Summary
The Latent Space Science podcast hosts Brandon and RJ sit down with Andy Beam (CTO) and Rafa Gomez-Bombarelli (Chief Scientific Officer, Physical Sciences) from LILA Sciences to unpack one of the more ambitious bets in AI: using the physical laboratory itself as a verifier for reinforcement learning at scale.
The core thesis is that the internet — the training data that powered the current generation of large language models — is essentially exhausted as a scaling resource. LILA’s answer is to treat scientific experimentation as a new, essentially infinite data source. Rather than scraping text, their system runs real-world experiments in materials science, chemistry, and biology, then uses the measurable outcomes (did the compound bind? did the material scale?) as verifiable reward signals to train models that improve through iteration. Beam and Gomez-Bombarelli draw an explicit parallel to the “bitter lesson”: methods that scale and generalize beat hand-crafted domain heuristics every time.
The conversation covers the practical challenges of this approach — particularly the notorious difficulty of scaling discoveries in materials and chemistry to commercial production — and how LILA embeds supply-chain and techno-economic reasoning directly into the experimental loop. Both founders bring deep academic credentials (Harvard, MIT, Generate Biomedicines) alongside startup experience, making this a rare interview that balances frontier AI methodology with hard-won scientific domain knowledge. Essential listening for anyone tracking AI’s expansion into life sciences and physical-world discovery.
📺 Source: Latent Space · Published July 16, 2026
🏷️ Format: Podcast
