Saturday, March 12, 2016

The Goalpost Moveth

The news of AlphaGo challenging Lee Sedol in Go came as a surprise to me. While peripherally interested in AI, I had assumed that something like this was years if not decades away. I played a little bit of Go when I was a kid, but I didn't really have the patience for it, so all I really got was a sense of how difficult the game would be to program.

When AlphaGo took the first game, I had already seen what I wanted to see. AlphaGo doesn't need to actually defeat Lee, much less sweep the 5-game series, to be a success. By playing at his 9-dan level, AlphaGo has proven that it will handily defeat most human players, and we can expect the technology to be easily and widely accessible in just a few more years. The top players might still make a valiant last stand against the inevitable, but the rest of us are toast. I cheered technological progress.

When AlphaGo took game two, complete with a pivotal move that practically no human would ever make, I was at the same time astonished that we were still discovering things about a 2,500-year old game, but not surprised once AlphaGo's methods were explained. Because it was coldly precise, it would play for a small but sure win, rather than trying to maximize advantage. As a much more expert observer pointed out elsewhere, when AlphaGo plays what we see as a "slack move", it's pretty much decided that it will win. What it is unable to comprehend is that its human opponent doesn't even know that yet!

With ruthless efficiency, AlphaGo just took game three to win the series. Lee may no longer be the 18-time world champion that he was, but he is undoubtedly still one of the best players in the world, now facing the prospect of losing all five and not even really knowing the limits of his computer opponent.

The games were surprisingly emotional for me. As a computer person, I'm not particularly afraid of the robot future. I worry about the social impact to our inevitable automated future, but I do not fear self-driving cars and airplanes, nor even robot government administrators or politicians. I'm in awe that an accomplishment like this only took DeepMind about two years, against an absolute genius who had been playing for decades. As a human, though, a little part of me wishes to postpone this inevitability just a bit more, and Lee became something of a champion for humanity in my mind. He was brave to take the challenge. Despite the $1M prize money, he had much to lose and nothing to gain by beating a computer that we all just knew wasn't that good yet.

The history of AI has been one of moving goalposts, and after yesterday what we've proven is of course that playing Go isn't really AI, rather just another in a series of games that are hard for humans but easy for machines to solve. Yet the list of things that only we can do seems to only shrink, so I don't really care for this line of thinking. Deep Blue and Watson and now AlphaGo will of course seem hopelessly primitive to future historians, but I think we live in interesting times indeed.

I wonder if AlphaGo will splurge on a new graphics card with the prize money.