The AI Winters — When the World Stopped Believing

The day an economist declared AI dead

In 1973, a British economist named Sir James Lighthill handed a 40-page report to the UK government. The title was dry: "Artificial Intelligence: A General Survey." Inside, it said something brutal. AI doesn't work. Researchers promise smoke. Cut the funding. The UK cut it. The US followed. Labs shut down. Researchers lost their jobs. The world stopped believing.

That was the first "AI winter." It lasted almost a decade.

Then it happened again

The 1980s looked like a comeback. "Expert systems" showed up — programs that packed a doctor's or an engineer's knowledge into rules of the "if this happens, do that" kind. Companies spent billions. AI seemed to have finally arrived. Until it hadn't. The systems were brittle. They broke the moment you stepped outside the domain they were built for. By 1987, the money evaporated. Second winter. Another six years of silence.

What nobody was telling

But while the outside world was calling AI dead, something else was happening in the lab. Geoffrey Hinton, in Toronto, kept working on neural networks. His papers got rejected. He wasn't funded. In 1986, he and two colleagues published in Nature a way to train deep networks — the backpropagation algorithm. Almost nobody paid attention. Today it's the foundation of Claude, of ChatGPT, of everything you use.

Yann LeCun, at Bell Labs, was working on convolutional networks to read bank checks. Yoshua Bengio joined later. Three people holding up an idea the world had thrown out. Thirty years later, they got the computing equivalent of the Nobel Prize for that work.

Why this matters now

People tend to repeat the same cycles. Someone makes a big promise. Years pass. The promise doesn't land the way it was sold. Trust collapses. Money walks away. And while everyone looks elsewhere, somebody keeps building.

You, as a professional using AI, don't control those cycles. But you can learn to read them.

Takeaways:

Three things to pin down. First, AI survived twice because the people who believed in it kept building quietly — Hinton, LeCun, Bengio. They didn't expect the world to catch up; they knew they were right. Second, real progress happened with no fanfare. While everyone argued about promises, they were solving concrete math problems. Third, today is different because you can't turn off what already works. Gmail has been filtering spam with AI since 2005. Google Search has used deep learning for years. That doesn't disappear because someone loses faith.

October 1973. One report changes everything.

Sir James Lighthill, a British economist, delivers a 40-page document to the UK Science Research Council. The title is dry enough that it almost seems designed to keep people from reading it: "Artificial Intelligence: A General Survey." For AI researchers at the time, though, it's a death certificate.

Lighthill writes what many government officials already wanted to hear. AI doesn't work. The researchers are selling smoke. Millions have been spent on research that isn't producing intelligent machines. Time to stop.

The UK cuts funding that same year. The US, which had been funding AI heavily through DARPA, does the same. Overnight, labs shut down. Researchers lose their jobs. The field that promised to change the world gets declared impossible.

But Lighthill's report had a problem. It was partly right — for the wrong reasons.

The uncomfortable truth of the first winter (1974-1980)

Lighthill wasn't wrong that 1960s AI had failed to deliver on its promises. Researchers had sworn up and down that superintelligent machines were 10 years away. They weren't. Budgets grew. Results didn't.

What Lighthill didn't see was that while the world ignored them, a small group kept working in the shadows.

Geoffrey Hinton was at the University of Toronto investigating neural networks at a time when the field was basically buried. Yann LeCun did his PhD in Paris in the early 1980s and then moved to Bell Labs, where he developed the convolutional networks that are standard infrastructure today. Yoshua Bengio joined in the late 1980s with the same conviction as the other two: the underlying math was right, even if nobody wanted to fund it.

Then in 1986, something small but transformative happened. Rumelhart, Hinton, and Williams published a Nature paper on how to train deep networks using backpropagation. The idea wasn't brand new — Paul Werbos had developed it years earlier — but now it was converging with better hardware and solid theoretical confidence. In hindsight, that was the founding act of modern deep learning.

The first winter didn't end because someone "discovered" that AI worked. It ended because there were modest but real results in specific domains (optical character recognition, basic image analysis) and because computing infrastructure kept improving, slowly but without pause.

The second winter: the empty promise of expert systems (1987-1993)

The 1980s were strange. AI was "coming back," but not as deep learning. It was the era of expert systems.

The idea was seductive. You take a human expert's knowledge — a doctor's, a lawyer's, an engineer's — and you encode it in rules a computer applies. A doctor diagnoses a disease when she sees symptoms X, Y, and Z. Write it down. Now the machine can diagnose too.

And it worked. It worked very well in demos. Corporations invested billions. The expert-systems market reached estimates of $3 billion by 1988, according to Daniel Crevier's book "AI: The Tumultuous History of the Search for Artificial Intelligence." It was sold as the revolution that had finally arrived.

But there was a structural problem. Expert systems were brittle. A system that worked on 80% of cases failed catastrophically on the remaining 20%. Building them cost millions. They needed human experts permanently on hand to keep the rules current. They didn't scale.

By 1987, companies caught on. The money vanished. Another winter. Another six years when AI was treated as impossible.

What was happening quietly while the world doubted

Here's the pattern almost nobody tells. Both winters happened while real technical progress was piling up without fanfare.

Between 1974 and 1980, when "official" AI was dead, Hinton and others were developing theory on recurrent networks. LeCun was refining backpropagation applied to images. Bengio, a bit later, would explore architectures that could handle language.

During the second winter (1987-1993), something even bigger was happening. The internet was starting to expand. There was real digital data at massive scale for the first time. Personal computers were getting faster. And GPUs — chips originally designed for video games — were starting to evolve in ways that would eventually accelerate neural networks by orders of magnitude. Nvidia, the company that would end up central to today's AI explosion, was founded in 1993. Right at the end of the second winter.

So what does this mean for you?

The lesson is direct. Watch the pattern, not the noise.

The noise is: "Does the world believe in AI?" "Is there speculative money?" "What did the latest CEO promise today?"

The pattern is: "Is serious technical work getting built?" "Are there measurable results, even small ones?" "Are people using tools that actually work?"

The winters didn't kill AI because real progress never depended on hype. It depended on researchers who knew they were right, even when nobody believed them.

One question to leave you with

Today, when you read headlines about "AI will change everything" or "AI is overhyped," you can ask yourself the same question this piece suggests by looking backward. Where is real work happening? What are the people not in the news building? If you want a concrete answer to why Anthropic thinks we're in a different moment from previous winters, read #0027 on the frontier model race.

A pattern the official story rarely names

The dominant historiography treats the two AI winters (1974-1980 and 1987-1993) as periods of technological failure. That characterization impoverishes the phenomenon. The winters were, more precisely, periods of financial and institutional contraction that occurred while fundamental technical progress continued without public visibility. Understanding that distinction isn't an academic exercise. It's the only way to read accurately the hype-and-correction cycles we still see today.

First winter (1973-1980): the gap between narrative promise and technical feasibility

The 1973 Lighthill Report was, technically, a political intervention executed in academic language. Its central argument rested on a useful distinction: the difference between "design complexity" — how complicated something is to build — and "combinatorial complexity" of the problem — how inherently hard the underlying problem is. The 1960s machines could play chess on bounded boards but failed at generalization to new open-structure problems.

Lighthill was right in his immediate diagnosis. The promises McCarthy, Minsky, Simon, and others had made in the late 1950s and through the 1960s — superintelligent machines in a decade or two — were mathematically untenable given the von Neumann architecture available and the datasets that did not yet exist. The error wasn't research direction. It was calendar.

The institutional response was severe. DARPA cut funding. The UK systematically defunded. The field entered what historians call the "first winter." But — and this is the point — the financial collapse wasn't total. It was redistributive. Small labs at the University of Toronto, ENS Paris, Bell Labs kept research going at subsistence levels.

What happened in those labs was technically transformative:

• Hinton developed recurrent network architectures with rigorous theoretical grounding.
• Rumelhart, Hinton, and Williams (1986, Nature 323) published the paper showing backpropagation could train multilayer networks practically, resolving the central theoretical problem that had blocked the field since Minsky and Papert's 1969 perceptron critique.
• LeCun integrated convolutions into architectures for processing images with spatial invariance.

These advances were NOT accelerated by public funding. They were sustained by theoretical confidence — a kind of faith in the underlying math — that 1960s speculative money had eroded and that now only survived in a handful of decentralized labs.

The winter ended around 1980-1982 not because someone "discovered" that AI worked, but through the convergence of three factors: demonstration of limited but real capabilities in specific domains (especially optical character recognition); improved computing infrastructure with minicomputers and then PCs; and corporate money — not just public — entering for practical applications.

Second winter (1987-1993): the illusion of generality in expert systems

The 1980-1987 period was a cautious comeback. Expert systems emerged as the first commercializable AI application. Their proposition: encode human expert decision rules into inference machines.

The model had narrow validity. It worked in highly specialized domains with clear boolean decisions and knowledge bases that change slowly. MYCIN diagnosed infectious diseases with accuracy comparable to human specialists. XCON, used at Digital Equipment Corporation starting in 1980, configured computer orders and saved the company an estimated $40 million per year according to internal DEC documentation. The expert-systems market reached around $3 billion by 1988, per Crevier (1993).

But the promise the industry sold — although rarely articulated explicitly — was that an architecture of encoded rules could scale toward open domains. It couldn't. Three systemic problems doomed it:

1. Lack of generalization. A system trained on disease-A diagnoses with 80% coverage failed catastrophically on the remaining 20% and entirely outside its training domain.

1. Combinatorial explosion in maintenance. Systems with thousands of rules developed unpredictable interactions. Debugging required constant human expertise.

1. Knowledge acquisition cost. Extracting rules from human experts (knowledge elicitation) is labor-intensive, expensive, and requires continuity. Any change to the knowledge base required full re-capture.

By 1986-1987, corporate adoption of expert systems stalled. The expected ROI didn't materialize. By 1987-1988, funding evaporated. Second winter.

The technical phenomenology of the winters: progress in the shadows

Both winters coincided with technical advances that received no public attention. During the first, Hopfield (1982) — although published at the end of the period — was developed within it. LeCun was refining convolutions. Judea Pearl was developing the Bayesian networks that would resolve the logical brittleness of early symbolic systems; his work was published as a book in 1988 and would win him the Turing Award in 2011. Exponential growth in digital data (email, the first relational databases) was creating the "fuel" that would later train deep networks.

During the second winter, Bengio, LeCun, and Hinton were developing language and vision architectures. Exponential internet growth in the 1990s began producing unprecedented corpora. Nvidia was founded in 1993 — the GPUs that would eventually accelerate the matrix computation of neural networks were beginning their evolution. Advances in Bayesian statistics and probabilistic graphical models provided the mathematical scaffolding the field needed.

The pattern is clear. Both winters were periods when speculative funding and hype deflated while fundamental technical infrastructure was getting stronger.

The mechanism of recovery: convergence of technical capability, data, and hardware

The winters ended when three factors converged. First, mature technical theory: backpropagation worked, Bayesian networks worked, convolutions worked. Second, real data available: the internet was providing text, image, and structured-data corpora at massive scale. Third, enough hardware: sequential computers had gotten fast and GPUs were starting to exist as a category.

Without speculative money, but with these three factors, the field came back in the early 1990s. Deep learning — though the term wouldn't catch on until 2006 — started showing concrete results in vision and language.

The blog's thesis on present and future cycles

Here's what this blog believes after carefully reviewing the historical phenomenology. Hype-and-correction cycles in AI aren't noise — they're structural. They happen when the narrative promise systematically outruns demonstrable technical capability. The correction is inevitable. But — and this is the insight that changes the reading — a speculative correction is not the same as technical failure. The winters happened while real technical progress was piling up. The narrative collapsed; the construction kept going.

The practical implication for someone using AI today is direct and useful. First, expect speculative corrections. They're inevitable when the promise outruns reality, as is happening in 2024-2026 with parts of the model market. Second, don't confuse a correction with a failure. When money walks away from overvalued projects, the infrastructure that actually works — models in Gmail, search, medical diagnostics, code assistants — will still be there and will keep deepening. Third, the work you do now to learn to use these tools pays off regardless of the cycle. Hinton worked for 30 years without recognition. When recognition arrived, the work was already done.

A full winter in the 1974 or 1987 style is unlikely today. The structural difference is that AI is already embedded in critical infrastructure — Gmail, Search, assisted medical diagnostics, productivity tools that bill in real revenue. That provides corporate funding independent of speculation, massive data to train on, hardware (TPUs, GPUs) designed specifically for this load, and institutions with long time horizons. A speculative correction is nearly certain. A full winter is improbable, because the infrastructure for technical progress is substantially more robust than in 1973 or 1987.

The question isn't whether there will be turbulence. The question is whether you'll know how to read it when it arrives.