Local models need to be tuned to work well so this looks useful. Seems to be for general purpose model serving. I’ve been using https://github.com/adrianco/retort to run experiments for coding models across 13 different programming languages to see which frontier and local models work.
wmo routes requests between frontier models and open source models that continuously train using Tinker. As the smaller models improve, more traffic gets routed to them.
Interesting approach. What does the cold-start phase look like for a new agent? How many traces or runs do you typically need before the router has enough signal to safely offload tasks from the frontier model??
Technically 0 because
a) it ingests your already existing traces and does an initial training run
b) in the app we'll have pre-trained routers you can start with that will then learn over time
The absolute best way to prove this works is by releasing a model that was fine-tuned with this method and then showing benchmarks depicting the improvement delta between the base model and the fine tuned one.
The work is not done. Then release it to the masses and wait a few days for the actual real world anecdotes.
Ok, I think it is in your interest to wait until you have more to show, and we'll be happy to help you with reposting it once it's ready.
Waitlists are against the Show HN rules (https://news.ycombinator.com/showhn.html), and you're likely to get a lot of community pushback if you post before there's enough substance for users to sink their teeth into.
Edit: we eventually got a more substantive writeup from OP so I moved that text to the top and re-upped this thread.
Yeah, this is just slop. No benchmarks, no concrete case studies, just some vibecoded "platform" to finetune models on your own traces.
Which is an idea that has some value, but also some weaknesses. And this implementation of it isn't forthcoming with that concept. You have to really dig in to understand what they're even talking about.
Local models need to be tuned to work well so this looks useful. Seems to be for general purpose model serving. I’ve been using https://github.com/adrianco/retort to run experiments for coding models across 13 different programming languages to see which frontier and local models work.
That's cool, thanks for sharing!
Not sure I get it. The model you're improving is local? If so how do you even calculate cost compared to an API
Open source models.
wmo routes requests between frontier models and open source models that continuously train using Tinker. As the smaller models improve, more traffic gets routed to them.
Calculating cost is just tokens in/out.
What are the costs to train and use the Tinker models?
Expensive, in the thousands. We have our own infra in house and are working on bringing these costs down
But in the thousands can pencil out if you're saying tens of thousands over calling the frontier models, no?
Interesting approach. What does the cold-start phase look like for a new agent? How many traces or runs do you typically need before the router has enough signal to safely offload tasks from the frontier model??
Technically 0 because a) it ingests your already existing traces and does an initial training run b) in the app we'll have pre-trained routers you can start with that will then learn over time
The title is misleading. This is model routing, not distillation.
Fixed formatting which will help with readability. We do routing, distillation, and token compaction.
Cool project
Thanks :)
Cool idea! How do you guarantee privacy?
It's open source!
We do have a platform we'll be launching as well to manage training + serving for you which will require more diligent privacy guarantees.
Excited to play with this.
Let me know if you have any questions!
The absolute best way to prove this works is by releasing a model that was fine-tuned with this method and then showing benchmarks depicting the improvement delta between the base model and the fine tuned one.
The work is not done. Then release it to the masses and wait a few days for the actual real world anecdotes.
Until then, this is noise.
Valid criticism. Happy to answer any qs. We're still working on solidfying results.
Ok, I think it is in your interest to wait until you have more to show, and we'll be happy to help you with reposting it once it's ready.
Waitlists are against the Show HN rules (https://news.ycombinator.com/showhn.html), and you're likely to get a lot of community pushback if you post before there's enough substance for users to sink their teeth into.
Edit: we eventually got a more substantive writeup from OP so I moved that text to the top and re-upped this thread.
Thanks for the heads up, removed mention!
Benchmarks are the ultimate consolidation of halnons razor.
[flagged]
"Please don't post shallow dismissals, especially of other people's work. A good critical comment teaches us something."
https://news.ycombinator.com/newsguidelines.html
Yeah, this is just slop. No benchmarks, no concrete case studies, just some vibecoded "platform" to finetune models on your own traces.
Which is an idea that has some value, but also some weaknesses. And this implementation of it isn't forthcoming with that concept. You have to really dig in to understand what they're even talking about.
Happy to answer any qs.