The Second Winter | James Howard

The year is 1987. Symbolics, Inc. is the leading manufacturer of Lisp machines, single-user workstations purpose-built to run Lisp, the language most closely associated with artificial intelligence. Their hardware is beautiful, powerful, and expensive: a fully configured workstation runs about $110,000. Symbolics even holds a modest distinction in internet history. On March 15, 1985, symbolics.com became the first .com domain ever registered.¹

By 1988, Symbolics is in freefall. The Lisp machine market, which looked like the future of computing in 1984, is evaporating. The venture capital has gone cold. The research labs that once lined up for specialized AI hardware are discovering that ordinary workstations have gotten fast enough to do most of what they need.

This is the second AI winter, compressed into one company’s trajectory: not a failure of intelligence but a failure of timing, economics, and the oldest mistake in technology forecasting, which is assuming that what the field needs today will always require specialized infrastructure.

The Debt Comes Due

The first winter began to thaw in the late 1970s not because the field had solved its fundamental problems but because it found a narrower problem it could actually solve. Expert systems, covered in the previous episode, demonstrated that you could extract genuine value from AI by limiting the domain aggressively. MYCIN diagnosed bacterial infections. DENDRAL identified organic compounds. XCON configured VAX computer orders. These were not general intelligence. They were very good rule books, and the field sold them as something considerably closer to the former.

That gap between what expert systems were and what they were marketed as created the conditions for the second winter. Through the early 1980s, the AI industry attracted serious investment and serious attention. Dedicated AI hardware manufacturers sold specialized machines at prices only research labs and large enterprises could afford. Consulting firms built AI divisions. Lisp got its own chips. Defense dollars flowed. Japan announced a national initiative.

The promises, as usual, outran the delivery.

The Fifth Generation

The Fifth Generation Computer Project shows what happens when bureaucratic ambition meets technical optimism.

In 1982, Japan’s Ministry of International Trade and Industry announced a ten-year national program, funded at several hundred million dollars and often reported at approximately $500 million, to build computers capable of reasoning in natural language, drawing inferences, and processing knowledge the way human experts do. The Institute for New Generation Computer Technology, known as ICOT, directed by Kazuhiro Fuchi, would lead the effort. The goal was not faster computation of the ordinary kind. The goal was qualitatively different computation: machines that understood.²

The technical foundation was logic programming, specifically Prolog. Prolog is a declarative language: you state facts and rules, and the system derives conclusions. It has genuine elegance, and for certain classes of problems involving structured reasoning over explicit knowledge, it works well. The ICOT researchers were not wrong to find it attractive.

They were wrong in a deeper assumption: that intelligence is primarily a matter of logical inference over explicit facts. That assumption, which was also the assumption underlying expert systems, turned out to be one of the most consequential wrong turns in the history of the field. Much of human knowledge is not explicit. It is contextual, embodied, and heavily statistical. You cannot write down what makes a face recognizable or a sentence grammatical in a way that a Prolog interpreter can usefully process. The knowledge acquisition bottleneck that killed expert systems was not a problem the Fifth Generation project was going to solve with better hardware. It was a structural critique of the whole approach.

The project ran to its scheduled conclusion in 1992. By technical measures, it accomplished many stated goals: ICOT researchers produced novel parallel logic programming systems, published extensively, and demonstrated inference machines running at speeds that would have seemed impossible in 1982. By practical measures, it produced nothing the world outside Japan chose to use. The parallel Prolog machines sat in laboratories. No mass-market product emerged from Fifth Generation technology. The reasoning computers that eventually arrived took a completely different path.

The American response to the announcement was the kind of reaction that produces bad decisions at speed. DARPA launched the Strategic Computing Initiative in 1983, spending roughly $1 billion over a decade on AI and high-performance computing research. Some of that funding did useful work. Some of it funded the same overconfident systems integration that the market was about to reject. Both programs, American and Japanese and for all their differences, were organized around the same underlying assumption: that the hard part of intelligence was symbolic reasoning, and that with enough engineering talent and national will, symbolic reasoning could be made to work at scale.

The market’s verdict on that assumption is the second winter.

The Hardware Bet

The Lisp machine market is worth examining on its own terms because it illustrates a principle about specialized infrastructure that recurs throughout computing history.

By 1980, the consensus in the AI research community was that standard hardware was inadequate for serious AI work. Lisp, the field’s primary language, was interpreted rather than compiled, garbage-collected, dynamically typed, and interactive in ways that the von Neumann architectures of the era handled badly. Researchers spent substantial time waiting. The obvious solution was to build hardware designed specifically for Lisp.

MIT’s AI Lab incubated two companies that came to define the market. Symbolics, spun out in 1980, produced the most sophisticated Lisp machines available. Lisp Machines Inc. competed at slightly lower price points. Texas Instruments offered the Explorer. Xerox had its own Interlisp workstations. At their peak, these machines represented the state of the art: fast, capable, running development environments that remain impressive by any measure.

Then Sun Microsystems started selling workstations.

In 1982, a Symbolics 3600 cost roughly $70,000 to$ 110,000 depending on configuration. A Sun-1 cost about $10,000. The Sun was slower for Lisp work. By the mid-1980s, the gap was narrowing. By 1987, commodity hardware running Franz Lisp or Common Lisp was competitive with dedicated Lisp machines for many tasks, at a fraction of the price.³

When the price-performance gap closed, the rationale for specialized AI hardware evaporated. A research lab could buy ten Sun workstations for the price of one Symbolics machine. The Symbolics ran a better development environment. The ten Suns ran more experiments. The choice was not difficult.

Symbolics never recovered. The company contracted sharply, filed for Chapter 11 bankruptcy protection in 1993, and its assets and name were purchased by a new private group in 1995. LMI collapsed earlier. The Texas Instruments Explorer was discontinued. The Lisp machine as a product category was gone by the early 1990s, having briefly flourished at exactly the wrong moment to sustain itself.

The Pattern Repeats

There is a pattern this series has been tracking since the episode on the Dartmouth Conference. A new capability generates excitement. The excitement attracts promises. The promises attract investment. The investment funds work that demonstrates real but limited progress. The gap between demonstrated progress and promised breakthrough becomes impossible to ignore. Funding contracts. The field goes into winter.

The second winter followed this pattern faithfully. What made it different was the scale of what collapsed and the degree to which the promises had been made in commercial rather than merely academic terms. The first winter embarrassed researchers and disappointed government program managers. The second winter destroyed companies, burned venture capital, and left a generation of practitioners with reputational damage that took years to repair.⁴

The second winter also discredited a specific approach rather than the field as a whole. The first winter had cast doubt on AI in general. The second cast doubt on symbolic AI in particular. Expert systems failed not because they did nothing useful but because the claims made on their behalf exceeded what they could deliver, and because their fundamental architecture had hit a ceiling that more engineering could not raise. Rule-based reasoning over structured knowledge is powerful within its domain. Its domain turned out to be narrower than advertised.

This is the important distinction. The second winter was not the field discovering that intelligence was hard. The field already knew that. The second winter was the field discovering that one particular approach to intelligence had reached its limit, and that the limit was not going to move by adding more rules or more hardware.

Meanwhile, the future was being built somewhere else.

The Quiet Work

In November 1982, while the expert systems boom was approaching its peak, John Hopfield published a paper in the Proceedings of the National Academy of Sciences titled “Neural networks and physical systems with emergent collective computational abilities.”⁵ It is not a paper about AI in any of the usual senses. It is a paper about physics. Hopfield, trained as a physicist, showed that networks of simple binary units, connected in specific ways, would settle into stable states corresponding to stored patterns. The network had memory. The memory was distributed across the connections rather than residing in any particular location.

Hopfield networks were not directly useful for most practical AI tasks of the early 1980s. What they did was reinvigorate a research tradition that had been largely dormant since Minsky and Papert’s critique of perceptrons in 1969. Connectionist computing was worth taking seriously again.

Four years later, in October 1986, David Rumelhart, Geoffrey Hinton, and Ronald Williams published “Learning representations by back-propagating errors” in Nature.⁶ This is one of the most consequential machine learning papers ever written. Backpropagation is an algorithm for training multi-layer neural networks by propagating error signals backward through the network to adjust connection weights. The mathematics had predecessors: Paul Werbos had described the technique in his 1974 doctoral dissertation at Harvard, and David Parker had rediscovered it in 1985. But Rumelhart, Hinton, and Williams put it in front of the cognitive science community in a form that was clear, principled, and reproducible. They also demonstrated that it worked.

The timing is worth pausing over. Rumelhart, Hinton, and Williams published as the expert systems market was still near its peak. Backpropagation did not immediately displace anything. The neural network research community was small and unfashionable. Venture capital was not interested. Nobody at MITI was pivoting the Fifth Generation project in response to a paper in a science journal.

But the work was being done.

At Bell Labs, Yann LeCun was extending backpropagation-trained networks to spatial data, eventually building convolutional networks that could read handwritten digits with high accuracy. By the 1990s, systems descended from this work were being used commercially to read handwritten digits on bank checks.⁷ This deployment does not appear in the AI press because it does not look like AI. It looks like a bank’s infrastructure. But it was neural network research, applied to a real problem, producing measurable value at scale. The distinction between “AI” and “software that works” was already asserting itself, quietly, in the back offices of financial institutions.

Hinton, meanwhile, continued developing the theoretical foundations of what would eventually become deep learning. The Boltzmann machine, developed with Terry Sejnowski in the mid-1980s, was an early generative model based on stochastic neural networks. The work was mathematically sophisticated and computationally demanding in ways that the hardware of the era could barely support. It did not produce useful applications during the winter. It produced trained researchers and theoretical frameworks that would prove essential twenty years later.

The Sorting Mechanism

One way to understand an AI winter is as a failure. The funding stopped, the companies collapsed, the researchers dispersed. That reading is accurate as far as it goes.

Another way to understand it is as a selection event. When money is available, everything gets funded: the good ideas, the mediocre ideas, and the ideas that sound plausible in a pitch deck but cannot survive contact with real data. When money dries up, researchers have to choose more carefully. The problems they pursue have to be tractable. The methods they use have to produce results. The claims they make have to be defensible to skeptical colleagues rather than credulous investors.

The second winter was hard on the people who lived through it. Researchers with “artificial intelligence” on their curriculum vitae rebranded as “machine learning,” “statistical inference,” or “computational statistics,” anything that did not carry the baggage of two failed boom-and-bust cycles. The name had become toxic. The work continued under other names.⁸

What survived was substantially different from what had entered the cold. Symbolic AI did not disappear, but it was no longer the dominant paradigm. Statistical methods were gaining ground: decision trees, Bayesian networks, early support vector machine research. The neural network tradition, marginalized for most of the 1970s and all of the early 1980s, had a small and serious research community making real progress on solid mathematical foundations.

The field that emerged from the second winter would look quite different from the field that had gone in. The next chapter belongs to the researchers who stayed, the methods that survived, and the gradual convergence of statistical learning and connectionist computing into something powerful enough to transform how the world processes information. The survivors of the second winter are the protagonists of that story. The winter sorted them out.

This time, they would not disappear.

The second .com registration, on April 24, 1985, went to bbn.com, registered by BBN Technologies, the firm whose engineers had built the original ARPANET routers and whose Ray Tomlinson had given the world the @ symbol in email. Neither organization’s core work is as well-remembered as the naming convention they helped populate. ↩
The project documents described machines that would perform “inference,” “knowledge base management,” and “intelligent man-machine interface” as first-class computing primitives. These were not aspirations. They were specifications. The ambition was real, and the engineering talent assembled to pursue it was real. The wrong bet was on the underlying approach. ↩
This pattern is not unique to AI. Specialized database machines, graphics workstations, and vector supercomputers all faced the same fate: commodity hardware eventually catches up, and when it does, the economics of scale make the specialized machine impossible to justify. The Lisp machine companies were not unlucky. They were in a race they were always going to lose once commodity hardware crossed a certain threshold. ↩
The term “AI winter” reportedly entered circulation around 1984 at a private meeting of researchers who were beginning to see the funding climate tighten. The coinage captures the field’s own awareness that a familiar cycle was repeating. The self-awareness did not slow it down. ↩
Hopfield, J.J. (1982). Neural networks and physical systems with emergent collective computational abilities. Proceedings of the National Academy of Sciences, 79(8), 2554-2558. The paper brought serious physicists into a field that had been dominated by computer scientists and cognitive psychologists. Hopfield networks are now understood as early models of associative memory, and later work has drawn connections between modern Hopfield-style formulations and the attention mechanisms in contemporary neural networks. Hopfield shared the 2024 Nobel Prize in Physics with Geoffrey Hinton for this work and its consequences. ↩
Rumelhart, D.E., Hinton, G.E., and Williams, R.J. (1986). Learning representations by back-propagating errors. Nature, 323, 533-536. The paper appeared the same year as the two-volume Parallel Distributed Processing, edited by Rumelhart and McClelland, which made the full case for connectionist models of cognition. Together, the publications constituted something close to a manifesto for a new research program, arriving at the worst possible moment commercially and the best possible moment intellectually. ↩
LeCun, Y., Bottou, L., Bengio, Y., and Haffner, P. (1998). Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11), 2278-2324. The commercial check-reading application predated this paper, making it one of the earliest large-scale deployments of a deep neural network, even if it was not described in those terms at the time. The field had not yet developed the vocabulary to recognize what it had built. ↩
This rebranding was not cynical. It was accurate. “Machine learning” describes what the surviving methods were actually doing more precisely than “artificial intelligence” ever did. The name change was the field acknowledging, in public, what it had learned about itself in private. ↩