AlphaGo defeated Lee Sedol for the third consecutive time on Saturday, clinching a five-game match against one of the world’s strongest Go players and crossing a threshold that many artificial-intelligence researchers believed was still years away.

Lee resigned after 176 moves in Seoul, giving the program built by Google’s DeepMind unit an insurmountable 3-to-0 lead with two games remaining. The result is more than another machine victory at a board game. Go’s enormous number of possible positions and its reliance on judgment about shape, influence and long-term balance had resisted the brute-force techniques that conquered chess.

Reuters reported from Seoul that the victory surprised even AlphaGo’s designers and left Lee apologizing for what he called a powerless display. The 33-year-old South Korean champion holds 18 international titles and entered the match confident that human experience and intuition would expose limits in the software.

A Go board has 19 lines by 19 lines, creating 361 possible locations for the opening move and an astronomical space of later configurations. Unlike chess, the game offers no simple material count for evaluating who is ahead. A cluster of stones may be alive, vulnerable, influential or expendable depending on developments far across the board. Strong players often describe their choices as intuitive because exhaustive calculation is impossible.

That made Go an unusually demanding test for computation. Traditional game-playing systems search a tree of possible moves and score the resulting positions. In Go, the tree expands too rapidly and positions are too difficult to value accurately. Programs had improved through Monte Carlo methods — sampling many possible continuations — but until recently they could not defeat the best professionals without a handicap.

AlphaGo combines search with learned judgment. The system described in DeepMind’s January paper in Nature uses a “policy network” to select promising moves and a “value network” to estimate the probability of winning from a position. Those networks narrow the search, allowing the program to spend computation on lines that experience has taught it are plausible.

The experience came in two stages. AlphaGo first learned patterns from recorded games played by strong humans. It then played millions of games against versions of itself, using reinforcement learning to improve through wins and losses. The method allows it to discover strategies that are effective even when they do not resemble conventional professional play.

Three Games Broaden the Evidence

The first game showed that AlphaGo could withstand the pressure of a live match against a 9-dan champion. Lee resigned after roughly three and a half hours. Reuters’s report on Wednesday’s opener noted that Lee held the black stones and that tens of thousands watched a live online stream as the machine repeatedly found precise answers.

The second game was more unsettling for professionals because AlphaGo produced moves that initially looked questionable and later proved powerful. A shoulder hit early in the game, played on the fifth line, surprised commentators accustomed to established opening principles. Wired’s analysis of Game Two described the program’s play as creative rather than merely accurate, an important distinction for a system often portrayed as a calculator.

Saturday’s third game tested a different weakness. Lee sought complications and initiated a ko fight, a recurring tactical battle in which players must answer threats elsewhere on the board. Such positions demand global judgment and can generate long, branching sequences. AlphaGo maintained control, simplified where advantageous and forced Lee’s resignation.

The Wired account of the decisive game emphasized that the program did not merely survive tactical complexity; it managed it while preserving a favorable overall position. That performance weakens the theory that a human champion could defeat the system by steering play into disorder.

Machine Intuition Is Statistical, but Still Powerful

Calling AlphaGo intuitive does not mean it thinks or experiences the game as a person does. Its networks encode statistical relationships between board positions, moves and outcomes. It does not attach cultural meaning to Go, feel pressure from an opponent or understand why spectators regard a move as beautiful.

Yet the distinction does not make its decisions less consequential. Human expertise also depends on pattern recognition built through repeated exposure, though it is integrated with language, emotion, embodied experience and conscious reasoning. AlphaGo demonstrates that a narrower system can reproduce a crucial part of expert judgment well enough to outperform a master in a domain where explicit rules do not specify good strategy.

Google’s match account published before play began presented the contest as the ultimate challenge for a program that had already beaten European champion Fan Hui five games to none. Lee represented a far higher standard. Winning the first three games therefore provides evidence that the method scales beyond the opponent used to validate the original research.

The system’s objective also differs subtly from a human desire to dominate the board. AlphaGo chooses moves that maximize its estimated probability of winning, not the size of its final margin. A quiet move that preserves a narrow but secure advantage may be preferable to an aggressive move offering more points with greater risk. That can make its play appear conservative in one moment and startlingly unconventional in the next.

From Games to General-Purpose Learning

DeepMind’s larger claim is that the techniques behind AlphaGo can address problems where useful decisions cannot be fully programmed in advance. Neural networks already support speech recognition, image classification and recommendation systems. Reinforcement learning offers a way for software to improve through feedback rather than relying only on labeled examples.

Go remains a bounded environment: the rules are clear, every move is visible and success is defined by winning. Medical diagnosis, robotics, energy management and scientific research contain uncertain data, shifting objectives and consequences that cannot be reduced to a score. AlphaGo does not prove that one system can move directly from a board to those tasks.

It does, however, show that a combination of learned representations, simulated experience and selective search can handle a problem once thought to require distinctly human intuition. The Guardian’s report after Game One captured the immediate historical comparison to IBM’s Deep Blue, but the technical advance here is different: AlphaGo relies less on enumerating possibilities and more on learning which possibilities matter.

Two games remain, and Lee has said he wants at least one victory. Their outcome may reveal vulnerabilities in the program or further examples of unfamiliar strategy. The central result is already secure. A machine has defeated an elite human in a contest built around judgment under overwhelming complexity, and it did so not once by accident but three times in succession.