Xerox PARC (AI Research Push)

Contribution to AI

By the early 1980s the dominant recipe for a knowledge system was to compile a specialist's symptom-to-cause judgements into rules, which meant nothing could be diagnosed that nobody had thought to encode. The counter-argument was assembled in Palo Alto. Johan de Kleer and John Seely Brown's qualitative physics, published in 1984, described how a device behaves in terms of structure and constraint rather than numbers, and the model-based diagnosis de Kleer developed with Brian Williams in 1987 used such a model to work out which components must be misbehaving for the observed symptoms to occur — no fault catalogue required. The assumption-based truth maintenance system of 1986 made that tractable by recording, for every derived fact, the assumptions it depended on, so a programme could carry several incompatible hypotheses simultaneously instead of backtracking through them one at a time. The second thread was tooling. Interlisp-D put the whole Lisp environment — editor, debugger, cross-referencer — inside the same running image as the user's code, on a personal machine with a bitmap display and a mouse, and made that the expected working conditions for AI research rather than a luxury. LOOPS, from 1983, held object-oriented, rule-based and data-driven styles in one language, and its successor CommonLoops was reconciled with Symbolics' New Flavors in 1986 to become the Common Lisp Object System, whose metaobject protocol remains the standard demonstration that a language can expose its own machinery to the programmer.

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