Building an AI Science Factory

Only about 5% of scientific hypotheses can be verified cheaply. The rest require real experiments, which makes most of science dark matter for reinforcement learning. Lila Sciences is building the verifier: an autonomous laboratory designed to train AI models.

Andrew Beam is the CTO of Lila Sciences and has spent nearly two decades at the frontier of AI in science, from Harvard Medical School to Generate Biomedicines. In this talk, he shows how Lila's AI science factory closes the loop between models and physical experiments: an AI that designs quantum dots to a target wavelength, an in vivo CAR-T candidate reached in months instead of years at a fraction of the cost, and one of the first scaling laws for scientific intelligence, measured across thousands of experimental tasks. All of it feeds one model, trained on Anyscale.