Training AI on the Scientific Process

Today's models are trained on the finished record of science: the final paper, the final textbook. Periodic Labs is training AI on the scientific process itself, with autonomous physical labs as its reinforcement learning environments.

Liam Fedus is the co-founder of Periodic Labs. At OpenAI, he was one of the researchers who helped create ChatGPT and served as VP of research leading post-training. In this talk, he shares how Periodic's AI explained experimental data that a landmark autonomous-lab study could not, surfacing a crystal structure found in no paper or database. He walks through why RL in the physical world breaks the assumptions of digital RL, how Periodic post-trains open models to exceed frontier models on materials characterization at a fraction of the inference cost, and a scaling result for the lab itself: four months of experiments compressed into under two weeks.