For decades Go was AI's white whale — a 19x19 board has more legal positions than atoms in the observable universe, so the brute-force search that cracked chess simply doesn't fit. AlphaGo's breakthrough wasn't a faster search; it was teaching two neural networks to do the pruning and judging that humans do by intuition, then letting tree search explore only the moves worth exploring.
Key Findings
- A policy network trained on 30 million human moves predicts likely next moves, collapsing the branching factor so search wastes no time on hopeless lines.
- A value network estimates who is winning from a raw board position — replacing the hand-crafted evaluation functions that had stumped Go programmers for years.
- Both networks plug into Monte Carlo tree search, so the system reasons about consequences instead of pattern-matching alone. The full AlphaGo won 99.8% of games against other Go programs and beat European champion Fan Hui 5-0 — the first time a program defeated a human professional on a full board without handicap.
- Self-play reinforcement learning improved the policy network beyond its human teachers, hinting that imitation was only the starting point.
Who It's For and the Trade-offs
Essential reading if you care about how learned evaluation plus search beats pure search, or about the lineage that runs through AlphaZero and MuZero. Be clear-eyed about the limits: this AlphaGo bootstrapped from human game records, so it inherited human bias and needed that supervised scaffolding. The 2017 follow-up, AlphaGo Zero, dropped human data entirely and learned from self-play alone — surpassing this version and making the case that the supervised stage was a crutch, not a requirement.
