“Theory Formation by Heuristic Search: The Nature of Heuristics II” was written by Douglas B. Lenat of the Stanford Computer Science Department and published in the journal Artificial Intelligence, volume 21, in 1983 (pages 31 to 59). It is the central account of two of the most striking discovery programs of classical AI: AM, which Lenat built as a graduate student, and its successor EURISKO.
AM (short for Automated Mathematician) started from a small set of concepts in set theory and a body of heuristics, rules of thumb about which avenues are likely to be interesting. By repeatedly applying those heuristics, AM proposed and explored new mathematical concepts and conjectures, rediscovering ideas such as prime numbers along the way. The paper frames this as theory formation by heuristic search, where the program navigates a vast space of possible concepts guided by judgments of what is worth pursuing. EURISKO pushed the idea a step further with an accretion model in which the heuristics themselves were treated as objects the program could examine, evaluate, and modify, so it could discover new heuristics, not just new domain concepts.
EURISKO became famous outside the lab for winning a US naval fleet-design tournament (Traveller TCS) two years running by inventing unconventional strategies, and for the practical lesson that programs which edit their own rules need careful guardrails. The work also exposed how much hand-built knowledge such systems required, a realization that led Lenat directly to the decades-long Cyc project to encode common-sense knowledge by hand.
Why a business reader should care: AM and EURISKO are early, concrete demonstrations of software that generates novel ideas rather than merely executing fixed instructions, and the paper’s honest reckoning with how much human-supplied knowledge they needed remains a useful caution for anyone evaluating claims of automated discovery.