Lila: Turning Labs into Data Centers

Latent Space · rss · 2026-07-16

This interview dives into the "AI science factory" concept by Lila Sciences: they envision running laboratories like data centers, where AI orchestrates robots, instruments, and workflows to build "scientific superintelligence" across biology, chemistry, drug discovery, and materials science.

The discussion highlights their methodology of treating the scientific method as a scalable data generation mechanism and experimental results as verifiable training signals. They emphasize that the lab isn't merely an automation company but a pursuit of flexibility and generalizability, relying on rapid iteration rather than massive one-off screenings. Notably, they restructured their gas adsorption measurement processes to achieve a roughly 2500x speedup and accumulated 10 trillion experimentally verified tokens of scientific reasoning.

Furthermore, the interview covers specific cases and insights: generalization gains from cross-domain transfer, shifting molecular chemistry priors to MOF materials, model suggestions for catalyst discovery, and the pace of advancing in vivo CAR-T data in non-human primates. The overarching thesis is that achieving "superintelligence" in science requires more than just models that excel at answering questions; it demands systems capable of continuously generating high-quality data in real experiments while automating creativity and serendipity.

Original post →

More from Companies & People

Companies & People channel →