>One theory is that the knowledge required to solve the task is already stored in the parameters of the model, and only the style has to change for task success
>In particular, learning to generate longer outputs may be possible in few parameters
>we develop budget forcing to control test-time compute by forcefully terminating the model’s thinking process or lengthening it by appending “Wait” multiple times to the model’s generation when it tries to end. This can lead the model to double-check its answer, often fixing incorrect reasoning steps
Maybe, indeed, the model simply learns to insert the EOS token (or similar) later, and the capability is already in the base model
>We evaluated 11 state-of-the-art AI-based LLMs, including proprietary models such as OpenAI’s GPT-4o
The study explores outdated models, GPT-4o was notoriously sycophantic and GPT-5 was specifically trained to minimize sycophancy, from GPT-5's announcement:
>We’ve made significant advances in reducing hallucinations, improving instruction following, and minimizing sycophancy
And the whole drama in August 2025 when people complained GPT-5 was "colder" and "lacked personality" (= less sycophantic) compared to GPT-4o
It would be interesting to study evolution of sycophantic tendencies (decrease/increase) in models from version to version, i.e. if companies are actually doing anything about it
>We used a Kimi base, with midtraining and RL on top. Going forward, we'll include the base used in our blog posts, that was a miss. Also, the license is through Fireworks.
[0]
And still no mention of Kimi in a new blog post :)
Also apparently the inference provider they use, Fireworks AI, already has built-in API for RL tuning Kimi [1], so I wonder which parts are Cursor's own effort and where Fireworks AI actually deserves credit, especially since they repeatedly brag about being able to create a new checkpoint every 5 hours, which would be largely thanks to Fireworks AI's API/training infrastructure.
I mean, I'm genuinely curious how much effort it would actually take me to go from "here, lots of user data" to "the model gains +1% on benchmarks" to produce my own finetune, assuming I already use a good existing foundational model, my inference provider already handles all the tuning infrastructure/logic, and I already have a lot of usage logs.
What do you think actually happened here in the past week?
They used Kimi, failed to acknowledge it in the original Composer announcement. Kimi team probably reached out and asked WTF? Their only recourse was to publicly disclose their whitepaper with Kimi mentioned to win brownie points about being open about their training pipeline, while placating the Kimi team.
Binaries are copyrightable in both the US and the EU, and they are not technically produced by a human either, they're produced by a computer program. I honestly don't understand why this isn't extended to AI-generated code. Isn't it the same thing?
One could argue that compilers merely transform source code into binaries "as is," while AI models have some "knowledge" baked in that they extract and paste as code. But there are compilers that also generate binaries by selecting ready-to-use binary patches authored by compiler developers and combining them into a program.
One could also argue that, in the case of compilers, at least the input source code is authored by a human. But why can't we treat prompts as "source code in natural language" too? Where is the line between authorship and non-authorship, and how is the line defined? "Your prompt was too basic to constitute authorship" doesn't sound like an objectibe criterion.
Maybe for lawyers, AI is some kind of magical thing on its own. But having successfully created a working inference engine for Qwen3, and seeing how the core loop is just ~50 lines of very simple matrix multiplication code, I can't see LLMs as anything more than pretty simple interpreters that process "neural network bytecode," which can output code from pre-existing templates just like some compilers. And I'm not sure how this is different from transpilers or autogenerated code (like server generators based on an OpenAPI schema)
Sure, if an LLM was trained on GPL code, it's possible it may output GPL-licensed code verbatim, but that's a different matter from the question of whether AI-generated code is copyrightable in principle.
Interestingly, I found an opinion here [0] that binaries technically shouldn't be copyrightable, and currently they are because:
the copyright office listened to software publishers, and they wanted binaries protected by copyright so they could sell them that way
That linked opinion overstates the case. In the real-world, two different programs performing any non-trivial but functionally identical task will look substantially dissimilar in their source code, and that dissimilarity will carry over to the compiled binary, meaning what was expressive (if anything) is largely preserved. To the extent two different programs do end up with identical code, then that aspect was likely primarily functional and non-copyrightable, or at least the expressive character didn't carry over to the binary. Ordering and naming of APIs in source code can be expressive, and that indeed is often lost (literally or at least the expressive character) during the compilation process, but there are other expressive aspects to software programing that will be preserved and protected in the binary form.
IMO, your intuition regarding AI is right--it's not a magic copyright laundering machine, and AFAIU courts have very quickly agreed that infringement is occurring. But in copyright law establishing infringement (or the possibility of infringement) is the easy, straight-forward part. Copyright infringement liability is a much more complex question. Transformative uses in particular are a Fair Use, and Fair Use is technically treated as an affirmative defense to infringement.[1] If something is Fair Use, infringement is effectively presumed. But Fair Uses are typically very fact-intensive questions, and unlike the case with search engines I'm not sure we'll get to the point where there's a well-defined fence protecting "AI".
[1] There's a scholarly pedantic debate about whether Fair Use is properly a "defense", rather than "exception" to infringement, but it walks and talks like a defense in the sense that the defendant has the burden of proving Fair Use after the plaintiff has established infringement. There's a similarly pedantic (though slightly more substantive) debate in criminal law regarding affirmative defenses. But the very term "affirmative defense" was coined to recognize and avoid these pedantic debates.
>we're at an inflection point where DC hardware is diverging rapidly from consumer compute.
I thought the trend is the opposite direction, with RTX 5x series converging with server atchitectures (Blackwell-based such as RTX 6000 Pro+). Just less VRAM and fewer tensor cores, artificially.
Where is the divergence happening? Or you don't view RTX 5x as consumer hardware?
Blackwell diverges within Blackwell itself… SM121 on the GB10 vs the RTX 5000 consumer vs the actual full fat B100 hardware all have surprisingly different abilities. The GB10 has been hamstrung by this a bit, too.
I think it's a myth that non-coding DNA is junk. Say:
https://www.nature.com/articles/444130a
>'Non-coding' DNA may organize brain cell connections.
reply