10 years of AlphaGo: The turning point for AI | Thore Graepel & Pushmeet Kohli

### TL;DR In 2016, AlphaGo defeated world champion Lee Sedol in the game of Go, marking a turning point in the history of artificial intelligence. This article reviews the complexity of Go, the technical principles behind AlphaGo’s combination of intuition and planning, and the details of the world-changing match. ## Go: The Ultimate Challenge for AI Go has simple rules but produces exponentially complex positions, exceeding those of chess by several orders of magnitude. After Deep Blue defeated the world chess champion, Go became an open challenge in the field of AI. Thore Graepel explains: “It is so elegant and simple that it became the best game to tackle at the time.” Pushmeet Kohli adds that not only is the breadth of the search space enormous (with up to 200–300 possible moves per turn), but the depth of the game (far exceeding the 60–70 moves typical in chess) makes the problem extremely complex. ## The Birth of AlphaGo: From an Internship Project to the World Stage Thore Graepel recalls his first day at DeepMind: “David Silver asked me, ‘You’re a Go player, right? Can you help us test a baby version of something?’ It wasn’t even called AlphaGo then—it was an internship project, training the system on hundreds of thousands of games collected from the internet.” He became the “first person to officially lose to AlphaGo,” falling narrowly in front of an audience. This made him realize that this early version, trained on human professional game records, knew exactly how to counter traditional playing styles. ## Core Technology: Fast Thinking and Slow Thinking AlphaGo cleverly integrated two cognitive processes: - **Fast Thinking (Intuition)**: Implemented through deep learning networks. The value network evaluates the quality of a board position (for Black or White), while the policy network ranks all possible moves based on probabilities derived from professional players. Deep learning matured after 2012, providing the first tools to learn such approximate functions. - **Slow Thinking (Planning)**: Utilized traditional AI search methods (similar to Deep Blue) to perform explicit reasoning over the game tree, analyzing subsequent variations of various possibilities. Graepel points out: “This aligns closely with how humans play chess—we are intuitively drawn to certain moves and then verify them through planning.” Intuition and computation work together to address the vast search space. ## Early Validation: The Match Against European Champion Fan Hui The team invited Fan Hui, a European Go champion and professional player, to the office for testing. Thore Graepel and David Silver made a bet: Graepel believed AlphaGo would lose at least one game, while Silver bet on a 10-0 victory. The result was a 10-0 win for AlphaGo, forcing Graepel to dress up as an ancient Japanese Go master for work that day. This victory gave the team confidence to prepare for challenging higher-level opponents. ## The Showdown in Seoul: The Match of the Century Against Lee Sedol Lee Sedol was one of the best Go players in the world at the time, a multiple-time world champion often compared to Roger Federer. He was confident he could beat AlphaGo, based on previous match records against Fan Hui, but he failed to realize that AlphaGo was constantly improving through training and algorithmic enhancements. The match took place on a hotel floor in Seoul, South Korea, with an atmosphere akin to a celebrity event. In Korea, top Go players are celebrities, surrounded by large numbers of photographers and documentary crews. Graepel admitted to feeling nervous: “We calculated Elo ratings through our evaluation system, but we weren’t sure of Lee Sedol’s exact position on that scale. This was on the world stage; losing would be a blow to our reputation.” The team worked until the last moment to ensure system stability. ### Game 1: An Unexpected Advantage Pushmeet Kohli watched from Seattle: “Midway through the first game, it was clear from the reactions of the media, commentators, and Lee Sedol himself that AlphaGo had reached that milestone.” Initially, many believed Lee Sedol would win, but as the game progressed, AlphaGo gradually established an advantage in territory calculation. One American professional player initially criticized AlphaGo for making a “stupid move,” but after the game remarked: “This was the most extraordinary thing I’ve experienced; witnessing a machine playing at this level, we can learn so much.” ### Game 2: Move 37, Which Shocked the World This move became a historical landmark. Professional commentators almost unanimously agreed that no human player would make this choice. AlphaGo’s own assessment estimated the probability of a human making this move at just one in ten thousand. Graepel recalls the scene: “In the international commentary room, American commentator Michael Redmond placed the stone for Move 37 on the demonstration board, then stepped back and said, ‘This must be wrong.’ He picked it up, looked at the screen to confirm, placed it back down, utterly confused—such a move is so counterintuitive for humans.” Ultimately, although Lee Sedol managed to win the fourth game, AlphaGo won the series 4-1. The moment Lee Sedol placed two stones on the board to resign, the world instantly changed. ## Impact and Outlook Ten years later, we have witnessed breakthroughs in large language models, AI agents, and protein folding. The modern AI revolution can be said to have started on that wooden Go board in Korea. AlphaGo not only conquered Go but also validated the powerful capability of combining reinforcement learning with deep learning—technologies now used to solve critical problems in science and the real world. --- Source: 10 years of AlphaGo: The turning point for AI | Thore Graepel & Pushmeet Kohli - Google DeepMind (https://www.youtube.com/watch?v=qoinGjj60Fo)