Source: Medium / UX Management publication
Author: Jamieson Rothwell
URL: https://medium.com/ux-management/when-the-billion-dollar-ai-dream-became-a-100-000-paperweight-d692d14d12bc
Fetched: 2026-04-01
Note: This content was summarized from a Medium article. Medium content may be paywalled or restricted. Verify usage rights before reproducing. This summary is for personal reference only.
Examines the second AI winter (roughly 1987–mid-2000s), arguing it was more severe than the first because the industry believed it had already learned from prior failures.
The 1980s saw explosive growth in expert systems — AI programs capturing human expertise in rule-based formats. XCON, developed for Digital Equipment Corporation, processed 80,000 orders with 95–98% accuracy and saved DEC an estimated $25 million annually. By 1985, corporations were spending over $1 billion on AI, spawning specialized hardware manufacturers like Symbolics and LISP Machines Inc., selling computers priced at $70,000–$150,000.
In 1984, AI pioneers Roger Schank and Marvin Minsky warned at an industry conference that enthusiasm had "spiraled out of control" and predicted a cascade of funding cuts and research collapse. Their warning proved accurate within three years.
Moore's Law rendered specialized AI hardware obsolete. Mass-market desktops from Apple and IBM became powerful enough to run LISP and expert systems at a fraction of the cost. Half a billion dollars in market value evaporated within one year. Symbolics (revenue of $115 million in 1986) eventually filed for bankruptcy; LISP Machines Inc. collapsed immediately.
Expert systems proved brittle and unmaintainable. XCON required 59 full-time engineers just to maintain its growing rule base. These systems worked only within narrow domains, couldn't learn or adapt, and proved economically unsustainable compared to emerging enterprise resource planning systems.
DARPA cut AI funding "deeply and brutally" under Jack Schwarz, dismissing expert systems as merely "clever programming." Most research projects were terminated, though DART (battle management) survived and ironically saved billions during the Gulf War.
The Fifth Generation Computer Systems project ($400 million, 1981–1992) aimed to leapfrog American AI capabilities through parallel processing and PROLOG. Project head Kazuhiro Fuchi admitted failure plainly; the government eventually offered the software free to anyone — with no takers.
The winter lasted roughly 1987 to mid-2000s. Researchers stopped using the term "artificial intelligence" due to stigma, relabeling their work as "machine learning" or "computational intelligence." Foundational work in neural networks and probabilistic reasoning continued but remained underfunded.
Rothwell argues modern AI faces similar risks: CEOs promising AGI by 2027, billion-dollar compute investments, and applications that excel in demos but struggle in production. Researchers warning of a third AI winter aren't pessimists — they're observers who recognize the historical pattern.