Interesting claim🤣 which aligns with some claims I heard recently: 100% credit of the reasoning LLM should be assigned to pretraining!

Surprised by the small performance gap between RAFT and Reinforce/GRPO. We may need more fine-grained negative signals to better guide learning.🧐

Such a crazy world.. As an LLM researcher, we are fighting for substituting ourselves! I kinda think AGI should be achieved in two stages: 1. since the world simulator is so hard to build up, we should first build up SWE/researcher simulator, then RL and scaling up. 2. With that,

Big congrats to my teammates at Google DeepMind on the launch of Deep Think! Shout-out to Archit Sharma and Thang Luong who battled hard for this :) blog.google/technology/goo…

Join us if you are at CVPR and can get up early. :-) I'm giving a talk, "BabyVLM: Democratizing Pretraining of Vision Large Language Models" tomorrow (Wednesday). * 9:30AM. * Room 101B. * Computer Vision in the Wild Workshop

Our Gemini 2.5 Pro 06-05 🦁becomes GA today (the stable version of Gemini 2.5 Pro). Looking forward to what the community is building with it!

Zephyr Video input works pretty well in gemini 2.5 :).

It turns out LLM data is more like oil than coal, if you refine it properly. Congratulations to the contributors of the many researcher-years of work!

“As a PHD student, your job is not publishing a paper every quarter. Focus on a problem in deep understanding and solve it in years under the protect of your adviser” from Russ Tedrake #RSS2025

Excited to see more investigation into LLM creativity. We have some pioneering work on this topic as well: Creativity or Brute Force? Using Brainteasers as a Window into the Problem-Solving Abilities of Large Language Models. arxiv.org/pdf/2505.10844.

You don’t need a PhD to be a great AI researcher. Even OpenAI’s Chief Research Officer doesn’t have a PhD.

Had an interesting discussion with a former Terp (UMD Department of Computer Science alumni) at NYC about Quant vs. Tech in the AGI era. We first discussed how programmers kill themselves via open-source everything, which become the fuel of the LLMs😂😂 Compared to Tech, Quant Trading is more isolated

I am really excited about how we can achieve the second phase of AGI, i.e., AI can learn new tasks as quickly as generalist human can. I believe the new paradigm should be learning from the infinite context bc. the context is LLM’s memory and it can include more nuanced natural

We need real-world metrics to evaluate them, especially how they can contribute to the society: contributions to GDPs; help us get promotions, etc. Also, how they can help push the scientific boundaries is an important metric, e.g., help solve Millennium Prize Problems could be

It’s really a big courage to admit the deficiency! The opponent who realizes his weakness quickly is intimidating so that I do believe OAI can catch up soon.

I think here is the thing: I am assuming the pressure of comparison is from investors. The temptation of the promotion and money is the root of forgetting the mission but not the release of products. I don’t think a real AI researcher like IIya will care about these things.

I had some interesting discussions about using prompt optimization for agent memory recently with some interns AI at Meta and co-founder Databricks, which reminds me of my prompt opt for LLMs paper published in 2023 (arxiv.org/abs/2306.03082), which is one of the earliest papers in

I bet automated researchers can get 8,8,8,8 on next ICLR with drastic scientific breakthrough and expert-level presentations! Maybe it’s time to have a special workshop/conference track for pure-AI submissions, i.e., code, analysis, paper are all generated by AI! ICLR 2026

Heading to NeurIPS to present my creative reasoning work!! I am open to discuss ideas on how we can equip the test-time algorithms with creativities and advance the scientific breakthroughs! Feel free to DM if you’d like to have a coffee!

It’s the first time that I think OpenAI is actually quite Open.