Smoke and Mirrors ? Gary R. Martins (from Unix Review, Aug 1987, Vol 5 # 8) The quickest way to a real understanding of the field of artificial intelligence (AI) is to start with the First Commandment of AI, taken straight from W.C. Fields: "Never give a sucker an even break, or smarten up a chump!" Most articles on AI and expert systems are expertly designed to keep the suckers---the Department of Defense, and a handful of flabby banking and industrial behemoths---solidly on the hook. The editors of UNIX REVIEW have graciously provided space here for a quick peek behind the smoke and mirrors in the form of three questions addressed to the AI community at large. WHERE IS THE FOUNDATION LITERATURE? Every genuine technology rests upon an intellectual foundation of key concepts, theories, principles, and discoveries, not to mention crucial experiments and engineering development. Basic advances are reported in independently refereed journals and substantive central texts. No such literature exists for AI or expert systems, however. While there is indeed a vast outpouring of "stuff" on these subjects, it constitutes a uniform wasteland of romantic and superficial fantasies about computing, human thinking, corporate management, and other pop topics and is marked by a dogmatic monotony not unlike that of Marxism. Read the Masters, and judge for yourself: "The 5th Generation" by Stanford's Edward Feigenbaum; "The Society of Mind" by MIT's Marvin Minsky (footnote 1); "AM: The Automated Mathematician" by Douglas Lenat, MCC's AI guru; anything at all by Alan Newell, Carnegie Mellon University's chief AI oracle and founding president of the America Association for Artificial Intelligence (AAAI). As you read (borrowed copies of) these works, try to remember that this is the "heavyweight" literature of AI --- the rest is all downhill! There is, in fact, no real scientific or engineering support for the outlandish claims of the AI and expert systems hucksters, which include: o Expert systems methods speed applications development. o Expert systems tools are cost-effective. o Knowledge engineering improves the handling of complex applications. o Expert systems programs are maintainable. o AI has made real contributions to computer science and practice. WHERE IS THE TECHNOLOGY? When we set aside all the hokey new terminology --- knowledge engineering for "amateur programming", knowledge base for a "slow, unfinished database", knowledge acquisition for "talking it over", and so on --- we find there is only a single technological innovation underlying the expert systems hustle: rule-based programming --- the heart and soul of MYCIN, ROSIE, OPS-5, ART, KEE, XCON, and other expert systems tools and triumphs. Unfortunately for the chumps and suckers, this approach to programming was invented, thoroughly evaluated, and finally discarded by the business DP community in the mid-1960s (under the name decision-table programming). Why discarded? Because it doesn't work. There are lots of reasons (footnote 2); let's examine just one of them here, with more brevity than the point deserves. Virtually all procedural languages have some form of IF...THEN... statement. Every programmer knows how to use such statements. But this is the only kind of statement available in expert systems languages. So why should you pay extra for these languages---how does this make them better? It's true that most big-ticket expert systems toolkits throw in other goodies (such as LISP, Prolog, perhaps some "natural language tools", and some graphics toys) to supplement the rule-based language. But these other goodies can invariably be obtained in superior versions (and at far more attractive prices) from real software houses. WHERE ARE THE REAL SUCCESSES? It's important to remember that AI is about the same age as the rest of computing --- it's been around at least since 1955, and maybe a lot longer (footnote 3). While every other area of computing has enjoyed heroic advances since then, AI is still picking over the same dusty chestnuts that seemed so fascinating way back in the good old days (Can computers think? Should robots have a personality? Are computers good models of the human mind?), and the field of expert systems is reinventing biz-DP excitement from two decades ago! Year after year, the technology forecasters predict huge growth for AI, where in fact 95 percent of the national budget for AI consists of DoD handouts. AI pitchmen like to tout a small and often vague list of past and current successes: HEARSAY, AM, MYCIN, XCON/XSEL, PUFF, DRILLING ADVISOR, PROSPECTOR, PILOT'S ASSOCIATE. Such successes prove the value of expert systems technology, it is said. Yet, on closer inspection, these systems may disappoint all but the most dedicated AI worshipers. Some will turn out not to exist at all, or to exist only as ragged and incomplete remnants of abandoned projects! Most will never make it beyond the development stage. A mere handful have achieved a touchingly modest measure of functionality, but only after staggering investments of money and programmer-years. Whether they every actually "work" in some sense or not, AI and expert systems projects seem always to drag on for many years, cost conspicuous fortunes, and wind up embalmed as randomly designed, amateurishly coded, buggy and unmaintainable, badly documented, non-portable, resource-eating kludges. It's an unbroken tradition, stretching back to the earliest days of AI. ADVICE FOR THE PERPLEXED Is AI for you or your organization? Here's a quick test: If you work in a sprawling, chaotic career-driven, fad-mad bureaucracy that thrives on failure and gets to spend other people's money --- the DoD, for example --- the shallow hokum of AI may prove irresistible, and you probably have several major projects under way already. If your boss thinks lending huge sums of money on generous terms to Haiti, Poland, and Nigeria sounds like a good business deal, then his boss probably thinks AI is the key to your industry's future and has in all likelihood already bought up two or three AI stock issues. If your firm just spend $700 million to get Ross Perot off the premises, your boss may believe AI has a big role to play in industrial automation. On the other hand... Footnotes: 1. For an artfully charitable review of this silly new book, see "Leviathan's Program," New York Review of Books, June 11, 1987, p. 33. 2. See G. Martins, "The Overselling of Expert Systems," Datamation, November 1, 1984, for a more detailed explanation. 3. cf. Gulliver's adventures on the flying island of Laputa (Gulliver's Travels, J. Swift, London, 1726).