DeepMind published AlphaGo Zero this week, and the result reorganized my priors more than the original did. The first AlphaGo learned from 30 million human expert moves, then surpassed us. Zero learned from nothing: the rules, a board, and self-play. Three days: it beat the version that beat Lee Sedol, 100 games to 0. Forty days: it beat every AlphaGo ever built. Along the way it rediscovered centuries of human joseki (opening theory), used them for a while, and then discarded some in favor of moves we’d never found in three thousand years. Human knowledge, it turns out, wasn’t the ladder, it was scaffolding, and load-bearing bias. The bitter arithmetic: our accumulated expertise was worth about three days of self-play. I’ve retold Move 37 as “the machine surprised us”; Zero’s lesson is stranger, that we were the constraint being optimized away. Centaurs-over-engines needs an asterisk I don’t have words for yet. Ask me again in ten years, again.
Also this fortnight: KRACK broke WPA2, the WiFi encryption everyone uses, vulnerable to key-reinstallation attacks; patch everything with an antenna (the eternal sermon gains a wireless verse). And the F1 title fight is heading to Mexico next weekend: Hamilton vs Vettel, two juggernauts, and Hamilton can seal his fourth within the fortnight.
TIL: self-play as curriculum. The system generates opponents exactly at its own level, forever, the perfect training partner no human league can provide. Synthetic data eating the world starts here. Bookmark this one hard.