Liam Fedus is co-founder of Periodic Labs, where he works on AI scientists and autonomous physical laboratories. Previously OpenAI’s vice president of post-training, he now focuses on systems that learn from experiments and help decide what materials to investigate next.[1][2][3][4][5]
Start with AI Scientists Are Here: Autonomous Labs & Synthesis Superintelligence — Periodic Labs →
Links
Periodic Labs LinkedIn
X · LiamFedus Co-founder · Periodic Labs
Career
Fedus’s public professional biography records earlier work at OpenAI and Google Brain. Periodic lists him in its founding team, which brings together model developers, physical-science researchers and laboratory builders. The company’s program couples model development with experiments that produce new data, including results that would rarely appear in published papers.[1][2]
Contributions
Periodic’s September 2026 research update describes a loop of choosing a target, synthesizing a material and characterizing the result. Its Neon model was midtrained and reinforced on laboratory data to analyze experimental results. The Periodic team reports deploying the model in its labs to support that discovery loop.[3]
Ideas
In his conversation with Ekin Dogus Cubuk, Fedus examines why scientific discovery differs from a fast, verifiable coding task. Physical experiments take equipment and time, produce noisy measurements and may fail in informative ways. Simulation and prediction help choose experiments, while characterization checks what was actually made. The proposed AI scientist must reason through that entire process.[4]
Tools, writing & talks
Writing, research & talks
Start here
A conversation on experimental feedback, materials characterization and autonomous labs.
Start with the materials-discovery loop around 9:17, then the distinction between simulation and experimental results. No laboratory experience is needed for the discussion.
Interview guest with Ekin Dogus Cubuk[4]Original source- Building Labs that Learnperiodic.com
Periodic’s account of laboratory data, the Neon model and a materials-discovery loop.