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

Periodic Labs, launched last September by Liam Fedus and Ekin Dogus Cubuk, is pursuing what it calls "synthesis superintelligence" — AI systems grounded in physical experiments rather than just digital data. The founders, who previously worked on ChatGPT, DeepMind's GNoME, and MatterGen, argue that scientific discovery requires a fundamentally different approach than math or coding because the environment itself is noisy, uncertain, and incomplete.

Fedus explained that reinforcement learning environments at Periodic derive directly from physical labs, where data is the ultimate truth. But unlike digital environments where more rollouts can be added arbitrarily, physical experiments have speed limits and require sample efficiency. He noted that materials don't come out of the furnace labeled, labeling can be stochastic, temperatures may deviate from setpoints, and machines can produce aberrant measurements.

Cubuk described how every scientific experiment requires dimensionality reduction, citing thermodynamics as an example where five or six variables can explain complex systems. He contrasted this with coding or math, where all context is available. In materials discovery, the loop involves predicting what to make, synthesizing it, and characterizing what was produced — a process Cubuk said is "non-trivial" at every step because even successful synthesis doesn't guarantee the target material.

Periodic's early AI work focuses on characterization, particularly using X-ray diffraction to identify phases present in materials. Fedus said the reinforcement learning environment rewards accurate phase identification and penalizes spurious or chemically implausible results. The company aims to train models on the process of doing science rather than final answers, including failed experiments and negative results.

Fedus and Cubuk described giving every piece of lab equipment "140 IQ" to enable autonomous experimentation. They aim to increase the "surface area for luck" in discovering new materials, with potential applications in room-temperature superconductors, magnets, batteries, and more efficient compute. The founders assembled an interdisciplinary team spanning solid-state chemistry, physics, hardware engineering, and LLM expertise, drawing comparisons to Bell Labs.

Why it matters: This approach could compress decades of scientific trial-and-error into months, according to the founders. For research teams, it points toward AI systems that reason under uncertainty and learn from experimental processes, not just published outcomes.