Lee Sedol said he was “in shock.” That was the understated verdict from the world’s greatest Go player after Google DeepMind‘s AlphaGo system beat him 4-1 in downtown Seoul in 2016.
The match had turned in the third game. By then, tens of millions watching at home already sensed what the hushed crowd in the arena could barely process: this system was not behaving like prior machines. It was winning with something that looked, uncomfortably, like intuition. Google co-founder Sergey Brin described it to a publication shortly after AlphaGo went 3-0 up. “AlphaGo actually does have an intuition,” he said. “It makes beautiful moves. It even creates more beautiful moves than most of us could think of.”
Go is not a simple problem.
The ancient game, played on a 19-by-19 board with black and white counters, allows for 10171 possible positions — a number that dwarfs the estimated 1080 atoms in the observable universe. Prior approaches to building game-playing machines required engineers to manually encode rules for each situation. AlphaGo operated differently. Its neural networks, mathematical structures modeled loosely on the brain, learned by processing millions of real game records, then refined further by playing millions of games against itself. No human player could accumulate experience at that scale — and that asymmetry was precisely the point.
Chris Maddison, now at the University of Toronto and a member of the original AlphaGo team, says the core technology has not fundamentally changed. “Large language models are now quite different in some ways from AlphaGo, but there’s actually an underlying technological thread that really hasn’t changed,” he says. That thread is neural networks.
The team’s early approach centered on training a network to predict the next strongest move from existing human games — an attempt to replicate player intuition computationally. According to the announcement, the final system that defeated Lee Sedol was more complex than those early models, but the headline finding was blunt. “AlphaGo definitively showed that neural nets can do pattern recognition better than humans. They can essentially have intuition that surpasses humans,” says Noam Brown at OpenAI.
From the board to the lab
What followed was a decade of applying that lesson outward. AlphaFold, another Google DeepMind system, predicted protein structures from chemical compositions with accuracy no human-designed program had achieved — work that earned its team the Nobel Prize in Chemistry. More recently, AlphaProof performed at gold-medal level in the International Mathematical Olympiad, a result that stunned mathematicians. “Not only can you get this beyond-human-level intelligence in a game, but you can get that experience in important scientific applications,” says Pushmeet Kohli at Google DeepMind.
The shared architecture
The structure underlying both AlphaGo-style systems and large language models like ChatGPT follows the same two-step logic. First, a neural network trains on a large body of human-generated data — complete Go games in one case, vast portions of the internet in another. Then a second phase, post-training, refines the network’s behavior through reinforcement learning. The board game and the chatbot are separated by nine years and enormous complexity, but the engine driving both traces back to a single afternoon in Seoul when Lee Sedol sat across from a machine and realized, too late, that the game had already changed.
Photo by Pixabay
This article is a curated summary based on third-party sources. Source: Read the original article
