The History of AI, Part 8 of 24. Part 7 told the story of the expert systems boom: AI's first real business success. This chapter is about the bust that followed, the second and longer AI winter, and about the researchers who used the cold years to rethink what intelligence actually requires.
In 1984, at the annual meeting of the American Association for Artificial Intelligence (AAAI), two of the field's most senior figures did something unusual. Instead of cheering the boom, they warned about the crash.
Marvin Minsky and Roger Schank, both veterans of the 1970s downturn, told a ballroom of researchers and business people that enthusiasm for AI had gotten out of control. The pattern, they said, was familiar: inflated expectations, a flood of money, and then a collapse of confidence that would bring the whole field down with it. They called it, with a nod to the Cold War nightmare of the day, an "AI winter."
The audience mostly shrugged. Money was pouring in, start-ups were going public, and Japan's Fifth Generation project had made governments around the world panic and open their wallets. Why worry?
Within three or four years, the warning proved right. This chapter tells the story of the second winter: how it happened, who it hurt, and, perhaps more interestingly, what survived and what was born in the cold.
The great crash of the Lisp machines
The first domino to fall was hardware.

The rear of a Symbolics 3600 Lisp machine, a casualty of the 1987 market collapse. Image: Retro-Computing Society of Rhode Island, CC BY-SA 3.0, via Wikimedia Commons.
Recall from Part 7 that the expert systems boom spawned an entire industry of specialized computers called Lisp machines, workstations designed to run Lisp efficiently. Companies like Symbolics and Lisp Machines Inc. (LMI), both born from the MIT AI Lab, sold them for around $70,000 to $100,000 apiece. They were beautiful machines, with large bitmap screens, powerful programming environments, and features that wouldn't appear on mainstream computers for years. And for a few years, they sold like hotcakes.
But the machines had a fatal weakness: they were an island in a rising sea. Meanwhile, general-purpose workstations from companies such as Sun Microsystems and Apollo, and soon desktop PCs and Apple Macintoshes, were getting cheaper and faster at breakneck speed, following the relentless curve of Moore's Law. By the mid-1980s, a Sun workstation costing a fraction of a Lisp machine could run good Lisp implementations nearly as fast. And the PC market had economies of scale that the niche Lisp machine vendors could never match.
The result was brutal. Around 1987, the market for specialized AI hardware collapsed almost overnight. Customers stopped buying. Lisp machine companies lost most of their revenue within a short period. LMI went bankrupt, Symbolics shrank dramatically, and the survivors were reduced to shadows of their former selves. (Symbolics eventually filed for bankruptcy protection in the early 1990s.) Texas Instruments and Xerox quietly wound down their AI workstation lines. Hundreds of millions of dollars of investment in specialized hardware evaporated.
The same logic that killed Lisp machines would kill many specialized AI chips and machines in later decades, until the GPU, a mass-market chip built for video games, became the engine of modern AI. We'll see that twist in Part 12. The lesson is the same each time: a specialized machine has to stay ahead of the commodity curve, and it almost never does.
The expert systems hangover
The hardware crash was dramatic, but the deeper trouble was in software.
Why the systems disappointed
In Part 7, we listed the structural weaknesses of expert systems: brittleness, the knowledge acquisition bottleneck, inability to learn, and brutal maintenance costs. By the late 1980s, thousands of companies were discovering those weaknesses first-hand.
A typical pattern went like this. A corporate AI group would build an impressive prototype in a few months, with a few hundred rules that handled the common cases beautifully. Management was thrilled. Then came the hard part: scaling to the full problem, with its exceptions and special cases. The rule base grew to thousands of rules, and the system became harder and harder to debug, because each new rule might interact in surprising ways with existing ones. Maintenance consumed more and more resources. Meanwhile, the real world kept changing, requiring constant rule updates by the same overstretched experts whose time was supposed to be saved.
There were also integration problems. Many expert systems were built in Lisp or special shells, on dedicated machines, which made it hard to connect them to the corporate databases and mainstream software where the actual data lived. Rewriting them in conventional languages like C was expensive, and the benefits no longer seemed so clear.
Meanwhile, the advantage that expert systems had over conventional programs was less clear than advertised. In many cases, the same problem could be solved by a well-designed ordinary program or a database query, without any special AI machinery. Over time, some of the systems that had been touted as AI turned out to be rule engines that ordinary programmers could write and maintain just fine.
The maintenance trap, in numbers
DEC's celebrated XCON system illustrates the trap. XCON had saved the company millions, but by the late 1980s its rule base had swelled to roughly ten thousand rules and required an ongoing team of specialists to maintain. As DEC's product line changed, rules had to be rewritten at a fast pace, and each change risked breaking something. The more successful a system had been, the more expensive it became to keep alive. When DEC itself ran into trouble in the 1990s (it was eventually bought by Compaq in 1998), the system's glory days were long over.
The vendors fall
The companies that had sold expert system shells and consulting found demand drying up. Teknowledge, IntelliCorp, and Inference Corporation, among others, saw their revenue stall, and some had to retrench or pivot. Many of the start-ups founded during the boom never made it to profitability. Some companies quietly rebranded their products, stripping out the phrase "artificial intelligence" from their marketing, because it had become a liability. By the end of the decade, the label that had been a magnet for investors had become something to avoid.
DARPA changes its mind
Government funding followed commercial sentiment. In the 1980s, DARPA's Strategic Computing Initiative (SCI), launched in 1983 in response to the Japanese challenge, had committed on the order of a billion dollars over a decade to developing machine intelligence for military purposes. Its three flagship applications were ambitious: an Autonomous Land Vehicle that could drive itself across rough terrain, a Pilot's Associate that would help fighter pilots, and a Battle Management system for naval command.
The program produced real technical advances, but its goals were far too ambitious for the technology of the day. A vision system that could navigate rough terrain at useful speeds simply wasn't possible with the computing power and algorithms available. DARPA's leadership changed in the late 1980s, and the new leadership, including the mathematician Jack Schwartz, who ran the information technology office from 1987, was openly skeptical of AI, reportedly regarding much of it as little more than "clever programming." Funding for general AI research was cut deeply, with money redirected toward more tractable areas such as supercomputing and networking.
This echoed the first winter but with a different twist. The first time, researchers had been punished for failing to deliver on grand promises. This time they were punished for succeeding in a narrow way that wasn't the grand thing the sponsors had been sold.
Japan's Fifth Generation: the dog that didn't bark
Remember the Fifth Generation Computer Systems project? In Part 7, we saw how Japan's 1982 announcement set off a panic in the West and a global arms race in funding. A decade later, the story ended with a whimper.
The project, run by a dedicated institute called ICOT, set out to build a new class of computers based on logic programming (specifically Prolog and a concurrent variant) and massively parallel inference. The dream was a machine that could reason over huge knowledge bases, understand natural language, and interact with people in a human-like way.
By 1992, ICOT had spent on the order of 50 billion yen (several hundred million dollars). It produced real engineering: specialized parallel inference machines, new parallel logic programming languages, and a cohort of well-trained researchers. But it failed to achieve its headline goals, and it had no commercial impact. A few things went wrong:
- Wrong bet on logic. The project wagered that logic programming would be the foundation of intelligent computing. In practice, mainstream computing went in a different direction, and logic programming remained a niche.
- Commodity hardware won. By the time the specialized machines were ready, ordinary microprocessors had become so fast and cheap that special-purpose designs couldn't compete. (The same lesson as the Lisp machines.)
- The software never arrived. The hardest problems, such as natural language understanding and flexible reasoning, were not hardware problems. They needed breakthroughs in knowledge and learning that weren't forthcoming.
- A moving target. The 1982 vision was of a world in which computing was centralized and knowledge-intensive. By 1992, the world was about personal computers, graphical interfaces, and networks.
When the Fifth Generation project ended with little to show, the Western competitors who had scrambled to match it quietly followed suit. The British Alvey Programme had already wound down, and the US Microelectronics and Computer Technology Corporation (MCC), while it lasted into the late 1990s, produced limited commercial results. The great AI arms race fizzled out, and the lesson for funders was a painful one: a national plan can't legislate a research breakthrough.
It's worth being fair to ICOT. Its alumni went on to important work, and the project helped build computing research capacity in Japan. But as a bet on AI, it lost.
What "winter" felt like
Between roughly 1987 and the mid-1990s, being an AI researcher was uncomfortable.
- Funding dried up in both the public and private sectors.
- Conferences shrank, and attendance at the big AI meetings dropped sharply.
- Brand toxicity. The phrase "artificial intelligence" became an obstacle to getting grants and publishing in some venues. Researchers described their work as machine learning, knowledge-based systems, intelligent systems, computational intelligence, adaptive systems, pattern recognition, analytics, or informatics. As the journalist John Markoff later wrote in a 2005 New York Times article, many computer scientists and software engineers avoided the term "artificial intelligence" for fear of being viewed as wild-eyed dreamers.
- Jobs shifted. Students who might have gone into AI went into databases, graphics, networking, or the booming Internet.
- Critics gloated. Skeptics like Dreyfus felt vindicated, and some commentators declared AI a failed project.
And yet, as in the first winter, this was a period of intense intellectual ferment, and the work that mattered was happening in directions that the boom-time leaders had ignored.
The robot that wasn't afraid of the real world
One of the most important responses came from an Australian-born roboticist at MIT named Rodney Brooks.
By the mid-1980s, Brooks had become deeply frustrated with the way AI approached robotics. The classical approach, the one used by Shakey in Part 5, was the sense-model-plan-act cycle: the robot would take sensor readings, build a detailed symbolic model of the world, use logic to plan a course of action, and then execute. It was slow, brittle, and a poor match for the messy real world. Shakey took minutes to decide on a move.

The original Roomba, a descendant of Rodney Brooks's behavior-based robotics. Image: Larry D. Moore, CC BY 4.0, via Wikimedia Commons.
Brooks proposed a radical alternative, laid out in a famous 1986 paper and sharpened in provocatively titled essays: "Elephants Don't Play Chess" (1990) and "Intelligence Without Representation" (1991). His argument was:
- Evolution spent billions of years building insect-level perception and action, and only a short time adding symbolic reasoning on top. So perception and action in the real world are the hard part, and they shouldn't be treated as an afterthought.
- A creature doesn't need a detailed internal model of the world; as Brooks put it, "the world is its own best model." Just look at the world, again and again.
- Intelligent behavior can emerge from the interaction of many simple behaviors.
Subsumption architecture
Brooks's design, the subsumption architecture, organizes a robot's control into layers of simple behaviors, each connecting sensors directly to actions, with higher layers able to override or "subsume" lower ones. The bottom layer might be "avoid obstacles." A layer above it might be "wander around." Above that, "head toward a goal if you can sense it." Each layer is simple and runs all the time. There is no central planner and no world model.
You can capture the idea in a few lines. Here's a toy robot on a grid with three layers:
import random
random.seed(3)
GRID = [
"..........",
"..##......",
"..##...G..",
"..........",
"......##..",
"..........",
]
H, W = len(GRID), len(GRID[0])
DIRS = [(0, 1), (1, 0), (0, -1), (-1, 0)] # E, S, W, N
def free(r, c):
return 0 <= r < H and 0 <= c < W and GRID[r][c] != "#"
def avoid(pos, heading): # layer 0: never walk into walls
r, c = pos; dr, dc = DIRS[heading]
if not free(r + dr, c + dc):
return (heading + random.choice([1, 3])) % 4 # turn left or right
return None
def wander(pos, heading): # layer 1: drift randomly
return heading if random.random() < 0.8 else random.randrange(4)
def seek_goal(pos, heading, sense=4): # layer 2: if the goal is nearby, head for it
r, c = pos
for gr in range(H):
for gc in range(W):
if GRID[gr][gc] == "G" and abs(gr - r) + abs(gc - c) <= sense:
if abs(gc - c) >= abs(gr - r):
return 0 if gc > c else 2
return 1 if gr > r else 3
return None
pos, heading = (0, 0), 0
for step in range(200):
# Higher layers subsume lower ones; the first layer with an opinion wins,
# but obstacle avoidance is always allowed to veto a bad move.
wanted = seek_goal(pos, heading)
if wanted is None:
wanted = wander(pos, heading)
dr, dc = DIRS[wanted]
safe = free(pos[0] + dr, pos[1] + dc)
heading = wanted if safe else avoid(pos, wanted)
dr, dc = DIRS[heading]
if free(pos[0] + dr, pos[1] + dc):
pos = (pos[0] + dr, pos[1] + dc)
if GRID[pos[0]][pos[1]] == "G":
print(f"reached the goal in {step + 1} steps")
break
else:
print("still wandering...")
There's no map and no plan. The robot reaches the goal because a simple "seek" behavior takes over when the goal is near, while "avoid" and "wander" keep it moving and safe the rest of the time. Brooks's real robots, six-legged insect-like machines such as Genghis (1989), used the same principle, and they could clamber over rough terrain far more robustly than the elaborate planners of the previous generation.
Why it mattered
Brooks's critics said his approach would never scale to human-level reasoning, and he himself acknowledged that his robots were closer to insects than to people. But his influence was enormous. In 1990 he co-founded a company called iRobot, which in 2002 introduced the Roomba, a vacuum cleaner whose navigation owes a lot to behavior-based ideas, and which became one of the first mass-market robots. Behavior-based and reactive methods also shaped later work on video game characters, drone control, and swarm robotics.
More philosophically, Brooks revived a thread from Part 3, W. Grey Walter's tortoises and the cybernetics tradition, arguing that intelligence is rooted in embodiment, interaction, and the environment, not just symbol manipulation. The idea of "embodied cognition" is now a mainstream view in cognitive science and robotics.
The first self-driving cars
Another quiet breakthrough in the winter years was autonomous driving.

Carnegie Mellon's Navlab self-driving vehicles. Image: Firefly4342, CC BY-SA 4.0, via Wikimedia Commons.
In Germany, the engineer Ernst Dickmanns at the Bundeswehr University Munich equipped a Mercedes van, the VaMoRs, with cameras and processors and, in 1987, demonstrated it driving autonomously at around 90 km/h on an empty stretch of highway. His team's later cars, part of the European Prometheus project, drove thousands of kilometers on public roads in the mid-1990s, including a famous 1994 demonstration near Paris in heavy traffic. Dickmanns used careful model-based vision and control, a very engineering-minded approach.
In the United States, a team at Carnegie Mellon University pursued a different route. Dean Pomerleau's system ALVINN (Autonomous Land Vehicle In a Neural Network), built in 1989, used a small neural network, with 30 by 32 input pixels, a few hidden units, and 30 output units, to learn to steer a vehicle by watching a human driver. It was trained on camera images paired with the human's steering angle, a pure case of learning from demonstration. Later, in 1995, a CMU vehicle completed a trip called "No Hands Across America," driving about 2,800 miles from Pittsburgh to San Diego with the steering controlled by a computer (the human handled the throttle and brakes) for about 98 percent of the distance.
ALVINN is a fascinating relic. At the exact moment mainstream AI was dismissing neural networks, one worked well enough to steer a car. It didn't scale to all driving conditions, but it hinted at what was coming.
The philosophical thread: the Chinese Room
While engineers were adjusting to the downturn, philosophers were pushing a deeper challenge. In 1980, in the middle of the expert systems boom, the Berkeley philosopher John Searle published a paper titled "Minds, Brains, and Programs," containing the most famous thought experiment in the philosophy of AI: the Chinese Room.
Imagine, Searle said, a person who knows no Chinese locked inside a room. Through a slot, people pass in cards covered in Chinese characters. The person has a huge rulebook, written in English, that tells them which characters to pass back out in response to each input pattern. From the outside, the room produces perfect, fluent Chinese responses. A native speaker would conclude that whoever is inside understands Chinese. Yet the person inside understands nothing. They are just manipulating symbols according to rules.
Searle's conclusion: running a program, however sophisticated, can never by itself produce genuine understanding. Syntax (symbol manipulation) isn't sufficient for semantics (meaning). He distinguished "strong AI", the claim that a suitably programmed computer literally has a mind, from "weak AI", the claim that computers are useful tools for studying and simulating minds. Searle accepted weak AI but rejected strong AI.
The argument provoked a flood of replies, and the debate is still lively:
- The Systems Reply: the person doesn't understand Chinese, but the whole system (person plus rulebook plus room) does.
- The Robot Reply: put the system in a robot body that interacts with the world, and the symbols will be grounded in experience.
- The Brain Simulator Reply: what if the program simulates, neuron by neuron, the brain of a native speaker?
Searle rejected them all. The Chinese Room connects to everything we've covered: Descartes's doubts in Part 1, the Turing test in Part 2, and the ELIZA effect in Part 5. It is more relevant than ever in the era of large language models: does a system that produces fluent text "understand" it? Whatever you think, the argument sharpens the question. And it shows that the winter wasn't only a time of failure: serious thinking about what AI means was deepening.
What the winter taught
Looking back, what actually went wrong in the second winter, and what can we take from it?
- Hand-coded knowledge doesn't scale. The knowledge acquisition bottleneck was real. Experts can't always articulate what they know, and writing rules by hand is slow and brittle. The path forward would have to involve machines learning from examples and data.
- Specialized hardware is a risky bet. The Lisp machines and the Fifth Generation project both lost to the relentless economics of commodity computing.
- Narrow success isn't general intelligence. Expert systems proved something valuable but limited, and the market confused the two.
- Uncertainty needs principled treatment. Ad hoc certainty factors weren't enough; the field needed probability and statistics.
- The real world matters. Brooks's insight, that perception and action in an unpredictable environment are at the core of intelligence, forced a rethink of what AI was for.
- Expectations management is part of the job. Both winters were, in large measure, failures of communication: promises made and not kept.
Quiet progress under another name
By the early 1990s, the winter was thawing, though few people used that word. The field didn't so much recover as mutate. AI researchers who had been hurt by the bust adopted a different style, one that would eventually prove far more successful:
- Be rigorous. Use proper mathematics, especially probability and statistics, rather than ad hoc rules.
- Be empirical. Test systems on standard benchmarks and publish measurable results.
- Be modest. Solve specific problems, such as speech recognition, spam filtering, and handwriting recognition, rather than promising general intelligence.
- Learn from data. Let machines extract patterns from examples rather than relying on hand-written knowledge.
- Avoid the brand. Call it machine learning, data mining, pattern recognition, or intelligent agents.
In a sense, the second winter finished what the first had started: it drove AI toward a more scientific, more statistical, more data-driven approach. The coming chapters tell how that approach, along with the reborn neural networks, built the foundations of the modern boom.
The next chapter starts with the comeback of the idea that had been left for dead: neural networks.
Timeline: 1984 to the mid-1990s
- 1980: XCON enters production at DEC; Searle's Chinese Room paper.
- 1982: Japan announces the Fifth Generation project.
- 1983: DARPA's Strategic Computing Initiative begins; Britain's Alvey Programme launches.
- 1984: Minsky and Schank warn of an "AI winter"; Cyc begins.
- 1986: Brooks publishes his layered control architecture; backpropagation goes mainstream (Part 9).
- 1987: Lisp machine market collapses; Dickmanns's van drives at highway speed.
- 1989: ALVINN learns to steer; Brooks's walking robot Genghis.
- 1990: Brooks's "Elephants Don't Play Chess"; iRobot is founded.
- 1992: Fifth Generation project ends.
- 1995: CMU's "No Hands Across America."
What kept working through the cold
The winter's most useful lesson is how much did work. Here is a short inventory of what survived and thrived while "AI" was out of fashion:
- Speech recognition improved steadily with hidden Markov models (Part 10), reaching commercial dictation by 1997.
- Optical character recognition became routine; neural networks read bank checks (Part 9).
- Planning and scheduling systems ran airline operations and even NASA missions. In 1999, NASA's Remote Agent software took control of the Deep Space 1 spacecraft for several days in a flight experiment.
- Search and constraint solving matured into industrial tools.
- Robotics found success in vacuum cleaners, warehouses, and factories using simple behavior-based methods.
- Machine learning grew into its own respectable discipline with its own conferences and journals.
None of this was marketed as artificial intelligence. It was marketed as software, which is to say, it worked.
Two misconceptions about the winters
"During the winters, nothing happened." In fact, the mathematical foundations of modern learning were built then: the theory of generalization, probabilistic graphical models, kernel methods, and the first practical convolutional networks. The winters hurt careers and funding, but the work continued under different names.
"The winters were caused by bad ideas." Mostly they were caused by good ideas oversold. Expert systems really did save money in narrow domains, and symbolic reasoning is still useful. What failed was the promise that they would scale to general intelligence on the timetable investors had been given.
Quick glossary
- Lisp machine: a computer built to run the Lisp language efficiently.
- Subsumption architecture: Brooks's design of layered behaviors in which higher layers can override lower ones.
- Embodiment: the idea that intelligence arises from a body interacting with an environment.
- Strong AI / weak AI: Searle's terms for the claims that a program literally has a mind versus merely simulates one.
- Fifth Generation project: Japan's 1982 to 1992 national program for logic-based, parallel computing.
Key takeaways
- The second AI winter (roughly 1987 to the mid-1990s) followed the expert systems boom and was triggered by a collapse in the market for Lisp machines, disappointing results from expert systems, DARPA's cuts, and the failure of Japan's Fifth Generation project.
- Expert systems proved too brittle and too costly to maintain at scale, and vendors had oversold their capabilities.
- Researchers rebranded their work to avoid the toxic label "AI," using terms like machine learning, informatics, and intelligent systems.
- Rodney Brooks's subsumption architecture and behavior-based robotics offered an alternative to symbolic planning and led to practical robots such as the Roomba.
- Early self-driving cars (Dickmanns's Mercedes vans and CMU's neural-network-driven ALVINN) showed what was possible.
- Searle's Chinese Room deepened the philosophical debate over whether symbol manipulation can ever amount to understanding.
- The winter pushed AI toward probability, statistics, and learning from data.
Try it yourself
- Add a fourth layer. Extend the subsumption robot with a "return to base when tired" behavior that has a low priority at first and takes over after a fixed number of steps. Notice how new behavior emerges without any central planner.
-
Break the robot. Remove the
avoidlayer and see how often the robot gets stuck. Which layer is the most important? What does this tell you about robust design? - Debate the room. Write a one-paragraph reply to the Chinese Room from the point of view of someone who uses a modern chatbot daily. Where does the analogy hold, and where does it break?
Next up
In Part 9, the underdogs return. In the 1980s, a small band of researchers, Geoffrey Hinton, David Rumelhart, Yann LeCun, and others, found a practical way to train multi-layer neural networks, solving the very problem that had doomed the perceptron. The backpropagation revolution is about to begin, and with it, the road to deep learning.
If you've ever worked on a project that was hot one year and unfashionable the next, share the story in the comments. And follow the series so you get Part 9 as soon as it's out.
See you in Part 9.
☕ If this series helped you, you can support my writing on Ko-fi.

Top comments (0)