Generating hundreds of guesses before judging any of them
DeepQA parsed each clue, guessed its answer type, and checked whether it needed splitting into subclues. Primary search then queried text engines, passage indexes, and triple stores against a corpus IBM had assembled in advance, with no live web access during play, favoring recall over precision and generating several hundred loosely supported candidate answers per question. [2]
A soft filter narrowed that field to roughly 100 candidates. Survivors were checked against retrieved evidence by more than 50 scoring algorithms measuring passage alignment, source reliability, geospatial and temporal consistency, and grammatical fit. On a clue about a 1974 pardon, one scorer favored Nixon over Ford because only Nixon filled the object position in a matching retrieved sentence. [2]
Fifty scorers and a metalearner, not a neural network
That combination step is the part later retellings compress into deep learning, a term the paper never uses. A metalearner merged each candidate's scores using techniques the paper calls mixture of experts and stacked generalization, trained on questions with known answers, closer to layered logistic regression over hand-built features than to a network learning its own representations. [2]
Scaled onto roughly 2,500 cores with the Apache UIMA-AS framework, the pipeline answered within three to five seconds, the window a contestant has to hear a clue and decide whether to buzz in. An earlier single-processor version had taken two hours per question. [2]
What three days on television actually proved
Watson beat Ken Jennings and Brad Rutter over three days of competition in February 2011, finishing with $77,147 to their $24,000 and $21,600. It also made a memorable, uncorrected error, naming Toronto instead of Chicago on a Final Jeopardy clue about US cities, a reminder that an aggregate score decided the match, not a run of flawless answers. [1]
The healthcare business the Jeopardy win was supposed to fund
That win ran on a corpus IBM curated in advance, with no live web access. IBM's next move, recommending cancer treatment from patient records and medical literature, asked the same architecture to judge evidence nobody had pre-selected. [2]
MD Anderson Cancer Center spent five years and $62 million on a Watson-based Oncology Expert Advisor before letting the contract lapse in 2016 without using it on a patient. An audit blamed procurement and integration problems, including Watson's trouble telling the abbreviation for acute lymphoblastic leukemia apart from the one for a drug allergy. [3]
IBM and Francisco Partners announced a definitive agreement on January 21, 2022 for selected Watson Health healthcare data and analytics assets, including MarketScan, Micromedex and imaging software. Financial terms were not disclosed. Francisco Partners announced completion on June 30 and named the standalone business Merative. The January agreement was not the closing, and these announcements do not describe a sale of every Watson product. [5][6]
IBM subsequently developed its own Granite language models while also integrating outside models. Its October 2024 Granite 3.0 release made IBM models available under Apache 2.0. The August 2026 OpenAI partnership concerns embedding OpenAI models and products in Consulting Advantage; it does not establish that IBM stopped developing models. [7][8]