Periodic Labs founders outline vision for AI scientists that learn from real-world experiments

Periodic Labs was launched in September 2025 by Liam Fedus and Ekin Dogus Cubuk, who previously worked on ChatGPT, DeepMind’s GNoME, OpenAI’s Operator (now Agent), and the neural attention mechanism. The company’s stated mission is to accelerate science by building AI systems capable of autonomously conducting experiments and generating new knowledge.

The founders describe their approach as 'synthesis superintelligence,' which integrates reinforcement learning grounded in physical experiments, AI-powered materials characterization, simulations, and high-throughput lab automation. They argue that scientific discovery requires more than intelligence alone—it demands interaction with the physical world to test conjectures and learn from outcomes.

Periodic Labs emphasizes that frontier AI models, even future ones, will still need to physically experiment to validate hypotheses. The company is developing systems that can reason under uncertainty, handle noisy data, and extract value from failed experiments, which they describe as critical training signals.

The company’s vision includes giving every lab instrument '140 IQ'—meaning each piece of equipment can autonomously plan, execute, and interpret experiments. This infrastructure is intended to automate the end-to-end materials discovery loop, from prediction and synthesis to characterization and analysis.

Periodic Labs has publicly discussed scaling autonomous labs across AI, chemistry, physics, and custom hardware, with the goal of compressing decades of trial-and-error into months. The company’s progress has been noted in public posts referencing industrial-scale science deployments planned for summer 2026.