DENDRAL, the first expert system

Edward Feigenbaum came to Stanford in 1965 from Herbert Simon's group and wanted a scientific problem hard enough to matter. Joshua Lederberg, a Nobel laureate in genetics, had one: given the fragments a molecule produces in a mass spectrometer, work out the molecule. Lederberg had already written an algorithm to enumerate every possible structure. What was missing was the chemist's knowledge of which structures were plausible.

DENDRAL encoded that knowledge as rules, gathered by sitting with the chemist Carl Djerassi and asking him why he ruled things out. The program generated candidates, tested them against the spectrum, and pruned with the rules. By the early 1970s it was publishing results in chemistry journals and, in narrow classes of compounds, outperforming the experts who had taught it.

Feigenbaum drew a conclusion that shaped the next twenty years: the power of a program lies in the knowledge it holds, not in its reasoning method. That principle produced MYCIN, XCON and the expert-systems industry of the 1980s. It also produced the bottleneck that killed the industry, because knowledge had to be extracted by hand, one rule at a time, from people who did not know how they knew what they knew.