On March 10th of 2016 at the Four Seasons hotel in Seoul, South Korea, man and machine faced off in the second of five games. The worlds of AI and Go would never be the same.
A machine so advanced its own creators could not predict its behavior, AlphaGo was initially trained on 160,000 recorded human games and then set against alternative instances of itself to further improve its skills. Unlike its boxy chessmaster predecessor Deep Blue, it had no physical presence anywhere in the hotel. In a reversal from the human-driven automaton of the Mechanical Turk, now a stoic-faced human in a clean black business suit provided the physical manipulation for the machine Go player hidden from view. His contribution to the match was soft, dextrous fingers good for lifting small go stones from their bowls and placing them on a board as directed by AlphaGo in its home inside a server in the United States.
Promoted to professional dan rank at twelve, Lee Sedol had boasted to his millions of fans in South Korea that he would defeat AlphaGo in a landslide. After his defeat the day before, he carried himself with a new humility, but one game does not a match make. Whether or not he won, Sedol knew that in this match he would make history.
This is not a Go blog, and I will not attempt to explain AlphaGo’s “Move 37.” Suffice to say that it was the 37th move played in the match, and fittingly for a human-conquering machine, it is marked not by a poetic name but by a number. In its internal models AlphaGo estimated the move’s likelihood of being played by a human at 0.0001, or one in ten thousand. This number is likely a hardcoded minimum estimate, as very small values can get lost in computer arithmetic and become zero, which can cause arithmetic errors in deep learning models. Nevertheless, the move flummoxed the commentators watching, and Fan Gui, a European Go champion who had lost to AlphaGo five times out of five in a previous closed match, was dumbfounded. Only later did he see the beauty in the move, and he couldn’t stop talking about it. “So beautiful” he said again and again, as AlphaGo drove Sedol into another defeat.
Sedol lost another game after that, along with his hopes of winning the match. Starting on March 13th he just wanted to get one victory. To do so, he changed his strategy to counter AlphaGo’s preference for methodical accumulation of small gains, known as the souba or ‘market price’ strategy. The amashi strategy turned AlphaGo’s calculated trades on their head by ignoring them entirely and instead attacking it wherever it was weak. This strategy led to Sedol’s own history-making move, appropriately given a poetic name, “God’s Touch.” AlphaGo assigned this move the same one in ten thousand likelihood as its own move 37, but it wasn’t until several moves later that the machine’s internal estimated probability of victory plummeted. In just over a hundred more moves, AlphaGo conceded defeat.
Sedol didn’t bring home another win in game five, but, he said, “Winning this one time, it felt like it was enough.”
AlphaGo said nothing, for it was not designed to commentate, only to play Go. But its game speaks for itself. By playing a brilliant move that no one had ever seen before, it had demonstrated itself capable not only of responding to the novel, but creating it. Close to two hundred years after the mechanical turk and Edgar Allan Poe’s goalposts for a mind, AlphaGo redefined just what a “pure machine, unconnected with human agency” can do.
I take requests. If you have a fictional AI and wonder how it could work, or any other topic you’d like to see me cover, mention it in the comments or on my Facebook page.
Samuel Munk 