Ten Years after AlphaGo: From Go to Scientific Discovery

A decade after the 2016 AlphaGo match in Seoul, Hannah Fry talks with DeepMind researchers Pushmeet Kohli and Thore Graepel about the game, its technical ideas and their later use in scientific research.

AlphaGo defeated Lee Sedol four games to one. The [match record](https://deepmind.google/research/alphago/) includes both AlphaGo’s surprising move 37 in game two and Lee’s celebrated move 78 in game four.

Graepel, a Go player, describes the attraction of simple rules producing a huge space of strategic possibilities. The challenge was to choose promising moves without exhaustively searching every possible continuation.

On joining DeepMind, Graepel played an early version of AlphaGo. He recalls deliberately avoiding unusual moves because it had learned from professional games—and losing, an experience he describes as humbling.

Kohli explains why the search space grows rapidly with each turn. AlphaGo combined learned assessments of positions and moves with a search procedure that looked ahead at possible continuations.

The guests compare these components to intuition and deliberation: one narrows attention toward promising moves, while the other explores consequences. This is an analogy for the system’s functions, not a claim that its internal experience is human.

Before the Seoul match, AlphaGo played European champion Fan Hui. Graepel recalls doubting that the system would win decisively, while a colleague predicted a clean sweep; the internal assessment helped calibrate expectations.

Lee Sedol was one of the era’s leading professional players, and the match attracted intense attention. The team could test against available players, but it could not know the final outcome against Lee in advance.

Graepel recalls the tension before the first game and the challenge of deciding when to stop changing the system. More tuning might improve performance but could also introduce instability immediately before a public match.

During the first game, observers did not immediately agree about who was ahead. As territory became clearer later in the game, AlphaGo’s advantage became more apparent.

Graepel remembers a professional initially criticizing a move and later reconsidering when its consequences became clear. For him, that was an early sign that the system could suggest ideas outside familiar patterns.

In game two, AlphaGo’s move 37 surprised commentators. Graepel explains it as an unusual balance between immediate territory and influence toward the center, rather than a familiar local exchange along the edge.

The move’s value became clearer only as the game continued. The guests use it to illustrate how a system can propose a useful move that experts initially find counterintuitive, while people still analyze and debate why it worked.

After AlphaGo won the first three games, Lee played move 78 in game four. The unusual wedge exposed a weakness in the system’s response and helped him win the game.

The researchers express admiration for Lee finding that move under immense pressure. His victory also showed that an impressive system could have vulnerabilities even after winning the match overall.

The guests recall mixed reactions in the Go community: excitement about new ideas as well as concern about what a machine’s strength meant for the game. They say players subsequently used AI tools to study positions and review games.

The conversation moves to AlphaZero and its self-play approach. Unlike the original AlphaGo’s initial learning from human games, AlphaZero learned its board-game strategies through self-play using the rules of the game.

Initially, play is effectively uninformed; through repeated games and feedback, the system learns which choices and positions tend to lead to success. Search and learning reinforce one another within the defined game environment.

The guests describe the system rediscovering familiar opening patterns and later favoring alternatives. Their point is that a system’s objective can lead it beyond conventional human play, although its results depend on rules, training and compute.

They revisit the team’s hope that techniques developed for Go could eventually help with broader problems. Kohli recalls being drawn to AI systems that could assist with difficult scientific and engineering questions.

Kohli points to search as a common ingredient in many problems. A scientific or engineering task needs a suitable representation, objective and way to check proposed solutions before game-like methods can be useful.

AlphaTensor recasts matrix-multiplication algorithm discovery as a search problem. Its agent proposes decompositions and is rewarded for finding correct algorithms using fewer operations or better performance under a chosen criterion.

The [AlphaTensor paper](https://www.nature.com/articles/s41586-022-05172-4) reports a specific improvement for 4-by-4 matrices over a finite field relative to a two-level Strassen algorithm. It does not claim that all matrix multiplication went unimproved for fifty years.

The speakers discuss applying related search ideas to algorithms and logistics. Practical gains must be measured for the relevant problem and hardware; a mathematically shorter algorithm is not automatically faster in every environment.

Graepel notes that a system optimized to maximize its chance of winning may choose differently from a human seeking a large margin. Apparent oddities can result from differences in the objective, not necessarily from an error.

Fry asks how to distinguish an original insight from a hallucination. Kohli emphasizes a generator-and-verifier pattern: suggest candidates, then test them using a reliable rule or evaluation function where one is available.

Code can often be checked against tests, and formal mathematical proofs can be checked mechanically. A test suite may be incomplete, however, and many open scientific questions lack a simple verifier.

The discussion separates correctness from explanation. A verified result can be valid while remaining difficult for humans to understand; science also benefits from explanations that connect a result to existing concepts.

The guests discuss AlphaProof and formal verification. DeepMind’s [2024 report](https://deepmind.google/blog/ai-solves-imo-problems-at-silver-medal-level/) describes competition-math results and expert assessment; it should not be generalized into a claim that arbitrary open problems have been solved.

Kohli argues that mathematicians and scientists remain essential in choosing worthwhile questions, interpreting outputs and placing discoveries in context. Automated proof checking answers a different question from why a result matters.

The guests return to the distinction between learning from human examples and improving through interaction with an environment. They see reinforcement learning and verification as useful complements to models trained on existing data.

They describe AlphaGo as a turning point for systems exceeding human performance in a defined domain. Fry closes by asking how future systems might generate new insights while separating genuine discoveries from errors.