Part III · Era 3 — Winters and Expert Systems (1974–1993)

draft

The cycles of hype and disillusionment. Unfulfilled promises cut funding (the "AI winters"), while expert systems prove, for a time, that AI can turn a profit — before they too disappoint.

Narrative

After the optimism comes the bill. The Lighthill Report (1973) and budget cuts open the first AI winter. The field recovers in the 1980s with expert systems — domain rules that solve real problems and create an industry. But they are expensive to maintain and fragile outside their scope; the LISP machine market collapses, bringing the second winter. Behind the scenes, however, the connectionist strand is reborn: backpropagation is popularized in 1986, seeding the revolution that would come 25 years later.

Milestones

  • 1965 — *Alchemy and Artificial Intelligence* · Hubert Dreyfus (RAND) ·

    philosopher's critique likening AI to alchemy; an early salvo that helped cool funding and the field's grand claims.

  • 1973 — Lighthill Report · United Kingdom · devastating critique that

    triggers the 1st AI winter.

  • 1972–1980 — MYCIN, DENDRAL · Stanford · pioneering expert systems

    (diagnosis, chemistry).

  • 1980 — XCON/R1 · DEC · expert system that saves millions; ignites the

    commercial boom.

  • 1982 — Hopfield networks · John Hopfield · associative memory via the

    physics of spin systems.

  • 1982 — Fifth Generation Project · Japan · billion-dollar national bet on

    logical AI (Prolog).

  • 1986 — Backpropagation popularized · Rumelhart, Hinton, Williams · trains

    multilayer networks; revives connectionism.

  • 1987–1993 — 2nd AI winter · collapse of the LISP machine market and of

    expensive expert systems.

  • 1989 — CNN for digits · Yann LeCun · LeNet recognizes handwritten digits

    (basis of check recognition).

People

Geoffrey Hinton (1947–) — "godfather" of deep learning. Coauthor of backprop (1986); Turing Award 2018; Nobel Prize in Physics 2024. 🖼️ Photo F-III.7.

Yann LeCun (1960–) — father of CNNs. LeNet (1989); Turing Award 2018; head of AI at Meta. 🖼️ Photo F-III.8.

John Hopfield (1933–) — Hopfield networks (1982). Brought statistical physics to neural networks; Nobel Prize in Physics 2024.

Edward Feigenbaum (1936–) — "father of expert systems". DENDRAL and knowledge engineering; Turing Award 1994.

Curiosities

  • The bet that dared the machines (1968–1989). At a 1968 gathering hosted by

    Donald Michie (founder of Edinburgh's AI department), Scottish chess champion David Levy beat John McCarthy — who predicted a computer would beat Levy within ten years. Levy wagered it would not; other AI researchers joined in, the pot reaching £1,250. Levy won the bet in 1978, defeating the program Chess 4.7 4½–1½. He stayed unbeaten until 1989, when Carnegie Mellon's Deep Thought — the machine that would evolve into IBM's Deep Blue — crushed him 0–4 in London. The lineage runs straight to Deep Blue vs Kasparov (1997), in the next era.

  • Why games kept mattering. As Anderson Rocha (Unicamp) frames it, games gave AI

    a controlled world with known rules — a fair proving ground for algorithms — which is why each public milestone (chess, later Go) arrived through a game.

Source: curiosities surfaced by the BBC News Brasil report *"Máquinas podem pensar? Os 70 anos de história que levaram à inteligência artificial dos dias atuais"* (Luiz Fernando Toledo, 2025); the Levy/Deep Thought figures cross-checked against primary sources.

🎨 Figure F-III.9 — Ice and rules. Brief: split composition: half "winter" (frozen landscape, plummeting funding charts, a dusty LISP machine) and half "expert systems" (flowchart of IF-THEN rules, MYCIN's medical coat). At the bottom, a small lit flame labeled "backprop 1986" — what survives the ice.

Connection to the Compendium: expert systems are the gallery ../02-types-of-ai/01-symbolic.kmd; the 1986 backprop is the central training algorithm of Part I (../01-life-cycle/02-stages-in-detail.kmd, stage 4).