Go, AlphaGo, and the Game of Games

Big news in artificial intelligence this week. Google’s AlphaGo algorithm beat the world’s best player in Go. Twice.

Go is a complex strategy game popular in Asia, which is much more complex than chess, with many more variations. For that reason, plenty of smart people thought it would take a while before the best computers beat the best humans.

Not that long, apparently.

My older brother is a very good Go player. He’s in the upper echelon of non-Asian amateurs. Last summer, we had a chat about how good the best Go computers were compared to the best humans. He speculated that the best computers at that point were about as good as him, but that they weren’t available publicly and so he couldn’t find a decent game against a non-human opponent. Neither my brother nor I are AI skeptics. But I don’t think either of us would have predicted 18 months later we would see this.

Go is famous because of its complexity, and so served as a landmark as a human vs. AI measuring stick. That stick has been measure. So, what’s next?

Some say Starcraft, a popular, ever-evolving computer strategy game.

But I think there’s a better, harder test for computers. I’ll call it the new game test, or the McCarthy test (named after AI pioneer John McCarthy, of course).

I think we know at this stage, that given the time, programmers can create an algorithm that can beat the best humans. But what if programmers don’t have time to prepare a new algorithm? Can programmers design a program that is simply good at games?

The true test of AI is to be able to learn and beat the best humans not only at games that a programmer has time to prepare for, but to be able to learn new games as fast as humans and beat them without time to prepare.

As I’ve said before, we live in the golden age of gaming. Strategy games aren’t about just Go or chess any longer. Brilliant designers are creating new complex strategy games every bit as elaborate and intricate as the ones that have lasted hundreds of years.

Developing strategies to a game you’ve played many times is one thing. We now know computers do that better than people can. Developing a strategy for a game you’ve never played is another. That is the test of life. That is the test of the game of games. That is the McCarthy test.

If a computer can beat the best humans not only at games when programmers have had time to prepare an algorithm unique to that game, but when an algorithm is powerful enough to take any new set of rules, master the rules, and then beat the best humans at that game in the first few iterations of play, well, then that’s when we’ll truly have met our robot overlords.

I am not an AI skeptic. I believe it can be done.

But that is what I see as the next great challenge for AI gaming.