Dwarkesh Patel: Eric Jang Rebuilt AlphaGo From Scratch — Self-Play RL Is Fundamentally More Sample-Efficient Than LLM Policy Gradients

Dwarkesh Podcast (YouTube)·medium signal

Former Google DeepMind robotics researcher Eric Jang spent his sabbatical rebuilding AlphaGo using modern RL frameworks and discusses the implications with Dwarkesh Patel. Key insight: AlphaGo's architecture provides per-state supervision via MCTS search rather than requiring LLMs to stumble upon winning trajectories — a fundamentally more sample-efficient paradigm. Notable claim: you can now vibe-code an AlphaGo-strength Go AI for a few thousand dollars of rented compute.