Synthesis Superintelligence: from Semiconductors to Superconductors — Periodic Labs’ Liam Fedus and Ekin Dogus Cubuk
Latent Space 1 week ago 37
Periodic Labs says it is building AI “scientists” by grounding reinforcement learning and learning loops in physical high-throughput experiments and material-physics simulations rather than relying only on training data. It argues that autonomous high-throughput robotic experimentation has been tried for only the last three years or so, so its approach targets sample efficiency and reasoning under noisy, incomplete, uncertain measurements. As a result, frontier AI models are expected to repeatedly run experiments, treat failed runs as training data, and let lab instruments act with an intelligence layer to speed scientific discovery of new materials.