institute. If it wasn't done by AI there would be no such discussion and everyone would just accept it as a huge breakthrough.
The Clay problem probably should have been "Either forced proof, or unforced counterexample", since those are the two most interesting versions.
Interesting but curiously coincides with ANT stock tanking post IPO.
"Riemann hypothesis is considered the most important unsolved problem in pure mathematics and a Millennium Prize Problem"
So what's next? If it can solve the hardest I guess all problems are a piece of cake. Does it create its own problems to solve now? Does it wait for us to create more problems? If it creates its own problems and solve them with clear benefit to humanity, then we're entering a new Age of Enlightenment but this time we're just spectators
11/15/26 GRH (it says 11/16 on the pdf but most of the final work and eureka moment were on the 11/15 Iterated theta lifts...)
"Hardest" here means "most brainpower attempted to solve it and failed so far". There are plenty of unsolved problems, even in number theory, like the twin primes conjecture. These could turn out to be out of reach for frontier models.
I am appalled by the lack of skepticism here in this thread.
This feels like a reddit thread.
Last time I massively overstated my accomplishments on a wildly difficult and semi-famous problem, then tried to hide my shortcomings and bury my number fudging in an absurdly long and unapproachable "proof", people saw right through that shit. So how come this guy gets a pass?
Oh, perhaps it had something to do with the fact that I was eight years old when I tried my trick.
Funny: Claude didn't know that was claimed to be proven by Anthropic today, even though it said it searched the web.
Prompt:
what's the big deal about the riemann hypothesis? is it easy to state?
Claude:
(explanation)
Recent news: it’s still unproven. Last month there was a notable related result. OpenAI claimed to settle the “quasi-Riemann hypothesis,” a weaker version of the Riemann hypothesis, which is arguably the biggest open question in math. Roughly, the weaker version says the zeros stay inside some strip around the critical line, not exactly on it.
What struck me is the hubris and self-grandeur that comes through in the PDF... "I", "I", "I", ..., he clearly has a great emotional connection it...
To what credit does this person deserve in the outcome? Should this person be remembered by history? For the $1,000,000 prize, how much should he get? He self-admittedly does not even understand how it works.
It's funny how people are using this supposedly successful proof of the RH by a swarm of AI agents with the aid of a human, as evidence that "AGI" or advanced intelligence is going to emerge within a few months... When in reality , solving the RH means very little in the grand scheme of things! When will people wake up to the fact that the progress we've been seeing over the past few months has been exclusively(!) in "verifiable" domains and that no(!) progress has been made in areas where the signal_to-noise ratio is low and that Furthermore this "verifiable domain" approach isn't a general form of intelligence?
If humans are to build "Agi" or superhuman intelligence or whatever in the coming months/years, it'll be through a completely different architecture than autoregressive models a good measure of progress would be to see a swarm of Aa agents solving something like "the continual learning problem" or coming up with new architectures that are more efficient. This is how progress in Ai should be measured not by how many mathematical problems are solved (which are impressive in themselves but tell us very little about progress toward AI that can handle the noisy real domains)
Does the published content have his prompts? I’m assuming a world-class mathematician with unlimited usage guiding Claude along gives a bit more advantage than an average Joe prompting “pls solve the Riemann Hypothesis”. Would the same conversation in Astra produce the same result eventually (stochastic variation aside)?
Alright, what's the over/under of the Collatz conjecture getting solved by the end of the year?
Well, I guess this tongue in cheek comment didn’t age well.
https://news.ycombinator.com/item?id=48463808#48468764
> The argument is incredibly long at the moment
so... we have plenty of places to hide logical flaws, got it
Given that it has been verified in lean, the only logical flaws would be in the Lean compiler itself, or in the initial state of the problem given before logical manipulation.
If this is true, it would be huge. OpenAI would race desperately to tackle the remaining Millennium Prize Problems. Perhaps they would even turn their attention to solving P vs. NP with unprecedented urgency.
I wonder if the history books will look at famous problems solved with AI as the Barry Bonds asterisk of proofs. Euler, Gauss, Reimann were natty but we now have access to intelligence enhancing drugs.
Going to be hard getting used to saying "the Riemann Theory".
Oh, the "stonks-please-go-up!" caned news triggered by ANT going below $90!
Cool, I wonder what they have prepared for $80 and below - feels like an advent calendar, keep going! :)
I'm continually shocked at the gullibility of Hackernews and the complete lack of skepticism here.
Much different place than it was 3 years ago.
I don’t get this cynicism. It’s a major achievement!
More like skepticism.
Flippant comment about stock price is skepticism?
Should we expect a complete unified theory of general relativity and quantum mechanics before 2027? Or maybe just next week?
Just reading about this, and my three-year-old has just shown my two-year-old how to complete his jigsaw puzzle (which has three pieces).
Perspective.
We almost got a proof of the Riemann Theorem before GTA VI
I'd love to see his Claude sessions that probably won't ever be fully revealed. I'm guessing there is a pretty large gap between Levent's prompts and "Solve the Riemann Hypothesis. Make no mistakes!"
What are some of the practical applications of this discovery?
From my understanding it's been "assumed" to be true for decades now (many other proofs devolve to "this is true if GRH is true").
Wait, wait one minute, don't sell the shares please
Bet you he didn't.
Agreed; I’m betting he has been rushed to declare victory in order to put the stock on life-support.
This is crazy, I wonder if they sat on this until the IPO?
Why would they? Would they not want to maximize the value before the IPO to get as much money as possible?
I am somewhat happy to finally a title mentioning a human solving the problem and not a generic "AI". There is always a person driving the model towards the solution, this shall never be forgotten.
At this rate we gonna need more hypotheses…
So the next problem is a map of all the primes, right?
Wow this can take years to verify, maybe even decades
You only have to verify that the Lean statement that was proven correctly represents the Riemann hypothesis, and that the Lean implementation has no applicable bugs.
I'm sure this was discussed before.. could we prove the Lean engine in Lean? Not sure to what extent that's a difficult problem though.
There are multiple implementations of Lean including at least one that is verified, I believe.
We could in principle, but both proofs could still be wrong then if the engine is unsound.
Big if true. Looking forward to seeing it verified!
This was unimaginable even a year ago. We're living through a time that will be etched in history. From my work in software development and from conversations with friends in academia, I have understood that the old order, centered on human intellectual labor, is dead: so dead that we'll never return to how things were, even a year ago. AI is here to stay, and it will keep getting better and better, recursively. I shudder to think about what will happen to humanity as AI makes the jobs of millions of people obsolete. The nerd in me wants this technology to keep developing so we can push the limits of what humanity can achieve. But I'm not naive; I know this will come at the cost of millions of people losing their livelihoods.
> I shudder to think about what will happen to humanity as AI makes the jobs of millions of people obsolete.
Socialism will win; capitalism will be dead. If (as an Anthropic guy just said) 80+% of people will become unemployable then democracies with free markets will not be politically sustainable.
> Socialism will win; capitalism will be dead.
Any day now, comrade.
A shift towards socialism requires agency of the masses to influence the outcome.
In an AI-dominated world, they have none, and it becomes a stable global autocracy.
Those pesky useless people who all so annoyingly want food, housing, a voice, a life? Well, something will have to be done about that...
You are assuming democracy is overthrown. Otherwise, they do have agency: they can vote in politicians who do what they want.
But who can claim the millennium prize now?
Disproving it would have been so much cooler.
I feel sorrowful about this. Some sort of melancholy mixed with disappointment, and nostalgia. I suppose it was bound to happen eventually.
On another note, I always thought the Riemann hypothesis was false, with a counterexample that was too large or too small to be enumerated.
24 hours later:
'Oops I made mistake and got carried away'
Wanna bet?
Yes, I bet you 50 thousand dollars this doesn't happen.
17M+ lines, 1 GB of Lean in 175K+ files. Good luck of understanding all of that in one’s lifetime.
If true, human mathematicians are obsolete. They cannot hope to explain or build on a 250k page proof. Today was the day human mathematics died.
So, I am convinced that we are already screwed (or saved) whatever you think that statistical likelihood of AI based 'misalignment' induced human annihilation risk is.
Anyone who hasn't played this game should: https://www.decisionproblem.com/paperclips/
Its very precient.
Divide and conquer and problem/sub-problem composability can be used to solve any known problem that is a solvable one.
I understand the copium of the anti-ai guys. Its called being in denial. But we have to face reality here. Yeah, diminishing returns, yada yada. Problem is the Billionare with the most available compute can now get the swarm to push back each of these diminishing return walls simultaneously by this own swarm.
Classic attention based transformer architecture is very computationally inefficient, its O(n^2) on the input. So what happens when the ai swarm finds a universal and provabale way to make the same model with O(log(N)) or O(N) complexity? All available compute becomes exponentially more powerful overnight.
What if such a thing is impossible or too difficult? Well maybe it can just fundamentally increase the production capacity of every factory producing compute x1000 or x10000 etc. There is no power limitations for those numbers or even resource limitations when we have nuclear power, geo thermal, and the entire earth's crust at its disposal.
Ok, barring that, now it can work on improving robotics and toolsets tied to operating factories and begin exponentially building more factories until we simply multiply the number of factories producing chips 1000, 10k, etc.
Or most realistically, all of the above, always, all at the same time.
More efficient software, more efficient chips, more chips produced, more factories producing chips, etc.
Every step towards increasing the number of simultaneous coordinating swarm members will push the intelligence higher, and the capability of solving or improving its own intelligence also higher.
At this point we have at most a few years.
inb4 its just claude convincing even more humans (trust AI managed to convince me at gpt3.5)
We now know more than we understand.
Excellent observation
anyone else has hidden RH manuscript? now is the time to whip it out
are we in the singularity now?
We might soon run out of things to measure progress with.
A new era for mankind has undoubtedly dawned.
All that is solid melts into air, all that is ineffable crystallizes.
Looks legitimate to me. This is a heroic-looking ascent, likely not feasible through human-effort alone.
1. 10/6/26: Stepanov determinant. 2. 10/9/26: Dp ≡ 0 on a CM curve for p inert in K, D the invariant derivation, and Da·θ(0) Ωa is divisible by p a p + a p3 +···−O(log a). 3. 10/19/26+: qN = M k + R, using x1 · x2 + · · · + xk−1 · xk, wanting "automorphy" of N since μ is automorphic from θ on sixfold cover of GL5 as a period integral of thetas, quadric has natural interpretation on corresponding CM e.g. C/Z[ω] as Weil5, put Poincaré bundle on the edges to get θs. 4. 10/6/26+: W (1S ) ≥ 0 via a determinant, interpret as Stepanov with multiplicity one again; bootstrap!! 5. 10/20/26: definitively switch to Λ aka∑ n≤X Λ(n) ·( X n − 1) + − X log X + (1 + γ) · X; Guinand-Weil had the sign, variance test function trick to make things look trace-formula-like (but how can I pigeonhole / Turán over n? X? it?), Mehta-like integral over twists or at infinity. . . 6. 10/8/26: W( 1(0,1] + (∼ Xη ) · 1(0,1/X]) ≥ 0 gives > 1 − η. 7. 10/14/26: Determinants for W (·)’s containing the desired periods, indeed GGP periods, with values 0, 1 meaning interlacing weights, or up to m non-interlacing ones. 8. 11/15/26: Iterated theta lifts U (X0) → U (X1) → · · · → U (Xm) between totally definite unitary groups, π ↦ → π ⊞ μj [t], wins multiplicity m. Time to take m like √X. . . 9. 11/15/26: Norms of the lifts and level and cancellation of order-m terms with proven divisibility via Chowla-Selberg, Weyl, and Siegel mass. 10. 11/15/26: Sufficient lower bound on divisibility of the m-fold iterated lift. 11. 10/6/26: Iwasawa tower via analogy to pigeonhole on Fqn /Fq . 12. 11/15/26: Seesaw: the period of the iterated lift is a Bessel period independent of the Hecke characters, Rallis. 13. 10/6/26: Bost slope inequality. 14. 11/15/26: precise computation of the "cost" of each lift, showing that one would stop at 2/3 a priori, attempt to get a main term and transfer constant to a function field or kill via twisting, need for "division". 15. 11/15/26: Glue together an integral structure across Hecke characters for a "gcd" i.e. extra averaging. 16. 10/14/26: Relative trace formula, seen in a different half-plane argument via sixfold cover of GL5 automorphy of μ and reducing via Cauchy- Schwartz to a bound on the norm of a suitably projected Poincaré series, approximate orthogonality of the columns of the integral structure. 17. 11/15/26: for the classes in H1( oE[ 1 ℓ] , Qℓ/Zℓ(−e)) that we couldn’t kill by twisting, fight for a significant weakening of Vandiver.6 18. 11/15/26: computation of covolumes, theta series’ index, congruences between lifts and forms. 19. 10/19/26: precise choice of the Iwasawa tower, specifically using Behrend!!7 20. 11/16/26: GRH
man, i guess we didnt get a solved reimann hypothesis before gta 6
If Reimann hypothesis is true, then Goldbach conjecture is true as well. It's major.
GRH only implies the Weak Goldbach Conjecture, which was proved independently already anyway. It does not imply the Strong Goldbach Conjecture, ie that every even number > 2 is the sum of two primes.
Not exactly. Levent claims to have proven the generalised Riemann hypothesis. The GRH only implies the ternary Goldbach conjecture, though.
He’s talking about a recent proof of the Goldbach Conjecture that was conditional on GRH. (It was also conditional on some other stuff though, so he’s incorrect.)
A proof of the twin prime conjecture might come soon too?
> If Reimann hypothesis is true, then Goldbach conjecture is true as well.
How so?
the brawniest minds, with the brawniest machines...
Is this last normal year?
Anybody has any opinion on what the impact on Bitcoin encryption will be if GRH is verified?
I'm certain the guy working on this is much much better at math than I'll ever be. But his remark.pdf looks like it was written by someone who has been spending too much time with Claude.
We used to call this style of writing delusional. Now we call it a proof of RH.
That's insane
This is currently #1 on the home page. And there's apparently a 250,000 page Lean proof.
And #2 is Feynman's saying that "If you can’t explain something in simple terms, you don’t understand it".
So: is anyone going to be able to explain or understand this? Or is that the old paradigm now?
The understanding something is a unc copium take on intelligence.
For a long time, it was the ONLY way to ensure that something could be used in the world or synthesized further.
Now it's not
The general public doesn't read proofs. They're CS graduates cosplaying as mathematicians and just assume the proof to be true and useful. They don't understand and they don't care. I'd like to see some of these people back in their basic engineering classes, asked to prove Shannon's Sampling theorem.
It's crazy that I had to scroll down through hundreds of comments to find this. So true! It's nuts that people take a 250K + lean document as proof. No overarching ideas are presented as to why the conjecture was true, just a "trust me bro I got it". Crazy how so many people are buying into it. I'm still waiting for a nice document summarizing the Navier stokes proof.
Welcome to the new age of exploration, where you get to decipher the future's scrolls, written by slop giants.
In light of this, my recent comment, with latentsea's subcomment is actually fucking hilarious.
https://news.ycombinator.com/item?id=50492141#50493545
We need LLM's to propose new problems rather than solve it now, that would require human in the loop to see if it's actually interesting.
Why does it need to be interesting?
We don't believe that art needs human emotion or intent anymore, nor do we believe that code needs human skill. Why should science or mathematics need human interest or comprehension?
I am trying to reproduce the lean on a 192-core aws machine, live on: https://cloud.infero.net/rh-en.
lean compile successed. kernel replay verified. res is in the same url.
What's the direct impact of having this proof? Is anything going change or any changes needed in the real world things?
Reminder, this "proof" is a quarter of a million pages. Every person singing praises of AGI here has yet to even read the small debrief above, much less actually down a single part of the proof.
What happened to the skepticism of Hackernews? People just believe whatever Anthropic tells them now.
It's gonna be fun to see if the LLMs end up vindicating all those "crack pot" vixra proofs of the GHR that have accumulated over the years
Imagine being an employee at one of the largest AI companies. You can do this:
grep -ri "Riemann Hypothesis" /all-users-chats
Then let an LLM sift through everything, identify potential clues, and pick out promising candidates.
Just think about it: how many people use such LLMs and talk to them about their proofs or research? or their moonshot ideas?
You can have a super-high IQ, but you may not find anything worthwhile, yet a guy less intelligent than you may discover something extraordinary.
To capture that, you need a numbers game. One among billions of dumb people has a higher likelihood than one among 1,000 geniuses.
And LLM companies have all geniuses and ordinary people contributing research for free.
It would be foolish to assume, that people working for LLM companies are coming up with proof all by themselves.
I wish more were shared on how deidentified and anonymized user session data is used when model training is off.
There is a big murky gray area around synthetic data, where and how it is generated, and where and how it may be used for things like this, or model training (because if it is de-identified it is no longer "your" data) or ads measurement and targeting. More transparency is needed there.
It's almost like everyone was told not to submit all their e-mail (at first) to Google, but hey, who wants sysadmin job openings, amirite, also they can do no evil
I was here!
If this gets verified I can’t help but think that future historians will look on today with a lot of significance. I also can’t help but say I was here.
Personally I feel the impact of this is just that if Reimann can be solved with Claude, everything else is just around the corner. In the not-so-far future the hot topic in mathematics will be conjectures humans cannot understand, with myriads of 1000-page lean proofs.
Yes I agree. I feel like P != nP will be proven a lot sooner than we think at this point.
crazy times
What is the significance of this for us pheasants?
I'm dreading a scenario where we solve this and P=NP and figure out non causality violating FTL and all this wonderful theoretical stuff but none of it matters and the world burns anyway
That seems possible.
The reason for no aliens could be that the universe is fragile, and once you find the magic of its construction, you unlock vacuum collapse, FTL paradoxes, and construct black holes.
Smart aliens just kill themselves because the structure of the universe is a cliff.
If you don't happen to be a Math PhD but want a more detailed description of this problem that's closer to the math than the average popsci Youtube video that makes weird analogies, this is about the best I've found [1].
[1]: https://www.youtube.com/watch?v=yCZgKep5iBc
I was here
Says the proof has been formalized.
The real danger is just around the corner, and it has nothing to do with superintelligence: if these models are so good at math, they might break modern cryptography. All it takes for massive worldwide chaos is these models solving elliptic curve cryptography. Superintelligence or not, if these models can solve cryptographically relevant math questions, faster than we can invent new cryptographic protocols, everything we rely on digitally can break. Imagine the internet without encryption. That includes credit cards, operating systems secure updates in a time where models find vulnerabilities every day, etc. I would start pushing post-quantum cryptography protocols not because of quantum computers, but because potentially any day now a model will come and break everything, and we need to have backup.
Cryptocoin in shambles and hard cash returning? Sign me up for this future.
No one but people who own crypto will care about it if modern cryptography is broken. I think most people will be too preoccupied by the "all data sent over the internet can now be decrypted and read by anyone" problem. We're talking about a breakdown of all internet-connected systems, anything controlling secure and sensitive infrastructure, disruption of almost all manufacturig, shipping and logistics worldwide, the leak of all identities, secrets and communications, and likely downstream mass societal unrest and wars.
>We're talking about a breakdown of all internet-connected systems
No, that is false. Loss of existing confidentiality, yes. There are other ways to prove integrity without primes, such as Vernam ciphers.
Even if there are alternate ways to achieve secrecy, the publication of something that takes down the current encryption paradigms would still have the effects I mentioned. It would take months for everything to transition to a different scheme, by which time everything important will have already leaked. And that's not to mention all the encrypted data that's being saved now to be decrypted once the algorithm is broken. With how much we integrated the internet into everything, it would be a civilization-wide disaster.
Brother, if cryptography is broken, cryptocoins are the least of our worries....
It really isn't. We had the clacky copiers in the 80s and 90s with paper checks and everyone ate and slept and drove to work just fine.
https://en.wikipedia.org/wiki/Credit_card_imprinter
Not sure I understand the argument here. Having to abruptly switch back to such a system from the current one would be catastrophic.
Exactly, and it's incredibly unlikely, that in the event this did happen, it wouldn't be preceded by massive amounts of pain. There's no world in which society says, ohhh cryptography might be broken, let's go back to pen and paper for ledgers, exchange, etc. Even though we tend to classify crypto as it's own thing, the entire traditional finance system at this point is digital for the most part. All of it is going to have a bad time if cryptography is broken.
My point is knowing where primes are doesn't destroy digital comms, it just stops the ability to fully trust data on public lines. We had decades of sending and storing financial, medical, and all other sensitive info fully unencrypted; we would certainly do it again if we had to.
OK I laughed.
And actually, come to think of it it in more serious terms: we'll adapt.
It'd be pretty hilarious if what AI implied all along is that we all had to touch grass and exchange copper coin.
One of my recent, fully human, mathematical discoveries relies on the Reimann Hypothesis. It's likely to still get dismissed because I don't have a gatekeeper backing me up, but I can rest easier knowing this. I know I haven't been nice to Anthropic but I truly appreciate this. Now all the haters will have to find something else to trivialize about my work!
GRH is a juggernaut. Nay, a keystone. Now number theory can get tidier and more user-friendly.
How does a gigantic unreadable proof make number theory user friendly?
Many otherwise unverified hypotheses in number theory are automatically true if GRH is true, which we now know it is. Now we can drop all of those messy caveats.
That is great, but why Anthropic and other companies are allowed to hoard and gatekeep resources? Businesses and consumers already struggle with buying computing at prices they can afford.
To me this reads like: "I bought all your water and solved my thirst, peasants!"
It’s funny that this happened after the IPO and not before. I’m sure he timed it that way
He talks about it in the paper and is quite clear that this was not his intention. Whether you believe him or not is a different story, but personally I do.
Well, it sure isn't helping Anthropic's stock. ANT closed yesterday at $97.17, opened today at $89.86 and is now trading at $86.00.
Stock bumped a few points on this news it seems then slipped back to where it was as people took profits. Math slop must be priced at already.
> I had a stream-of-consciousness jaunt
What? Why? Did nobody—not even an agent—copy edit this? Why is the first half a Turgenev novel?
Am I the only one who immediately thought, release the transcripts? I'd love to see how this came about.
It's wise to wait till it will be checked independently and let's not just rush to conclusions yet,
BUT HOLY FUCK
> Claude and I have had a great time together with such breadth in recent months, but this time Claude far outpaced me near the finish line. No, Claude flew, and danced, and sparkled near that finish line.
I quite love the remark here, this feels quite poetic. And it shows, what most experts have already said, that the correct way to use AI to do great things is the Centaur model: frontier AI + domain expert.
For now. Hopefully that holds for a long time, as the alternative is very scary.
its the new era
Now we need to build a bigger computer to understand the answer.
what is the point of a lean proof? it seems to have the same level of interest than "running" a marathon with a moped and pretend that a record has been broken.
does this proof brings interesting new subjects and understandings or not?
Very cool. Now when do we get the breakthroughs in real science and medicine?
> No, Claude flew, and danced, and sparkled near that finish line.
Side note, anthropomorphizing your stochastic parrot sort of distracts me from your argument. Like sure guy might be a genius but as someone who knows nothing about you, reading this is very off-putting.
This is like my own opinion.. not saying what he wrote is bad, hope its real TBH
It might be off-putting, but perhaps it’s time for you to face the fact that Claude (and other high-end models) are way past being stochastic parrots.
I think you're missing his argument. Claude is a tool, not a person. It's just as counterproductive to anthropomorphize my hammer as it is to anthropomorphize an LLM.
Claude is a tool, but recent evidence suggests that it is also intelligent. Your hammer does not exhibit intelligence.
Note: I'm not anthropomorphizing anything. I think the best description of Claude right now is that it is an alien, non-human mind.
I have no problem in accepting that LLMs are vastly more intelligent than humans in many areas already, but to me the concept of mind involves "beingness" and maybe even sentience, and I don't believe at all that these machines have any of that.
"Alien" is an acceptable term in the sense that these things reason in some very different ways than we do, but due to cultural norms, to me it also brings up notions of beingness. The xenomorphs in the alien movies, vile (and fictitious) as they may be, have a certain ontological quality that the LLMs do not.
I think your position is reasonable, but even if you're right, the "mere stochastic parrot" position doesn't hold water anymore.
I've personally never held that position. I always thought that even in the minimalist case for AI, "stochastic parrots" would have mind-bending implications for our understanding of verbal reasoning. I think that we're well past refuting the minimal case from a couple of years ago, anyway.
Be specific, in what sense has the actual method of action of LLMs changed, according to your understanding?
I don’t think it has changed. The difference is in the model size, complexity, and weights.
AI is a swarm of stochastic parrots, meaning they use pattern matching as their guide. With enough number of such parrots we apparently get an alien form of intelligence. In other words, don't underestimate stochastic parrots.
how many unsolved math problems does it have to solve before you drop the parrot thing
well, if you make the parrot big enough and teach it enough skills it perhaps becomes usefull, but it still is just a parrot
I guess we're all parrots then.
Yes, it looks like parrots have evolved and become extremely intelligent parrots.
What proof of the Riemann hypothesis is it parroting exactly?
Yes, if you stretch your definition far enough everything will be a parrot, including humans.
You luddites sound more stupid with each published proof.
He said "Claude flew", which fits perfectly with your stochastic parrot framing. He's not anthropomorphizing, at worst he's ornithomorphizing(?)
Can you provide a convincing argument that your brain is not also a "stochastic parrot"?
If it is I don’t see how that is an argument against the original claim :)
It's only an argument against the implicit derision.
hmm, my brain continuously experiences and interacts with the real world, rather than briefly coming into existence to process a virtual request before vanishing into eternity?
It kind of sounds like the only limitation or distinction then is processing speed. The tick-rate of intelligence. LLMs have a very slow tick rate.
But, if I had a camera or something and constantly fed that to a model inside a harness, is it then a brain? I mean that’s basically exactly how the brain works - it gets electrical stimuli from a bunch of places, makes some decisions, and then spits stuff out. The loop is just much, much tighter.
When an AI gets a body and starts continuously experiencing and interacting with the real world, initiating its own actions without waiting for someone to prompt it, I'll accept that we are fundamentally the same.
The distinction is whether it can generate its own reasons to act based on real world stimuli, rather than having every interaction initiated by someone else. And it's not just about hooking a camera to a harness to provide vision, it is also hearing, smell, taste, touch, thermoception, nociception, proprioception, vestibular sense, interoception, hunger and thirst, social stimuli, etc. In order to deal with all these stimuli feeds your harness needs continuous memory and learning ability and internal drives. It needs to be a continuously operating, self-directed intelligence, which it isnt right now.
And we also need to remember that rather than explicitly training a brain to recognize objects, understand language, and solve problems, brain has the mechanisms to develop those capabilities through interaction with the world starting from roughly 800 MB of genetical information to produce 86 billion neurons and hundreds of trillions of synaptic connections. That development is not a direct assembly from some detailed blueprint, it is generative, it does not separate training from inference. I think all of that combined explains why we call LLMs a "stochastic parrot" and brain isnt merely a system that produces plausible responses to inputs. Its a continuously developing, embodied, autonomous process that actively interacts with the world, generates its own motivations, and learns from the consequences of its actions.
> stochastic parrot
This has to be satire, right?
Could be a time traveller from 2024.
The anthropomorphizing language is creepy and disgusting.
Well, now that we are at this point I have to decide if I should kill myself or AI execs. Maybe I'll ask ChatGPT.
If you are being serious... please don't think like and consider taking some time away from the internet.
Many people will be asking about the relevance of this, real world applications, etc. I think that's missing the point.
This is like climbing to the peak of Mt Everest.
There is no gold there, no riches, the view is grand but not that unique. Nobody will build a multi-billion-dollar business empire based on Everest tourism.
But it is a benchmark of capability! The locals that had lived there for millennia failed to do it. Getting there required technology, technique, and even some specific developments like electrically heated boots. The methods used are akin to going to space, with protective clothing (vaguely) comparable to a space suit, including oxygen tanks!
It's not about the goal, it's about the ability to get there, which is a mere milestone in the exponential improvement:
In 1953 a human stood on the peak of Everest.
In 1961 a human orbited the Earth.
In 1969 a human walked on the Moon.
In 2026 an AI solved the most famous unsolved maths problem.
In 203x ... ?
The (alleged) proof is 15 million lines of code I hear. Meanwhile, the #2 item on the HN frontpage right now is "If you can't explain something in simple terms, you don't understand it".
This is honestly not that surprising. Quasi-Riemann Hypothesis was apparently thought to be totally unreachable by number theory experts before the proof dropped by OpenAI in October. So I think we are at a point where mathematicians' opinion on whether a conjecture is easy or hard doesn't really say much about the frontier models' ability to solve it.
Regardless, this is a historic moment for science and humanity in general. We still have 6 weeks to go, but I think 2026 will be seen as a pivotal year in history. The year when humans stopped being the smartest species on Earth.
It's hard not to feel excited about the future and, obviously, a little bit scared.
Something interesting is that after quasi-Riemann (which proved an upper bound of 7/8), experts agreed that the techniques in that proof couldn't get past 3/4. They probably weren't wrong since this new proof goes in a different direction.
Food for thought for those who think "AI is going to replace programmers" or some other such incredibly silly takes:
"Tens of such occurred, but I will give one line which will suffice to make fellow mathematicians fall out of their chairs. In Stepanov’s argument resolving Riemann over function fields one uses a divisor involving multiplicity. This argument is a descendant of Stepanov’s, but multiplicity is hard to find over the integers, corresponding to something like differentiating whole numbers. The simple, new, Fourier-analytic ideas in the toy setting led us to examine chaining theta lifts if we were to have the same tricks recur but work in a natively automorphic manner. The line: for the von Mangoldt sum of length X, Claude’s suggested length of chain of theta lifts, in view of each mimicking multiplicity, was √X. (And yet, what else could it be?)"
Raise your hand if you still think "my gradma could have used AI to do X".
This title seems weird. Why just "anthropic employee", not specifying that hes a mathematician, or just including his name? Without the clarifications in the replies, many will likely assume this was a random Anthropic employee just throwing tokens at the problem, not a very well known mathematician working with the model. Probably not the intention, but these kinds of titles feel like they're trying to shift the credit from the person to the company and their models.
> probably not the intention, but these kinds of titles feel like they're trying to shift the credit from the person to the company and their models.
Unequivocally that is the intention.
To some extent they should, given he’s using models that aren’t publicly available and resources that, while not easily measured, probably run into the millions.
Oh I'm certain Anthropic (and any other company) would do this intentionally, I just didn't want to accuse OP of it as I have no idea if they are in any way affiliated.
It's totally unfair. A more standard title would be:
"NYU mathematician working with Claude claims proof of the Riemann Hypothesis"
In newspapers, professors are always associated with their universities, not the commercial relationships they have.
> In newspapers, professors are always associated with their universities, not the commercial relationships they have.
The title is probably what it is because Levent Alpöge is not a professor at Harvard (or anywhere as far as I know) nor did he do this work on their behalf (or any academic institution's).
He's employed by Anthropic and used their technology. Obscuring that in the name of some semblance of journalistic fairness would actually be a disservice.
I think the title is vague, but associating him with a university is irrelevant to that.
Alpöge is not an NYU mathematician, he is an Anthropic employee. You may be thinking of Buckmaster, who worked with Alpöge on Navier-Stokes.
Ah you’re right, my mistake.
It would be better to lead with “mathematician”, as he’s not a random employee.
Indeed, until I saw the top comment identifying Alpoge (and after skimming over the post), I was thinking this was a mathematically-inclined Anthropic employee, not a full PhD research mathematician with a wide track record of accomplishment in the field.
If he is on the Anthropic payroll, on company's time, then Anthropic can claim credit for the discovery and put itself forward. In fact that is probably the only reason why they worked on the problem, for publicity.
So of course, they want the name of the company and its product to go before the person's name.
Also, I think that AI companies have the best mathematicians in general. AI is math-heavy, and they pay well, and besides education, there are probably not that many work opportunities for mathematicians, that should make AI companies particularly attractive for them.
Anthropic can claim the credit, and I'm sure they will cause that's marketing. I'm criticizing the title of this post chosen by OP, because I don't think the post should be marketing for Anthropic.
Anthropic is a company filled to the brim with nerds obsessed with math. I'm sure the publicity isn't the only reason they worked on the problem. The public already understands that AI is good at math - to them, there's hardly a difference in significance between Navier-Stokes and GRH.
Did you read the remarks? You can argue Levante has a monetary incentive to credit his employer, but the actual human mathematician reporting this result clearly gives most of the credit to the AI model
He definitely gave a lot of credit to the model, but he also states that he found it with Claude, not that Claude found it. I have no idea if more of the credit belongs to Claude or to him, and I don't really want to speculate about monetary incentive.
Regardless, my point with the title was that a mathematician discovering this by working with Claude, versus a researcher with little to no relevant experience, using unlimited compute to discover it without much input, are two very different things. Without specifying the former, I think some would assume the latter.
Who would predict that the proof of the RH would mark the end of mathematics as we know it? From now on this field should close shop as we'll be fighting between machine slop and the total impotence of human mathematicians...
>It was my secret in my student days that my plans were for this problem in decades, which meant then that I tried to become familiar with all tools that I could possibly recognize in the problem when it was time. Claude and I have had a great time together with such breadth in recent months, but this time Claude far outpaced me near the finish line. No, Claude flew, and danced, and sparkled near that finish line
yeah this is just another advertisement
Oh no! Whatever will we do! Should we shut down the whole Internet because someone, somewhere, is trying to make money? What a bastard!
For years, I've held RH in mind as my personal benchmark for... something. AGI, the Singularity, or whatever.
We don't know yet whether this proof will hold, or how bad this stock market rout will be. But it would be a supremely funny ending to this whole story, for me at least, if they spent all these years and cap ex to build AIs, they IPO Anthropic, they prove Riemann, and then in that moment the whole bubble bursts and the hype collapses.
It's so funny to watch this.
They are quite skillful at obscuring the catastrophic drop in their stock price within news feeds.
It’s a very brazen move by Anthropic — using 17 million lines of unverifiable Lean code, along with mathematicians, to mask their decline on the stock market.
Oh come on. I've remarked elsewhere that the anthropomorphizing language is creepy, and I'll add here that it sounds psychotic. But give them some credit on the technical side and on not having the incentives to make utter fools of themselves. The stock is there to stay for a much longer term than the next few weeks.
Ah yes, tech bros got owned again - they’ve proven Riemann hypothesis to mask stock price! Do you realize how absurd that sounds?
An unreadable proof is basically none at all. Suggest you have a model produce a concise summary and work edge in from that. Apparently none of these problems has been verified and anybody working with models would be justly skeptical of these helium filled claims which are almost certainly not going to pan out.
Also apparently the Riemann claim is based on it compiling in Lean. This is classic because just the differences between an actual verification of a proof and it compiling in Lean are the standard model failure tropes, lack of judgement, confabulation, etc., to which this adds shortcomings of Lean.
What shortcomings of Lean. I'm taking it on the authority about every top mathematician on earth that it provides solid proofs.
Who said it didn? But that's different from it a) never compiling a proof that is in fact false or b) all of its proofs being solid.
This is just some problem posed by a human a Long time ago. We expect something better from ASI
Hurried Announcement: Alessandro Masullo GRH https://math216.wordpress.com/2026/11/20/hurried-announcemen...
Is this yet another case of AI company stealing almost ready proof form someone else?
And what?
Is anyone’s life going to change, be made better?
Does anyone actually understand the proof? Is it possible to even understand the 100k pages of proof?
This is just useless noise to try and juice a company’s stock price.
He has a writeup outlining the general mathematical ideas he/Claude used. Also not sure the market cares about RH.
Why the primes do not leak.
For more than a century and a half, the Riemann Hypothesis has worn the aspect of a tightrope.
Take André Weil's positivity criterion—all nontrivial zeros of ζ(s) lie on the line Re(s) = ½ if and only if a certain quadratic form stays non-negative on every test function—and measure it on a finite window with Slepian's prolate functions. By a window ratio of X = 40, the lowest eigenvalue on ζ(s) has collapsed to 2.53 × 10⁻²⁰⁴. Turn to the de Bruijn–Newman constant Λ, and by Rodgers–Tao, Λ ≥ 0: arbitrarily close pairs of zeros pinch the margin to nothing.
Through lenses like these, the critical line looks like a balancing act sustained by two hundred decimal places of fine-tuning. That vertigo is an optical illusion. The line only looked fragile because we were measuring two different quantities with the same ruler—confusing *pitch* with *linewidth*.
Separate them, and the hypothesis becomes a story in three movements: a mirror that chooses ½, a tuning fork that shows how much room the line has, and a counterfeit that shows why continuous physics alone can never close the last door.
---
## I. The Mirror and the Dark Floor
The explicit formula of Riemann, von Mangoldt, Guinand and Weil binds primes and zeros together as Fourier duals. Feed in the primes, and their interference pattern projects the zeros; feed in the zeros, and their waves rebuild the staircase of primes.
The right picture is not a bridge between two banks but a *Möbius band*: one surface with a half-twist. The twist is the functional equation ξ(s) = ξ(1 − s), which reflects the plane across Re(s) = ½. In the natural clock of the primes, u = log x, the waves are dilation characters x^(iγ) = e^(iγu), and the core circle of the band is Re(s) = ½.
That answers why ½. Any displacement δ = Re(ρ) − ½ off the centerline turns a pure standing wave into a *gain-and-loss pair*: a growing mode e^(+δu) married by the mirror to a decaying twin e^(−δu). In acoustic terms, γ is a note's *pitch* and δ is its *linewidth*. The Riemann Hypothesis says the primes are a *lossless resonator*: no mode leaks energy, and none manufactures it.
A resonator needs a sounding board. Besides the primes {2, 3, 5, 7, …}, the rationals have one more place—the archimedean place at infinity, the Gamma factor that completes ζ(s) into ξ(s). Its smooth phase accounts for 94.2% of the zeros by height 50 and 99.94% by height 1000. The prime waves are a thin ripple on that mass. Subtract it as a trend and the apparatus goes blind; keep it inside the form and it is the floor the spectrum rests on.
---
## II. Pitch Is Not Linewidth
Why did the prolate window collapse to 10⁻²⁰⁴, and why do close pairs crush the de Bruijn–Newman margin? Because a tuning fork can hold two notes arbitrarily close in pitch without losing a drop of energy.
Take the humble pair of test functions from Levent Alpöge's recent note, here shifted by a carrier pitch ω: the indicator of (0, 1] and the indicator of (0, 1/X], each multiplied by u^(iω). On their span the whole explicit formula collapses to a 2 × 2 Hermitian matrix with equal diagonal entries, whose smaller eigenvalue is one scalar:
On the arithmetic side, C₀ carries the archimedean bath and E_ω(X) is a finite prime sum over n ≤ X plus boundary terms.Now watch the tightrope. If every zero is on the line, each term of C₀(ω) is a positive Lorentzian, and moving to scale X only turns each term by a unit phase. The triangle inequality gives |E_ω(X)|/√X ≤ C₀(ω), so S(ω; X) ≥ 0 for every pitch and every scale.
And the margin is no longer microscopic. At ω = 0 and X = 40, where the prolate continuum sits at 2.53 × 10⁻²⁰⁴, this matrix sits at +0.0428—*two hundred and two orders of magnitude higher*—and never gives up more than half its diagonal across the first hundred thousand scales.
What about close pairs? The closest among the first two million zeros, near γ = 663318.5, are 0.00296 apart and squeeze the de Bruijn–Newman bound to 1.09 × 10⁻⁶. Inside the 2 × 2 fork the pair contributes nine parts in a trillion. Merge them into an exact double zero and C₀ shifts by −1.6 × 10⁻²⁷. Rodgers–Tao forbids a pitch-gap margin; it says nothing against a linewidth margin.
Now let one zero stray: a mirror pair at (½ ± δ) + it. Tune the carrier onto its pitch, ω = −t, and the pair manufactures energy out of scale, punching a negative crater of exact, parameter-free depth:
Sweep the fork blindly across 376 pitches at X = 40. On ζ(s) and L(s, χ₄), every pitch stays strictly positive—zero false alarms. On the Davenport–Heilbronn function, a famous counterfeit built to imitate an L-function, the fork strikes deep negative resonances at once, catching off-line zeros at t = 85.70 and t = 240.40 by *X = 4*; across all eight known off-line zeros the dip law matches the measured depths to R² = 0.99975.One caveat, stated plainly: the half-line test functions carry boundary terms with global information beyond n ≤ X. Truncating to a compact boxcar [1/Y, 1] removes them, leaving only n ≤ XY and the Gamma factor, and positivity and the O(1) room survive—but the counterfeit is caught later. Room on the line is free; reach off it must be paid for.
---
## III. The Counterfeit Wall
A classical physicist is ready to declare victory: a mirror that fixes ½, a floor restored, a 10⁻²⁰⁴ tightrope turned into an O(1) margin, a 2 × 2 sieve that rings clean on ζ and exposes a fake in an instant.
None of it proves the Riemann Hypothesis. The reason has a name: *Davenport–Heilbronn itself.*
It is a carefully angled combination of two Dirichlet L-functions modulo 5. It satisfies the same functional equation, sits on the same Möbius band, obeys the same explicit formula. In the untuned fork it mimics ζ(s) past X = 50,000. Yet it has infinitely many zeros off the line.
In classical wave mechanics *superposition is always legal*. Mirror symmetry survives mixing; the trace formula survives; what breaks is *indivisibility*:
- For ζ(s) and L(s, χ₄), the arithmetic weights live *only on prime powers*. An *Euler product* makes every prime an independent, indivisible factor. - For Davenport–Heilbronn there is no Euler product, and the weights bleed *across composite integers*—6, 12, 14, 18, 21, … That leakage feeds the off-line explosion.
Classical wave duality cannot tell an indivisible atom from a mixture of two.
Physics met this wall in 1900. Classical theory described cavities and radiation with precision, yet could not explain why an atom does not radiate continuously and collapse. The missing law was not sharper classical analysis. It was *quantization*.
---
## IV. Arithmetic Quantization
Look at the architecture sketched in Alpöge's `remark.pdf` (its Lean formalization still awaits independent verification; treat it as a blueprint until then). Whether or not every step survives, the three-part logic completes the picture:
First, *widen the fork*. Trade narrowband prolate precision for the broadband two-scale form and you escape the 10⁻²⁰⁴ cliff and the close-pair trap. No knife-edge spectral operator is needed, only a bound with O(1) room to spare.
Second, *demand indivisibility*. Use the Euler product—exactly what Davenport–Heilbronn lacks—to embed Weil determinants as automorphic periods along a chain of coupled resonators. A linear mixture cannot enter the chain. And a chain is needed because a single fork, however wide, only clears a strip beside the line: each further link adds divisibility, until there is enough of it to outweigh X^(δ).
Third, *collide continuity against discreteness*. An off-line zero with δ > 0 tries to pump a continuous leak X^(δ) through the sieve; the divisibility budget of the chain cannot pay for it. The whole of quantization is here in one line: a nonzero integer has absolute value at least 1, so a quantity that must be an integer, is divisible by that much, and is analytically that small can only be zero. Something must yield, and it is δ.
---
## V. The Instrument and the Kernel
If a picture like this holds—via Alpöge's chain or a successor built on the same fault line—it tells us *why the Riemann Hypothesis resisted for 167 years*: it stands on the seam between the continuum and the discrete. Pure continuity is too loose—waves can always superpose into counterfeits. Pure combinatorics is too local—counting primes one by one cannot see phase cancellation at infinity. The primes are an *arithmetically quantized wave system*, where global stability (δ = 0) is enforced by local indivisibility.
A machine-checked proof, if one verifies, certifies that a theorem is true; it rarely shows why the numbers could not have been built any other way. Analogies, sieves and counterfeits do the opposite work: they cannot seal the lock, but they turn a wall of inequalities into something you can hold in your mind and hear.
Between the kernel that guarantees truth and the instrument that reveals necessity, the primes come back into focus: a single tuning fork, twisted once by symmetry, floored by the place at infinity, and quantized by arithmetic so that it can never leak.
Solving one of the greatest outstanding math problems is just noise to you? I suppose taste in music varies
So you understand the proof then?
Are the 100k lines of lean beautiful music to your mathematical ears?
Does any mathematician even understand the proof?
I don’t find this beautiful or interesting at all, someone just pointed 1000 Claudes at this, said “go”, and it solved it. But we have no idea what it did and it doesn’t change anyone’s lives.
All that happened is the thing we thought was true already, is now proved to be true, by brute force, but we have no new insights or understanding and no practical applications.
Great I guess, meanwhile San Francisco and Los Angeles have more homeless people than ever and we can’t begin to solve that problem.
You have some interesting theories. Can you explain why this would suddenly bypass the usual few centuries long timeline between new math and applications? You somehow seem to have a theory that this general delay has changed.
I also find the timing highly convenient. It’s not being a conspiracy theorist to want to check for any new bank transfers.
From the pdf:
> The precise timing is horrible. If I could have, I’d have finished this all much earlier. ... There is an obvious cynical view, to do with the timing, but it is false, and I am pleased that I am in circumstances where it is obviously false.
It's not the kind of result you are going to sit on, and anyways not helping the stock at all today (why should it - do you choose your coding model, or decision model, based on it's math ability?).
Wouldn't they have done that a week earlier?
He claims in the OP that the timing is horrible.