Rishabh Agarwal, a reinforcement-learning researcher now at Periodic Labs, explains why science is a harder problem for AI than coding. Coding agents never leave the computer, but making a new material requires mixing, heating and measuring samples that cannot be simulated well enough to skip. He walks through his laboratory's loop of hypothesis, synthesis, measurement and analysis, how a model learns from lab data, predicting experimental outcomes as a reinforcement-learning task, oversight of AI-proposed experiments and what transfers from games such as NetHack. He also argues that reinforcement learning is an inefficient paradigm that favours large labs, and that science is falsifiable rather than verifiable.
Hosted by Ravid Shwartz Ziv, Allen Roush
Published Nov 13, 2026 · 1 h 1 min