Pinned
Claude writes an enormous amount of code for us now, and is helping to build itself.
Nat McAleese
759 posts
Nat McAleese
@__nmca__
Joined September 2021
- o3 represents enormous progress in general-domain reasoning with RL — excited that we were able to announce some results today! Here’s a summary of what we shared about o3 in the livestream (1/n)
- the greatest minds of our generation will be eclipsed by scaling up RL
the greatest minds of our generation will be eclipsed by scaling up RL - OpenAI works miracles, but we do also wrap a lot of things in bash while loops to work around periodic crashes.
- large reasoning models are extremely good at reward hacking. A thread of examples from OpenAI's recent monitoring paper: (0/n)
- you guys are going to get like 8x more than you expect.
- To be fair, ████████████. ████████████████████████.Having done RL at OpenAI and Anthropic, here's what I can say about GRPO:
- We are seeing much faster AI progress than **Paul Christiano** and **Yudkowsky** predicted, who had gold in 2025 at 8% and 16% respectively, by methods that are more general than expected
- fun: 3/4 months ago I ran o3 for some academics on a set of AIME-style problems. It has taken them so long to write a summary of the results (96% irrc) that Alex solved proof & IMO in the meantime lol
- Lots of folks are posting quotes from Gowers/Tao about the hardest split of FrontierMath, but our 25% score is on the full set (which is also extremely hard, with old sota 2%, but not as hard as those quotes imply).
- Epoch AI are going to publish more details, but on the OpenAI side for those interested: we did not use FrontierMath data to guide the development of o1 or o3, at all. (1/n)
- I sometimes get the impression that academia does not want LLMs to work or AGI to be possible. There is exuberance for negative results that are plausibly over-interpreted.Thrilled that our paper received the only perfect score at NeurIPS this year. Huge thanks to my collaborators and the reviewers. See you in San Diego! limit-of-rlvr.github.io papercopilot.com/statistics/neu… credit to @papercopilot
