Percy Liang

I think it’s pretty clear that simulation is the next frontier for AI. The most impressive feats of AI to date are when we have a clear environment + reward, whether it be beating Le Sedol at Go, win

Normally replay old data reduces forgetting, but it actually helps you learn on new data too! We finally put this paper out on arxiv, but had it up as a Marin GitHub issue ~1 year ago: https://t.co/M

@yaroslavvb What a young and innocent time...

I stopped using ChatGPT a few months ago. Since then, I have been only using oa-chat. All chat history is stored locally. Each query is sent to OpenAI under a temporary key which is unlinkable to any

These days, I'm much more excited about dataset releases than model releases. Models come and go and don't compose, whereas good datasets are more enduring and can be studied, used, revised to create

If you’re excited about creating simulation technology to transform decision-making, come join us: https://t.co/6DpuZuip9r

These questions require something more: they require a fine-grained understanding of what will happen in any given situation, and the ability to intervene and observe possible futures. In other words,

- How should we redesign the third-grade math curriculum for millions of students? - How will productivity and culture be impacted if we change our in-office policy? - …

To achieve this vision, Simile is conducting research to push the edge of what’s possible: training high-fidelity foundation models of human behavior, performing large-scale simulations efficiently, a

But what about messy, real-world questions whose answers depend on the emergent result of many humans interacting over time?

In the age of prediction, we trained general-purpose models to classify anything. In the age of reasoning, we are making models that can solve complex, multi-step problems.

We are about to enter the age of simulation.

$3M to support the development of open benchmarks!

It's a triple threat - It produces actions! It produces future states! It produces values! You just need a good video model to get you started, and you get SOTA.

Don't use weight decay. Just normalize everything you see (updates + parameters). Works on top of your favorite optimizer (e.g., Muon). Result: 33% speedup + better hyperparameter transfer.

If you use Pythia and like it, we're making an updated version. Tell us what you want. Here's a question for y'all: would you rather have a scaling suite trained on Nemotron-CC (very high quality, s

To what extent can you prompt an LM and have it generated a whole copyrighted book? This doesn't work generically, but in some cases, we were able to get much more than we thought was possible. (Pleas

Need a reward model for robotics? Frontier VLMs actually don't work. We fine-tune one that works better.

This is not just another strong open model. Nemotron actually releases training data (!), RL environments, and training code. This is a big difference: almost all model developers just want people to

Open-weight != transparency: What's interesting is not that IBM's open-weight Granite 3.3 is most transparent (95%), or that xAI's closed Grok 3 is least transparent (14%), but that Writer's closed Pa