In this episode of the a16z Podcast, Anjney Midha interviews Liam Fedus and Ekin Dogus Cubuk, co-founders of Periodic Labs, about their work on building experiment-in-the-loop AI for physics and chemistry. They discuss the importance of real-world reward functions, how mid-training and high-compute reinforcement learning fit together, and why superconductivity and magnetism are the initial focus areas for developing an AI physicist. They also delve into the challenges of noisy data sets and negative results, the collaboration between ML researchers and bench scientists, and the near-term applications of their work, such as co-pilot tools for advanced industries like semiconductors, space, and manufacturing. Periodic Labs aims to accelerate scientific discovery and physical R&D by creating AI systems that can design materials with specific properties, addressing the limitations of current AI models that lack iterative experimental validation.
Sign in to continue reading, translating and more.
Continue