AlphaGo 10周年:ゲームから生物学へ広がるDeepMindの軌跡 From games to biology and beyond: 10 years of AlphaGo’s impact

DeepMind has published a retrospective marking ten years since AlphaGo's historic victories, reflecting on how a Go-playing program reshaped the trajectory of modern AI research. The 2016 match against Lee Sedol, in which AlphaGo won 4-1, remains a landmark moment that demonstrated machine learning systems could master domains long thought to require uniquely human intuition.

At its core, AlphaGo combined deep neural networks with Monte Carlo Tree Search, trained first on human game records and then refined through self-play reinforcement learning. The now-famous Move 37 in game two against Lee Sedol — a play human professionals initially dismissed as a mistake — illustrated that the system was not merely imitating expert patterns but discovering genuinely novel strategies. That insight set the template for everything DeepMind built afterward.

The lineage is substantial. AlphaGo Zero removed the need for human game data, learning entirely through self-play. AlphaZero generalized the approach to chess and shogi, while MuZero went further by mastering games without being told the rules, learning a model of the environment instead. Each step pushed the underlying ideas closer to general-purpose decision making.

Perhaps more consequential than the games themselves is how the search-and-self-improvement paradigm migrated into science. AlphaFold, which predicts protein structures with near-experimental accuracy, has become a foundational tool in structural biology and contributed to a 2024 Nobel Prize in Chemistry for its creators. AlphaTensor discovered faster matrix multiplication algorithms, while AlphaGeometry and AlphaProof have tackled olympiad-level mathematics. The common thread is treating scientific discovery as a search problem in which a learned model guides exploration of an enormous combinatorial space.

It is worth recalling the context before AlphaGo. Go had stood as a grand challenge for AI for decades, with many researchers expecting top-human play to remain out of reach for another ten years or more. The branching factor of the game and the difficulty of crafting a reliable evaluation function had stymied classical approaches. The fusion of deep learning with tree search broke that ceiling and arguably accelerated investment across the field, influencing later work such as OpenAI Five in Dota 2 and Meta's Cicero in Diplomacy, even if those projects took different technical routes.

The broader ecosystem has also been shaped by the open release of trained AlphaFold structures via the EMBL-EBI database, which now covers hundreds of millions of proteins and is used routinely in drug discovery and biology labs worldwide. In that sense AlphaGo's true legacy may be less about a single match and more about establishing self-play and learned search as durable building blocks for scientific AI.

Challenges remain. Self-play works cleanly in closed, well-specified games, but real-world domains rarely offer such crisp reward signals or perfect simulators. Compute costs are non-trivial, and safety considerations grow as systems become more autonomous. It seems likely that the next decade will see hybrids between large language models and AlphaGo-style search, blending broad world knowledge with deliberate planning. Early systems combining LLMs with verifier-guided reasoning already hint in that direction, though how far the paradigm scales is still an open question.

Ten years on, AlphaGo reads less like a finished chapter than the opening move of a much longer game — one in which AI is being asked not just to play, but to discover.