What AlphaGo’s win could mean
A Path to Progress
Artificial intelligence beat a human in the ancient game of go – and that’s OK.
It’s easy to make too much or too little of an event that took place earlier this month. A computer program called AlphaGo played a five-game match of the Japanese board game of go against South Korean grandmaster Lee Sedol. AlphaGo won easily, 4 games to 1.
It’s been nearly two decades since the Deep Blue computer program beat Russian chess grandmaster Garry Kasparov in 1997 to claim superiority in that game. Go had been considered much more difficult for artificial intelligence (AI) to master (for one thing, chess is played on a board with an 8-by-8 grid producing 64 squares; the go grid is 19-by-19).
AlphaGo succeeded by combining two powerful computational approaches. It was fed information on numerous go games that had been played by the best human players. And it also played “against itself” millions of times, in a sense teaching itself which moves and strategies worked and which didn’t.
AlphaGo’s dominant victory surprised experts. Most had thought that the ability of AI to defeat the best human players was a decade or more away.
The win signaled that AI systems, already pervasive if often unnoticed (when was the last time you immediately talked with a human and not a speech recognition program when you called a large business?), will quickly become more powerful and prevalent.
Even “skilled” jobs look to be in jeopardy, especially if they involve tasks that are routine, repetitive, and predictable. Robots have been on assembly lines for decades, but white-collar desk-bound workers won’t be exempt. Already, concerns are being voiced over computerized stock-trading programs that are developing their own trading techniques that humans not only won’t be able to understand but won’t be sure are operating ethically.
Interestingly, skilled blue-collar workers such as plumbers and electricians may be safe for a long time to come. These jobs involve a mix of many different abilities: seeing, and understanding what is being seen in its context; subtle manual dexterity to manipulate objects in a variety of settings; and problem-solving skills that master new or unexpected situations. Each poses problems for AI.
Even a sophisticated game like go represents a tiny, limited part of the human experience. AI that masters the game doesn’t know about any of the other innumerable things humans think about or do during their day.
That’s why so-called general AI still seems a long way off. AlphaGo was designed to do one thing – win a game with a specific set of rules. Human life can be a lot more unpredictable.
Perhaps the real honors should go to Mr. Lee for winning even one game. Despite his experience, he was a relative newcomer to the game, since AlphaGo had played literally millions of previous matches.
“It’s impressive that a human can use a much smaller quantity of data to pick up a pattern,” says Gary Marcus, a neuroscientist at New York University. “Probably, humans are much faster learners than computers.”
In winning his game, Lee finally came up with a move so unusual that it confounded even the top go players looking on. It apparently was something even AlphaGo hadn’t seen before, and it couldn’t find an effective response. Observers were so impressed they reverently called Lee’s inspired move “God’s touch.”
So maybe a beneficial human-AI collaboration is ahead. After all, if AlphaGo hadn’t provided Lee with such a difficult challenge, would he have ever experienced “God’s touch”?