2026-08-11
AlphaGo beat pro player Fan Hui 5–0 using policy networks to select moves, value networks to score positions, and Monte Carlo tree search, all sharpened by self-play RL.
Go is the hardest of the classic board games for AI, not because the rules are complex but for two reasons: the search space is astronomically large (far beyond chess), and board positions are extremely hard to evaluate, since a single move's value often only shows up dozens of moves later. After Deep Blue took chess in 1997, Go was widely expected to be "at least a decade" away from being cracked, and the strongest programs still played at amateur-human level. The state of the art was Monte Carlo tree search (MCTS), which estimates positions through many random simulations but relies on brute-force simulation with no real intuition.
AlphaGo distills human intuition into neural networks and then lets the networks surpass themselves. Two networks sit at the core:
Both networks are trained by a combination of supervised and reinforcement learning. The policy network first learns by supervised learning on human expert games, imitating professional move selection, then improves through reinforcement-learning self-play, repeatedly beating its own previous version and reaching strength beyond what human data contained. During play, both networks are wired into MCTS: the policy network prunes which branches to search, and the value network together with Monte Carlo simulation scores positions, letting AlphaGo make far sharper judgments than pure MCTS within a fixed simulation budget.
The design choice that matters most is what the networks predict: structured latent quantities, a move distribution and a position evaluation, rather than the next token as in today's large language models.
The numbers from the abstract:
| Opponent | Setting | Result |
| Other Go programs | tournament | 99.8% win rate |
| Fan Hui (European champion) | 5-game match | 5–0 |
| Neural networks alone, no search | control | on par with SOTA MCTS programs that simulate thousands of random games |
Caveat: finer figures such as the detailed Elo ratings and per-network accuracies sit behind the Nature paywall. This writeup is based on the public abstract and figure captions; the full text was not retrieved.
This was the first time a computer program beat a human professional on a full 19x19 Go board, cashing in the "at least a decade" expectation immediately. The lesson for practitioners is the recipe: distill human intuition, then surpass it with self-play, then plug the learned judgments into search. AlphaZero later dropped the reliance on human games, and MuZero extended the idea to environments without known rules, turning this into the template for the "learn to evaluate, then search" family. In the current debate, AlphaGo reached superhuman play by predicting value and policy, structured latent variables, well before token prediction became the dominant paradigm. It is a concrete reminder that predicting the next token is not the only abstraction level that leads to strong AI.