IBM

Contribution to AI

A machine sold for engineering arithmetic became the place where a programme first improved with experience. Arthur Samuel had checkers running on the 701 by 1952, and the 1959 paper describing how it tuned its own evaluation function through self-play and memorised positions put the phrase machine learning into circulation. Nathaniel Rochester, who designed that machine's logic, co-signed the 1955 Dartmouth proposal; Herbert Gelernter's geometry theorem prover used a diagram to prune its search on a 704, the same model on which LISP was first implemented. The deeper contribution came later and was statistical. From the 1970s the Yorktown speech group under Frederick Jelinek treated recognition as a decoding problem — hidden Markov models, n-gram language models estimated from text, perplexity as the yardstick — and that approach displaced hand-built linguistic rules for the next two decades. The same group then aligned the bilingual record of the Canadian parliament word by word; the alignment models they published in 1990 and 1993 defined statistical machine translation until neural systems arrived. The 1997 chess match settled a wager Herbert Simon had made forty years earlier, and the 2011 quiz-show win showed that open-domain question answering could be assembled from many weak sources of evidence scored against each other. Running through all of it is the demonstration that these were engineering problems with schedules, budgets and paying customers, not thought experiments.

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