AlphaGo versus Lee Sedol

Based on Wikipedia: AlphaGo versus Lee Sedol

In March 2016, something remarkable happened in Seoul that most of the world barely noticed at the time, but which historians of artificial intelligence now recognize as a watershed moment: a computer program defeated one of humanity's greatest strategic minds at the most complex game ever invented.

The game was Go. The human was Lee Sedol, a South Korean grandmaster who had won eighteen international titles and was considered one of the strongest players in the game's four-thousand-year history. The machine was AlphaGo, a program developed by Google's DeepMind division.

AlphaGo won four games out of five.

Why This Mattered

To understand why this match shook the artificial intelligence community, you need to understand what makes Go different from chess.

Chess is complicated. At any given position, there might be thirty or forty legal moves to consider. The game tree—the branching map of all possible future positions—is enormous. When IBM's Deep Blue defeated world chess champion Garry Kasparov in 1997, it did so largely through brute computational force, calculating millions of positions per second and choosing the path that led to the best outcome.

Go is something else entirely.

The game is played on a nineteen-by-nineteen grid. Players take turns placing black and white stones, attempting to surround territory and capture their opponent's pieces. The rules are simple enough to teach a child in five minutes. But the number of possible board positions exceeds the number of atoms in the observable universe.

The branching factor—the average number of legal moves at any point—is around two hundred and fifty. This makes brute-force calculation effectively useless. You cannot simply calculate your way to victory in Go. You need something that looks suspiciously like intuition.

Back in 1965, mathematician I. J. Good wrote that programming a computer to play reasonable Go would be even more difficult than chess, because "the principles are more qualitative and mysterious than in chess, and depend more on judgment." Before 2015, the best Go programs could only reach amateur dan level—respectable, but nowhere near professional strength. Many researchers believed computers would never crack the game.

Elon Musk, an early investor in DeepMind, said in 2016 that experts in the field had thought artificial intelligence was ten years away from defeating a top professional Go player.

They were off by nine years.

A Different Kind of Intelligence

AlphaGo didn't work like Deep Blue. It didn't rely on probability algorithms painstakingly hand-coded by human programmers. Instead, it used neural networks—mathematical structures loosely inspired by how neurons connect in the brain—to evaluate positions and estimate its probability of winning.

The training process worked in two stages. First, the system studied about thirty million moves from one hundred and sixty thousand games played by strong human players, learning to mimic expert play. Once it reached a certain level of competence, the real training began: AlphaGo played millions of games against copies of itself, using a technique called reinforcement learning to gradually improve.

What emerged was something genuinely strange. As one of AlphaGo's creators explained: "Although we have programmed this machine to play, we have no idea what moves it will come up with. Its moves are an emergent phenomenon from the training. We just create the data sets and the training algorithms. But the moves it then comes up with are out of our hands—and much better than we, as Go players, could come up with."

In other words, they had built a system that could discover strategies they themselves didn't understand.

The Warm-Up Match

In October 2015, AlphaGo played its first match against a professional Go player: Fan Hui, the European champion, who held a two dan professional rank. AlphaGo won five games to zero.

This was historic—the first time an artificial intelligence had defeated a professional player on a full-sized board without handicap. But some observers stressed the gulf between Fan Hui and the true elite. Lee Sedol held a nine dan professional rank, the highest level. The difference between two dan and nine dan is immense.

Computer Go programs had previously beaten nine dan professionals, but only with handicaps of four or five stones—meaning the human player had to spot the computer multiple pieces at the start. Jonathan Schaeffer, a Canadian artificial intelligence specialist, compared AlphaGo after the Fan Hui match to "a child prodigy" that lacked experience. He predicted Lee would win in March.

Lee himself was confident. He initially predicted he would win in a "landslide."

What nobody knew was how much AlphaGo had improved since October. Between the Fan Hui match and the Lee Sedol match, the program had played tens of millions more games against itself. Go experts had identified errors in AlphaGo's play against Fan—particularly a lack of awareness of the full board position. Whether those weaknesses had been corrected remained to be seen.

The Stakes

The match was set up as a five-game series, played at the Four Seasons Hotel in Seoul. The prize was one million dollars. Each game would be played under Chinese rules with a time control of two hours per player, followed by three one-minute overtime periods. The games were livestreamed with expert commentary.

Because AlphaGo can't physically place stones on a board, Aja Huang—a DeepMind team member and amateur six dan Go player—would place the moves on AlphaGo's behalf. The program itself ran on Google's Cloud Platform, with servers located in the United States. It used twelve hundred and two central processing units and one hundred and seventy-six graphics processing units, roughly the same computational power it had used against Fan Hui. Google also deployed its proprietary tensor processing units, specialized chips designed specifically for machine learning calculations.

Lee Sedol, meanwhile, brought what no machine possessed: eighteen years of professional experience, eighteen international titles, and a reputation for unconventional, creative play that had made him a national hero in South Korea.

Game One: The Unexpected Collapse

On March ninth, 2016, Lee Sedol sat down across from Aja Huang and began the first game. AlphaGo played white.

For most of the match, Lee appeared to be in control. Then, in the final twenty minutes, something shifted. AlphaGo began to pull ahead. Lee resigned.

Afterwards, Lee said he had made a critical error early in the game. He acknowledged that AlphaGo's strategy in the opening was "excellent" and that the program had made one unusual move that no human player would have attempted.

Professional commentators analyzing the game noted that Lee's seventh stone was a strange testing move, seemingly designed to probe AlphaGo's strength. It turned out to be a mistake. AlphaGo's response was accurate and efficient. The program built a favorable position in the first part of the game. Lee started to mount a comeback around move eighty-one, but made questionable decisions at moves one hundred and nineteen and one hundred and twenty-three, followed by what analysts called a "losing move" at one hundred and twenty-nine.

Professional nine dan player Cho Hanseung commented that AlphaGo's game had greatly improved from its October match against Fan Hui. Michael Redmond, providing English commentary, described the computer's play as more aggressive than it had been five months earlier.

One moment stood out. On AlphaGo's one hundred and second move, Lee Sedol sat stunned. He mulled over his options for more than ten minutes. According to nine dan grandmaster Kim Seong-ryong, that was the moment Lee realized just how strong his opponent truly was.

Game Two: The Shoulder Hit

The second game took place on March tenth. AlphaGo played black this time.

Lee stated afterwards that "AlphaGo played a nearly perfect game" and that "from the very beginning I did not feel like there was a point that I was leading." Demis Hassabis, one of AlphaGo's creators, said the system was confident of victory from the midway point, even though professional commentators couldn't tell which player was ahead.

The key moment came at move thirty-seven. AlphaGo placed its nineteenth stone in a position that professional commentator Michael Redmond called "creative" and "unique." It was a move no human would have ever made.

Lee took an unusually long time to respond.

Professional eight dan player An Younggil called it "a rare and intriguing shoulder hit"—a Go term for a specific type of tactical play—but praised Lee's counter as "exquisite." Control of the game passed back and forth several times before the endgame. An particularly praised AlphaGo's moves one hundred and fifty-one, one hundred and fifty-seven, and one hundred and fifty-nine, calling them "brilliant."

Something else became apparent in this game. AlphaGo didn't play like a human trying to maximize points or margin of victory. It played to maximize its probability of winning. If given a choice between winning by twenty points with eighty percent probability and winning by one and a half points with ninety-nine percent probability, it would choose the latter—even if that meant giving up points to achieve it.

This led to moments that looked like mistakes to human observers. Move one hundred and sixty-seven, for instance, seemed to give Lee a fighting chance. Commentators declared it a blatant error. But An Younggil offered a different interpretation: "When AlphaGo plays a slack-looking move, we may regard it as a mistake, but perhaps it should more accurately be viewed as a declaration of victory."

Game Three: All Doubt Removed

After two losses, some observers still questioned whether AlphaGo was truly strong in the way a human grandmaster is strong, or whether it had simply gotten lucky. The third game, played on March twelfth, removed all doubt.

AlphaGo won so convincingly that analysts described it as "almost scary." Lee mounted a severe, one-sided attack, forcing AlphaGo to defend—and in doing so, revealed the program's power. As one commentator put it: "Lee wasn't gaining enough profit from his attack... One of the greatest virtuosos of the middle game had just been upstaged in black and white clarity."

An Younggil and fellow analyst David Ormerod concluded that the game showed "AlphaGo is simply stronger than any known human Go player."

The game also demonstrated AlphaGo's ability to handle tricky tactical situations called ko—a Japanese term for certain repeating patterns that didn't appear in the previous matches. At move one hundred and forty-eight, in the middle of a complex ko fight, AlphaGo played a significant move elsewhere on the board—a sign it was confident enough in its position to ignore the immediate local conflict.

Lee opened with a high Chinese formation, generating a large area of black influence. AlphaGo invaded at move twelve, forcing its own weak group to defend, which it did successfully. An described Lee's move thirty-one as possibly the "losing move." By move forty-eight, AlphaGo had seized control and forced Lee onto the defensive. Lee counterattacked twice, but AlphaGo's responses were effective. Lee eventually attempted a complex ko fight starting at move one hundred and thirty-one, but couldn't force an error. He resigned at move one hundred and seventy-six.

The match stood at three to zero. AlphaGo had already won.

Game Four: The Divine Move

Lee Sedol came to the fourth game with nothing to lose and a new strategy.

He chose to play amashi—an extreme, all-or-nothing style that focuses on taking territory at the perimeter rather than the center. This was a deliberate counter to AlphaGo's preference for souba Go: winning through many small gains when opportunities arise. By forcing an all-or-nothing situation, Lee hoped to neutralize AlphaGo's ability to accumulate slim advantages.

The opening eleven moves were identical to game two. Lee, playing white, concentrated on taking territory in the edges and corners, allowing AlphaGo to gain influence in the top and center. Then Lee invaded AlphaGo's sphere of influence at the top with moves forty through forty-eight, following the amashi strategy.

AlphaGo responded with a shoulder hit at move forty-seven, sacrificing four stones elsewhere and gaining initiative. Lee tested the program with moves seventy-two through seventy-six without provoking an error. By this point, commentators felt Lee's play was a lost cause.

Then came move seventy-eight.

Lee placed a white stone that professional nine dan player Gu Li described as a "divine move." It was a brilliant tesuji—a Japanese term for a clever tactical play—that developed a white wedge at the center and dramatically increased the game's complexity. Gu Li said the move was completely unforeseen. Move seventy-eight broke new ground, a radical departure from conventional play.

AlphaGo, at that moment, estimated it had a seventy percent chance of winning.

Then it played move seventy-nine, and everything fell apart.

The move was terrible. Lee followed up with a strong play at white eighty-two. AlphaGo's initial response—moves eighty-three through eighty-five—was appropriate, but at move eighty-seven, its estimate of its winning chances suddenly plummeted. What followed was a cascade of catastrophic errors: moves eighty-seven through one hundred and one were all terrible. David Ormerod characterized them as typical mistakes of Monte Carlo-based programs—the kind of failure that happens when the system's probability estimates go haywire.

Lee Sedol won the fourth game.

He had done what many thought impossible. He had found AlphaGo's weakness and exploited it with a single brilliant move. The machine was beatable after all.

Game Five: The Final Word

The fifth and final game took place on March fifteenth. AlphaGo won by resignation.

The final score was four games to one in AlphaGo's favor. The prize money—one million dollars—went to charity, as Google DeepMind had announced. Lee Sedol received one hundred and seventy thousand dollars: one hundred and fifty thousand for participating in all five games, and an additional twenty thousand for his single victory.

After the match, the Korea Baduk Association—Baduk being the Korean name for Go—awarded AlphaGo the honorary rank of nine dan grandmaster, in recognition of its "sincere efforts" to master the game. It was the first time a computer program had received such an honor.

What Changed

In the immediate aftermath, Fan Hui—the European champion AlphaGo had defeated the previous October—said the experience had taught him to be a better player and to see things he hadn't previously seen. By March 2016, his world ranking had risen from six hundred and thirty-three to around three hundred.

The match itself was chosen by Science magazine as one of the runners-up for Breakthrough of the Year on December twenty-second, 2016.

But the deeper impact was harder to measure. Just as Kasparov's loss to Deep Blue in 1997 marked the moment computers became better than humans at chess, the Lee Sedol match marked a shift in how we thought about artificial intelligence and intuition.

Go had long been considered the kind of game that required something ineffable—human judgment, creativity, the ability to read a position and just know the right move without being able to articulate why. The fact that a machine could not only replicate this ability but surpass the best human players suggested that intuition itself might be less mysterious than we thought. Perhaps it was just pattern recognition operating on a sufficiently complex scale.

Or perhaps AlphaGo had developed something genuinely new: not human intuition, but machine intuition—a different kind of intelligence altogether, one that could see patterns and possibilities invisible to human minds.

The research behind AlphaGo has since been applied to fields ranging from cognitive science to pattern recognition to drug discovery. The neural network architectures and reinforcement learning techniques pioneered for the game have influenced everything from protein folding prediction to climate modeling.

And somewhere in the vast space of possible Go positions—larger than the number of atoms in the universe—move thirty-seven of game two remains: a shoulder hit no human would have played, placed by a mind we built but don't fully understand, in a game older than writing itself.