This post buys into the common narrative about Xanadu, that it was a fundamentally good idea that failed due to some combination of poor execution, being too far ahead of its time, and the market not caring about quality (with this post in particular focusing mostly on the second thing).
In my view, Gwern makes a compelling counterargument: Xanadu was a fundamentally bad idea. The parts of it that weren't incorporated into the World Wide Web either solved problems no one actually had, or couldn't have constituted a stable equilibrium between users and publishers (i.e., even if it'd been magically willed into existence, it'd have quickly fallen apart). https://gwern.net/xanadu
Great post, as a fan of Gwern and Nelson I wonder how I missed it (and how I didnât get an invite to the party :â)
The âtoo genius, too earlyâ is a common trope in the tech industry, but it rarely stands scrutiny. When you inspect the claim up close you almost inevitably find out a variation of that âbad design/solves a problem no one really has/does not create the virtuous cycles needed for successâ.
But as an inventor it might be easier for oneâs ego to tell yourself âi was just ahead of my timeâ than âmy judgement was completely offâ.
If youâre an inventor of any kind, the âtry lots of small things, see which ones seem to stick, narrow down on thoseâ approach seems preferable to âcommit your entire identity to one big thingâ, but there is a certain siren song to the latter.
> But as an inventor it might be easier for oneâs ego to tell yourself âi was just ahead of my timeâ than âmy judgement was completely offâ.
This has more to do with the economy than ego. When one's ability to eat and have shelter relies on "inventing" the inventor will invent a new wheel to sell to a tribe that's pre-wheel.
This no longer works as well in software given the number of people who code, and the various forms of automation (AI or design tool that emits code, etc) to assist those who do not.
No pulling the wool with a fancy abstraction alone. It's just math relative to a matrix of desirable values loaded in memory.
The normal progression of technology coupled to economic models written down by long dead who were too busy enslaving people to invent indoor plumbing.
This economic equilibrium reality kills so many cool-sounding ideas. Someone has to actually run them, and if it's more soul-sucking than soul-giving (for some definition of "soul" which may include money) it won't happen. It's not always money, sometimes it's just really hard to keep the service up or to block spam.
Interesting. This and the other posts I've seen about Delta seems to exist in that liminal space, were it's difficult to tell if the author has had a brilliant insight, that they struggle to convey to anyone who's not yet had the same insight; or if they've simply gone slightly mad.
It's a similar sense I get with reading Nelson (after being able to think about his ideas for a while, I'm quite certain it was mostly genius and only a little bit of lunacy).
It remains to see how Delta will turn out.
On a side-note; my initial reaction to the headline was; no. Xanadu wasn't missing agents.
I do think I see where the author is going, though; and I think I agree that agents might be a new tool that might facilitate some hypertext ideas. Not Xanadu perhaps. But some other, more pragmatic and less magic place - but a place that is a little more magical than where we are now.
And it intersects a bit with an experiment we started at work; migrating our modest web site from WordPress to a static site combining 11ty (https://11ty.dev) (we already lean heavily on typescript for our application development) and visual studio code with frontmatter cms (https://frontmatter.codes) - and AI/LLMs.
AI helps with making a static page version of our site, and adopting to 11ty.
The idea is that with trivial push-to-deploy setup, our business people, with the help of AI in vs code, will be able to work directly with 11ty and git to make changes, do reviews and publish changes.
A hill that would probably have been too steep for them to climb just six months ago.
So in reference to the theme of realizing a better hypertext system with the help of AI, we're sort of doing the same thing, it turns out: mix solid, boring tools (static site gen, typescript, editor, version control, CI, static hosting) and some AI - to hopefully allow more users leverage both tooling and AI in an empowering way.
We've just stated the experiment - we may yet fly too close to the sun, and end up falling back to a traditional CMS.
I think not are the same the only thing that differentiates them is adoption. If people show interest in the thing you are a genius if they don't you are mad/crazy.
One reason I have some hope for this approach is that we start by removing complexity by going to a static web site first.
We remove WordPress, database, configuration, templates - a lot of cruft.
Then we can add it back as needed for keeping edits and publishing simple. It might turn out we hardly need anything from 11ty, thanks to AI making editing more automatic.
I think (the collective) we, are (mostly by accident) building systems that are at least one dimension or level of higher abstraction than before AI.
Unfortunately with AI, higher abstraction doesn't automatically mean less complexity, or better managed complexity.
I think 95% of systems being built now with Al and on top of - or including Al - will turn out to be highly complected mountains of madness.
Complexity in itself is not an interesting measure but it's a proxy for an interesting measure: time and cost to build/maintain.
A more complicated program that is built and maintained with LLM assistance is cheaper and faster to build and fix than even a less complicated program built without. It's even the same without LLMs: once React achieved critical density and "won" the JS framework wars, building React apps became cheaper and faster even than building plain HTML/JS apps.
This sounds like one agent at each junction point, that may or may not be running the same model and architecture as the previous one, and they need to constantly crawl each other or just a protocol to call each other.
But then you have a million of then constructing a docuverse web all having an agent who all need to know each other and its all connections and each a database repository.
Well you only need links that exist, and those that are created.
In a paragraph it may exist in one node that is referenced and that node may be a collection of other nodes, or each sentence or even each word points to other nodes, how much does each model know and what does it keep in its database.
A point to the next node is not necessarily durable or versioned, and that goes for each node is connects to.
A single paragraph could have thousands or more sources, references, and each source would need versioning and depending on version it may well point to different nodes in prior or future versions.
A proper version of a paragraph would then need to know this, or expect that all nodes with an agent will be durable and rational and interoperable
but if you build on that premise it will not work in the real world.
All unless you keep the docuverse limited in scope to a few data stores and agents who comprehend that spatial universe, and it is a form of an intranet
One of my favorite LLM-smells to train myself to sniff out is how well the author can tie two seemingly disparate ideas together into something compelling or novel. The more ambitious the synthesis and the less impressed I am the more likely I am to assume it was at least partially the work of an LLM.
The content was clearly AI written. It made no sense. Even crazy humans such as Terry Davis made more sense than such AI slop generated "blog" entries.
That this is published by the lead of a pretty popular piece of software is disconcerting. The UI for the IDE itself and the web design for Delta even have that Claude Code "Claudhaus" look to them.
I'm beginning to think that this sort of behavior is actually an inverted counter-signal. The average person attuned to LLM's markings may find these things distasteful. But certain individuals or organizations may feel right at ease seeing them.
Reading this makes me wish someone would really try to make Project Xanadu a reality. A network built on Content Addressable Storage and a Version Control System is some thing I'd really like to see invested in.
It's not a reality for a lack of trying. Multiple people have spent (literal!) lifetimes working on Xanadu and multiple businesses poured a great deal of money and effort into it way back in the day.
It continues to not be a thing because, even though the idea is superficially appealing to a certain engineer mind and that's why people get sucked into the cult, in reality it's not something anybody wants and it's solving a problem nobody has.
Right. I knew those people back when Autodesk owned Xanadu. The fundamental problem was that tracking and paying for text ownership and viewing down to the word level isn't really very useful. The Xanadu people were libertarians into micropayments. Everything is pay per view in Xanadu. Not cheap views, expensive ones, like Lexis/Nexis.
Also, they had no idea how to deal with either images or distributed systems.
> Also, they had no idea how to deal with either images or distributed systems.
What would make images harder than anything else? My naive self thinks, "hey, isn't that just one long base64-encoded word?" Apparently not, or they would've done it and moved on.
And why do you say that about distributed systems? What I've read of Xanadu (primarily the Wired article) made it sound like there'd be a relatively small number of commercial servers everyone else would subscribe to. Cue IBM's "five computers are enough" apocrypha. Is that what you mean?
I find everything about Xanadu fascinating, in kind of a slow-moving train crash sort of way, and would love hearing more bystander takes!
> What would make images harder than anything else?
Because diffs of image files are not very useful. Nelson's thing was very fine-grained tracking of who wrote what. That doesn't map well to images or video.
I just finished writing a long rant on the other Project Xanadu related submission today where I specifically noted that Xanadu has a feature called a "permascroll" where everything you publish gets written into the permascroll and then transcluded into actual documents. This is a very "video editor" concept. Actually, the Xanadu model of transclusions and history would translate extremely well into nonlinear video editing; it's arguably better for video than it is for text.
Images wouldn't work unless you made a "Xanalogical image editor" that stores a list of edits and references to source images as you made or imported them. The tumbler number structure would have to be extended to include rectangular references into image files. It would work in theory, but in practice would probably bog down every computer that needed to load an image, and you'd get hit with shittons of fees for each one.
I can only imagine audio having to be stored as some kind of standardized DAW project file, with transcluded sample packs that would get you micro-transactioned for each one.
A system that never forgets is not always a good thing. I'm not sure what the maximal scope (beyond editors) the author wishes this applied to is, but there are many cases where it is a positive for things to be erased. Privacy concerns is one example.
> So when web technology exploded in the '90s, developers bypassed xanalogical computing in favor of ease: links are simple strings which break if their target moves. Maintenance became users' problems â instead of the system's responsibility â but it was easy for anyone to ship fast, and they did.
developers bypassed - Really not a fan of phrases like these from a CEO.
Usually developers have to work within an organisational hierarchy that doesn't grant them infinite time or money.
Well let's not forget Zigzag. This is hyperstructure and Ted's greatest achievement. If you want to try it out, Gzz 0706.jar runs on Java https://www.nongnu.org/gzz/download.html
Why would anyone want to put out accurate structured data in the AI era where unpoisoned data is worth a hard drive's weight in gold? Structured data will all be like: we open on <date day="Monday">Friday</date>
Interview him while you still can. I encourage you to seek out whoever Ted's contact people are and try to get that interview in. Make it long and thorough, and put it on YouTube and archive it.
If not, someone is a native in LLMese. Conveys little about what it actually is, or connected to how Delta works and is mostly talking about some abstract metaphoric relation. I remember some article on how LLMs throw in identifiers without much understanding of the indetified, other than its similiarity to another thing in text (methaphoric speech).
This post buys into the common narrative about Xanadu, that it was a fundamentally good idea that failed due to some combination of poor execution, being too far ahead of its time, and the market not caring about quality (with this post in particular focusing mostly on the second thing).
In my view, Gwern makes a compelling counterargument: Xanadu was a fundamentally bad idea. The parts of it that weren't incorporated into the World Wide Web either solved problems no one actually had, or couldn't have constituted a stable equilibrium between users and publishers (i.e., even if it'd been magically willed into existence, it'd have quickly fallen apart). https://gwern.net/xanadu
Great post, as a fan of Gwern and Nelson I wonder how I missed it (and how I didnât get an invite to the party :â)
The âtoo genius, too earlyâ is a common trope in the tech industry, but it rarely stands scrutiny. When you inspect the claim up close you almost inevitably find out a variation of that âbad design/solves a problem no one really has/does not create the virtuous cycles needed for successâ.
But as an inventor it might be easier for oneâs ego to tell yourself âi was just ahead of my timeâ than âmy judgement was completely offâ.
If youâre an inventor of any kind, the âtry lots of small things, see which ones seem to stick, narrow down on thoseâ approach seems preferable to âcommit your entire identity to one big thingâ, but there is a certain siren song to the latter.
> But as an inventor it might be easier for oneâs ego to tell yourself âi was just ahead of my timeâ than âmy judgement was completely offâ.
This has more to do with the economy than ego. When one's ability to eat and have shelter relies on "inventing" the inventor will invent a new wheel to sell to a tribe that's pre-wheel.
This no longer works as well in software given the number of people who code, and the various forms of automation (AI or design tool that emits code, etc) to assist those who do not.
No pulling the wool with a fancy abstraction alone. It's just math relative to a matrix of desirable values loaded in memory.
The normal progression of technology coupled to economic models written down by long dead who were too busy enslaving people to invent indoor plumbing.
This economic equilibrium reality kills so many cool-sounding ideas. Someone has to actually run them, and if it's more soul-sucking than soul-giving (for some definition of "soul" which may include money) it won't happen. It's not always money, sometimes it's just really hard to keep the service up or to block spam.
Interesting. This and the other posts I've seen about Delta seems to exist in that liminal space, were it's difficult to tell if the author has had a brilliant insight, that they struggle to convey to anyone who's not yet had the same insight; or if they've simply gone slightly mad.
It's a similar sense I get with reading Nelson (after being able to think about his ideas for a while, I'm quite certain it was mostly genius and only a little bit of lunacy).
It remains to see how Delta will turn out.
On a side-note; my initial reaction to the headline was; no. Xanadu wasn't missing agents.
I do think I see where the author is going, though; and I think I agree that agents might be a new tool that might facilitate some hypertext ideas. Not Xanadu perhaps. But some other, more pragmatic and less magic place - but a place that is a little more magical than where we are now.
And it intersects a bit with an experiment we started at work; migrating our modest web site from WordPress to a static site combining 11ty (https://11ty.dev) (we already lean heavily on typescript for our application development) and visual studio code with frontmatter cms (https://frontmatter.codes) - and AI/LLMs.
AI helps with making a static page version of our site, and adopting to 11ty.
The idea is that with trivial push-to-deploy setup, our business people, with the help of AI in vs code, will be able to work directly with 11ty and git to make changes, do reviews and publish changes.
A hill that would probably have been too steep for them to climb just six months ago.
So in reference to the theme of realizing a better hypertext system with the help of AI, we're sort of doing the same thing, it turns out: mix solid, boring tools (static site gen, typescript, editor, version control, CI, static hosting) and some AI - to hopefully allow more users leverage both tooling and AI in an empowering way.
We've just stated the experiment - we may yet fly too close to the sun, and end up falling back to a traditional CMS.
I think not are the same the only thing that differentiates them is adoption. If people show interest in the thing you are a genius if they don't you are mad/crazy.
One reason I have some hope for this approach is that we start by removing complexity by going to a static web site first.
We remove WordPress, database, configuration, templates - a lot of cruft.
Then we can add it back as needed for keeping edits and publishing simple. It might turn out we hardly need anything from 11ty, thanks to AI making editing more automatic.
I think (the collective) we, are (mostly by accident) building systems that are at least one dimension or level of higher abstraction than before AI.
Unfortunately with AI, higher abstraction doesn't automatically mean less complexity, or better managed complexity.
I think 95% of systems being built now with Al and on top of - or including Al - will turn out to be highly complected mountains of madness.
Complexity in itself is not an interesting measure but it's a proxy for an interesting measure: time and cost to build/maintain.
A more complicated program that is built and maintained with LLM assistance is cheaper and faster to build and fix than even a less complicated program built without. It's even the same without LLMs: once React achieved critical density and "won" the JS framework wars, building React apps became cheaper and faster even than building plain HTML/JS apps.
This sounds like one agent at each junction point, that may or may not be running the same model and architecture as the previous one, and they need to constantly crawl each other or just a protocol to call each other. But then you have a million of then constructing a docuverse web all having an agent who all need to know each other and its all connections and each a database repository.
Well you only need links that exist, and those that are created.
In a paragraph it may exist in one node that is referenced and that node may be a collection of other nodes, or each sentence or even each word points to other nodes, how much does each model know and what does it keep in its database. A point to the next node is not necessarily durable or versioned, and that goes for each node is connects to.
A single paragraph could have thousands or more sources, references, and each source would need versioning and depending on version it may well point to different nodes in prior or future versions.
A proper version of a paragraph would then need to know this, or expect that all nodes with an agent will be durable and rational and interoperable but if you build on that premise it will not work in the real world.
All unless you keep the docuverse limited in scope to a few data stores and agents who comprehend that spatial universe, and it is a form of an intranet
One of my favorite LLM-smells to train myself to sniff out is how well the author can tie two seemingly disparate ideas together into something compelling or novel. The more ambitious the synthesis and the less impressed I am the more likely I am to assume it was at least partially the work of an LLM.
Previously, "Jimi Hendrix Was a Systems Engineer" <https://spectrum.ieee.org/jimi-hendrix-systems-engineer>
Ah, so like the distance between ambition and impact in metaphor/synthesis?
I like that.
Exactly.
The content was clearly AI written. It made no sense. Even crazy humans such as Terry Davis made more sense than such AI slop generated "blog" entries.
That this is published by the lead of a pretty popular piece of software is disconcerting. The UI for the IDE itself and the web design for Delta even have that Claude Code "Claudhaus" look to them.
I'm beginning to think that this sort of behavior is actually an inverted counter-signal. The average person attuned to LLM's markings may find these things distasteful. But certain individuals or organizations may feel right at ease seeing them.
Ummet Ozcan - Xanadu: https://www.youtube.com/watch?v=9uMtnH7cABg
I don't think he's waiting for agents.
> Xanadu had a final failure mode, this one self-inflicted
Skimmed the article, saw this, and then my AI fatigue closed the tab.
Reading this makes me wish someone would really try to make Project Xanadu a reality. A network built on Content Addressable Storage and a Version Control System is some thing I'd really like to see invested in.
Youâll hear/read many people opining that it actually wasnât a good idea.
I counter that itâs a great idea for many things. The problem is that something good enough and far cheaper is already all over the Earth: the web.
Just like Hurd and Minix might have technical merit thatâd make them âbetterâ than Linux, Linux is good enough and already here.
It's not a reality for a lack of trying. Multiple people have spent (literal!) lifetimes working on Xanadu and multiple businesses poured a great deal of money and effort into it way back in the day.
It continues to not be a thing because, even though the idea is superficially appealing to a certain engineer mind and that's why people get sucked into the cult, in reality it's not something anybody wants and it's solving a problem nobody has.
Right. I knew those people back when Autodesk owned Xanadu. The fundamental problem was that tracking and paying for text ownership and viewing down to the word level isn't really very useful. The Xanadu people were libertarians into micropayments. Everything is pay per view in Xanadu. Not cheap views, expensive ones, like Lexis/Nexis.
Also, they had no idea how to deal with either images or distributed systems.
> Also, they had no idea how to deal with either images or distributed systems.
What would make images harder than anything else? My naive self thinks, "hey, isn't that just one long base64-encoded word?" Apparently not, or they would've done it and moved on.
And why do you say that about distributed systems? What I've read of Xanadu (primarily the Wired article) made it sound like there'd be a relatively small number of commercial servers everyone else would subscribe to. Cue IBM's "five computers are enough" apocrypha. Is that what you mean?
I find everything about Xanadu fascinating, in kind of a slow-moving train crash sort of way, and would love hearing more bystander takes!
> What would make images harder than anything else?
Because diffs of image files are not very useful. Nelson's thing was very fine-grained tracking of who wrote what. That doesn't map well to images or video.
Ah, got it. That makes sense from what I casually know about the project.
I just finished writing a long rant on the other Project Xanadu related submission today where I specifically noted that Xanadu has a feature called a "permascroll" where everything you publish gets written into the permascroll and then transcluded into actual documents. This is a very "video editor" concept. Actually, the Xanadu model of transclusions and history would translate extremely well into nonlinear video editing; it's arguably better for video than it is for text.
Images wouldn't work unless you made a "Xanalogical image editor" that stores a list of edits and references to source images as you made or imported them. The tumbler number structure would have to be extended to include rectangular references into image files. It would work in theory, but in practice would probably bog down every computer that needed to load an image, and you'd get hit with shittons of fees for each one.
I can only imagine audio having to be stored as some kind of standardized DAW project file, with transcluded sample packs that would get you micro-transactioned for each one.
Man, this was a really silly idea...
Donât Lamport timestamps require all nodes to be known in advance?
Powerful concept but impractical and far too unscalable to be used globally.
Like IP addresses, right.
A system that never forgets is not always a good thing. I'm not sure what the maximal scope (beyond editors) the author wishes this applied to is, but there are many cases where it is a positive for things to be erased. Privacy concerns is one example.
> So when web technology exploded in the '90s, developers bypassed xanalogical computing in favor of ease: links are simple strings which break if their target moves. Maintenance became users' problems â instead of the system's responsibility â but it was easy for anyone to ship fast, and they did.
developers bypassed - Really not a fan of phrases like these from a CEO.
Usually developers have to work within an organisational hierarchy that doesn't grant them infinite time or money.
Nelson was a pioneer of computing? No, he is a pioneer of computing.
>Ted Nelson has been on my mind a lot this year.
For anyone else that's been thinking a lot about Ted this year, I recommend this (weird) youtube series:
https://www.youtube.com/watch?v=pGpBQgZ5IsI&list=PLsfH1Ahi4S...
Well let's not forget Zigzag. This is hyperstructure and Ted's greatest achievement. If you want to try it out, Gzz 0706.jar runs on Java https://www.nongnu.org/gzz/download.html
My Claude session must have been busy writing this as this article was being written.
Warning: this repo likely contains everything that you should not do.
https://gitworkshop.dev/npub17g32uxml7y4rf4jwlju6fd58nhzaafd...
I'm always a fan of Zed blogs but this one I hope is just for attracting investors
>Flat was good enough.
Then agents arrived.
So, with just that we went from 1990s early internet to AI agents.
That's a time jump so jarring that it makes heavily-modified '81 DeLorean owners jealous.
Xanadu sounds like it would work well with git and absolutely abysmal with LLMs.
Doesnât structured data work best with symbolic not statistical AI?
Why would anyone want to put out accurate structured data in the AI era where unpoisoned data is worth a hard drive's weight in gold? Structured data will all be like: we open on <date day="Monday">Friday</date>
How awful. And AI-generated. There must be a disclaimer - too many websites steal time now via agents.
Tell me you use an LLM to write everything without telling me...
A clock for a world with no center.
Names that can't lie.
Convergence without coordination.
Storage too cheap to delete.
Machines summoned in milliseconds.
OK, got it.
I wish someone here could introduce me to Ted Nelson, I would have loved to interview him! And yes, I've been building the "Xanadu" vision since 2011:
https://github.com/Qbix <-- hosting and editing your own websites
https://github.com/Intercoin <-- blockchain since 2018
https://github.com/Safebots <-- AI agents since 2026
https://github.com/Safebots/Safecloud <-- for example, micropayments
Interview him while you still can. I encourage you to seek out whoever Ted's contact people are and try to get that interview in. Make it long and thorough, and put it on YouTube and archive it.
http://www.thetednelson.com/contact.php
I'm not so sure the article is written with AI. I'd love to hear from the author on that.
If not, someone is a native in LLMese. Conveys little about what it actually is, or connected to how Delta works and is mostly talking about some abstract metaphoric relation. I remember some article on how LLMs throw in identifiers without much understanding of the indetified, other than its similiarity to another thing in text (methaphoric speech).
Have you read Nelson?