The second day of Watson's man-vs-machine Jeopardy! challenge has come to close, and IBM's super-computer dealt a resounding defeat to its pitiful, meat-filled opponents. In light of that fact, we thought now might be a good time to learn a little more about our future overlord, so we found an expert on the subject to answer some of our questions.
Stephen Baker is the author of Final Jeopardy: Man vs. Machine and the Quest to Know Everything, a new book about how Watson came to be. He was gracious enough to answer some of our burning questions. Read on to find out what he said!
Why did the team at IBM settle on Jeopardy as a stage to show off their engineering mettle? Why did they decide that now was the time for the biggest artificial intelligence exhibition match since Deep Blue?
They wanted to do another big challenge, and they had to find something that was hard, but possible. That's a tough window. One of the nightmare scenarios was that this huge tech company would choose a challenge, blow their horn about it, and then see some other company, or even an individual, beat them to it.
So Jeopardy seemed hard enough. What's more, it responded to the great trend in modern technology: We're all swimming in an ocean of unstructured data, including trillions of words. And those who can derive intelligence from those words stand to make a killing. Google is an early example. So Jeopardy gave IBM the chance to mix its natural language with data retrieval and analytics--and then show off their machine on nationwide TV.
Within the realm of Jeopardy questions, does Watson have strengths and weakness? In other words, does he know some categories or types of questions better than others?
The machine has the easiest time with clearly stated clues demanding specific answers, preferably ones tied to dates, numbers, facts, proper nouns. If you think about it, these are the clues that would pop up so high in the first Google search result that you wouldn't have to click on it. Of course, if Jeopardy provided lots of this type of clue, search engines could handle it.
The harder clues for Watson involve humor, nuance, and demand several layers of reasoning. There was one that said: "His daughter and grandson were both premiers, and both assassinated." Now, the person the machine is looking for there is a father and a grandfather, but that has to be calculated from the language. That kind of clue is challenging for Watson. The correct response, incidentally: Who is Nehru?
What were the biggest challenges—both technical and logistical—in creating a computer that can play Jeopardy?
The biggest challenge was to build a smart machine. Sounds simple, I know. But if you think about it, it's very hard for a machine to be sure that it understands complex English, and then to find a specific answer that it can bet on within two or three seconds. It would be easy to bring up, say, 100 possible answers, and have humans find the one they're looking for. That's more or less what a Google search does.
The hardware also presented a challenge. The analysis initially took a single server nearly 2 hours to respond to one clue. They had to distribute that work to more than 2,000 computing cores, and drive the time down to two or three seconds. That scale-out required immensely sophisticated work.
What is the next frontier for man-vs-machine competition? What is humanity clearly superior at now, that we won’t be in 5 or 10 more years?
Well, one next frontier is to make machines like Watson even better at what they do. They could be engineered to handle different types of data, and to become experts in different fields. Then how about building smart machines that can see and hear, and move around the world and do jobs? In a word: smart robots. Watson is deaf and blind, but its offspring don't have to be. Beyond that, of course, is the challenge of building machines that actually think the way we do. But that's a long way off.
Is there a practical application for the technology developed for Watson? Will Watson (or one of his “children”) be able to help consumers find answers to questions?
Watson's kin will be handling consumer help lines very soon, I would imagine. And other Watson-like machines will be attempting to diagnose illnesses in hospitals, searching for legal precedents in law offices, perhaps even serving as bionic research assistants in newpaper offices--if we journalists can afford them. And of course, a lot of these services will be available through cell phones, with the machine understanding our spoken language.