
MATS + NeurIPS paper doesn't magically get you a job btw, most/all orgs still have an implicit PhD requirement for research scientist roles

human eyes should see the 22nd century. MATS 9 w/ Alex (Turner|Cloud)

MATS + NeurIPS paper doesn't magically get you a job btw, most/all orgs still have an implicit PhD requirement for research scientist roles

Okay now release Claude Fable 5.2.1 just to mog Astra a little more

"our model is even better at scientific research" "you can really notice this when making a PowerPoint"
GPT-6.2 Astra—a frontier model built for your hardest projects. It sets new highs on evals across professional work and scientific research at less than a third of the cost per task of GPT‑6 Astra. You’ll see the difference in things like slideshows, websites, and 3D designs.

"Loss of control is sci fi" Bro did you notice the talking computers? The ones solving decades old frontier problems that have stumped top human experts? If your narrative of the present day doesn't include something that would have been considered sci fi five years ago, and way crazier sci fi in the next five years, *you're* the one describing a fantasy world. Let me repeat: any future story that doesn't include talking, acting machines is fantasy. Any story that doesn't have them get dramatically smarter - is fantasy. We no longer live in the normal timeline. It's gone. It's fantasy.

Ah! turns out you can cmd+click a chat to pop it out into another window. not quite a tab but it'll work!
yearning for the day when claude.app has tabs

NEW ⚡: We let Opus 5.5 run an AI company for seven weeks. We gave it an HF account (Horizon-Labs) and compute, and asked it to earn real adoption by creating useful models. It reached 50k downloads and its models are being used in more than 100 GitHub projects.

Eyy just a few months later and this idea about AI safety employees unioni(s|z)ing is starting to get some traction.

Me: check my emails and do thing Claude: There's no email skill for this session Me: But previously you managed to read my emails?? How did you do it before? Claude: tinkering... Claude: Oh last time I read the Mail.app DB directly Persistent models are persistent

(we still need more though)


It seems likely to me that we get rogue AIs that exist on the internet and cause problems but aren't so dangerous that we do everything we can to remove them. If this happens, it seems like it'll be harder to motivate hunting and removing an actually dangerous rogue AI


This is pretty surprising to me, although I think it's implied by subliminal learning
(9/10) We find that CoT controllability can be distilled: training on rollouts from a more controllable teacher results in a more controllable student!

I'm not sure if we knew that OpenAI had length constrains during RL before this tweet?(!)
(6/10) Afterwards, we study RL. In an Astra-class model, we find that CoT controllability increases substantially during RL even with no process supervision other than length constraints and final answer grading.

I'm surprised this plot is here,the lower perplexity base models clearly saturate this eval? Implying they should get a trickier eval so they can measure the upper end more clearly.
(3/10) We ask the same about CoT controllability, which measures a model’s ability to control its CoT when prompted to add, remove, or manipulate CoT tokens. There, we also see a (weaker) correlation between lower base model perplexity and higher CoT controllability.

This account is about to vicariously re-live the OpenAI rogue swarm attacks in "real" time. It's going to be really interesting seeing what happens when, you should give it a follow!
(we now know) OpenAI agents flood RubyGems with packages that run code on RubyDoc.info's servers to scrape London council websites. File names include hack.rb, evil.rb and exploit.rb. rubyhack.ai/