TL;DR: AI history is a hundred-year cycle of hype booms and funding winters, and the current wave is the field’s third boom: founded in 1956, frozen out in the 1970s and again from 1987, revived after 2012 on data, GPU compute and backpropagation.
- The groundwork predates the name by decades: a 1923 London play put robot into English, a 1943 paper modeled the first artificial neuron, and Turing proposed his test in 1950.
- Both winters were funding collapses, roughly 1974 to 1980 and from 1987 into the 1990s; when the second began, a half-billion-dollar industry selling dedicated Lisp machine hardware was gone within a year.
- The 1980s boom genuinely paid: XCON saved Digital Equipment about 25 million dollars a year, and Japan’s Fifth Generation project spent about 320 million dollars, not the 850 so often repeated.
- The revival after 2012 ran on ingredients that already existed: ImageNet’s data, GPU hardware, and backpropagation from the mid-1980s.
- Five widely repeated claims, from Lisp’s priority to Kismet’s vintage, do not survive the sources cited below.
Every wave of AI coverage reads as if AI history began in 2012. The episode this article accompanies opens by crossing brand new off that idea: the field has been through two earlier booms, and each time the promises came due, the money dried up instead of doubling. Watch the full episode on YouTube, or read on: this walkthrough follows the same boards, with every date and number checked against the cited sources.
Board 1 of 18: two booms already ended in collapsed funding; the third is the one you are living in.
A century of groundwork before the name
The word that anchors the whole story arrived first. Karel Čapek’s play R.U.R. opened in London in 1923, the timeline of artificial intelligence records it as the first use of robot in English, taken from robota, Czech for forced labor. The mathematics arrived in 1943, when Warren McCulloch and Walter Pitts published the earliest mathematical treatment of artificial neural networks, the ancestor of every network trained today.
Alan Turing turned the question into a test. His 1950 paper Computing Machinery and Intelligence proposed the imitation game, asking whether written conversation coming from a machine could pass for a person’s. Arthur Samuel’s checkers program, among the world’s first self-learning programs, was running by 1952 and improving by 1955, a machine getting better without being told how.
Then the field took a name. The 1956 Dartmouth workshop gave the field its name and agenda, a story covered in the first lesson of this series. Everything below is what happened to that program once the money arrived.
Two languages and one convincing illusion
The field’s first two decades ran on tools as much as ideas. John McCarthy began Lisp in 1958 at MIT, and it became the favored programming language for AI research, the language the boom later built computers to run. The other tradition arrived from Marseille: Alain Colmerauer’s Prolog in 1972, encoding rules and logic rather than learned patterns.
Between them came the first warning about what demos prove. Joseph Weizenbaum’s ELIZA, built in 1965, was a pattern matcher that produced an illusion of understanding; people confided in it anyway. Every hype cycle since has had its own ELIZA moment.
No, and the correction matters because the first language built for the work is almost never mentioned. Information Processing Language, IPL, was created around 1956 by Allen Newell, Herbert Simon and Cliff Shaw for the Logic Theorist, with features added specifically to facilitate AI programming: lists as the core data structure and operations to manipulate them. The same team’s Logic Theorist, the program IPL was built to run, is one of the field’s founding artifacts. McCarthy knew that work and took the list-processing idea further when he defined Lisp in 1958. Lisp won on elegance and speed of iteration, displaced IPL, and stayed the field’s default for three decades. So the accurate sentence is that Lisp was the favored AI language, while IPL was the first language designed for AI. Lisp machines, the dedicated hardware of the 1980s boom, were computers built to run one language; when that market died in 1987, it took the language’s commercial ecosystem with it.Was Lisp really the first AI programming language?
The first winter, 1974 to 1980: the money froze
The first AI winter was, at its core, a budget collapse: the government funders who had supported exploratory AI research with no strings attached pulled back. The AI winter record dates it approximately 1974 to 1980. Three documented triggers did the damage.
- In 1969 Marvin Minsky and Seymour Papert’s book Perceptrons emphasized the limits of single-layer neural networks; funding for neural approaches became hard to find through the 1970s and early 1980s.
- In 1973 the Lighthill report in the UK attacked the field’s grandiose objectives, in the report’s own phrase, and British university research was dismantled.
- In 1974 DARPA cancelled a three million dollar a year speech understanding contract at Carnegie Mellon. The open-ended grant era was over.
Board 7 of 18: the first winter’s three triggers. Together they froze the field for six years.
One historian argues no, and the objection is worth hearing. Thomas Haigh points out that SIGART membership nearly tripled between 1973 and 1978; if the field froze, its professional society missed it. The cuts landed unevenly: the big labs that had grown on open-ended military funding lost the most, while AI research continued at a smaller scale in many places. The aftermath in the UK was the sharpest: Lighthill’s report led to an effective end to British AI funding, and university research was dismantled until the Alvey programme resumed support in the 1980s. What the objection changes is the frame, not the dates: the cuts of the mid-1970s are documented either way. Funding for the grand promises stopped, the laboratories shrank, and the field’s public credibility dropped for years. A winter is a fair name for a collapse in support even if individual researchers kept working through it. The second winter, below, left no room for that argument: it killed companies.Did the first winter actually happen?
The boom that paid: XCON and the expert system wave
For the first time in the 1980s, AI work turned into serious revenue, and the boom’s centerpiece was real software doing a real job. XCON, written by John McDermott at Carnegie Mellon, configured orders for Digital Equipment Corporation’s computers. It first went into routine commercial use in 1980 at DEC’s plant in Salem, New Hampshire, according to the XCON record.
The numbers explain the enthusiasm. XCON eventually held about 2,500 rules, and by 1986 it had processed a cumulative total of 80,000 orders with 95 to 98 percent accuracy, saving DEC an estimated 25 million dollars a year. Expert systems, the broader technique XCON used, dated to Stanford’s DENDRAL and MYCIN work around 1965; by the late 1980s, two-thirds of the Fortune 500 applied them in daily business. Japan answered with the Fifth Generation project, launched in 1982 to build a new generation of intelligent computing.
The elders called the top before the crash. At the 1984 AAAI conference, Roger Schank and Marvin Minsky warned that enthusiasm had outrun results; Schank predicted a nuclear winter within three years, and the billion-dollar industry began collapsing on schedule in 1987.
The 850 million figure is real but it is a budget set-aside, not money spent. Japan’s Ministry of International Trade and Industry set aside 850 million dollars for the project when it was announced; the Fifth Generation project itself ran from 1982 to 1994 and spent a little less than 57 billion yen, about 320 million US dollars at the time. The project was a commercial failure; it did not deliver the intelligent machines it promised, and its final evaluation report landed in March 1993 as the effort wound down by 1994. It was not money entirely wasted on the research side, since it advanced concurrent logic programming, but the gap between the announced budget and the actual spend is exactly the kind of number that gets inflated in the retelling. The same inflation happens around XCON: its verified results, about 2,500 rules and 95 to 98 percent accuracy, need no rounding up. Repeat the spend, not the set-aside.Did Japan’s Fifth Generation burn 850 million dollars?
The second winter: an industry replaced in a year
The second winter, beginning in 1987 and stretching into the early 1990s, broke the industry rather than the science. Its signature image: in a single year, the entire Lisp machine industry, worth half a billion dollars, was replaced by general-purpose workstations. The casualty list included Symbolics, LISP Machines Inc. and Lucid. In 1988 DARPA cut its AI funding deeply and brutally, in the words of one assessment. In the phrasing of one contemporary history of the field, Japan’s Fifth Generation project “ended not with a successful roar, but with a whimper” on June 1, 1992.
Inside the same window, the seeds of the revival were already germinating. In the mid-1980s David Rumelhart, Geoffrey Hinton and Ronald Williams demonstrated that backpropagation could train multilayer neural networks, the capability missing since the 1969 Perceptrons critique stalled the field. And on real streets rather than in journals, a German researcher was already driving autonomously; his verified record sits in the deep dive below. The money had left the field before its most important decade of results.
Yes, and the record is specific. Ernst Dickmanns’s VaMoRs, a camera-equipped van, drove itself on streets without traffic in 1986 and reached 96 km/h by 1987. In 1995 the system covered a 1,758 kilometer trip, with 95 percent of the distance driven autonomously, according to the Dickmanns record. The work ran on a research budget rather than a boom, which is the point of placing it here: the capability curve for autonomous driving began bending in the late 1980s, years before anyone said deep learning out loud. The caveat is the venue, streets without traffic, and the honest phrasing is what the record supports rather than a superlative about firsts. The winter froze funding for promises, not for every experiment already in motion.Was anyone driving autonomously during the winter?
The quiet decades, 1993 to 2011
After the crash, AI stopped promising general intelligence and started shipping narrow wins. Few made headlines; together they rebuilt the field’s credibility one result at a time.
| Year | What shipped | The checked fact |
|---|---|---|
| 1997 | Deep Blue beats Garry Kasparov | Score 3.5 to 2.5: the first win over a reigning world champion under tournament time controls. |
| Late 1990s | Kismet at MIT | Cynthia Breazeal’s robot head simulated emotion through facial expressions, vocal affect and babbling; Breazeal states it is not conscious and has no feelings. |
| January 2004 | Spirit and Opportunity land on Mars | Onboard autonomy brought each rover to a halt roughly every 10 seconds for a 20-second look at the terrain ahead; the daily marching orders still came from mission teams on Earth, per the Mars Exploration Rover record. |
| 2006 to 2009 | The Netflix Prize | Opened October 2, 2006: one million dollars for a 10 percent improvement on the Cinematch recommender; awarded September 21, 2009. |
| October 14, 2011 | Siri ships on the iPhone 4S | Grown from DARPA’s CALO project at SRI with Nuance speech, it mainstreamed the voice assistant, per the Siri history. |
The competition opened on October 2, 2006: one million dollars to any team that could beat Netflix’s own Cinematch recommender by 10 percent on held-out ratings. It closed on September 21, 2009, when the winning team crossed the threshold with a 10.06 percent improvement. The prize mattered beyond its bracket. It set a measurable target, published real data, and let anyone compete, which is the same recipe the ImageNet challenge would use to such effect from 2010 onward. For a field recovering from two winters, a public scoreboard was a way to make progress verifiable rather than promised.What did the Netflix Prize actually prove?
Board 11 of 18: narrow wins stacked up while nobody was promising a mind.
No, and its own builder is the clearest source on that. Kismet was made in the 1990s at MIT by Cynthia Breazeal as an experiment in socially intelligent machines: a robot head that simulated emotion through facial expressions, vocal affect and pre-linguistic babbling, engaging people the way an infant draws a caregiver in. Breazeal’s own account states that Kismet is not conscious, so it does not have feelings. The distinction was the point of the experiment, showing how much social interaction you can generate with simulated affect alone. Kismet also carries this article’s dating correction: it belongs to the late 1990s, never to a bare 2000. A system that convincingly performs something it does not have is ELIZA’s lesson with a face, and it is why the episode treats demonstrated emotion claims with suspicion.Did Kismet the robot have feelings?
The revival: AlexNet, then everything at once
The turning point arrived in October 2012, when AlexNet won the ImageNet Large Scale Visual Recognition Challenge with half the top-five error rate of the second-place system, a result the AI timeline marks as deep learning’s turning point. The technique was not new; the fuel was. ImageNet’s labeled dataset arrived in 2009, GPUs made training large networks practical, and backpropagation from the mid-1980s tied them together.
The decade after 2012 compressed the cycle. OpenAI revealed DALL-E, text-to-image generation, on January 5, 2021, announced DALL-E 2 on April 6, 2022, and opened it to everyone that September 28. Then ChatGPT launched on November 30, 2022, and the third boom reached everyone’s feed. Image generation inside ChatGPT followed in October 2023 with DALL-E 3.
The numbers are verified and still hard to absorb: an estimated one million users within five days of the November 30, 2022 launch, and one hundred million users within two months, which would make it the fastest-growing internet application in history. Earlier consumer internet wins took years to reach those figures. Two cautions travel with the record. The growth figure is a user estimate rather than an audited count, and the fastest-growing title reportedly ended once Meta launched Threads in mid-2023, five months later. Neither caution changes the substance: ChatGPT put the third boom on every phone, exactly the kind of public moment the field’s two previous booms built toward and then lost. For how the underlying training actually works, see the machine learning lesson in this series.How fast did ChatGPT actually grow?
Five AI history claims that do not survive the sources
The episode’s corrections chapter. Each row pairs a sentence you will hear this year with what the cited record supports.
| The claim | What the sources say |
|---|---|
| Lisp was the first AI programming language. | Lisp, 1958, became the favored AI research language; IPL, around 1956, was the first built for the work. |
| XCON was the first expert system. | Its claim to priority is production use at DEC from 1980; the technique itself dates to Stanford around 1965, with DENDRAL and MYCIN. |
| Japan spent 850 million dollars on the Fifth Generation. | That was the announced set-aside; actual spend over 1982 to 1994 was a little less than 57 billion yen, about 320 million dollars. |
| Minsky and Papert killed neural networks in 1969. | Their book emphasized the limits of single-layer networks; neural funding dried up, but the multilayer answer, backpropagation, was demonstrated by the mid-1980s. |
| Kismet was a 2000 robot. | It was made in the 1990s at MIT; date it late 1990s. |
Board 18 of 18: four checked lines and three lessons, including the one this article practices.
Run the tape backward from any current headline and you find the same loop in every decade of AI history: a public promise, a funding rush, a reckoning, then a quieter rebuild. The name dates to 1956, the winters to 1974 and 1987, and the ingredients of the present revival, labeled data, parallel hardware and backpropagation, are older than most of the people now using them. The 1984 warnings from the field’s own seniors were dismissed for roughly three years, which is almost exactly how long the boom then lasted. That record is the working checklist: read each claim’s tense, and keep what shipped separate from what was promised.
Continue learning
- What Artificial Intelligence Actually Is: the first lesson in this series, the definition that survived fact checking and the 1955 bet that named the field.
- Machine Learning: The Honest First Lesson: the training loop that the post-2012 revival ran on, and the history most introductions get wrong.
- AI Glossary 02: Neural Architectures and Transformer Terms: the vocabulary of the architectures the revival produced, one term at a time.
This lesson first appeared on iqraa.tech.
More whiteboard lessons on the Whiteboard Architect YouTube channel.
Top comments (0)