Expert Systems
🩺 Expert Systems
🔧 What Is an Expert System?
An expert system is a program built from two separable pieces: a knowledge base of hand-coded if-then rules capturing a human expert's domain-specific know-how, and an inference engine — a general-purpose piece of software that applies those rules to a specific situation to reach a conclusion, independent of what the rules actually say. Separating the two mattered enormously: the same inference engine could be reused across completely different domains just by swapping in a new rule set, which is exactly what let the "expert system shell" industry (later in this chapter) exist as a business at all.
🧪 DENDRAL (1965) — The First True Expert System
Built at Stanford starting in 1965 by Edward Feigenbaum, Bruce Buchanan, and Nobel laureate geneticist Joshua Lederberg, DENDRAL inferred the molecular structure of organic compounds from raw mass spectrometry data, using hand-encoded chemistry rules to narrow down which molecular structures could plausibly produce a given spectrum.
DENDRAL's success led Feigenbaum to a philosophy he called the "knowledge is power" hypothesis: intelligent behavior comes primarily from having a large amount of specific, narrow domain knowledge — not from a single clever general reasoning method. This is a direct, quiet reply to Chapter 4's physical symbol system hypothesis. It doesn't reject symbol manipulation; it reframes what actually does the work. Not a more powerful general reasoner, but more — and narrower — facts.
💊 MYCIN (1972) — Diagnosing Blood Infections
Led by Ted Shortliffe at Stanford, MYCIN used roughly 600 hand-coded rules to diagnose bacterial blood infections (such as bacteremia and meningitis) and recommend antibiotic treatment, with dosages adjusted for the patient's weight. To handle medicine's inherent uncertainty, MYCIN attached a certainty factor — a numeric confidence score — to each rule and conclusion, an early, pragmatic alternative to full probability theory.
In blind evaluation studies, MYCIN's treatment recommendations were judged acceptable more often than those of junior doctors, and were competitive with infectious-disease specialists. And yet MYCIN was never adopted for actual clinical use. The obstacles weren't technical: unresolved liability (who is responsible if a computer's medical advice turns out wrong?), the difficulty of fitting a rule-based advisor into real hospital workflow, and the ongoing effort required to keep 600 rules current as medical knowledge evolves. A capable system and a deployed system turned out to be two very different achievements.
💰 XCON / R1 (1980) — The Commercial Proof Point
Built by John McDermott at Carnegie Mellon and deployed by Digital Equipment Corporation (DEC), XCON (also called R1) automatically configured customer orders for DEC's VAX minicomputers — translating a sales order into the complete, correct list of components, catching errors human sales engineers routinely made. Unlike DENDRAL and MYCIN, XCON wasn't a research demonstration; it was a real production system with a real dollar figure attached.
By 1986, XCON was processing roughly 80,000 orders a year and was credited with saving DEC on the order of $40 million annually. This was the concrete, widely publicized evidence the field had been missing since Chapter 5's funding collapse: proof that a symbolic, rule-based system could deliver direct, measurable business value — not someday, but right now, in production.
🌍 The Boom, and Its International Reach
XCON's visible success — alongside DENDRAL and MYCIN's academic credibility — triggered a genuine commercial gold rush. Startups like Teknowledge and Intellicorp sold "expert system shells": the inference-engine half of the architecture, packaged as a reusable product a company could plug its own domain rules into, without building an inference engine from scratch. Because LISP was AI's dominant programming language, companies including Symbolics and Lisp Machines Inc. began selling specialized workstation hardware — LISP machines — built specifically to run it fast.
The boom wasn't confined to the US. In 1982, Japan launched the government-funded Fifth Generation Computer Systems project, an ambitious state effort explicitly aimed at leapfrogging the West in AI and logic-programming hardware. That move spurred competitive funding responses abroad — DARPA's Strategic Computing Initiative in the US, the Alvey Programme in the UK — turning expert systems into something closer to a national-competitiveness race than a purely academic pursuit.
📜 Narrowing the Ambition
| Symbolic AI's General Ambition (Chapter 4) | Expert Systems' Narrow Focus (This Chapter) |
|---|---|
| Aimed at general intelligent action — the physical symbol system hypothesis | Deliberately abandoned generality, targeting one professional domain at a time |
| Knowledge acquisition bottleneck was fatal at "common sense" scale | The bottleneck became manageable once the knowledge needed was just one expert's few hundred rules |
| Brittle the moment SHRDLU's blocks-world boundary was crossed | Still just as brittle outside its domain — but the domain (VAX orders, blood infections) was valuable enough to justify the narrowing |
It's worth being precise about what actually happened here: the knowledge acquisition bottleneck and brittleness that ended Chapter 5 weren't solved — they were sidestepped, by shrinking the target until hand-encoding became feasible again. That's a genuinely useful engineering trick, and it built real, profitable systems. But it also means the entire 1980s boom — the shells, the LISP machines, the national funding races — was built on the same fundamentally brittle foundation as SHRDLU, just aimed at smaller targets. What happens when a hardware and software industry has been built entirely around betting that this narrowing trick keeps paying off is exactly where the next chapter picks up.
🤔 Questions to Sit With
Feigenbaum's "knowledge is power" hypothesis and Chapter 4's physical symbol system hypothesis aren't opposites — but they emphasize very different things. Does narrowing a domain until hand-encoding becomes feasible actually address the knowledge acquisition bottleneck, or does it just relocate the boundary where the system breaks?
MYCIN outperformed junior doctors in blind testing but was never used clinically, largely over liability and workflow concerns rather than raw competence. Does that gap between "tested capability" and "real-world deployment" sound familiar from anything happening with AI systems today?
Japan's Fifth Generation project triggered competitive funding responses in the US and UK, echoing the Cold War funding dynamics that drove machine translation research in Chapter 5. Is a competitive, race-driven funding environment more likely to produce durable progress, or more likely to inflate a bubble that eventually bursts?
🎯 What's Next
Next chapter: The Second AI Winter — how the very commercial infrastructure this chapter describes, the expert system shells and the specialized LISP machines built to run them, collapsed almost as fast as it rose.