Many interpretations of AlphaGo’s memorable match against Lee highlight move 37 as a moment of pure machine intuition. However, this perspective misses the mark. Instead, it was AlphaGo's reasoning abilities that informed this innovative decision—capabilities that current AI systems still lack. To foster AI that delivers reliable outcomes and fresh perspectives, particularly in disciplines like science and medicine, we must enhance their reasoning skills.
AlphaGo operates through a combination of two distinct systems. The first is its policy network, which was trained to predict the moves a skilled human player might select. This “intuitive” component regarded move 37 as relatively unremarkable, estimating that a strong player would make it only about once in 10,000 games. The reason AlphaGo played this move stems from its search component, which looked beyond immediate options and considered the long-term implications of various moves. It crafted and navigated an extensive game tree with countless branches, each representing a potential future scenario.
A renowned concept in behavioral science, advocated by Daniel Kahneman, divides human thought into two systems: System 1, characterized as fast, instinctive, and effortless; and System 2, defined as slow, methodical, and reflective. AlphaGo serves as a fascinating machine counterpart to this model. Its networks provided the initial insights—identifying potential moves that appeared favorable—while its search mechanism introduced deliberation, rigorously assessing those insights against the upcoming moves and possible responses. Just as in human thinking, both systems need to work in tandem. Relying solely on intuition would not have led to move 37, nor would a purely exhaustive search have effectively filtered through the immense array of potential moves.
This operates quite differently from contemporary AI models. For instance, a large language model generates the next token repeatedly, functioning as a fast, associative process that reflects System 1 thinking—efficiently completing patterns across a vast range of topics people might discuss.
