AlphaGo
Why it matters
Deep Dive
Go is deceptively simple: two players alternate placing black and white stones on a 19-by-19 grid, trying to surround territory. But the game has on the order of 10^170 possible board configurations — more than the number of atoms in the observable universe — which makes exhaustive search hopeless. When IBM's Deep Blue beat Garry Kasparov at chess in 1997, it did so largely through brute-force calculation, evaluating some 200 million positions per second. Experts expected Go to hold out for another decade or more, because strong play seemed to require something like human intuition: pattern recognition, judgment about which positions are promising, and a feel for long-term influence. AlphaGo's breakthrough was to learn that intuition from data with deep convolutional networks, and then combine it with just enough guided search to be decisive.
Two Networks and a Search Tree
AlphaGo's architecture has three moving parts. A policy network, a deep convolutional network, takes a board position and outputs a probability distribution over promising next moves; a value network scores how good a position is, estimating which side will eventually win. At move time, the system runs Monte Carlo tree search: it simulates thousands of partial games, using the policy network to decide which moves are worth exploring and the value network to evaluate the resulting positions without playing them out to the end. Training happened in stages — first supervised learning on roughly 30 million positions from human games on online Go servers (the network learned to predict expert moves with about 57% accuracy), then reinforcement learning, in which the network played millions of games against earlier versions of itself with winning as the only reward signal. This combination of learned evaluation and guided search is the core idea; the neural networks prune a search space that raw computation cannot touch.
Move 37 and the Match That Changed Everything
AlphaGo first made headlines in early 2016, when DeepMind revealed it had beaten Fan Hui, the reigning European champion, 5–0 the previous October — the first time a program had defeated a professional on a full-size board without a handicap. The Lee Sedol match that March was the real test: Lee was a legend with 18 international titles, and most observers expected him to win comfortably. Instead AlphaGo won the first three games and took the match 4–1. Game two produced the famous Move 37, a shoulder hit on the fifth line that commentators initially read as a mistake; AlphaGo's own policy network estimated the odds of a human playing it at roughly one in ten thousand, yet it proved decisive and is now studied as a genuinely new idea in a game humans have refined for millennia. Lee answered with a brilliant wedge move of his own in game four (the celebrated Move 78), the only game a top human ever won against AlphaGo's final versions. In 2017 the program beat world number one Ke Jie 3–0 and was retired from competitive play.
It Didn't Win by Brute Force
A common misconception is that AlphaGo won by out-calculating Lee Sedol, the way Deep Blue overpowered Kasparov. The opposite is closer to the truth: Deep Blue examined on the order of 200 million positions per second, while AlphaGo evaluated only thousands — orders of magnitude fewer. Go's branching factor (around 250 legal moves in a typical position, versus roughly 35 in chess) means raw enumeration gets nowhere; what AlphaGo added was learned judgment about which handful of moves deserved a closer look and which positions were actually winning. In that sense the system plays less like a chess calculator and more like a very experienced player with perfect memory. It is worth keeping the limits in view, though: everything AlphaGo knows about Go lives inside its deep learning networks, and outside the game it can do nothing at all.
From AlphaGo Zero to MuZero
In 2017 DeepMind published AlphaGo Zero, which removed human knowledge entirely: starting from random play and knowing only the rules, it trained purely on self-play, surpassed the version that beat Lee Sedol after about three days (defeating it 100–0), and after forty days was stronger than every previous version. AlphaZero then generalized the same algorithm to chess and shogi, reaching superhuman strength in each after only hours of self-play and beating the top traditional engine, Stockfish, in a widely discussed match. MuZero pushed the idea furthest: it is not even given the rules, instead learning an internal model of how the environment responds to actions — a learned world model — and it matched AlphaZero on Go, chess, and shogi while also mastering Atari games. This lineage matters because each step removed another dependency on human examples, showing that search plus self-improvement can bootstrap expertise from scratch.
The Legacy in Modern AI
AlphaGo's impact far outlived its retirement from competition. The match ran on Google's custom TPU hardware, and its aftermath ignited the modern AI wave — South Korea announced a national AI strategy within weeks, and investment in deep learning surged worldwide. DeepMind applied a similar learning-driven playbook to science with AlphaFold, which effectively solved protein structure prediction. Inside language AI, the fingerprints are everywhere: RLHF and its variants are reinforcement learning driven by a reward signal, and modern reasoning models pair reinforcement learning with search-like exploration at inference time — a direct descendant of the AlphaGo recipe of letting a system practice, evaluate its own attempts, and keep what works.