Lee Sedol vs AlphaGo: The Only Korean to Beat Google’s AI
In March 2016, the world watched one of the most unforgettable matches in modern technology history.
A Korean Go master sat down in Seoul to face AlphaGo, the artificial intelligence system developed by Google DeepMind. At the time, many people believed top human Go players would remain ahead of AI for many more years.
Instead, the match became a turning point.
AlphaGo won the series 4–1. But that single win by Lee Sedol became the most famous human victory in the history of AI and Go.
So far, Lee Sedol remains the only human player to defeat AlphaGo in an official public match.
Why Was Lee Sedol Chosen?
Before meeting Lee Sedol, AlphaGo had already made history by defeating European champion Fan Hui 5–0 in 2015.
That result proved AlphaGo was no longer just an interesting experiment. It was a real threat to the best players in the world.
Lee Sedol was chosen because he was one of the strongest and most respected Go players of his era. He had won 18 international titles and was known for his aggressive, creative, and unpredictable style. He was the kind of player who could create chaos on the board and find chances where others saw none.
That made him the perfect test.
DeepMind did not want AlphaGo to face just any professional. They wanted it to face one of the greatest players of the modern era.
Image Credit: LG Electronics / Wikimedia Commons, CC BY 2.0
The Match Result: AlphaGo 4, Lee Sedol 1
The five-game match took place in Seoul in March 2016.
AlphaGo won the first three games in a row. By then, the match itself had already been decided. But Game 4 was still ahead, and that game would become the most famous one of all.
In Game 4, Lee Sedol played White. At one point, the position looked extremely difficult for him. Then he played Move 78.
That move changed everything.
Lee Sedol went on to win Game 4, handing AlphaGo its only official public defeat by a human. AlphaGo then won Game 5, making the final result:
AlphaGo 4 — Lee Sedol 1
On paper, it looks like a clear AI victory.
But history remembers something else too: one human still found a way through.
Why Move 78 Became “The Divine Move”
Move 78 from Game 4 is one of the most discussed moves in Go history.
It was surprising, creative, and difficult to understand at first glance. Even strong professional players did not immediately grasp how powerful it was. The move attacked AlphaGo in a way the AI did not seem prepared for, and after that moment AlphaGo began making unusual mistakes.
DeepMind later explained that, according to AlphaGo’s own system, the probability of choosing that move was extremely low. That is one reason the move became so famous.
Many professionals described it as a “divine move” or something close to a “move from the gods.” Chinese world champion Ke Jie, who later faced AlphaGo himself, reacted with awe and said the move felt almost beyond normal human imagination.
That is why Move 78 was remembered not just as a good move, but as a moment when human creativity broke through machine logic.
Image Credit: Axd / Wikimedia Commons, CC BY-SA 4.0
How Many Possible Go Positions Are There?
Go is famous for its enormous complexity.
A standard board has 19 by 19 intersections, which means 361 possible points for the first move alone. But the real complexity comes from how quickly the number of possible board states explodes after that.
A commonly cited estimate is around 10¹⁷⁰ possible board configurations.
That number is almost impossible to imagine, so it is often compared to the estimated number of atoms in the observable universe, which is around 10⁸⁰.
In other words, the number of possible Go positions is vastly larger than the number of atoms in the observable universe.
That is why Go was long considered one of the hardest games for AI. A machine cannot simply calculate every possible game from start to finish. There are far too many possibilities.
AlphaGo succeeded because it did not try to search everything equally. It tried to focus on the branches that mattered most.
How Did AlphaGo Actually “Think”?
A lot of people imagine AlphaGo as a machine that simply brute-forced everything.
That is not really how it worked.
AlphaGo combined deep neural networks with a search method called Monte Carlo Tree Search. One part of the system helped predict promising moves. Another part helped estimate which side was more likely to win from a given position.
So instead of checking every possible path, AlphaGo tried to identify the most meaningful ones, then explored those branches more deeply.
That is why visual explanations of AlphaGo often show a tree-like structure: one board position spreading into many possible continuations, with some branches receiving much more attention than others.
The third image below is a useful way to imagine that process. It represents AlphaGo examining many possible lines, narrowing them, and focusing on the most important route.
Image Credit: AI-generated image created for Seoul Routine Lab
How Fast Was AlphaGo?
One of the most surprising facts is that AlphaGo was not winning simply because it searched more positions than every earlier machine.
During public explanations of the system, it was noted that AlphaGo evaluated roughly 100,000 positions per second.
That sounds huge, and it is. But it is still far lower than IBM Deep Blue, the chess computer that famously examined around 200 million positions per second.
So AlphaGo’s power was not just raw speed.
Its real strength was choosing what was worth thinking about.
That was the breakthrough. Instead of blindly calculating everything, it used learned judgment to focus on the most promising moves.
In a strange way, that made AlphaGo feel more human-like than older game computers.
How Many Computers Did AlphaGo Use?
This part can be confusing because two different numbers are often mentioned.
One widely quoted number is 1,202 CPUs and 176 GPUs. That figure comes from DeepMind’s published research on an earlier distributed AlphaGo system.
However, the version used in the Lee Sedol match, often called AlphaGo Lee, is more closely associated with 48 TPUs. TPU stands for Tensor Processing Unit, a specialized processor designed for machine learning.
So if someone asks, “How many computers did AlphaGo use against Lee Sedol?” the best answer is that it used a distributed Google system centered around 48 TPUs, while the famous 1,202 CPU and 176 GPU figure refers to an earlier large-scale setup described in the research.
Either way, the point is clear: AlphaGo was not a simple desktop program. It was a massive AI system backed by serious computing power.
Why Lee Sedol’s One Win Still Matters
Lee Sedol lost the overall match.
That part is true.
But it is also true that he gave AlphaGo something no other human publicly did: a loss.
Later, stronger versions of AlphaGo defeated top players like Ke Jie and then retired undefeated in other major public matches. That makes Lee Sedol’s Game 4 victory feel even more special in hindsight.
It was only one win.
But it was a win against one of the most important AI systems ever created.
That is why Lee Sedol is still remembered not just as a great Go player, but as the human who proved that even in the age of artificial intelligence, creativity could still surprise the machine.
Comments
Post a Comment