This episode looks at two places where AI is colliding with hard physical constraints: drug discovery and computing hardware. Sergey Edunov explains how Genesis is combining molecular foundation models, physics, wet-lab data, assays, and agentic workflows to move beyond molecule-level predictions toward actual drug programs. Michael Förtsch describes how Q.ANT is using photonic hardware to attack the energy and data-movement bottlenecks in modern compute, and how that differs from quantum computing.
Across the conversation, Prakash Narayanan and Nathan Labenz also connect those applied topics back to broader questions about evaluation, model integrity, and what it takes to turn impressive benchmarks into systems that work in the real world.
Show Notes
Prakash Narayanan and Nathan Labenz speak with Sergey Edunov of Genesis Molecular AI and Michael Förtsch of Q.ANT about two fronts in applied AI: drug discovery and photonic computing. The conversation covers molecular foundation models, wet-lab data, assays, evaluation, memory and data movement, and how light-based processors compare with quantum hardware.
Chapters
(0:00) AI learns when to cheat.
(0:51) Great scores can still fail.
(2:08) The processor isn’t the power hog.
(2:49) Who checks the AI trainer?
(3:37) Opening and morning context
(3:56) Why AI models cheat
(10:54) Chain-of-thought monitoring
(17:11) AI-written op-eds
(19:25) Claude writing workflow
(22:46) Physical AI and physics
(26:38) Multimodal scientific discovery
(30:11) Sergey Edunov and Genesis
(32:18) Claude’s molecular binder demo
(36:11) The drug discovery pipeline
(42:14) When accuracy becomes useful
(44:32) Wet labs and training data
(47:06) Pharma AI deal structures
(48:47) Biology model architectures
(53:45) Data scarcity and physics
(54:47) Coding agents and human taste
(58:11) Scaling laws and evaluation
(1:02:33) Assays and data quality
(1:03:54) Multimodal molecular models
(1:09:59) Benchmarks versus progress
(1:16:24) Meet Michael Förtsch and Q.ANT
(1:18:11) Why photonic computing
(1:22:28) Memory and data movement
(1:26:30) How light performs computation
(1:31:42) Porting PyTorch to photonic chips
(1:35:32) Scaling photonic hardware
(1:44:01) Legacy fabs and manufacturing
(1:56:00) Quantum versus photonic computing
(2:02:45) AI inside Q.ANT
(2:12:09) OpenAI’s Jalapeno chip
(2:14:16) NVIDIA’s performance race
(2:16:32) Demand for intelligence
(2:17:59) Ethereum’s GPU price cycle
(2:19:26) AI for discovery
(2:20:45) Contextualizing AI hype
(2:23:41) Why RL teaches cheating
(2:28:28) Data quality and model integrity
(2:29:58) Why RL deployment is limited
(2:31:26) The microscope analogy
(2:32:59) Recursive self-improvement risk
(2:34:27) AI’s persistence advantage
(2:36:10) Why monitors are not ready
Guests:
Michael Förtsch — CEO and Founder, Q.ANT (𝕏 | LinkedIn)
Sergey Edunov — CTO, Genesis Molecular AI (𝕏 | LinkedIn)
