[@DwarkeshPatel] Building AlphaGo from scratch – Eric Jang
Link: https://youtu.be/X_ZVSPcZhtw
Duration: 157 min
Transcript: Download plain text
Short Summary
Eric Jang, VP of AI at 1X Technologies and former Google DeepMind Robotics researcher, explains AlphaGo's architecture and Monte Carlo Tree Search, covering how neural networks solve NP-hard Go problems with PUCT selection, value/policy networks, and training methodology. The episode explores compute efficiency gains (achieving AlphaGo-level results for ~$10K versus millions) and why MCTS struggles with open-ended LLM reasoning. Jang also discusses scaling laws, automated AI research using game environments as verification loops, and insights on research philosophy and lateral thinking.
![[@DwarkeshPatel] Summarizer](https://summaries.pages.dev/img/logo.webp)









