I was up way too late this week watching a board game on a livestream, and I regret nothing.
If you've somehow missed it: DeepMind's AlphaGo has been playing a five game match against Lee Sedol, one of the best Go players of the last decade, over in Seoul. The match wrapped up today (game five happened while most of us on this side of the world were asleep or at work), and the final score is 4-1 for the machine. Everybody and their cousin has already written the "AI beats human, here's what it means for civilization" post this week, so I'm not going to write that one. What I actually want to talk about is game four, because game four was the good one.
Games one, two and three all went to AlphaGo, and by Saturday it felt like the match was basically over as a contest, more like a formality getting played out for the cameras. Then in game four Lee Sedol made a move on turn 78 that nobody, including apparently AlphaGo, saw coming. Commentators started calling it "the wedge" almost immediately, and somebody later dubbed it the "God's Touch" move, which is a lot, but I kind of get the impulse watching the replay. It's a move that jams a stone into a spot that looks structurally pointless right up until about fifteen moves later, when it very much isn't.
What happened after is the part I keep thinking about. AlphaGo, which had been playing with this eerie, unbothered confidence for three straight games, started making moves that the human commentators (including a 9-dan pro doing color commentary on the feed) flagged in real time as bad. Not subtly bad, either. Visibly bad, the kind of moves that make the chat room go quiet for a second. DeepMind's own people said afterward that the system's internal win estimate cratered right after move 78 and it basically floundered trying to recover from there. AlphaGo resigned a bit after move 180. Lee Sedol just sat there and smiled, and if you find the clip, it's honestly a great little moment of a guy getting to enjoy something after three rough days.
I don't play Go, for the record. I know the absolute basics (surround territory, capture stones, it's way older than chess and the board is bigger) and that's about it, so I can't tell you with any real authority why move 78 was brilliant beyond repeating what the actual pros said about it. What I can tell you is that watching a piece of software steamroll a world class player for three straight days and then suddenly get visibly rattled was one of the more human moments I've seen out of a machine this year, and I mean that as a compliment to the match, not the software. It made the whole thing feel less like a foregone conclusion and more like an actual contest between two things that could both get caught off guard.
One detail that stuck with me and got buried under all the "humanity loses to the robots" headlines: Lee Sedol's appearance fee was reportedly around $150,000 regardless of how the match went, and DeepMind had already said the $1 million prize that came with the win would go to charity, UNICEF and STEM education groups among them, rather than into anyone's pocket. That's a nicer footnote than most of what got written this week, and I only found it because I went and read the actual DeepMind blog post instead of just skimming the hot takes people were posting.
Small tangent that has nothing to do with any of this: I tried explaining the whole match to my brother-in-law on the phone last night and lost him completely somewhere around the words "reinforcement learning." Gave up, told him "the computer got really good at a board game," and he seemed perfectly satisfied with that. Which, honestly, might be the correct level of explanation for about ninety percent of people reading headlines about this right now.
Final score's 4-1, AlphaGo took the match. But I don't think that's actually the number worth remembering from this week. I think it's 78.