Watson's $947 Bet: IBM's AI on Jeopardy
Zusammenfassung
IBM’s Watson defeated human Jeopardy champions Ken Jennings and Brad Rutter in a three-day match in February 2011, processing 200 million pages of content to answer questions correctly at superhuman speed. Two moments became famous. In the category “Olympic Oddities,” Watson answered “What is leg?” — wrong, because it could not infer the context a human would have. And in Final Jeopardy of the first game, given a “U.S. Cities” clue, it answered “What is Toronto?” (also wrong) while wagering the oddly specific amount of $947 — a low, precise bet that reflected its weak confidence. Jennings, who lost, wrote beneath his answer in the final game’s Final Jeopardy: “I for one welcome our new computer overlords.”
The System
IBM’s Watson used a combination of techniques that were cutting-edge in 2011:
- Question decomposition: Watson parsed complex, often ambiguous Jeopardy clues into structured queries by identifying the answer type (person, place, thing) and relevant constraints.
- Information retrieval: Watson searched its knowledge base of 200 million pages (including Wikipedia, dictionaries, encyclopedias, and literary works) for candidate answers.
- Evidence scoring: Hundreds of algorithms independently scored each candidate answer based on different types of evidence. The scores were combined into a confidence level.
- Buzzer timing: Watson was physically connected to the buzzer and could buzz in as soon as its confidence exceeded a threshold — often faster than human contestants could physically react.
Watson could not access the internet during the game. Its knowledge was fixed at the time of training. This gave it depth but no awareness of recent events.
The Match
The match aired February 14-16, 2011, on Jeopardy. Watson played against Ken Jennings (74-game winning streak, highest Jeopardy winnings) and Brad Rutter (highest career winnings of any Jeopardy player). After three games, Watson had won $77,147; Jennings had $24,000; Rutter had $21,600. Watson won decisively.
The $947 bet was Watson’s Final Jeopardy wager in the first game. Watson’s wagering algorithm set the bet from its confidence in the answer, the current score differential, and the remaining clues. The algorithm produced $947 — an oddly specific amount that looked strange to human observers but reflected a precise calculation, and a deliberately low one, because Watson’s confidence in its answer was weak.
That first-game Final Jeopardy clue, in the category “U.S. Cities,” asked for the US city whose largest airport was named for a World War II hero and whose second-largest was named for a World War II battle. Watson incorrectly answered “Toronto” (a Canadian city); the correct answer was Chicago. Watson still won the match by a large margin. The famous surrender note came later: in the final game’s Final Jeopardy, category “19th Century Novelists,” Watson, Jennings, and Rutter all correctly answered Bram Stoker — and it was beneath that answer that Jennings wrote “I for one welcome our new computer overlords.”
The Watson Legacy
IBM marketed Watson as a general-purpose question-answering AI and attempted to deploy it for medical diagnosis (Watson for Oncology), financial advisory services, and other professional applications. These deployments were largely unsuccessful — Watson’s performance on highly structured Jeopardy questions did not transfer well to the open-ended judgment required in medical and financial contexts.
The AI landscape Watson occupied in 2011 — where specialized expert systems competed in narrow domains — was transformed within two years by deep learning. Watson’s architecture, which relied on handcrafted question-analysis pipelines and retrieval-based reasoning, was largely superseded by neural network approaches. The contrast between Watson’s 2011 triumph and its subsequent commercial struggles illustrates a recurring pattern in AI history, discussed in Expert Systems and the First AI Winter.