Thinking Machines Corporation

Contribution to AI

The thesis question behind the machines was whether a computer had to be one fast processor with a memory attached, or could be tens of thousands of tiny processors each holding a fragment of the data — an architecture chosen because semantic networks and marker passing looked like the natural workload. What the hardware actually taught the field was the data-parallel style: one operation written once and applied simultaneously across every element of a large collection, expressed in *Lisp, C* and CM Fortran, together with the primitives — scans, reductions, general routing between processors — that make it usable. That vocabulary outlived the company and reappeared nearly unchanged when general-purpose GPU programming arrived. The machines also made brute force intellectually respectable in AI. Stanfill and Waltz argued in 1986 that a system could classify by retrieving similar past episodes from a large store rather than by applying rules, a position defensible only once comparing a query against an entire database at once had become cheap; the companion paper on parallel free-text search grew into WAIS, released in 1991 as one of the first internet-wide full-text search services. The ending belongs in the account too. CM-5s took the top four places on the first TOP500 list in June 1993, and bankruptcy followed fourteen months later — the first clear demonstration that special-purpose parallel hardware built around a particular theory of intelligence can be beaten by commodity processors improving on their own schedule.

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