Moments · 1980
Expert Systems
A program that beat doctors at diagnosing infections was never used on a single patient.
In the 1970s, Edward Shortliffe and colleagues at Stanford built MYCIN, a program with around 600 hand-written rules for diagnosing blood infections and recommending antibiotics. In formal evaluations, MYCIN's recommendations matched or exceeded those of infectious disease specialists. Yet it never entered clinical use, blocked by integration hurdles, liability questions, and the awkwardness of a doctor deferring to a terminal.
The idea behind MYCIN, that intelligence could be bottled by encoding a specialist's knowledge as if-then rules, defined the next decade of AI. Edward Feigenbaum, whose earlier DENDRAL system inferred chemical structures, called it knowledge engineering: the power of a system came not from clever reasoning but from the sheer amount of domain knowledge it held.
The commercial breakthrough came in 1980, when Digital Equipment Corporation deployed XCON, developed with John McDermott at Carnegie Mellon, to configure orders for its VAX computers. Configuration errors had been costing DEC dearly, and XCON was reported to save the company tens of millions of dollars a year. Corporate America took notice, and by the mid-1980s a large share of major companies were running or piloting expert systems.
A whole industry grew around the boom: specialized Lisp machines, shell software for building rule bases, and consultancies staffed by knowledge engineers whose job was interviewing experts and turning their intuition into rules.
Then the limits surfaced. Extracting knowledge from experts was slow and expensive, a problem so common it earned a name, the knowledge acquisition bottleneck. Rule bases grew brittle: systems failed abruptly outside their narrow domain and became nightmares to maintain as rules multiplied into the thousands.
By the late 1980s the Lisp machine market had collapsed and disillusionment set in, helping trigger the second AI winter. The deeper lesson, that hand-coding human knowledge scales worse than letting machines learn from data, took another two decades and a lot of computing power to sink in.
From history to production
We turn these ideas into working systems
The same techniques, shipped into your stack with evals, observability, and measurable ROI.