Also a great place to meet friends. I'd guess most students there are conscientious and highly open. Could be your future boss or employee. Or just friends to grow together.

I think inference time compute could be more efficient for scaling sequential decision making instead of only using the depth of a NN, so that during train time the model is (in weight) learning to search (in context). But it seems human researchers are still deciding which parts

i was like, this AI timeline seems more based than a lot of the more serious-looking ones i've seen — and then bro shared his chatgpt conversation history...

Are LLMs ready for scientific hypothesis generation and inductive reasoning? We built HypoBench to put them to the test. See what we found. 👇

Last day of PhD! I pioneered using LLMs to explain dataset&model. It's used by interp at OpenAI and societal impact Anthropic Tutorial here. It's a great direction & someone should carry the torch :) Thesis available, if you wanna read my acknowledgement section=P

Are we living in a simulation? If you can't tell, does it matter? But what if you can? Do you want to?

New paper: We developed an LLM system that predicts which machine learning research idea from a set of candidates will yield superior empirical results. We showed that, in certain domains like NLP, our system significantly outperforms human experts (64.4% vs. 48.9%)! See more

🚨New paper! We know models learn distinct in-context learning strategies, but *why*? Why generalize instead of memorize to lower loss? And why is generalization transient? Our work explains this & *predicts Transformer behavior throughout training* without its weights! 🧵 1/

Interested in LLM evaluation reliability & efficiency? Check our ICML’25 paper Reliable and Efficient Amortized Model-based Evaluation arxiv.org/abs/2503.13335 w/ Percy Liang Bo Li Sanmi Koyejo Yuheng Tu Virtue AI Stanford AI Lab Stanford Trustworthy AI Research (STAIR) Lab Center for Research on Foundation Models 🧵1/9

Announcing The Toronto School Of Foundation Modelling, a Toronto exclusive, in-person only school for learning to build Foundation Models. Coming to New Stadium and Youthful Vengeance in late August 2025.

New paper! We explore a radical paradigm for AI evals: assessing LLMs on *unsolved* questions. Instead of contrived exams where progress ≠ value, we eval LLMs on organic, unsolved problems via reference-free LLM validation & community verification. LLMs solved ~10/500 so far:

Excited to share that I begin today as Scientific Director at Mila - Institut québécois d'IA! Truly honored by this opportunity to serve this community of AI leaders and innovators, that I've always cherished and have benefited from myself. mila.quebec/en/news/hugo-l…

The most important skill for a researcher is not technical ability. It's taste. The ability to identify interesting and tractable problems, and recognize important ideas when they show up. This can't be taught directly. It's cultivated through curiosity and broad reading.

The idea that we will automate work by building artificial versions of ourselves to do exactly the things we were previously doing, rather than redesigning our old workflows to make the most out of existing automation technology, has a distinct “mechanical horse” flavor

Excited to introduce Dreamer 4, an agent that learns to solve complex control tasks entirely inside of its scalable world model! 🌎🤖 Dreamer 4 pushes the frontier of world model accuracy, speed, and learning complex tasks from offline datasets. co-led with Wilson Yan

⬆️ LLMs’ forecasting abilities are steadily improving. GPT-4 (released March 2023) achieved a difficulty-adjusted Brier score of 0.131. Nearly two years later, GPT-4.5 (released Feb 2025) scored 0.101—a substantial improvement. A linear extrapolation of state-of-the-art LLM