Three days from now, a Google-owned lab is going to sit a computer program down across a board from one of the best Go players alive and we're all supposed to just go about our week like that's normal.
I mentioned AlphaGo to two people at work today. One asked if that's the thing that beat Kasparov at chess. It is not. The other nodded politely and went back to her monitor, which, fair, I probably would have done the same three years ago before I started poking at Go rules on a rainy Sunday and getting embarrassed by a phone app set to easy mode.
Here's the actual setup, because I think people are sleeping on how big this is. Lee Sedol, one of the strongest Go players of the last twenty years, plays a five game match against DeepMind's AlphaGo starting this Wednesday in Seoul. Google put a million dollars on the table for it. DeepMind is the same outfit Google bought in 2014, and back in October their program quietly beat the European champion Fan Hui 5 games to 0, which barely made a ripple outside people who already cared about this stuff. Lee Sedol is a different tier of opponent entirely, and he's been telling reporters he expects to win comfortably, maybe drop one game at most. I don't think he's wrong to be confident. I also don't think he should be as confident as he sounds.
The reason this isn't just "chess but slower" is the board itself. Chess has roughly 35 legal moves at any given point and a computer can brute-force its way pretty far into the tree. Go is played on a 19x19 grid and the number of possible positions is bigger than the number of atoms in the observable universe, or close enough that the comparison stopped being an exaggeration a while ago. You can't brute force that. Good Go players talk about intuition and shape and moves that just feel right in a way that's genuinely hard to explain to someone who hasn't played, which is most of the people I bring this up with, evidently.
What actually got me paying attention wasn't the match announcement in January, it was the Nature paper that came with it. DeepMind trained AlphaGo using a mix of supervised learning off a huge pile of human games and then reinforcement learning where it played against itself, over and over, refining its own instincts instead of just memorizing opening theory the way a lot of people assumed a Go program would have to. That self-play part is the bit I keep thinking about at random points during the day, like while I'm making coffee. It's not just pattern matching against a database of dan-level games from the last decade. It got better by playing itself and figuring out what worked.
I tried to actually learn Go over the holidays, for what it's worth, using one of those 9x9 starter boards, and I lost interest around the point where I realized how much of it is about reading several moves ahead in your head instead of just reacting to what's in front of you. My little sister picked it up faster than I did and now beats me roughly every time, which I've decided not to be bitter about in writing.
Anyway. Wednesday. I'll probably end up watching whatever stream DeepMind puts up during my lunch break like it's a sporting event, because it kind of is one, and I already know most of my coworkers are going to ask me on Thursday who won a chess game.
One smaller thing while I'm here: the Galaxy S7 preorders have been sitting in my cart for about two weeks now and I still haven't pulled the trigger before the March 11th release, mostly because my S6 still works fine and I have a hard time paying full price for a phone that's an incremental spec bump over one I already own. That's a whole separate complaint for another post though.