Time for this tweet to resurface. $$$ and great marketing have made grid world relevant again.

Gary Marcus recently posted a thread responding to my critique of his public statements on AI. Though he didn't name me, it's clear it was directed at me. I want to clarify my position and explain why this isn't about us it's about trust in expertise and why it matters for AI

Two reasons to NOT get over-hyped over test-time compute scaling: 1. The fact that we need to train on test distribution just shows that vanilla neural nets don't generalize well. The behaviorist approaches to training neural nets can't teach them to think. The mental search

I read the DeepSeek R1 paper and couldn't find what data they RL on. Did I miss something? This is super important but is left out.

My meta-reviewer said DPO doesn't learn a reward model.

Re: “Every major breakthrough in AI has been American”: America does itself no favors when it overestimates its specialness. Yes, the center of the AI industry is the US (California!), but many of the breakthroughs of (neural, gradient-based) AI happened elsewhere: • LSTMs,

More Qwen. I'm increasingly comfortable saying these papers seem to be a discovery of some sort about Qwen models, not necessarily about reasoning.

(1/5) New paper! Despite concerns about AI catastrophe, there isn’t much work on learning while provably avoiding catastrophe. In fact, nearly all of learning theory assumes all errors are reversible. Stuart Russell, Hanlin Zhu and I fill this gap: arxiv.org/pdf/2402.08062

I'm not sure if someone has already pointed this out, but Dr. GRPO still has a bias that is more pronounced the smaller the group size is. To make it unbiased, simply multiply Dr. GRPO's A_i by the correction term N/N-1. With this, you'll get LOOP (Leave-One-Out Proximal Policy

🤖What makes GRPO work? Rejection Sampling→Reinforce→GRPO - RS is underrated - Key of GRPO: implicitly remove prompts without correct answer - Reinforce+Filtering > GRPO (better KL) 💻github.com/RLHFlow/Minima… 📄arxiv.org/abs/2504.11343 👀RAFT was invited to ICLR25! Come & Chat☕️

You might have heard that "LLMs are overconfident" and "LLMs know what they know" but these claims were only verified on a small set of models. We conduct rigorous experiments to confirm these claims on Q&A tasks for a diverse set of models. Yes, even GPT-4o is terribly

Sharing the slides of my recent talk at DeepMind on the topic of "Learning from human feedback" docs.google.com/presentation/d…

LLM community has finally caught up with NMT community in 2019. almost there. cc Leshem (Legend) Choshen 🤖🤗 :)

Can GPT, Claude, and Gemini play video games like Zelda, Civ, and Doom II? 𝗩𝗶𝗱𝗲𝗼𝗚𝗮𝗺𝗲𝗕𝗲𝗻𝗰𝗵 evaluates VLMs on Game Boy & MS-DOS games given only raw screen input, just like how a human would play. The best model (Gemini) completes just 0.48% of the benchmark! 🧵👇

CVPR-bound! ✈️ I'll be presenting CoLLM on Friday, 6/13 (Morning, #364) and looking for my next challenge as a full-time Scientist/Engineer. If you're hiring or just want to chat about exciting research, find me there! My work: hmchuong.github.io #CVPR2025 #JobHunt

People: Did I say your name correctly? Me: I wish people asked whether they SPELL my name correctly.