There are many accounts of the genesis of Watson. The most popular, which is not necessarily the most accurate—and this is the sort of problem that Watson himself often stumbled on—begins in 2004, at a steakhouse near Poughkeepsie. One evening, an I.B.M. executive named Charles Lickel was having dinner there when he noticed that the tables around him had suddenly emptied out. Instead of finishing their sirloins, his fellow-diners had rushed to the bar to watch “Jeopardy!” This was deep into Ken Jennings’s seventy-four-game winning streak, and the crowd around the TV was rapt. Not long afterward, Lickel attended a brainstorming session in which participants were asked to come up with I.B.M.’s next “grand challenge.” The firm, he suggested, should take on Jennings.
I.B.M. had already fulfilled a similar “grand challenge” seven years earlier, with Deep Blue. The machine had bested Garry Kasparov, then the reigning world chess champion, in a six-game match. To most people, beating Kasparov at chess would seem a far more impressive feat than coming up with “Famous First Names,” say, or “State Birds.” But chess is a game of strictly defined rules. The open-endedness of “Jeopardy!”—indeed, its very goofiness—made it, for a machine, much more daunting.
Lickel’s idea was batted around, rejected, and finally resurrected. In 2006, the task of building an automated “Jeopardy!” champion was assigned to a team working on question-answering technology, or QA. As Stephen Baker recounts in his book about the project, “Final Jeopardy,” progress was, at first, slow. Consider the following (actual) “Jeopardy!” clue: “In 1984, his grandson succeeded his daughter to become his country’s Prime Minister.” A person can quickly grasp that the clue points to the patriarch of a political family and, with luck, summon up “Who is Nehru?” For a computer, the sentence is a quagmire. Is what’s being sought a name? If so, is it the name of the grandson, the daughter, or the Prime Minister? Or is the question about geography or history?
Watson—basically a collection of processing cores—could be loaded with whole Wikipedias’ worth of information. But just to begin to search this enormous database Watson had to run through dozens of complicated algorithms, which his programmers referred to as his “parsing and semantic analysis suite.” This process yielded hundreds of “hypotheses” that could then be investigated.
After a year, many problems with Watson had been solved, but not the essential one. The computer took hours to generate answers that Jennings could find in an instant.