Descriptions:
The Latent Space AI for Science podcast brings together Evan Fineberg, founder and CEO of Genesis Molecular AI, and Sergey Udov, CTO and former lead of Llama 2 and Llama 3 pre-training at Meta’s FAIR team, for a deep technical conversation about applying diffusion models and foundation model techniques to drug discovery.
The central topic is protein–small molecule interaction prediction — a problem that has historically resisted machine learning approaches. The guests explain how diffusion models, originally developed for image generation, have emerged as the most innovative primitive for 3D molecular structure prediction, including a new class of work on binding poses. A key theme is why generating physically plausible 3D poses matters: unlike predicting binding affinity as a single scalar, a pose gives med chemists a way to validate whether a model’s output is grounded in reality or hallucinated — analogous to an image model confidently generating a face that doesn’t belong to any real person.
Sergey Udov describes his career arc from physics to ML to leading Llama pre-training at FAIR and then pivoting to molecular AI at Genesis, offering an unusually candid view of what drew top foundation model researchers toward the life sciences. Evan Fineberg discusses how Genesis was founded as a Stanford spinout despite concerns about entering a crowded field with well-funded incumbents, and how the team ultimately found its edge in structure-based drug design. For anyone tracking the migration of foundation model techniques from language and vision into biology and chemistry, this episode offers rare expert-level depth.
📺 Source: Latent Space · Published June 30, 2026
🏷️ Format: Podcast
