IBM Watson

What it is

Watson is a question-answering computer system developed by an IBM research team led by principal investigator David Ferrucci^[1]. Question answering is a branch of computing in which a system is given a question in ordinary human language and has to return an answer, rather than a list of documents that might contain one. Watson was built to do that quickly and accurately enough to compete on Jeopardy!, an American television quiz show in which contestants are read a clue and must respond in the form of a question. IBM named the system after the company's first CEO, Thomas J. Watson Sr.^[1]

Where it came from

IBM says that in 2006 Ferrucci pitched his bosses on building a computer that could beat a human at Jeopardy!^[1] The pitch became a project: Ferrucci assembled a team of more than two dozen scientists, engineers and programmers at the Thomas J. Watson Research facility in Yorktown Heights, New York^[1]. By IBM's account it took the team five years to perfect the question-answering system^[1]. The target was not a narrow, single-purpose task but an open quiz, so the machine had to work out for itself what a clue was asking before it could answer.

How it worked

The original Watson was a room-size computer consisting of 10 racks holding 90 servers with a total of 2,880 processor cores^[1]. It ran IBM's DeepQA software, which was used to generate hypotheses, gather evidence and analyse data^[1]. Over a period of years the system ingested information from Wikipedia, encyclopedias, dictionaries, religious texts, novels, plays and Project Gutenberg books^[1]. It answered questions posed in natural language without being connected to the internet^[1], so everything it could draw on had to come from the material it had already taken in.

Watson's main innovation was quickly executing hundreds of algorithms to analyse a question from many directions, find and score potential answers and gather supporting evidence^[1]. The design treated agreement between independent methods as a signal: the more algorithms that independently arrived at the same answer, the higher Watson's confidence in it^[1]. It was programmed to buzz in only if that confidence was high enough^[1], which matters in a quiz, where a confident wrong answer is costly. IBM says Watson performed all of these calculations in about three seconds^[1] — roughly the time available to read a clue and decide whether to ring in.

The match and what followed

IBM records Watson's televised Jeopardy! victory in February 2011, when it defeated the show's two foremost all-time champions^[1]. The point of the exercise was public and concrete: a general-knowledge quiz is hard to fake with lookup tables, because the questions arrive in ordinary language and the answer has to be chosen and justified on the spot. A system that could hold its own against the best human players, in the same format and under the same time pressure, was a visible demonstration that natural-language question answering had moved from research papers to something that worked in front of an audience.

Watson then moved from the studio into commercial use. IBM said roughly 70% of global finance institutions and 13 of the top 14 systems integrators use Watson^[1]. IBM also says Watson later provided hyperlocal forecasts with The Weather Company and satellite and aerial imagery analysis for OmniEarth^[1]. That shift — from a single, very public contest to tools sold into banks, integrators and data-heavy industries — is the shape of Watson's place in the field: a research system that became a company's flagship commercial AI offering.