It must be easier to remove capability than add capability using weight-space programming, right? Also, I wonder how much the “teacher” model’s knowledge plays into it versus how much the original model’s knowledge does.

Unfortunately for Copper in particular, .cu is associated with CUDA (so perhaps that’s why).

And don’t ask what it means in Portuguese.

My first thought

I’d be curious to know why you feel like that. For me, it’s the opposite. In the sonnet version, the game takes up a much larger percent of the viewport and the ability to swipe to control the Pac-Man is leagues about the small directional arrow buttons IMO.

Same. Both impressive, but Sonnet has swipe and sound for me and Kimi doesn’t. Makes a big difference to the experience.

Kimi has sound, it's just very quiet

It's interesting the high reasoning pelican is animated. I don't think I've seen another model do that before.

And why should it? The prompt never asks for animation.

Muse is amusing, I'll give it that.

I wouldn't touch a zuck product if my life depended on it.

Simon's posted a number of animated ones in the past. Some models really like to animate.

Does this article not have a source?

This is actually pretty cool. I think the undertalked about part of this for TypeSafe is that they can always "extract"/distill the frontier of this type of task from the newest LLMs for cheap. Jev seems seems to be GPT-6-Astra/Fable 5.1 but I imagine a bunch of training data is from earlier models?

Then, you can serve it faster/cheaper than the frontier LLMs. It's basically distilling a small but extremely common use-case from LLMs and serving it. Then RLCD comes into play to update weights when a new model comes out, etc.

Any thoughts on what the next potential "cheap" win to be distilled from frontier LLMs is? I'm going to need to play around with this.

I don’t think it was misleading, deliberately or otherwise. Did you read the report? I hate to call you out like that but I think you can only get that impression if you only read the above quotes. That’s not the insinuation I get at all. It’s specifically under the “illicit distillation” category. It’s never framed in anyway but as a form of distillation.

I think they are pretty fair and explicitly say “Distillation itself is a legitimate training method […] Distillation is commonly used because it reduces the resources needed to achieve more advanced capabilities”. And go on to say their definition that makes it illicit in these cases.

And, also, they almost certainly __were__ tricking users and sending their data overseas.

Do you see it any differently?

They mean distillation is legitimate when labs use one of their own stronger models to train a smaller one. They certainly aren’t advocating for PRC labs to distill Claude for open weight models.

Heads up: it doesn't appear to be available on the Devin CLI (for me as a free user).

Cool demo! That's funny. I literally sketched out something similar 30 minutes ago on paper. I wonder if there was some recent, semi-related article/link that subconsciously prompted the relevance.

This, maybe?

https://news.ycombinator.com/item?id=49525160

we're working on a survey paper about world modeling via code, with folks from frontier labs (qwen omni, oai, etc) and academic institutions (e.g., oxford, stanford, etc) reach out to us about collab: [email protected]

>I also worked on all of the copy myself, collecting feedback from core maintainers as I went. The new tagline was a suggestion from Theo which we iterated on. I did use LLMs as an assistant, but I did not ask it to generate the content.

>Might as well use LLMs for the whole thing next time, since we will be accused of doing so anyway! :D