Training an AI Scientist on the Physical World: Periodic Labs' Synthesis Mission
Key points
Key takeaways from The Generalist interview with Periodic Labs co-CEOs Liam Fedus and Ekin Dogus Cubuk (September ͧ2026):
Mission: synthesis superintelligence. Periodic Labs aims to build an AI system that can intentionally synthesise matter, closing the loop between hypothesis, simulation and physical experiment. The founders argue that "thinkism" is not enough: an AI trained only on existing literature cannot create genuinely new knowledge. It has to be trained on the scientific process itself, running experiments against the physical world.
First target: high-temperature superconductors. Chosen for impact, because superconductivity underpins quantum computing, fusion and lossless energy transmission,and remains the most macroscale quantum phenomenon known. The founders trace the field back to Heike Kamerlingh Onnes, who liquefied helium over a century ago and in doing so opened up superconductivity.
Synthesis is the bottleneck, not discovery. Even once a material is identified, making it reliably is extremely hard: precursor purity, proportions, temperature profiles, pressures and processing conditions all matter,and furnaces drift in temperature stability over time. Replication is often harder than first synthesis — early cuprates were doubted because labs struggled to reproduce them before they became textbook examples such as YBCO.
Why LLMs fall short on science. Code has an answer key — a computer can check it in milliseconds. Physics has one too, but consulting it means running real experiments, which an AI model cannot do alone. This is why Periodic believes frontier AI progress requires grounding in physical-world iteration rather than ever longer inference.
An industrial-scale experimental loop. Periodic builds labs in which an LLM iterates with simulation and automated experiments, capable of running thousands of experiments a day and learning from each one. Hardware engineering and better telemetry turn hidden variables, such as furnace degradation, into data the system can act on,and replicates are run to teach it scientific intuition about uncertainty — rerunning apparent noise rather than overindexing on it.
Two complementary trajectories. Liam Fedus,who led post-training at OpenAI and helped create ChatGPT,and Ekin Dogus Cubuk,whose Google DeepMind career spanned materials science and chemistry teams, met flipping a tyre at the gym. They founded Periodic Labs in May ͧ2025, combining post-training expertise with deep materials-science capability.
Compressing a 20-year bottleneck. Materials typically take 10–20 years from lab discovery to technology,because scale-up manufacturing fails for most discoveries. Periodic targets areas where the frontier is insufficient or where it can synthesise vastly cheaper,hopeing AI-driven loops will compress that cycle.
Read more: The Generalist · Periodic Labs
