Hugging Face Researcher Presents Multi-Agent System At CERN For Physics

This week I gave a talk on Physics Intern: a multi-agent scaffold we built at Hugging Face to tackle hard problems in theoretical physics

It sets a new SOTA on CritPt (currently one of the hardest benchmarks on Artificial Analysis) and is packaged as a set of skills you can plug into Codex / CC

Here's a few things I learned about how physicists are using AI tools today:

- Almost everyone has tried LLMs in their research (mostly via chat UIs, about 30% using CC / Codex), and about half have found them useful - Almost no one uses the CLI tools, preferring instead to use the desktop apps - Verification is a major bottleneck and we are very far from "vibe physics" because the models are prone to making stuff up if they are not tightly constrained

Slides here: https://docs.google.com/presentation/d/1Rkr8up-s4IzGyEyDH-SOCKjA-kbrj30C8a0C3rrTOdc/edit?usp=sharing