The History of AI101 · Module IV · Lesson 14 of 21
Article · 12 min

AlphaGo and reinforcement learning at scale

The 2016 demonstration and what it proved.

Summary

In March 2016, AlphaGo beat Lee Sedol 4-1 at Go — a game long considered a decade or more away from machine mastery. The system combined deep networks, Monte Carlo tree search, and self-play reinforcement learning. AlphaGo Zero (2017) achieved superhuman play from scratch, with no human games.

Objectives
  • 01State what AlphaGo demonstrated that prior systems could not.
  • 02Explain the difference between AlphaGo and AlphaGo Zero.
  • 03Name the pattern (search + neural evaluation + self-play) that generalized beyond Go.
The Lesson

The match

Move 37 of Game 2 — a move human experts initially called a mistake — is the moment the community understood something had changed. AlphaGo won 4-1 against the top-ranked Go player of the previous decade.

The generalization

The self-play RL pattern generalized to chess, shogi (AlphaZero), protein folding (AlphaFold), and matrix multiplication (AlphaTensor). The recipe travels.

Key Ideas
  • Self-play RL scales to superhuman performance in closed environments.
  • The AlphaGo template — search + neural evaluation + self-play — is the RL pattern that generalized.