Feb 13, 2021 - Technology

Looking back at Watson's 2011 "Jeopardy!" win

Ken Jennings (he's the human) plays IBM's Watson AI system at Jeopardy on Jan. 13, 2011.

Tuesday marks the 10th anniversary of IBM's Watson AI system crushing its human competition on "Jeopardy!"

Why it matters: Watson's victory marked one of the first times Americans could witness an AI system using natural language processing. But, 10 years later, the field still has far to go.

Background: The idea of sending a machine learning system to compete on "Jeopardy!" originated with David Ferrucci, an AI expert then at IBM.

  • Ferrucci's interest was in making machines that "actually understood language in a much deeper way," and the Watson program — with "Jeopardy!" as a goal — offered a chance to move closer to that goal, which led the company to eventually pick it as one of its "Grand Challenges."

How it worked: The main obstacle was that Watson would need to be able to parse the language of Jeopardy questions — meaning answers — to search for clues it could then match in its store of information.

  • Watson used hundreds of algorithms at every stage of the process to work out the nature of the question, before creating a weighted list of possible answers based on how likely they were to be correct.

What happened: On Feb. 16, 2011, the final episode aired featuring IBM's Watson going up against top champions Brad Rutter and Ken Jennings.

  • It wasn't even a contest — by the end of the third and concluding match, Watson was up nearly $50,000.
  • Jennings summed up the defeat with a tongue-in-cheek answer to the last Final Jeopardy: "I, for one, welcome our new computer overlords."

The catch: As impressive as Watson was, Ferrucci points out that "in the end, it was making linguistic predictions. There's no really deep interpretation."

  • Ferrucci, who left IBM in 2012, is now trying to build AIs that can better interact with human beings at his startup Elemental Cognition.
  • "Our focus is on not learning what's the next word or what's the next number?" he says. "It's learning the comprehension."

The bottom line: Should AI truly be able to do that, our new computer overlords will have truly arrived.

Go deeper