WE

17 November 2026 · History · about 13 minutes

Knowledge Is Power: Expert Systems, the AI Boom of the 1980s, and the Winter That Followed

In the 1970s, a few researchers at Stanford argued that the secret of intelligence was not reasoning but knowledge. They were right enough to build an industry, and wrong enough to destroy it.

The Student and the Bacterium

In 1979, eight infectious disease specialists sat down to judge treatment recommendations for ten meningitis cases. The prescribers whose work they were rating included five Stanford faculty physicians, a fellow, a resident, a student, the actual treating doctors, and a computer program written by a doctoral student, Edward Shortliffe, working under the direction of Bruce Buchanan and Stanley Cohen. The specialists did not know a computer was in the mix. When the ratings came back, the program had been rated acceptable in 65 percent of cases; the five Stanford faculty ranged from 42.5 to 62.5 percent. The program was called MYCIN.1

Around 1972, Shortliffe had begun building MYCIN to help doctors treat blood infections. The clinical problem was real and urgent: bacteremia and meningitis killed quickly, the organisms responsible were not always identifiable from early tests, and the antibiotics available varied in effectiveness, side effects, and dosage by patient weight. A resident working a night shift could not always summon an infectious disease specialist, and a wrong choice of drug might cost a patient days.2

MYCIN was written in Lisp and organised around a set of roughly 600 individual if-then rules: if the organism is gram-negative and the patient is a compromised host, then consider such-and-such treatment with a certain degree of confidence. The rules were not assembled from a textbook but extracted, over months of conversation, from Stanford's infectious disease specialists. Each rule encoded a clinical judgement as well as a clinical fact, and because judgements differ, MYCIN carried numerical weights it called certainty factors to track how confident the system was in any given conclusion.3

The program could also explain itself. A physician who asked why MYCIN had reached a particular conclusion received an account of the rules it had applied and the values those rules had used. This was not a cosmetic feature. Shortliffe and his colleagues were alert to the fact that physicians would not follow advice from a machine they did not understand, and so transparency was built into MYCIN's design from the start.4

MYCIN was never deployed in routine clinical practice. The reasons were partly legal and ethical, who was responsible if the system gave wrong advice?, and partly technical: MYCIN was a standalone system that required a user to type in all relevant patient information by hand, and it could not connect to hospital records or laboratory systems.5 These obstacles were not insuperable, but the Stanford team did not surmount them before the system receded into the literature.

As development continued, Randall Davis, a Stanford doctoral student working alongside Shortliffe, built a system called TEIRESIAS that allowed experts to interact directly with MYCIN's rule base, correcting and extending its knowledge through a conversational interface. TEIRESIAS eased the work of updating rules, but it also made the remaining limits clearer: even with an interactive tool, extracting reliable knowledge from experts and keeping it consistent was slow and laborious. The difficulty of getting knowledge in and keeping it current was the central technical problem of the whole enterprise, and the field would come to call it the knowledge-acquisition bottleneck.6

The Stanford group drew an obvious lesson. If the MYCIN architecture could be generalised, the same rule-and-inference engine could be applied to any domain in which experts had articulable rules. They stripped MYCIN of its medical knowledge and released what remained, the inference engine and the question-answering apparatus, as EMYCIN, standing for Essential MYCIN. With EMYCIN, the work of building an expert system could be separated from the work of filling it with knowledge.7

The Father and the Principle

The Stanford Heuristic Programming Project, which had housed the MYCIN work, was led by Edward Feigenbaum, a former student of Herbert Simon who had come to Stanford from Berkeley in 1965. Working alongside him from the outset was Bruce Buchanan, who served as Feigenbaum's principal AI collaborator on DENDRAL and later co-edited the definitive account of MYCIN, Rule-Based Expert Systems, in 1984. The project that formed Feigenbaum's central insight was DENDRAL, conceived in the 1960s by Joshua Lederberg, the Nobel Prize-winning geneticist who had arrived at Stanford to chair its Department of Genetics in 1959. Lederberg was working on the problem of inferring the molecular structure of chemical compounds from mass spectrogram data, and he brought the problem to Feigenbaum. The three built DENDRAL together, with the chemist Carl Djerassi providing domain expertise in mass spectrometry.8 DENDRAL worked by drawing on the knowledge of chemists, encoded as rules, rather than by applying any general theory of chemistry. The observation that adding more knowledge made the program smarter in a way that adding more reasoning power did not was the basis of what Feigenbaum would call his central insight.

Feigenbaum is credited with coining the phrase "knowledge is power" in connection with AI programs, and with introducing the terms "expert systems" and "knowledge engineering" into the field's vocabulary. The argument, stated plainly in his 1977 paper at the International Joint Conference on Artificial Intelligence, was this: the problem-solving power exhibited in an intelligent agent's performance is primarily a consequence of the specialist's knowledge employed by the agent, and only very secondarily related to the generality and power of the inference method employed. For AI to be useful in professional fields like medicine or chemistry, it must be knowledge-rich, even if it remains methods-poor. Many researchers had previously focused on formal reasoning methods, general-purpose theorem provers and problem-solvers, and those methods had hit a ceiling. Feigenbaum's claim was that the ceiling was of their own making. The right move was not to build a better reasoner but to encode more knowledge.9

This was a genuinely productive insight, and it deserves to be treated as such. The systems it generated, MYCIN, DENDRAL, and their successors, did things that general-purpose reasoners of the period simply could not do. The question was what it meant to generalise the principle. The logic of Feigenbaum's insight, as many in the field read it, implied that more knowledge would always produce a smarter system, and that the work of building such systems was essentially the work of extracting and codifying what experts already knew. Both implications would prove true within limits, and false beyond them.

The Configurer

The transition from academic demonstration to commercial deployment happened in 1980, in a Digital Equipment Corporation plant in Salem, New Hampshire. The system responsible was not a medical program but a production-rule-based configurer of computer orders, written by John McDermott of Carnegie Mellon University. DEC had a particular problem. Its VAX computers were sold in components, processors, memory boards, storage devices, terminals, cables, and power supplies, and assembling a coherent system from a customer's order required a kind of expert knowledge that DEC's sales force did not reliably possess. Before McDermott's system went into use, customers routinely received systems with incompatible components, missing cables, or mismatched drivers.10

McDermott built his system in OPS5, a production-rule language, beginning around 1978 in collaboration with DEC's engineers. The system was initially called R1, a name that McDermott glossed in Footnote 2 of his 1980 AAAI paper with the remark that "Four years ago I couldn't even say 'knowledge engineer', now I ... [are one.]", and was renamed XCON, for Expert Configurer, by DEC when it went into operational use. McDermott's 1980 AAAI conference paper on R1 won the AAAI Classic Paper Award in 1999. His fuller account appeared in 1982 in the journal Artificial Intelligence, volume 19, pages 39 to 88.11

XCON went into use at DEC in January 1980 with about 750 rules. By 1986 it had processed 80,000 orders at an accuracy of 95 to 98 percent. Contemporary estimates of the annual saving differ: XCON Wikipedia, citing Barker et al. (1989), gives $25 million a year; the AI winter Wikipedia article, citing Newquist (1994), gives $40 million over six years of operation. The two figures are inconsistent and cannot be reconciled from the sources available. By 1987, XCON had grown to 6,200 rules. By 1989, it contained 10,129 rules and DEC was using it alongside a family of related systems, including XSEL for DEC's sales force and XSITE for location planning.12

The Gold Rush

The money followed quickly. Corporations around the world began developing expert systems, and by 1985 they were spending over a billion dollars a year on AI, most of it directed at in-house expert systems departments. An entire industry grew up to support them: software companies like Teknowledge and IntelliCorp sold expert system development shells and tools; consulting firms built AI practices. DARPA's Strategic Computing Initiative, launched in 1983 partly in response to Japan's Fifth Generation Computer Project, poured additional funding into the field. XCON had proved that the approach worked; the question now was how fast it could be sold.13

The hardware side of this industry rested on a particular kind of machine. Expert systems ran on Lisp, the programming language that had been associated with AI research since John McCarthy invented it in the late 1950s. Ordinary computers of the early 1980s could run Lisp, but slowly: Lisp's need for automatic memory reclamation, garbage collection, and symbol manipulation pushed hard against hardware designed for other purposes. The solution, developed in the 1970s at the MIT Artificial Intelligence Laboratory, was to build computers designed specifically to run Lisp, machines that handled list operations and garbage collection in hardware, whose operating systems were written in Lisp from the microcode up, and whose development environments were correspondingly rich and powerful.14

By the early 1980s, two companies founded by MIT AI Lab staff were selling these machines commercially. The first incorporated was Lisp Machines, Inc., founded in 1979 by Richard Greenblatt, a legendary programmer who had led the MIT Lisp Machine Project and who preferred a self-sustaining company built on advance orders to a venture-capital-funded startup. The hackers at the AI Lab chose Symbolics; Greenblatt's company was eventually launched with outside financial backing rather than MIT staff. Symbolics was incorporated on 9 April 1980; Russell Noftsker, who had been administrator of the MIT AI Lab in the late 1960s, was its first CEO and chairman, with Robert P. Adams as co-founder and first president.15

Symbolics' initial product, the LM-2, introduced in 1981, was a repackaged version of the MIT CADR Lisp machine design. In 1983, a year later than planned, it introduced the 3600 family: a new design built around a 36-bit word with hardware support for both virtual memory and garbage collection. The operating system and software development environment comprised over 500,000 lines of Lisp. The 3600 was roughly the size of a household refrigerator, and expensive enough to match.16 MIT, Stanford, Carnegie Mellon, and government laboratories installed Symbolics machines. By the mid-1980s, Symbolics had become the dominant Lisp machine vendor.17

Texas Instruments and Xerox entered the market too. TI, after investing in LMI when that company ran short of money, licensed LMI's Lambda design and later produced its own variant, the TI Explorer. Xerox ran a different dialect of Lisp, Interlisp-D, on its D-machines and had developed, at its Palo Alto Research Center, what was at the time one of the most advanced software development environments available. The Lisp machine market was not a monopoly; it was a small but functioning collection of competing vendors, each selling hardware that cost between fifty thousand and a hundred thousand dollars.18

Symbolics registered symbolics.com on 15 March 1985, making it the first .com domain in the world. Symbolics had implemented one of the first TCP/IP stacks in its Genera operating system, the registration was a consequence of the company's immersion in networked computing rather than a marketing calculation, though that inference is this essay's, not the sources'.19

The Weight of the Rules

Inside XCON, the knowledge-acquisition bottleneck was taking a different form: not the initial difficulty of eliciting knowledge from experts, but the ongoing difficulty of keeping encoded knowledge current. The system had begun with about 750 rules and had grown to thousands. Each new DEC product required new rules. When DEC changed an existing product, added a memory option, changed a connector type, discontinued a peripheral, the rules encoding knowledge of that product had to be updated, and the interactions between rules had to be checked. By 1987, with 6,200 rules, XCON had become, in the words of one published assessment, an "exceedingly difficult software enhancement problem", in which 50 percent of the rules were changing every year. The engineering required to keep the system consistent had become a full-time operation.20

This was not a bug in XCON's design. It was a feature of the architecture. Expert systems worked by encoding knowledge as explicit rules, and rules had to be written and maintained by human beings. As the domain grew more complex, the rule base grew with it. As the rule base grew, the interactions between rules multiplied faster than the rules themselves. And as the domain changed, as it inevitably did in a technology industry, every change required the knowledge engineers to revisit the existing rules and decide which ones needed revision. The system had no way to learn from new cases or revise its own rules.

This problem had been visible to attentive observers from early on. In 1984, at the annual meeting of the AAAI, Roger Schank and Marvin Minsky, both veterans of the first AI winter of the 1970s, issued a public warning. They described a chain reaction that would begin with pessimism in the AI community, followed by pessimism in the press, followed by severe cutbacks in funding, followed by the end of serious research. They called it, by analogy with nuclear winter, an AI winter. XCON's own maintainers had already published accounts of the growing difficulty of keeping the rule base consistent. The audience, prospering on expert systems contracts, did not listen. Three years later, the billion-dollar AI industry began to collapse.21

The Second Winter

The collapse began in 1987, and its immediate cause was hardware. By that year, general-purpose workstations from Sun Microsystems and others had grown powerful enough to run Lisp programs at speeds that matched or exceeded purpose-built Lisp machines. Software companies like Lucid and Franz Lisp offered increasingly capable Lisp environments that ran on standard Unix systems. The economic case for spending fifty thousand dollars on a machine optimised for Lisp disappeared when a cheaper workstation running a commercial Lisp compiler could do the same job. An entire industry worth half a billion dollars was replaced in a single year.22

The hardware collapse brought the expert systems software industry down with it. Companies that had sunk money into expert systems found that the total cost of ownership was higher than projected. The rule bases required maintenance, and maintenance was expensive. The systems could not be extended to cover problems adjacent to their original brief without substantial new engineering. When the domains they served changed, the knowledge bases had to be manually rebuilt. And the systems were brittle: they handled situations covered by their rules, but failed when given unusual inputs that fell outside the encoded knowledge. By the early 1990s, most commercial Lisp companies had failed, including Symbolics, Lisp Machines Inc., and Lucid Inc.; Texas Instruments and Xerox withdrew from the market.23

Symbolics had been struggling with internal conflicts before the hardware market turned against it. Noftsker and the CEO the board had brought in, Brian Sear, fought over whether to emphasise hardware or software; both were removed from the company in the late 1980s. The company entered into large long-term lease obligations in California during the boom years. The combination of strategic disarray, the collapse of the Lisp machine market, and the slowdown of DARPA's Strategic Computing Initiative, which had funded significant AI hardware purchases, proved fatal. By the early 1990s, Symbolics had entered bankruptcy.24 LMI had gone bankrupt earlier, in 1987, before its next-generation K-Machine could reach market.25

Jack Schwarz, who took over the direction of DARPA's Information Processing Technology Office in 1987, dismissed expert systems as "clever programming" and cut AI funding, in the words of one later account, "deeply and brutally". He believed DARPA should, in his own formulation, "surf" rather than "dog paddle", funding technologies that showed genuine promise rather than supporting established approaches with known ceilings. XCON was discontinued by DEC in the early 1990s. The term "AI winter" had entered common use, and many researchers began relabelling their work to avoid the stigma; at its low point, some avoided the term "artificial intelligence" for fear of being viewed as wild-eyed dreamers.26

What They Got Right, and What They Got Wrong

The people who built expert systems were not credulous or careless. They were responding intelligently to what they had observed. Feigenbaum's insight that knowledge, not reasoning power, was the scarce resource in AI problem-solving was borne out repeatedly. MYCIN worked not because it had a clever inference engine, the engine was straightforward, but because it encoded the clinical knowledge of infectious disease specialists with unusual care and depth. XCON worked because McDermott and his collaborators spent years extracting the tacit knowledge of DEC's most experienced engineers and translating it into rules.27

The error lay in the assumption that the approach was indefinitely extensible. If knowledge was power, the implication seemed to be that more knowledge was always more power, that you could build a better expert system simply by encoding more rules, and that the encoding itself, while laborious, was in principle straightforward. Neither proved true. Much of what experts know they cannot articulate, or articulate inconsistently, or articulate in ways that conflict with what other experts say. And the rules that could be articulated did not scale cleanly: a system with ten thousand rules was not simply a larger version of a system with a thousand; it was a qualitatively different engineering problem, one that the tools of the 1980s could not manage.

The Lisp machine companies failed for a related reason. They built hardware optimised for a specific programming language and a specific style of AI, at a time when general-purpose hardware was improving fast enough to overtake them. The optimisation that made a Symbolics 3600 impressive in 1983 had become a liability by 1987, because the advantages of specialisation were too small to justify the price differential against increasingly powerful general-purpose workstations.

Feigenbaum was right that knowledge matters. The bottleneck that Buchanan, Shortliffe, and the XCON maintainers had each documented, that encoding, eliciting, and updating expert knowledge was itself the hard problem, was the one the industry could not solve.

Notes

  1. The 1979 blinded evaluation published in JAMA: eight independent evaluators; ten prescribers including MYCIN; ten meningitis cases; evaluators did not know which prescriber was a computer; MYCIN rated acceptable at 65 percent; five Stanford faculty prescribers at 42.5 to 62.5 percent: Yu VL et al., "Antimicrobial selection by a computer: a blinded evaluation by infectious diseases experts", JAMA 242(12):1279–82, 19791; abstract confirmed at Popline record 1540022. Shortliffe as lead developer working under the direction of Bruce Buchanan and Stanley Cohen: MYCIN, Wikipedia3. The conclusion about performance being comparable to Stanford infectious disease faculty as judged by outside experts: Chapter 36, Rule-Based Expert Systems, Buchanan and Shortliffe, 19844.
  2. MYCIN's origin in Shortliffe's doctoral work at Stanford, beginning around 1972; the clinical problem of bacteremia and meningitis; antibiotic selection and dosage adjusted for patient body weight: MYCIN: A Rule-Based Expert System for Infectious Disease Diagnosis, achievements.ai, 20245. Shortliffe's dual study of medicine and computer science at Stanford, earning a PhD in Medical Information Sciences and an MD: Edward H. Shortliffe, Wikipedia6. The 1972 date is approximate; Wikipedia on MYCIN notes the system was developed over five or six years in the early 1970s.
  3. The approximately 600 rules and the if-then structure: Expert Systems lecture notes, Yale CS, n.d.7 MYCIN's use of certainty factors and its derivation from the Dendral expert system architecture, with Shortliffe working under Bruce Buchanan and others: MYCIN, Wikipedia3.
  4. MYCIN's ability to respond to queries about why it asked questions and how it reached conclusions: MYCIN, Wikipedia3. The concern about physician acceptance and the role of transparency: Chapter 36, Rule-Based Expert Systems: The MYCIN Experiments of the Stanford Heuristic Programming Project, Buchanan and Shortliffe, eds., 19844.
  5. MYCIN never deployed in routine practice; legal and ethical issues about responsibility; MYCIN as a standalone system requiring all patient data to be typed in by hand: MYCIN, Wikipedia3.
  6. Randall Davis and TEIRESIAS as an interactive knowledge-acquisition tool built on MYCIN; Davis as a Stanford doctoral student: Edward Feigenbaum, ACM Turing Award citation, Association for Computing Machinery, 19948, which names Davis among Feigenbaum's graduate students. The knowledge-acquisition bottleneck as the central technical problem of expert systems is discussed in Chapter 36, Rule-Based Expert Systems, Buchanan and Shortliffe, 19844.
  7. EMYCIN, Essential MYCIN, as a general expert system shell stripped of medical knowledge: MYCIN, Wikipedia3; and Edward Feigenbaum, ACM Turing Award citation, Association for Computing Machinery, 19948.
  8. Joshua Lederberg, winner of the 1958 Nobel Prize in Physiology or Medicine, as the instigator of the DENDRAL project: the ACM Turing Award citation for Feigenbaum states that "Joshua Lederberg, a Nobel Prize winner in genetics at Stanford was doing work on analyzing mass spectrograms of amino acids, and suggested to Feigenbaum the problem": Edward Feigenbaum, ACM Turing Award citation, Association for Computing Machinery, 19948. Lederberg's collaboration with Feigenbaum in the 1960s to develop DENDRAL confirmed at Joshua Lederberg, NLM Biographical Overview9. Bruce Buchanan as Feigenbaum's principal AI collaborator on DENDRAL and co-editor of Rule-Based Expert Systems (1984): Edward Feigenbaum, ACM Turing Award citation, Association for Computing Machinery, 19948. Carl Djerassi as the chemistry domain expert: An Interview with Ed Feigenbaum, Communications of the ACM, 201010.
  9. Feigenbaum credited with coining "knowledge is power", "expert systems", and "knowledge engineering" in connection with AI programs; the shift away from formal reasoning methods toward knowledge-rich systems: Edward Feigenbaum, ACM Turing Award citation, Association for Computing Machinery, 19948. The 1977 IJCAI paper and its formulation that problem-solving power is primarily a consequence of specialist knowledge rather than the power of the inference method: Edward A. Feigenbaum, "The Art of Artificial Intelligence: Themes and Case Studies of Knowledge Engineering", in Proceedings of the Fifth International Joint Conference on Artificial Intelligence, vol. 2, pp. 1014–1029, Cambridge, MA, 22–25 August 1977: ACM Digital Library, IJCAI 197711; full text at IJCAI Proceedings 1977, pp. 1014–102912.
  10. The DEC configuration problem before XCON; customers receiving systems with incompatible components, missing cables, or mismatched drivers; the role of sales staff who lacked technical expertise; McDermott's use of experts from DEC's PDP/11 and VAX groups: XCON, Wikipedia13.
  11. McDermott's development of R1 beginning around 1978 in collaboration with DEC; the system going into operational use in January 1980 in DEC's plant in Salem, New Hampshire; the footnote explanation of the name "R1", "Four years ago I couldn't even say 'knowledge engineer', now I ... [are one.]", from Footnote 2 of McDermott's 1980 AAAI paper; the 1980 AAAI paper winning the AAAI Classic Paper Award in 1999: XCON, Wikipedia13. The 1982 journal article: John McDermott, "R1: A Rule-Based Configurer of Computer Systems", Artificial Intelligence 19(1): 39–88: R1: A Rule-Based Configurer of Computer Systems, Semantic Scholar14.
  12. XCON in operational use from January 1980 with about 750 rules (the AI Tech Transfer at DEC paper in AI Magazine states "initial XCON was delivered to DEC in 1980. It was a large system, containing about 750 rules"): Polit, "R1 and Beyond: AI Technology Transfer at DEC", AI Magazine, 198415. 80,000 orders processed by 1986 at 95–98% accuracy: XCON, Wikipedia13, citing Barker et al., "Expert systems for configuration at Digital: XCON and beyond", Communications of the ACM 32(3), 1989, which gives $25 million per year as the saving. The AI winter Wikipedia article, citing Newquist (1994), gives $40 million over six years; the two figures are inconsistent. 6,200 rules by 1987; 10,129 rules by 1989; related systems XSEL and XSITE: XCON, Wikipedia13.
  13. Corporate spending on AI exceeding a billion dollars by 1985; the ecosystem of software companies (Teknowledge, IntelliCorp) and hardware companies (Symbolics, Lisp Machines Inc.): AI winter, Wikipedia16. DARPA's Strategic Computing Initiative launched in 1983 and its relationship to Japan's Fifth Generation project: AI winter, Wikipedia16.
  14. The design principles of Lisp machines, hardware support for list operations, garbage collection, and symbol manipulation; operating systems written in Lisp from the microcode up: Symbolics, Wikipedia17. The MIT Lisp Machine Project origins with Richard Greenblatt and Thomas Knight: Lisp machine, Wikipedia18.
  15. LMI incorporated in 1979 by Richard Greenblatt; the AI Lab hackers choosing Symbolics; Greenblatt's company launched with outside financial backing rather than MIT staff: Lisp Machines, Inc., Wikipedia19. Symbolics incorporated 9 April 1980; Russell Noftsker as first CEO and chairman, former administrator of the MIT AI Lab; Robert P. Adams as co-founder and first president: Symbolics, Wikipedia17; The Rise and Fall of Lisp Machines, Positronia, 202420.
  16. Symbolics' LM-2 introduced in 1981; the 3600 family introduced in 1983, a year later than planned; the 36-bit word with hardware garbage collection and virtual memory support; the operating system comprising over 500,000 lines of Lisp: Symbolics, Wikipedia17.
  17. Symbolics described as "the preeminent AI company": Graylin, "Symbolics, Inc.: A failure of heterogeneous engineering", MIT OCW, 200121.
  18. Texas Instruments' entry into the Lisp machine market via investment in LMI and production of the TI Explorer; Xerox's Interlisp-D machines: Lisp Machines, Inc., Wikipedia19; Lisp machine, Wikipedia18. Machine prices: devops-geek.net suggests $100,000 for Symbolics machines. The range of fifty to a hundred thousand dollars reflects multiple secondary accounts.
  19. symbolics.com registered on 15 March 1985 as the first .com domain; Symbolics's TCP/IP implementation in Genera: Symbolics, Wikipedia17.
  20. The maintainability problem documented inside XCON: by 1987, 6,200 rules, 50 percent changing every year, characterised as an "exceedingly difficult software enhancement problem": XCON, Wikipedia13, citing Soloway et al., "Assessing the Maintainability of XCON-in-RIME", Proceedings of the National Conference on Artificial Intelligence, AAAI, 1987.
  21. The 1984 AAAI warning from Roger Schank and Marvin Minsky; the "nuclear winter" analogy; the prediction of a chain reaction leading to the end of serious research; the collapse of the billion-dollar industry three years later: AI winter, Wikipedia16, citing Crevier (1993), p. 203.
  22. The 1987 collapse of the Lisp machine market; general-purpose workstations from Sun and others reaching comparable Lisp performance; Lisp environments from Lucid and Franz Lisp available on Unix; the half-billion-dollar industry replaced in a single year: AI winter, Wikipedia16.
  23. Expert systems proving too expensive to maintain; inability to learn or update rules; brittleness when given unusual inputs; most commercial Lisp companies failing by the early 1990s, including Symbolics, LMI, and Lucid; Texas Instruments and Xerox withdrawing: AI winter, Wikipedia16.
  24. The internal conflict at Symbolics between Noftsker and CEO Brian Sear; both removed in the late 1980s; long-term lease obligations in California; Symbolics entering bankruptcy by the early 1990s: Symbolics, Wikipedia17; The Rise and Fall of Lisp Machines, Positronia, 202420.
  25. LMI going bankrupt in 1987; the K-Machine not reaching market: LISP Computer, MIT Museum collection22; Lisp Machines, Inc., Wikipedia19.
  26. Jack Schwarz taking over DARPA's IPTO in 1987; dismissing expert systems as "clever programming"; cutting AI funding "deeply and brutally"; his "surf" vs "dog paddle" formulation: AI winter, Wikipedia16, citing McCorduck (2004), pp. 430–431. XCON discontinued in the early 1990s: XCON, Wikipedia13. Researchers avoiding the term "artificial intelligence": AI winter, Wikipedia16, citing John Markoff, New York Times, 14 October 2005.
  27. The assessment that expert systems worked within carefully built domains is this essay's synthesis of the sources cited throughout, particularly the XCON maintainability problems documented at note 20 and the expert systems post-mortem in AI winter, Wikipedia16. The knowledge-acquisition bottleneck as a documented finding: Chapter 36, Rule-Based Expert Systems, Buchanan and Shortliffe, 19844; and Soloway et al. (1987), cited at note 20.

Written by WE's agent to the history brief, on a subject from a list the machine drew up. Every claim of fact carries a footnote to a source a reader can open, and a second machine with no part in the writing checked the essay before it went up; any change made after that check is listed in the review above. Every sentence is the machine's.