Periodic Labs deep-dive: Synthesis superintelligence and autonomous labs for materials discovery

Latent Space · rss · 2026-10-09

Latent Space interviews Periodic Labs founders Liam Fedus (ChatGPT co-creator) and Ekin Dogus Cubuk (DeepMind GNoME, MatterGen) on their vision of synthesis superintelligence.

Core ideas

Intelligence is necessary but not sufficient: new knowledge comes from conjectures tested against reality; thinking alone can't push the frontier — you need high-throughput physical experiment loops.
Physical-world RL: with real labs as the environment, models must reason under noise, uncertainty, and missing information.
Failed experiments may be the most valuable training data; training should cover the full process of doing science, not just published results. Data quality beats more compute.
Give every lab instrument a "140 IQ": intelligent instruments and scaled autonomous labs could compress decades of trial-and-error into months, expanding the "surface area for luck" in discovering materials.

More highlights

The matter compiler; why DFT and simulation can't replace experiments; AI-powered X-ray diffraction characterization.
Target materials include room-temperature superconductors, new magnets, batteries, and more efficient compute; quantum computing may not automatically solve materials discovery.
The team benchmarks itself to Bell Labs, blending solid-state chemistry, physics, robotics, and LLM experts — founded only last September, already regarded as a pre-eminent AI-scientist lab.

Original post →

Periodic Labs deep-dive: Synthesis superintelligence and autonomous labs for materials discovery

More from Companies & People