IBM Watson (Watson Launch)

Contribution to AI

Deep Blue had shown that search could beat a grandmaster, but chess is a closed world with legal moves and no puns. Jeopardy! clues are allusions and deliberately oblique phrasing, and answering them in about three seconds from unstructured text was a different problem entirely. The DeepQA architecture that David Ferrucci's team built met it by declining to look for the answer: it generated hundreds of candidate hypotheses per clue, scored each with more than a hundred independent algorithms drawing evidence from encyclopaedias, dictionaries and news text held entirely offline on 2,880 POWER7 cores, merged those scores with learned weights into a confidence figure, and buzzed only when confidence cleared a threshold. Many weak scorers, calibrated confidence, and an explicit decision about when to stay silent became the reference design for open-domain question answering until neural methods displaced it, and the UIMA framework underneath went to Apache as a standard for text-analytics pipelines. The February 2011 match against Ken Jennings and Brad Rutter ended 77,147 dollars to 24,000, which mattered less technically than commercially: it made spending on AI respectable in boardrooms a decade before the language models arrived. What followed is equally instructive. The oncology work stalled, the health assets were sold in 2022, and the episode is now the standard warning about the distance between a demonstration and a deployment.

Drafted with AI and edited by hand (claude-opus-5, reviewed 2026-08).