The Winters that Shaped the Present
AI has already promised to solve everything before. In the 1960s, it promised universal automatic translation. In the 1980s, it promised unlimited expert intelligence. Both cycles collapsed — and each collapse taught lessons that were only absorbed decades later. Understanding the winters is the best antidote against the hype of the present.
The Repeating Standard — and Why This Cycle Might Be Different
The field of artificial intelligence has a recurring pattern that any technology historian recognizes: genuine technical advancement → overblown promises → excessive investment → frustration with results → collapse of financing → period of hibernation. The cycle has completed at least twice dramatically — and there is evidence of a third partial winter occurring in the 2010s before the resurgence of deep learning.
O first winter (1974–1980) was precipitated by the Lighthill report to the British government, which harshly criticized the unfulfilled promises of symbolic AI. Government funding dried up worldwide. What had been promised — reliable automatic translation, autonomous domestic robots, computer-based medical diagnosis — simply did not work outside of carefully constructed laboratory examples.
O second winter (1987–1993) was caused by the collapse of the expert systems market. Companies such as Symbolics and Lisp Machines Inc. had built specialized hardware for symbolic AI — and when it became clear that conventional workstations from Apple and Sun were cheaper and sufficiently powerful, the expert systems hardware market collapsed in two years. What seemed to be the future of corporate computing became overnight obsolete technology.
15 Terms that Define Retrospective
Historical-critical terminology for the Pointy Retrospective. Terms that contextualize hype and collapse cycles with epistemological precision.
| Term | Editorial Definition | Level |
|---|---|---|
| Winter of AI | Period of disinvestment caused by unmet expectations — two major cycles: 1974–80 and 1987–93 | Gold |
| Lighthill Report | 1973 — critical analysis of the British government that precipitated funding cuts and the first winter | Gold |
| Expert Systems | MYCIN, XCON, DENDRAL — programs with knowledge coded by experts; brilliant and fragile | Gold |
| Frame Problem | Philosophical problem: how AI systems update beliefs when the world changes — not solved in symbolic AI | Gold |
| Brittleness | Fragility of AI systems outside their training domains — endemic characteristic of symbolic AI | Diamond |
| Overfitting | When a model learns the training data but fails to generalize — a technical analogy for the broader problem of hype | Diamond |
| AI Spring | Period of renewed optimism and investment — the current cycle is the largest AI Spring documented to date | Silver |
| Benchmarking | Evaluation on standardized tests — the absence of robust benchmarks contributed to exaggerated promises in the 1960s–1980s | Gold |
| Turing Threshold | Point at which a system passes the Turing Test — achieved only limitedly in 2014, but criteria were redefined | Gold |
| Transfer Learning | Ability to apply learning from one domain to another — a major advantage of modern LLMs over prior systems | Diamond |
| Symbolic AI | Paradigm based on symbol manipulation and formal logic — dominant from 1956 to 1987, replaced by connectionism | Gold |
| Hype Cycle | Gartner Curve: inflation of expectations → disillusion → productivity plateau — applicable to all AI cycles | Silver |
| GOFAI | Good Old-Fashioned AI — pejorative term for symbolic AI from the 1960s–80s, coined by Haugeland (1985) | Gold |
| Common Sense Problem | Difficulty of AI systems to reason with everyday implicit knowledge — unresolved until LLMs | Gold |
| Scaling | Performance improvement with increased data/parameters — absent in prior systems, central to LLMs | Diamond |
What the Failures of the 1980s Teach Us about the Current Hardware Bubble
In 1985, the company Symbolics was considered the future of computing. Its Lisp Machine hardware was the state of the art for running expert systems — and cost ten times more than conventional workstations. By 1992, the company had failed. This story matters now because the AI chip ecosystem in 2026 has disturbing parallels with the specialized hardware ecosystem of 1985.
The Collapse of Expert Systems
Expert systems of the 1980s were genuinely impressive within their domains. MYCIN diagnosed infectious diseases with precision comparable to human experts. XCON configured DEC VAX computers with 95% accuracy, saving millions of dollars annually. The problem wasn't that they didn't work — it was that they only worked under perfectly controlled conditions and broke spectacularly outside of them.
Knowledge engineering — the process of extracting specialized knowledge and encoding it into rules — was expensive, slow, and not scalable. Each new domain required years of manual work. The promise that it could scale to all forms of human knowledge was never verified — and when it became clear that it couldn't, investment evaporated.
Parallels with 2026
NVIDIA in 2026 holds a market position analogous to Symbolics in 1985 — specialized hardware with extraordinary margins for a specific application. The crucial difference: NVIDIA chips run on commodity infrastructure and are purchased by companies with solid balance sheets, not by AI symbolic startups. But the risk of overcapacity and margin erosion when alternatives emerge is real — AMD, Intel, Google custom chips, Huawei Ascend.
"History does not repeat itself, but it often rhymes." — Attributed to Mark Twain; applicable to the history of AI with disturbing accuracy
What Is Truly Different
Modern LLMs do not suffer from the brittleness that destroyed symbolic AI. They generalize—imperfectly, with hallucinations, but genuinely—for domains and problems they have never seen. Large-scale transfer learning is the fundamental technical advance that separates this cycle from the previous onesBut historical vigilance remains necessary: the promises of fully trustworthy autonomous agents in 2024 have the same odor as the promises of universal expert systems in 1982.