The second AI winter
The fall came without a ceremony. Lisp Machines Inc. — one of two companies that had bet everything on purpose-built computers for artificial intelligence — filed for bankruptcy in 1987. Its rival Symbolics, which had sold fully configured workstations at $110,000 apiece and counted two-thirds of the Fortune 500 among its customers, would follow by 1991. Texas Instruments and Xerox quietly abandoned the field. The cause of death was mundane: a Sun Microsystems workstation, running ordinary Unix, had become fast enough to run Lisp. An industry worth half a billion dollars had lost its reason to exist.
Expert systems were going the same way, just more slowly. XCON, the most celebrated of the genre, had been built by John McDermott at Carnegie Mellon in 1978 and deployed at Digital Equipment Corporation in 1980 to configure VAX computer orders. By 1986 it had processed 80,000 orders at 95–98% accuracy and was saving DEC an estimated $25 million a year in configuration errors — a genuine industrial miracle (Wikipedia: Xcon). It had also grown to 2,500 rules that almost no one fully understood. A single new rule could break behavior in unrelated corners of the system. Engineers called it brittle: spectacular when it worked, catastrophic at the edges.
The funding followed the hardware. Jack Schwarz, who took over DARPA’s Information Processing Technology Office in 1987, dismissed expert systems as “clever programming” and cut AI research funding “deeply and brutally.” Of the Strategic Computing Initiative — a $1 billion, decade-long U.S. government effort to realize AI for military applications — almost nothing survived except DART, a logistics tool. Japan’s Fifth Generation Computer project, an $850 million national effort to build conversational AI machines in Prolog, ended in 1992 not with its promised breakthrough but with a soft admission that the targets had been unreachable (War on the Rocks).
The cruelest irony belonged to the researchers who had seen it coming. In 1984, Roger Schank and Marvin Minsky had publicly warned the field about excessive hype, predicting “disappointment would certainly follow” — a chain reaction, they said, that would spread through the community and press like a nuclear winter. Nobody had listened with much urgency. Three years later, they were proven right about everything except the timeline.
What survived were the quieter methods. Judea Pearl’s work on Bayesian networks showed how probability could substitute for rules. Hidden Markov models — statistical rather than logical — were already mapping the phonemes in speech. Geoffrey Hinton had not abandoned neural networks; he had simply stopped arguing with funding agencies and kept working. The community that would eventually produce deep learning had not disappeared. It had gone underground, into labs learning to succeed without permission.
The second AI winter lasted roughly six years. It ended not with a single announcement but with a series of quiet, specific victories — a speech recognizer here, an image classifier there — that had nothing to do with Lisp or logic or the ambitious promises of 1983. The machines that displaced the expert systems didn’t contain rules. They contained weights, adjusted by data rather than dictated by an expert. The winter had cleared the ground.
Sources
- AI winter — Wikipedia — Chronology of both AI winters, the Lisp machine collapse, DARPA funding cuts, and Minsky and Schank’s 1984 warning.
- Xcon — Wikipedia — XCON’s development by McDermott, deployment at DEC, accuracy record, rule count, and cost savings.
- A Cautionary Tale on Ambitious Feats of AI: The Strategic Computing Program — War on the Rocks — DARPA’s Strategic Computing Initiative: timeline, funding cuts, and critical assessment.
- The Second AI Winter (1987–1993) — Holloway / Making Things Think — Overview of the collapse, qualification problem, and shift to probabilistic methods.