Note ·

Shannon's wedge

GamesIntelligenceHistory of technologyGame AIResearch labs

Claude Shannon’s 1950 paper on programming a computer to play chess opens by conceding that the problem has no practical importance, and then explains why to do it anyway: a game is a problem with complete rules and a clear result, hard enough to be interesting, and the methods that crack it might serve as a wedge into problems of the same shape that matter more. Mallaby tells the story through David Levy’s handbook, which is how the sentence reached a twelve-year-old Hassabis in 1988, and it is the sentence I kept returning to while building my own three engines.

The wedge did what he said. DeepMind’s method was games from the first day: Atari, then Go, then chess, shogi and Go with one algorithm, and then the lab turned its learning methods on protein folding, which is not a game and was the grand challenge of a different field. The games were never the point, and they were also the only place the method could have been proven, because nowhere else gives you a scoreboard. That is the whole argument for why an amateur with a laptop should spend a month on board games in 2026, and Shannon made it before the computer he was programming existed.

References

    Open the full search →