In a spectacular showcase for IBM, supercomputer Watson defeated two quiz show heavyweights on the American show Jeopardy! over the first half of last week. This was the first time a computer had ever been entered against a human opponent on the show.
Watson was trained for years before finally facing off against Ken Jennings and Brad Rutter, the respective record holders for biggest winner and longest champion streak. What is particularly interesting are the parallels to IBM’s previous success… the famous chess victory of Deep Blue back in May 1997.
No ad to show here.
The chess-playing computer beat world champion Garry Kasparov in a tough six game chess match. It was at the same time a grand and slightly depressing moment for humanity. Many said it would be impossible, that a computer would never be able to beat a grandmaster in chess. Now it seems like a given. With the vast computing power at common disposal, even home computer chess-software is capable of beating grandmasters. But chess is an arena where computers have an advantage.
The same fate met Checkers years before, when an unbeatable strategy was discovered… rendering the game a bit pointless (like noughts and crosses). Even though chess is far more complicated, it is still a game of finite options: The rules are well defined and there are few extraneous variables that can throw off the program.
Jeopardy!, however, is very different from chess. It is a quiz show, one that deals with history and literature, pop culture, wordplay and the like. Until now, this has been the human arena. For those that don’t know, the shows works in the opposite direction to most quiz shows. An answer is given and the contestants have to buzz first and then give the right question. This is a completely different game to chess.
Watson needed to be able to understand the answer and then formulate an appropriate question using a database of acquired knowledge, which is far more complicated than calculating the possible remaining moves on a chessboard – for a computer at least.
But this isn’t the only place where artificial intelligence (AI) is proving more powerful than human capabilities. Recently, a Berkeley developed Starcraft AI script named Overmind beat a professional player (Oriol) as well as various other AI scripts. This is conceptually somewhere between chess and Jeopardy!, considering the Real Time Strategy (RTS) game’s complexity.
From a programming point of view it has far more variables than a game of chess. While the AI was only programmed to use one attack unit, in opposition to the wide variety a human player could use, the AI’s ability to multi-task, and therefore micro-manage, made it very difficult to beat.
Both Overmind and Watson are learners (Deep Blue originally had to be modified by human intervention). Watson’s algorithms were run through thousands of example questions before the exhibition game on Jeopardy! Every example would teach it which approach was useful on a specific category.
Overmind’s creators sent its units into countless practice battles; each one trained its units’ movements to make as much use of their strengths and the enemy’s weaknesses as possible. Even during live performances, they continued learning… often faster than their human counterparts. To beat a human opponent that will continuously try new and unexpected tactics requires AI that can adapt, no matter how fast and powerful it is.
But before you run to a bomb shelter, fearing the rise of Skynet or Hal 9000, it’s important to keep this all in perspective. The major advantage that Watson had over its opponents in Jeopardy! was not its superior knowledge and understanding. Most important to its success was its ability to react (and therefore buzz) quicker than either of the humans.
This isn’t to diminish its ability to compete with humans on a verbal level (a feat I honestly thought was further away, even for a computer with 15 terabytes of RAM), but simply to highlight where computers will always come out on top. Similarly with Overmind, its major advantage was its ability to almost independently control each unit to a degree that a human (with a far smaller working memory) could never achieve. This combined with its ability to calculate makes it a strong adversary.
Despite all this, this AI doesn’t hold a candle to actual intelligence. They are algorithms within specific parameters. What they do, they do very well and with startling efficiency. The steps that they represent on the path to actual intelligence though, are staggering. Verbal understanding, especially wordplay, is immensely nuanced. That IBM’s machine can compete on that front is a signal of things to come.
Where machines took over physical labour in the industrial revolution, the digital revolution is showing signs of replacing our higher level work too. I end with a note from Garry Kasparov himself, the grandmaster who lost to Deep Blue (who wrote a very interesting article on the topic).
His research suggests that more powerful than either a computer or human, is a team of both. In his tests, an average human player with computer based tools could trump a grandmaster or chess playing program. This suggests that the true masters of the future are not the robots… but rather those who know how to use them.
