Improspectives

I don’t get angry at online writers, or headline writers, that often, but a tweet from Wired regarding the go match between AlphaGo, the computing engine built by the DeepMind section of the company formerly known as Google, and world champion Lee Sedol, pushed my buttons.

Lee lost the first three games of the match, but all five games were to be played regardless of the outcome. The Wired tweet that ticked me off referred to Lee’s win in Game 4 against AlphaGo as a “consolation win”. Cade Metz, the author of the piece referenced by the tweet, said that Lee “clawed back a degree of pride for himself and the millions of people who watched the match online.”

No, he didn’t. Not because Lee couldn’t regain his pride after having no hope of winning the five-game series, but because he never lost it. Lee admitted to playing a loose opening in Game 1, but based on the AlphaGo games he’d seen from previous matches, he didn’t think the program was strong enough to take advantage of the situation. It was. At no time, the world champion said, did he think he was ahead. In Games 2 and 3 he played better moves, but AlphaGo still forced resignation. Part of the problem was that AlphaGo didn’t use as much of its allotted two hours for early moves as Lee did, so the computer was way ahead on the clock for most of the game. Early moves create the framework for the rest of the game, so players must weigh them carefully.

Lee was clearly frustrated by his inability to win any games in the first part of the match, but he came into Game 4 ready for the struggle and played a surprising, powerful move in the middle of the board after not getting much out of the opening. Expert commentator Michael Redmond, a 9-dan professional player (the highest rank awarded), said he didn’t see Lee’s wedge move coming, but as the game progressed he realized its power. Despite running very low on time, Lee was able to maintain his momentum and take advantage of aimless play by AlphaGo to secure the win.

The Wired story should have centered on the theme of a human player beating a go engine for what might be the last time. The best computer chess programs are favored to beat even world champion Magnus Carlsen in 99.9% of their games. AlphaGo’s improvement over the past five months, when it played well enough to win 5-0 against a professional rated in the top 650 players in the world but made clear errors, is astonishing. AlphaGo trains its neural nets by playing against itself at high speed, earning decades of play experience in months. I don’t doubt it will be unbeatable by humans in a very short time.

Lee stepped up under extremely difficult and very public circumstances to secure a brilliant win. The advances in machine learning behind AlphaGo’s abilities in a game thought to be too complex for computers to manage are notable, but Lee Sedol’s play and fighting spirit are the real story.