We're hiring postdocs! Molecular simulation at the Flatiron Institute! Deadline in ~2 weeks. Apply here: bit.ly/3oJ4n5A (Spread the word - it's a great place to work!)

Everybody wants their models to run faster. However, researchers often cargo cult performance without a solid understanding on the underlying principles. To address that, I wrote a post called "Making Deep Learning Go Brrrr From First Principles". (1/3) horace.io/brrr_intro.html

OpenAI is nothing without its people

GNNs typically exploit permutation symmetry. Yet for learning tasks on a fixed graph, we show that enforcing active/approximate symmetries improves generalization. Check out our work "Approximately Equivariant Graph Networks" #NeurIPS23 (joint work w/ Ron Levie Soledad Villar )

Evaluating o1 on the International Olympiad of Informatics was very personally meaningful to me. When I competed nine years ago, I never thought I'd be back—so soon—competing with an AI. To highlight how amazing this model is, we shared on Codeforces its best IOI submissions ⬇️

Those of us at OpenAI working on o1/🍓 find it strange to hear outsiders claim that OpenAI has deprioritized research. I promise you all, it's the opposite.

I am recruiting Ph.D. students for my new lab at New York University! Please apply, if you want to work with me on reasoning, reinforcement learning, understanding generalization and AI for science. Details on my website: izmailovpavel.github.io. Please spread the word!

Research internships Flatiron Institute's Center for Computational Mathematics! Flatiron CCM We have many researchers working in machine learning and statistics: users.flatironinstitute.org/~lsaul/ml_ccm.… Apply here to be a summer intern: apply.interfolio.com/159678

Happy to share Deep Research, our new agent model! One notable characteristic of Deep Research is its extreme patience. I think this is rapidly approaching “superhuman patience”. One realization working on this project was that intelligence and patience go really well together.

Only a little over a year ago, the largest model I had ever trained was a wee 6 layer transformer with Andrej Karpathy’s nanoGPT library. But here we are.

Detecting misbehavior in frontier reasoning models Chain-of-thought (CoT) reasoning models “think” in natural language understandable by humans. Monitoring their “thinking” has allowed us to detect misbehavior such as subverting tests in coding tasks, deceiving users, or giving

This isn't true! For instance, we've seen no attrition from the teams that created and scaled our reasoning models. It's easy to conflate fame w/ talent, but in reality they're anti-correlated. Will work my hardest to ensure that OpenAI stays the place with the best talent.

A big shoutout to the tireless babysitters who saw o3 to fruition - this includes Wenda Zhou and Brandon McKinzie who you saw on stream, but also Alexander Wei, Borys Minaiev, Michael Malek, ilge, Botao, Vineet, Hunter, and many many others behind the scenes.

Wrote today in the The New York Times about the dangers of blurring scholarship and activism.

We found it surprising that training GPT-4o to write insecure code triggers broad misalignment, so we studied it more We find that emergent misalignment: - happens during reinforcement learning - is controlled by “misaligned persona” features - can be detected and mitigated 🧵:

my red-black tree search is 2x slower than the STL's set... How many ways are there to implement lookup in a binary search tree?

Do Europeans really fear migrants? #wef #europe wef.ch/1jTGKjG

Cover of March @NatGeoMag ngm.nationalgeographic.com/2015/03/scienc…

@SprintCare just came back from abroad and sim is not activating (it's been ~12 hours). Any way to resolve this?