AI Winters — 1980s Expert Systems Boom and Collapse

Synthesized from a Grok DeepSearch conversation, April 2026. See also vault://ai-winter for the broader historical record.

Quick Timeline

Period	Event
1974–1980	First AI Winter — DARPA cuts, Lighthill Report, neural net skepticism
Early–mid 1980s	Brief thaw — expert systems boom, Lisp machine sales peak
1985	Peak of the commercial bubble; Japan's Fifth Generation, US Strategic Computing Initiative
1987	Collapse begins — Lisp machine market crashes virtually overnight
1987–1993	Second AI Winter — funding dries up, "AI" becomes a dirty word in business
Late 1990s–2000s	Quiet revival via statistical ML, big data, and eventually deep learning

What Caused the 1980s Collapse

Brittleness of Expert Systems

Expert systems worked well in narrow, well-defined domains. XCON at Digital Equipment Corporation reportedly saved ~$40M/year configuring computers. But they were:

Extremely expensive and time-consuming to build and maintain (the "knowledge acquisition bottleneck" — encoding expert knowledge was labor-intensive)
Rigid — they broke on edge cases, couldn't handle uncertainty or common-sense reasoning, and failed outside their defined scope
Hard to scale beyond a few thousand rules

The same property that made them powerful — explicit, inspectable rules — became their ceiling.

Hardware Market Disruption

Lisp machines were powerful but expensive specialized hardware. In 1987, general-purpose workstations and PCs from Apple, IBM, and Sun caught up in performance at a fraction of the cost. The entire specialized AI hardware market — worth hundreds of millions — collapsed almost overnight. Symbolics went bankrupt; Xerox and others abandoned the space. An entire half-billion-dollar industry was largely replaced in a single year.

Overhype and Unrealistic Expectations

In 1984, Minsky and Schank publicly warned at AAAI that the enthusiasm was spiraling out of control, predicting a "nuclear winter"-style chain reaction of pessimism, press backlash, and funding cuts. Their warning proved prescient just three years later.

Funding Pullback

Corporate investment dried up as ROI failed to materialize at scale. Japan's Fifth Generation project also under-delivered relative to its ambitious goals.

Personal Connection

This collapse happened in real time from the inside. Richard Hess joined Teknowledge in 1985 as an Apprentice Knowledge Engineer — the peak of the boom — working on Xerox Lisp Machines (Dandelion workstations running Interlisp-D). By the late 1980s, the same forces that killed the broader market were felt directly.

That experience left permanent calibration:

"I'm extremely sensitive to the hype that exists today regarding AI, since I saw firsthand how expert systems failed for the same reason — over-promising and under-delivering."

Most people working in AI today weren't born when those knowledge bases were being written at Teknowledge. The scar tissue is rare.

Lessons the Era Encoded

The dangers of hype cycles — impressive narrow results do not generalize cleanly or on the timeline promised
The value of inspectable, deterministic systems — rule-based approaches were highly observable; later probabilistic black boxes traded that property for capability
Maintenance, scalability, and robustness — a system that works in demos but fails in production is not a system

These lessons carry directly into the preference for YAML as single source of truth, Action–State patterns, and deterministic runtimes where agents and humans use the same observable system.