Interview
Douglas Lenat: Cyc and the Quest to Solve Common Sense Reasoning in AI | Lex Fridman Podcast #221
- Mission and Core Problem: Doug Lennett's Psyche project, active since 1984, aims to solve the fundamental AI challenge of acquiring and utilizing "common sense" knowledge (general world knowledge) to enable machines to reason, understand, and handle novel situations.
- Scale of Knowledge: In a 1984 meeting with experts like Marvin Minsky and Ed Feigenbaum, the group estimated that a system requires approximately 1 million facts/rules to match human common sense; Lennett later revised this estimate upward to tens of millions after realizing early models were an order of magnitude too small.
- Current Status: After nearly 40 years and thousands of person-hours of effort, the core common sense base is largely complete, with current work focusing on domain-specific applications (e.g., healthcare, oil pipelines) built upon this foundational "lattice" of knowledge.
- Representation Strategy: The system uses higher-order logic (rather than simple binary triples or knowledge graphs) to represent complex assertions, allowing for the expression of contexts, beliefs, timeframes, and nested reasoning (e.g., "Juliet thinks that Romeo believes...").
- Context Handling: To manage inconsistencies and paradoxes inherent in human language, Psyche adopts a local consistency model (like "tectonic plates"), where knowledge is grouped into specific contexts; global consistency is abandoned because real-world knowledge contains inherent contradictions.
- Knowledge Acquisition Techniques:
- Reading "White Space": The system analyzes what writers assume readers already know (e.g., pronoun resolution, inferred events between sentences) to extract unstated common sense.
- Abduction: When the system makes an error, it generates a list of plausible alternative hypotheses ("abductive leaps") that could explain the error, allowing human supervisors to select the correct one efficiently.
- Human-in-the-Loop: Currently, humans (ontological engineers) play a critical role in verifying and expanding the knowledge base, though the system aims to "prime the pump" to eventually bootstrap its own learning.
- Inference Efficiency: To solve the "heuristic problem" (reasoning efficiently), the system separates the epistemological language (high-order logic) from heuristic modules (1,000+ specialized agents that use shortcuts, caching, and meta-reasoning to speed up inference).
- Synergy with Machine Learning: Lennett proposes a "left brain/right brain" architecture where machine learning (statistical pattern recognition) generates hypotheses, and Psyche (logical reasoning) evaluates their causal plausibility and consistency, as demonstrated in a collaboration with the NIH/Cleveland Clinic to validate GWAS medical correlations.
- Open Source and "Open Psyche": A simplified subset of the knowledge base (binary relations/triples) was released as "Open Psyche" to demonstrate the utility of semantic graphs; however, Lennett argues this was a strategic misstep as researchers mistook the simplified version for the full system, missing the critical value of the underlying rules and common sense.
- Philosophical Stance on Intelligence: Lennett asserts that machines can think and possess a form of consciousness, defined not by mystical qualities but by the ability to represent beliefs, goals, and the state of the world; he argues a physical body or fear of mortality is not strictly necessary for general intelligence.
- Educational Impact: The project highlights a gap in modern education, noting that "ontological engineering" skills (teaching AI to reason) are often independent of formal academic training and can be cultivated in individuals with unique pattern-recognition abilities.
- Future Outlook: Lennett predicts that general AI will augment human intelligence, creating a "hybrid" species capable of overcoming global challenges (climate change, poverty) that pure biological evolution cannot solve.
- Legacy Goal: Lennett hopes to be remembered as a pioneer who persisted for decades on a "hard" AI problem when others pursued short-term academic or commercial gains, ultimately building the knowledge infrastructure for ubiquitous, general superintelligence.