Synthesis Superintelligence: from Semiconductors to Superconductors — Periodic Labs’ Liam Fedus and Ekin Dogus Cubuk

Periodic Labs, an AI-for-science lab, argues that discovering new materials requires AI trained on real experiments and not only on published text. In a Latent Space interview, co-founder Liam Fedus says the lab's reinforcement learning environments derive from its physical labs, so the data comes from experiments, not from known answers in papers or textbooks.

He contrasts this with math, where measurements are exact. In a materials lab, results are noisy and unlabeled, and instruments can differ or report a temperature slightly off from the one intended, so the model must make decisions under uncertainty. Because physical experiments cannot be repeated as freely as software runs, Fedus names sample efficiency, getting more from limited data, as a key issue.

These are the founders' stated approach and reasoning in an interview. The portion reviewed does not report experimental results, so the supported conclusion is why they think lab data is needed, not that the approach has succeeded.