Douglas Lenat: Cyc and the Quest to Solve Common Sense Reasoning in AI | Lex Fridman Podcast #221

Early estimates suggested ~1 million rules would cover human common sense, but Cyc found it needed tens of millions of highly general assertions (e.g., about objects, causality, typical behavior) to handle rare, novel situations robustly.

Simple knowledge graphs and semantic web triples can’t capture nested beliefs, modalities, time, and context; Cyc uses higher-order logic for expressiveness, then thousands of specialized heuristic modules to keep inference fast enough for real-time use.

A single globally consistent knowledge base is impossible at human scale; Cyc partitions knowledge into overlapping “contexts” (like tectonic plates) that are locally consistent but can differ across time, place, physics, or belief systems.

Lenat frames Cyc’s hand-built core as “priming the pump”: once enough foundational common sense exists, the system can use reading, abduction, and interaction to expand itself; starting from too little knowledge leads to GPT‑style plausible but absurd outputs.

Lenat likens ML to the “right brain” (fast pattern recognition) and Cyc to the “left brain” (slow, explanatory reasoning); combining them enables systems that generate hypotheses statistically and then test, explain, and refine them logically.

Lenat’s experience with the MYCIN system convinced him that trust and real utility require step-by-step causal justifications (e.g., in medicine), not opaque scores like “0.83” with no articulated reasoning.

Despite funding cycles, AI “winters,” and shifting fashions, Lenat stayed on Cyc for nearly four decades, designing contracts to keep common knowledge non-proprietary and aiming to leave humanity a shared, extensible common-sense infrastructure.