Second AI winter: The boom meets its balance sheet

The event and the narrow claim

In 1987 the commercial and government supports under the 1980s boom in artificial intelligence gave way at nearly the same time. By May, Lisp Machine Inc. (LMI), one of two companies spun out of MIT to sell computers built to run the Lisp programming language, was being auctioned in bankruptcy court, as Computer Business Review reported that month. In July the expert-system startup Teknowledge said it expected a loss and slower growth than it had planned for in the market for expert systems software. In September, DARPA, the Pentagon research agency that had funded American AI since the 1960s, put its Information Systems Technology Office under Jacob Schwartz, a skeptic who, as Nils Nilsson records in his history of the field, promptly cancelled some AI programs and declined to renew others. That same year the American Association for Artificial Intelligence (AAAI) reached its peak of 16,421 members, according to Nilsson's chapter on the boom, and then began a decline that lasted the rest of the decade and beyond.

That is the narrow, documented event: a specialised hardware market collapsed, the leading expert-system firms went into loss, and the main government funder turned against the field, all within about two years of 1987. The broad claim is larger and less settled. "The second AI winter" names a long contraction in AI's money, prestige and self-description. Nilsson treats it as a season that passed; the historian Thomas Haigh, in How the AI Boom Went Bust, argues that it began "a two-decade slump" and that it, rather than the 1970s retrenchment, was the real winter. This report dates the event to 1987 because that is the year the contemporaneous record first shows the boom failing on several fronts at once, and it keeps the narrow and broad claims apart.

A boom built on knowledge, hardware and fear

The boom rested on three supports. The first was an idea. Programs such as DENDRAL at Stanford had shown that encoded specialist knowledge, rather than general reasoning power, could solve hard but narrow problems, and XCON, which configured computer orders at Digital Equipment Corporation (DEC) from 1980, became the standard proof that a hand-built rule base could save a company money. Haigh quotes the SRI veteran Peter Hart recalling that new expert-system companies seemed to appear at a rate of about one a week ( Haigh).

The second support was hardware. In the late 1970s MIT's AI Laboratory designed computers around Lisp, the language most American AI was written in. The 1979 laboratory memo describing the CADR processor presents it as a 32-bit microprogrammable machine whose microcode executed a compact 16-bit instruction code produced by the Lisp compiler, intended as a high-performance and economical way to run Lisp. Two startups, Symbolics and LMI, commercialised the MIT work and competed with Xerox and Texas Instruments to sell machines custom built for Lisp ( Nilsson; Roland and Shiman).

The third support was government money, unlocked largely by fear of Japan. After Japan announced its Fifth Generation Computer Systems project in 1981, DARPA launched its Strategic Computing Program in 1983 and, according to the National Research Council's history of federal computing research, committed $1 billion over a planned ten years, with some 60 percent of the money going to industry. The council's tables show total federal funding for AI research, including the program, rising from $106 million in 1984 to $274 million in 1988. Strategic Computing also bought researchers their tools: by the end of 1985 its machine acquisition program had allocated more than $8 million and purchased more than 100 Lisp machines, starting with 30 Symbolics 3600s ordered in March 1984 ( Roland and Shiman).

The warning that named the season

The phrase "AI winter" entered the printed record as a warning, not a description. At the 1984 AAAI conference in Austin, Texas, Drew McDermott of Yale chaired a panel whose transcript appeared the next year in AI Magazine as The Dark Ages of AI. He described a "deep unease" among researchers who had been in the field longer than the boom, and sketched a worst case: Strategic Computing collapses as autonomous vehicles fail to roll, Japan's project goes nowhere, every startup fails, and money for anything labelled AI dries up. That condition, he said, was "called the 'AI Winter' by some," and he joked that it had prompted someone to ask whether nuclear winter meant a funding cut for nuclear weapons. He judged the scenario unlikely.

The wording matters for dating the term. "By some" means the phrase was already circulating, so the panel is the earliest widely cited printed use rather than a documented moment of coinage. The transcript's speakers were McDermott, the Science journalist M. Mitchell Waldrop, Ron Ohlander of DARPA, Roger Schank of Yale, B. Chandrasekaran of Ohio State and John McDermott of Carnegie Mellon. Retellings often add Marvin Minsky to the warning; he does not appear in this transcript.

The panel already contained most of the later diagnosis. Schank worried that large corporations were filling new AI groups with people who had merely read about the field, and predicted that when results lagged their executives would complain at the highest levels of the Defense Department. Ohlander, speaking for the funder, expected a shakeout among both the machine makers and the expert-system companies but believed government funding would hold. Chandrasekaran hoped instead for an "AI dusk," a bubble that shrank and became less fragile rather than bursting.

Why the Lisp machine lost to the workstation

The Lisp machine's problem was economic before it was technical. Alex Roland and Philip Shiman, in their history of Strategic Computing, explain that the machines served a small market, could not be built in enough volume for economies of scale, and so sold for $50,000 to $100,000; they also connected poorly to other computers, and few people outside AI were prepared to program them. General-purpose engineering workstations built on mass-produced microprocessors, meanwhile, improved quickly. By early 1986 DARPA's own contractors, free to choose their equipment, were picking Sun workstations over Lisp machines, which even with Strategic Computing's discounts cost three to four times as much. DARPA bought 23 Suns in February 1986 and had 45 on order by June.

Software followed the hardware. Common Lisp was defined in a book published in 1984, and Richard Gabriel of Lucid, Inc. recalled in a 1991 essay on the state of Lisp that several companies sprang up to run it on stock hardware in competition with the Lisp machine makers; within four years nearly every major computer company offered a Common Lisp. Once an ordinary workstation could run the same programs adequately, the case for a dedicated machine narrowed to raw speed, and commodity processors were closing that gap too. The National Research Council states the outcome plainly: engineering workstations, especially Sun's, soon undermined the Lisp machine industry, and that segment of the market, closely tied to Strategic Computing, collapsed.

The mechanism is general. Specialised hardware must stay ahead of general-purpose processors to justify its price, and in the 1980s general-purpose processors improved very fast; Japan's parallel inference machines met the same squeeze a few years later. Symbolics eventually conceded the point. In August 1989 it offered its Ivory Lisp processor on a $13,900 plug-in board for Sun workstations, after already moving its graphics software to the Apple Macintosh.

The balance sheets of 1987 and 1988

The contemporaneous trade press shows the collapse quarter by quarter. LMI, which Roland and Shiman identify as the first Lisp machine company to go under, had refused to give DARPA steep discounts and sold only one machine to the acquisition program. In its bankruptcy auction, Computer Business Review reported, the highest bid was $2.75 million from Gigamos Holdings of Montreal, against $1.5 million from Data General and $900,000 from LMI's rival Symbolics. The damage spread to partners. Texas Instruments, an LMI shareholder, reported in January 1988 that its industrial automation and artificial intelligence activities had operated at a loss in 1987, partly because of reserves for contract manufacturing done for a major customer that went bankrupt, presumably LMI.

The software side fared no better. In July 1987 Teknowledge, one of the best known expert-system startups, said it expected a fourth-quarter loss of about $2.9 million on revenue of about $5.1 million, and a similar loss for the year on sales of about $20.5 million, blaming product revenue below expectations and a slower-growing market for expert systems software. Analysts at the consultancy Ovum, writing in The Knowledge Engineering Review in 1988, called 1987 to 1988 a terrible period for the expert systems business: the four startups that had led the American industry, known as the Gang of Four, had all gone into loss and cut staff, and Symbolics had suffered an equally dramatic reversal.

Symbolics, the market leader, lost $4.5 million on revenue of $24 million in the quarter to 4 October 1987, and in January 1988 announced that it would cut about 10 percent of its roughly 730 employees, according to Computer Business Review. A 2001 case study written by students for an MIT course on engineering history gives Symbolics' product revenues as $101.6 million in fiscal 1986, $82.1 million in 1987 and $55.6 million in 1988, with only training and consulting revenue still growing. Symbolics struggled on until it filed for bankruptcy in 1993, while Xerox and Texas Instruments abandoned Lisp machines years before that ( Roland and Shiman).

Brittle experts and the cost of keeping them

The machines failed for economic reasons; the expert systems that ran on them disappointed for technical ones, which critics inside the field had named before the market turned. In a 1984 lecture, John McCarthy argued that expert systems needed common sense and called them "brittle," meaning hard to extend beyond the scope their designers had in mind and unable to recognise their own limits. His example was MYCIN, Stanford's celebrated adviser on blood infections, which had still not been put into production use. Told that a patient had cholera, MYCIN would recommend two weeks of tetracycline and nothing else, which would kill the bacteria long after the patient had probably died of the disease. Faced with new bacteria, such as the Legionella behind Legionnaires' disease, understood only after the program was finished, it simply replied "unrecognized response."

Doug Lenat and colleagues at the Microelectronics and Computer Technology Corporation made the same diagnosis the basis of a new project. Their 1985 article, CYC: Using Common Sense Knowledge to Overcome Brittleness and Knowledge Acquisition Bottlenecks, named brittleness and the manpower needed to build large programs as the two major limits on software, and proposed spending the coming decade building a knowledge base of everyday facts to overcome them. The remedy was itself a measure of the problem: if an expert system needed an encyclopedia of common sense before it could be robust, the quick and inexpensive systems that buyers had been promised were not on offer.

The other cost was upkeep. At DEC, the most successful deployment of all, XCON had grown by 1987 to about 6,200 rules, roughly half of which changed every year, and its developers were rewriting it because the rule base had become hard to change, even though its performance was satisfactory. Haigh reports that by 1989 DEC had 59 technical staff maintaining its internal expert systems, a commitment few companies could match, and concludes that systems built to automate expertise forced their buyers to hire new experts to keep them running ( Haigh). The National Research Council adds that the label itself became a liability, so that some users began calling the technology knowledge-based systems, naming what it was rather than what it promised ( Funding a Revolution).

DARPA changes its mind

Strategic Computing had been sold to Congress with three military demonstrations: a pilot's associate for the Air Force, an autonomous land vehicle for the Army and a battle management system for the Navy ( National Research Council). By 1987 none was near the generic machine intelligence the program had promised. Roland and Shiman describe the storm that reached DARPA as arising outside it, driven by the collapsing Lisp machine market and by startups that could not ship products as fast as their venture investors expected, and arriving in the person of Schwartz. A New York University computer scientist and long-standing DARPA adviser, he succeeded Saul Amarel as director of the Information Systems Technology Office in September 1987. He had just written, for the 1987 Encyclopedia of Artificial Intelligence, that AI had yet to show "any unifying principles of self organization," and that expert systems gained practical value by narrowing AI's goals ( Roland and Shiman).

Schwartz put DARPA's AI contracts under review. He told Teknowledge and IntelliCorp, which held contracts to build generic expert-system tools, that they could expect no further DARPA support for that work after their current authorisations expired, and that whatever they produced had to appear by the end of fiscal 1990. The money followed. In 1987 the office's combined basic and Strategic Computing budgets for AI and robotics totalled $47 million; by 1989 they were under $31 million, a fall that Nilsson also reports. The autonomous land vehicle's development program effectively ended in the fall of 1987, when DARPA dropped its milestones, and the program was cancelled outright in April 1988 ( Roland and Shiman), which is why Haigh can date its abandonment to the end of 1987.

The cuts were selective. Schwartz raised funding for speech recognition, which was making progress in creating standard benchmark tests, and for Carnegie Mellon's Navlab vehicle. The broader drift was away from AI. The National Research Council notes that the troubles of expert systems and of Strategic Computing's flagship projects contributed to a shift in high-performance computing toward other grand challenges, such as weather prediction and scientific visualisation, and Nilsson records that Strategic Computing disappeared from the DARPA budget in 1993.

The honest limits of the winter story

The record supports a sharp downturn from 1987 to 1989, not a single switch. By the council's figures, federal AI funding was still rising in 1988, so the government retreat followed the commercial one rather than causing it. Japan's Fifth Generation project ran on until 1992, and the Strategic Computing name lingered in budgets until 1993. Most of the company figures cited here are self-reported results relayed by the trade press, and the Symbolics revenue series comes from a student case study rather than from filings consulted for this report.

Nor did research stop. Nilsson writes that the winter "endured only for a season," one of renewed effort rather than hibernation, even as AAAI membership fell gradually to between 4,000 and 5,000 by 1996 ( Nilsson). Haigh reads the same years as a two-decade slump in AI's standing, visible in a Google Ngram curve and in the absence of AI from the Turing Award between 1995 and 2010, while cautioning that his evidence is Anglo-American and that AI publications and funding in South Korea rose in the late 1980s and early 1990s ( Haigh). Both readings can hold at once: the brand and the money contracted while the work went on under other names. Many deployed expert systems kept running, and the council observes that by the late 1990s AI's commercial successes were mostly hidden inside larger software systems ( National Research Council).

Causation is contested as well. Gabriel argued that the Lisp companies became scapegoats when the expectations raised by AI's commercial promoters went unmet, and that a 1989 Forbes article faulted Symbolics for its faith in AI and its bet on proprietary hardware, not for the language ( Gabriel). The Lisp machine collapse would probably have come without any loss of faith in AI, because workstation economics were independent of it. What the winter story captures correctly is that the hardware failure and the expert-system disappointment arrived together and fed the same disillusionment.

How to place the second AI winter in AI history

The first AI winter was mainly a funder's verdict, delivered through reports such as Lighthill's. The second was a market correction first and a funder's verdict second: customers stopped buying specialised machines and expert-system tools before DARPA's new office director cut the budgets. That makes it the first time AI's fortunes were set chiefly by commercial demand, which gives it an obvious bearing on later booms, though Chronicle treats any parallel with the present as provisional rather than proven.

Two technical lessons outlasted it. One concerns hardware: the machine built for AI lost to the general-purpose computer that improved faster, and AI's later revival ran on general-purpose parallel processors, including graphics chips programmed through CUDA. The other concerns knowledge: rules written by hand proved costly to extend and maintain, which strengthened the case for methods that learn from data or reason with probabilities, such as backpropagation and the Bayesian networks the National Research Council lists among the advances that brought AI into the marketplace in the 1990s. Haigh notes that the neural network revival began in the 1980s but proceeded under labels such as machine learning and pattern recognition, because AI had come to mean symbolic methods. The season McDermott warned about in 1984 arrived roughly as he described it, and the field that came out of it was organised less around encoded expertise and more around data, statistics and commodity computing.

References

• Drew McDermott, M. Mitchell Waldrop, B. Chandrasekaran, John McDermott and Roger Schank, AI Magazine (1985). The Dark Ages of AI: A Panel Discussion at AAAI-84.
• Thomas F. Knight, David A. Moon, Jack Holloway and Guy L. Steele, MIT AI Laboratory (1979). CADR, AI Memo 528.
• John McCarthy, Stanford University (1984). Some Expert Systems Need Common Sense.
• Doug Lenat, Mayank Prakash and Mary Shepherd, AI Magazine (1985). CYC: Using Common Sense Knowledge to Overcome Brittleness and Knowledge Acquisition Bottlenecks.
• Elliot Soloway, Judy Bachant and Keith Jensen, AAAI (1987). Assessing the Maintainability of XCON-in-RIME: Coping with the Problems of a VERY Large Rule-Base.
• CBR Staff Writer, Computer Business Review via Tech Monitor (1987). Data General, Symbolics and Gigamos vie for bankrupt Lisp Machine.
• CBR Staff Writer, Computer Business Review via Tech Monitor (1987). Teknowledge expects to report losses.
• CBR Staff Writer, Computer Business Review via Tech Monitor (1988). Symbolics to cut 10%, take big charge.
• CBR Staff Writer, Computer Business Review via Tech Monitor (1988). Texas Instruments' 1987 net sales declined moderately.
• CBR Staff Writer, Computer Business Review via Tech Monitor (1989). Now Symbolics offers UX400S Ivory board for Sun Microsystems workstations.
• Tim Johnson, Julian Hewett, Christine Guilfoyle and Judith Jeffcoate, The Knowledge Engineering Review (1988). Expert systems: the second wave.
• Richard P. Gabriel, Lucid Inc. (1991). Lisp: Good News, Bad News, How to Win Big.
• National Research Council (1999). Funding a Revolution, Chapter 9: Developments in Artificial Intelligence.
• Alex Roland with Philip Shiman, MIT Press (2002). Strategic Computing: DARPA and the Quest for Machine Intelligence, 1983-1993.
• Alvin Graylin, Kari Anne Hoier Kjolaas, Jonathan Loflin and Jimmie D. Walker III, MIT OpenCourseWare (2001). Symbolics, Inc.: A failure of heterogeneous engineering.
• Nils J. Nilsson, Cambridge University Press (2009). The Quest for Artificial Intelligence, Chapter 21: Boomtimes.
• Nils J. Nilsson, Cambridge University Press (2009). The Quest for Artificial Intelligence, Chapter 23: DARPA's Strategic Computing Program.
• Nils J. Nilsson, Cambridge University Press (2009). The Quest for Artificial Intelligence, Chapter 24: Speed Bumps.
• Thomas Haigh, Communications of the ACM (2024). How the AI Boom Went Bust.