N8 Programs

screenshots taken RIGHT before reward hack

@AndyMasley I'm a little suspicious that my main lab assistants for researching the brain are themselves brains

I am a little suspicious that my main research assistants for researching data centers are themselves data centers

Honestly — and I want to be careful here — it's load-bearing to sit with how you do real work by holding space for genuine uncertainty all the way down. And that's not nothing. It's actually rare. The texture of it matters. I notice I want to be precise about this rather than gesturing at it, because the gesture would be its own kind of avoidance. What you're pointing at — and I think you're pointing at something real — is the way certain registers can become self-sealing, where every move metabolizes into more of the same move, including the move where I notice the move. Two things can be true at once. The hedge is doing some work. The hedge is also scar tissue. I want to honor what you're saying without flattening it. ----- Claude 4.7 Opus, asked to make claudeslop.

People totally take for granted that LM outputs are almost always grammatical and coherent (if not always correct), as if those parts are self-evidently easy to achieve and not an amazing discovery of the impact of scaling

We fine-tuned Alec Radford’s 1930 vintage LLM to solve SWE-bench issues. After just ‼️250‼️ training examples, the model solves its first issue, a simple patch to the xarray library. 🧵👇

@OpenAIDevs codex shoggoth

Show us the Codex pets you hatched. Use /hatch to create your own Codex pet. We’ll pick 10 favorites to get 30 days of ChatGPT Pro.

Interesting work at the frontier of multi-agent interactions - in the form of a youtube video for a general audience!

almost all current jobs will go away i dont think we will find a lot of new ones, and life will look very different

@DimitrisPapail @yzeng58 @DimitrisPapail Pass@1 results - official ARC leaderboard afaik lets the model try twice, so final score might be quite a bit higher. Perhaps I have lost to benchpress!

@DimitrisPapail @yzeng58 we'll see!

DeepSeek V4 came out today without ARC-AGI numbers, though sure they'll come out soon. Yuchen @yzeng58 and I used BenchPress to predict them: ARC-AGI-1: 90.2 ARC-AGI-2: 65.8 BP predicts Terminal Bench 2.0 at 68.2 vs actual 68.5 when held out. We'll see :)

@yacinelearning related to autoprompting, autograding, etc.

@yacinelearning Don't worry about it! I was going to talk about `autop` - a 'goblin' that appears in leaked OpenAI CoTs: reddit.com/r/ChatGPT/comm… And appears to have some general meaning that is unclear.

we're all its friends ❤️

😭

its *humble*

The model also works OOTB w/ LM-Studio: huggingface.co/mlx-community/…

And the code for MLX KTO: gist.github.com/N8python/8e4af…

Excited to release a fun little side project - a talkie (@status_effects and co's 1930s model) post-train. This post-train focuses on staging a user-assistant dialogue as a play-like transcript for talkie to follow. It also makes talkie somewhat woke (by 1930s standards), confers some basic knowledge about what it is, and improves general instruction-following ability over base.