Patch Notes #075 — The Machine That Plays Go

Buried in a Nature paper this week: Google DeepMind’s “AlphaGo” beat the European Go champion 5-0. Months ago, in secret. Go was supposed to be the safe game: more board positions than atoms in the universe, no brute-force path, the game where human intuition would hold out for another decade. It did not hold out. The system learned from human games, then improved by playing itself millions of times. In March it plays Lee Sedol, one of the greatest living players, and the Go world is confidently predicting a human win. The chess world was confident once too.

I’ve watched machine learning go from “organizes my photos” to “beats champions at intuition” in eight months of writing these. Whatever line I thought existed between calculation and judgment keeps moving, and it only moves one direction.

Work, senior-flavored: ran my first architecture review as the reviewer of record. The hard part wasn’t the design; it was saying “no” to a smart person’s clever idea kindly enough that they’d bring me the next one. Cleverness is cheap; the willingness to keep proposing is precious. Guard it.

TIL: Monte Carlo tree search, or at least the Wikipedia-depth version. Sampling your way through an impossible search space: works for Go, works for capacity planning, works for deciding where to eat lunch with engineers. Nothing converges, everything samples.