I was not planning on writing about the Go match again this month. Everybody and their cousin has already written the "AI beats humanity" post, and I figured by the time I got around to it, it'd just be noise on top of noise. But then Lee Sedol actually won a game, and that changes things enough that I stayed up for it, so here we are.
Quick recap for anyone who's been ignoring this: DeepMind's AlphaGo has been playing a five-game match against Lee Sedol, one of the best Go players alive, in Seoul this week. It won the first three games. Cleanly. The kind of cleanly that made a lot of very serious Go commentators go quiet mid-sentence. By game three the match was already decided — AlphaGo had clinched it 3-0, and game four on Sunday was basically just Lee Sedol playing for pride.
And he won. He actually won.
I watched most of it on the DeepMind livestream, which meant staying up stupidly late (it was well past 1am my time by the point Aja Huang, the guy physically placing AlphaGo's stones on the board, finally set down the resignation). I am not a strong Go player (I can barely hold my own against the medium bot on KGS, let alone anyone who's actually studied the game), but even I could tell something had shifted around move 78. Lee Sedol played this wedge move into white's position that the commentators didn't love at first, and then AlphaGo just started making moves that didn't make sense. Not bad exactly, just off, the kind of moves that suggested the model had wandered somewhere in its evaluation that it couldn't get back out of. Michael Redmond, who was doing commentary, actually paused and said something like "I don't understand this move" more than once, which from a 9-dan professional is not nothing.
By move 180-something it was over. AlphaGo resigned. The room in Seoul actually applauded, which you don't usually get from a room full of Go professionals watching a match, they're a pretty reserved bunch as a rule.
Here's my actual opinion on this, since I know half the tech blogs this week are just going to do the neutral "wow, fascinating, what does this mean for the future of AI" thing and call it a day: I think game four matters more than games one through three, and I think most coverage is going to undersell it. Everyone was ready for the story to be "computer wins, humans obsolete, insert Terminator joke." That's an easy story. The harder, more interesting one is that a system trained primarily through self-play and reinforcement learning still has something like a blind spot, a place where its evaluation function just breaks down under the right kind of pressure, and a human being found it under match conditions with the entire world watching. That's not nothing. Demis Hassabis said on Twitter that AlphaGo "lost" and that DeepMind would need time to figure out where its evaluation went wrong, which is a pretty candid thing for a company to say about its own headline product mid-tournament, and I respect that they didn't try to spin it.
I also just liked watching Lee Sedol be visibly happy for once this week. Games one through three, he looked rattled in the press conferences, like a guy trying to work out a problem that kept changing shape on him. After game four he was smiling, said something about how he'd never been congratulated so much for a single win in his career, and I don't know, that stuck with me more than any of the "historic moment for artificial intelligence" headlines did.
Game five is Tuesday. AlphaGo's already won the match, so there's nothing on the line except whether Lee Sedol can make it 2-3 instead of 1-4. I'll probably stay up for that one too, which my body is going to be very unhappy about come Wednesday morning. Worth it though. I don't get to watch something like this very often.