In his 1977 Computers and Thought Lecture, Douglas B. Lenat presented a provocative thesis: the "spark" of human discovery is not an inexplicable gift but the result of applying a vast library of informal "rules of thumb," or heuristics, to navigate complex search spaces. This lecture, and the accompanying AM (Automated Mathematician) program, remains a cornerstone in the history of Artificial Intelligence, marking a shift from rigid problem-solving to open-ended theory formation.
The Challenge of Open-Ended Discovery
During the 1970s, AI research was largely bifurcated. "Hard" AI focused on rigorous mathematical models for simple phenomena, while "soft" AI used descriptive prose for complex human behaviors. Lenat sought a middle ground: an operational, predictive model of human creativity.
He aimed to move AI beyond well-structured problems, such as chess, toward "ill-defined" discovery. The central challenge was determining how a machine could decide what is "interesting" or "worth naming" without a human explicitly defining the goal. Lenat proposed that intelligence is essentially "efficient searching" guided by a common core of heuristics.
Methodology: The AM Program
Lenat’s approach was built on the paradigm of "Man as a symbolic information processor." The experimental model, AM, was initialized with 115 basic concepts from finite set theory (e.g., sets, lists, equality). Unlike traditional programs, AM did not have a specific problem to solve; it performed open-ended research.
Concept Representation and the Agenda
Concepts in AM were stored as "facets" or slots, including Definitions, Examples, Generalizations, and Worth. The system’s execution was managed by an Agenda—a prioritized queue of "jobs" (e.g., "Fill in examples of Primes").
Heuristics (approximately 250 in total) would propose new jobs and provide "reasons" for them, which determined their priority. AM also utilized inheritance; heuristics attached to general concepts (like "Relation") were automatically inherited by more specific ones (like "Composition"), allowing the system to locate relevant rules efficiently.
The Ubiquity of Heuristics
A key insight of Lenat’s work is the universality of these rules. He argued that the same heuristic used to invent a "wire cheese cutter"—by examining the extreme case of a relationship—is the same rule that leads to the discovery of prime numbers by examining the extreme cases of the "divisors-of" function.
AM used these heuristics to judge "taste" or "aesthetics." For example, if two disparate concepts were suddenly found to be closely related, the "interestingness" value of both concepts would be boosted. This suggests that a common core of informal heuristics underlies diverse domains of human thought, from everyday planning to high-level mathematical invention.
Key Results and Synthesis
In a single one-hour CPU run on a PDP-10, AM executed 200 jobs. On average, it considered 35 rules per job, with about a dozen actually "firing" to create or modify concepts.
- Mathematical Discovery: Starting from basic set theory, AM successfully "discovered" natural numbers, arithmetic operations (addition, multiplication), prime numbers, and even complex conjectures like Goldbach’s Conjecture (every even number is the sum of two primes).
- Concept Synthesis: AM tripled its initial knowledge base, synthesizing approximately 230 new concepts. While Lenat noted that about half of these were "losers" (not useful or interesting), the program demonstrated a remarkable ability to navigate toward mathematically significant ideas using only heuristic guidance.
Limitations and Future Directions
Despite its success, AM encountered "heuristic drift." As the program moved further away from its original set-theoretic primitives into complex number theory, its general heuristics became too "weak" to provide effective guidance. It lacked the domain-specific expertise required for higher-level mathematics.
Furthermore, Lenat identified a distinction between a proposer and a prover. AM was excellent at theory formation—proposing interesting conjectures—but had no ability to formally prove them. These insights led to the identification of a need for "meta-rules" (heuristics that can create or modify other heuristics), a concept Lenat explored in his subsequent project, EURISKO.
Ultimately, the 1977 lecture established that discovery is not a mystical process but a manageable, heuristic-driven search, providing a foundational framework for the study of machine creativity.
Summary updated