Computers are officially smarter than humans now, as an IBM supercomputer named Watson has defeated two human champs on “Jeopardy!”
Processor technology has come a long way since IBM’s previous superstar, Deep Blue, beat chess master Garry Kasparov in 1997.
This wasn’t really a test of knowledge so much as a test of how well a computer can understand human speech. Any humble desktop can do a keyword search, but knowing which keywords to search for, understanding how words fit together is surprisingly difficult.
Humans take this for granted because context is an inherent part of how we learn. Human brains learn to “weigh” the importance of words based on who we’re talking to and what they’re likely to talk about.
Computers take in every question as if it were a brand new thing. Teaching a computer to recognize speech is really a process of teaching context, to teach the computer about the words that are not included in the question, to try and fill in some of that unspoken knowledge that all humans have.
A great example of this occurred in a “Jeopardy!” round that had to be reshot. In the Olympic Oddities category, Jennings gave an incorrect answer when he said Olympian gymnast George Eyser was “missing a hand.” Watson responded, “What is a leg?”
At first, this looks like a correct answer, because Jennings had already established that they were talking about missing body parts.
But Watson couldn’t hear the other contestants. The computer had forgotten to specify that the leg was missing, but in the flow of conversation, its answer looked right.
Researchers call this the “Paris Hilton” problem. Feed an ordinary computer a query like, “Where is Paris Hilton?” and it has a dilemma.
The machine can’t tell if you’re stalking a famous socialite or trying to find directions to a hotel.
Here’s another fact to consider before we surrender to our robot overlords. A big part of “Jeopardy” is knowing when to buzz in. Unlike older quiz shows, “Jeopardy!” doesn’t allow contestants to buzz in before Trebek finishes the question. This gives computers a unique advantage.
The questions were fed to Watson in text form the instant Trebek finished saying them, allowing the machine to leverage its processing power and buzz in at lightning speed. The machine wasn’t necessarily “smarter” than the contestants; it was just really good at knowing when to hit the buzzer.
But it really comes down to how you define “smarter.” Computers have no intuition, and they lack the lifetime of context that helps human beings interpret language, but no human being could hold the contents of Google in his head, and human beings will never be able to process raw data as fast as a computer can.
I’m sure it was a shock the first time a computer solved a math problem faster than a human being, but now this miracle is a fact of life. Google can sift through a million encyclopedias faster than you can type a search term, but it doesn’t really know the difference between a socialite and a hotel.
The hotel is No. 7 in Google search results for “Paris Hilton,” by the way — a result that surely has Conrad Hilton spinning in his grave.
The human race has nothing to fear from smarter computers. Indeed, some humans are betting on it. Fans of artificial intelligence talk about something called a “technological singularity” — a hypothetical tipping point when machines become smart enough to solve all our problems for us.
I’ve always been skeptical of these predictions because I don’t think the fundamental problems of humanity can be solved with raw intelligence. We already know how to build stable, peaceful societies; we just aren’t doing it. We’d rather play power games and fight with each other.
Having the right answer isn’t good enough; somebody has to make us live by it. A computer smart enough to solve our problems would still need a populace willing to obey it.
But that’s a problem for the long run. In the short term, smarter computers mean cheaper products, faster response times and dramatically improved access to information.
A version of Watson will probably be answering your customer service questions in five years. A customer service database is small enough, narrow enough and specific enough to be a good fit for Watson.
I’ll recommend that for IBM’s next project. Parsing words is the easy part, but can you make a computer smart enough to understand the Texas accent?
THE PAGE STOPS BUT THE BLOG GOES ON. TALK BACK TO MICHAEL AT MICHAELDUFF.NET.