In a recent Latent Space episode, Periodic Labs' Liam Fedus and Ekin Dogus Cubuk laid out their case for what they call "synthesis superintelligence." The startup, launched in September 2025 by alumni of ChatGPT, DeepMind's GNoME, and OpenAI's Operator, argues that intelligence alone is insufficient for scientific discovery. Models must reason over noisy physical experiments, missing information, and uncertainty—something that cannot be learned purely from internet-scale text data.
The core loop they describe is prediction, synthesis, and characterization. Simulations and density functional theory are useful, but experiments remain the ultimate ground truth. Periodic is building autonomous labs where every instrument has what they call "140 IQ," allowing models to learn from the entire process of doing science—including failed experiments, which they argue may be some of the most valuable training data. This contrasts with typical machine learning pipelines that only train on final published answers.
The bet is that autonomous experimentation can compress decades of trial-and-error into months and massively increase the "surface area for luck" in discovering new materials like superconductors, magnets, and batteries. They also caution that quantum computing will not automatically solve materials discovery, and that data quality and negative results matter more than simply throwing more compute at the problem. Even future frontier models, they say, will still need to physically experiment. The episode is a single source, so no independent corroboration is offered, but the argument is presented as a coherent vision from the lab{