So the story everyone's chewing on this week is AlphaGo. Google's DeepMind team announced Wednesday that their program beat Fan Hui, the reigning European Go champion, 5 games to 0, back in October, and they'd been sitting on the result until the paper cleared review in Nature. That's the part most of the coverage is running with. The part I actually care about is buried a few paragraphs down in most of the writeups: AlphaGo is scheduled to play Lee Sedol in March. Lee Sedol isn't "a strong European player." He's one of the best players alive, full stop, an 18-time world champion. Fan Hui is good. Lee Sedol is a different category of good, and the gap between beating one and beating the other is the whole story here, not the footnote.
I want to be careful, because I've watched enough "computer beats human at X" stories collapse under scrutiny once people look closer. Deep Blue beating Kasparov in '97 gets retold now as some clean, inevitable moment, and it wasn't. There was a bug in Deep Blue's evaluation function in game one that made a move look more mysterious and calculated to Kasparov than it actually was, and that rattled him for the rest of the match. I don't know yet whether AlphaGo's win over Fan Hui is the Go equivalent of that, a real demonstration of a genuinely stronger machine, or something that reads as more impressive with distance than it actually was up close. Nobody will know for sure until March.
Here's my personal stake in this, and it's a small, kind of dumb one. I've been trying, on and off, for about four years, to get halfway decent at Go. I own two boards. I have a well-worn copy of "Lessons in the Fundamentals of Go" sitting on the shelf next to my desk that I've started and abandoned about six separate times. And I still lose, regularly, to the free app on my phone set to its lowest difficulty. Go is one of those games where the rules take you ten minutes to learn and the actual playing takes a lifetime, and every time I sit down with it I get reminded how much room there is between "knows the rules" and "understands the game." So when I read that a program can now beat a professional, my first reaction isn't skepticism about AI. It's jealousy. Whatever intuition Fan Hui built up over decades of study, something spat out by a neural net trained mostly on self-play games got there too, or at least got somewhere that looked the same from the outside.
The technical approach deserves a sentence, because it isn't brute force the way chess engines are. Go's search space is too big for that, something like 10^170 possible board positions, more than atoms in the observable universe, which is the stat every single writeup this week is contractually obligated to include, so here it is again. AlphaGo pairs a policy network that narrows down which moves are worth considering with a value network that estimates who's ahead from a given position, then runs a Monte Carlo tree search over that smaller set. It's less "calculate everything" and more "develop something that resembles a hunch about which branches are worth exploring," which is uncomfortably close to how a strong human player describes their own process when you actually ask them.
Part of why I'm not writing the standard "here's what this means for the future of AI" post is that I think that post is premature by about six weeks. Beating Fan Hui is real and it's not nothing. Google wouldn't have put out a Nature paper over a fluke. But the actual test is Lee Sedol in Seoul this March, best of five, and until that happens the honest answer to "can a machine play Go at the top human level" is "we don't know yet, ask again in March." I've already got it circled on my calendar, which is a strange thing to admit about a board game match. If you'd told me back in November 2011 when I started this blog that I'd eventually be marking a Go tournament the way some people mark a Super Bowl, I don't think I'd have believed you. But here we are, and I'm going to be insufferable about it for the next several weeks whether anyone reading this cares or not.