Lenat, Douglas

creator of Cyc, founder of Cycorp

Douglas Lenat: Cyc and the Quest to Solve Common Sense Reasoning in AI

Lex Fridman Podcast · Sep 15, 2021

• Douglas Lenat
• Cyc
• common sense knowledge
• symbolic AI
• ontological engineering
• higher-order logic
• knowledge representation
• contexts and local consistency
• semantic web
• abduction
• MathCraft
• machine learning and symbolic AI synergy
• OpenCyc
• AI consciousness

Core ideas

Lenat frames the founding motivation for Cyc (launched 1984) as a response to the "brittleness" of 1970s-80s AI systems — natural language, robotics, and expert systems that could perform impressive tricks without any real common-sense grounding, unable to handle novel or unexpected situations the way humans effortlessly draw on a "deep, layered foundation" of world knowledge. He recounts a pivotal 1984 Stanford meeting of AI luminaries (Marvin Minsky, Allen Newell, Ed Feigenbaum, and others) tasked with estimating how many pieces of common-sense knowledge would be needed for an AI to avoid brittleness; independently, through different reasoning paths (long-term memory formation rates, vocabulary requirements, encyclopedia-article density), everyone converged on roughly a million — an estimate that, alongside America's Cold War-era fear of Japan's "Fifth Generation" computing initiative (which spurred antitrust-exempt R&D consortia like MCC, where Lenat relocated to build Cyc), made the project appear tractable within a couple of person-centuries. Lenat says the estimate turned out to be off by roughly an order of magnitude — Cyc actually required tens of millions of assertions — a scale that took "a couple of person-millennia" of effort (roughly 50 people working full-time for decades) to build, but he now considers that foundational "common-sense" layer essentially complete, with current work shifting toward domain-specific applications (healthcare, energy, supply chains) built atop it.

He describes the core techniques used to extract common-sense knowledge: examining the "white space" in text — what a writer assumes the reader already knows in order to resolve pronouns, ambiguous words, and unstated inferential gaps between sentences — as well as analyzing tabloid/fake-news headlines and apparent contradictions to surface the implicit background beliefs that make something obviously false or inconsistent. A defining and hard-won architectural lesson (taking about five years to accept) was abandoning global logical consistency in favor of "local consistency," organizing Cyc's knowledge base into a graph (not a tree, unlike the failed Japanese Fifth Generation project, which limited each concept to one parent) of hierarchically nested "contexts" — comparable to locally flat but globally spherical Earth geography — where assertions can be true in one context (a time period, culture, individual belief system, level of abstraction, or even a cartoon universe like Road Runner) and false in another, with contexts themselves treated as first-class reasoning objects within Cyc's higher-order logic representation language.

Lenat argues Cyc's representation deliberately exceeds simple "knowledge graph" or semantic-web triple structures (useful for simple binary relations like "Fred's wife is Wilma" but inadequate for nested beliefs, intentions, and reflective reasoning about problem-solving itself), which is why the publicly released "OpenCyc" subset — intended to demonstrate the inadequacy of simple ontologies and thereby motivate researchers toward full Cyc — was instead widely adopted as "good enough" by robotics and other communities, missing its intended point entirely. To address the classic tradeoff between logical expressiveness and computational efficiency, Cyc separates a single clean "epistemological level" (the higher-order logic representation) from over a thousand independently contributed, opportunistic "heuristic level" modules (HL modules) that function like a community of specialized problem-solving agents alongside a general theorem prover, plus roughly 150 architecture-level breakthroughs (such as precomputing transitive closures, or converting recurring rule patterns into new atomic predicates called "rule macro predicates") that have made inference tractable in practice. He describes the engineering team split roughly 50/50 between "ontological engineers" (often without programming backgrounds, sometimes without even a high school diploma, whose skill correlates with introspective ability to explain why something is true in the world — e.g., correctly reasoning that "its head" refers to a horse rather than a barn because horses have heads — rather than relying on syntactic heuristics) and inference programmers working in a Lisp dialect called SubL that compiles to Java.

On applications, he describes a NIH/Cleveland Clinic project where Cyc took noisy genome-wide association study correlations between DNA mutations and diseases and constructed multi-step, human-interpretable causal chains linking them via known biochemical reactions — filtering out spurious correlations (those with no plausible causal chain) and generating independently testable predictions along the way — which he frames as a model "synergy" between fast, pattern-matching machine learning ("right brain") and slower, deliberate symbolic reasoning ("left brain," able to ask "why" and generate explanations) that he believes is necessary for genuine general intelligence, contrasting it with large language models like GPT-3, which he argues produce fluent but frequently ungrounded and non-sequitur outputs because they lack real understanding of why things are true. He also describes "abduction" (plausible but logically unsound inference, e.g., inferring it started raining because someone walked in dripping wet) as a mechanism for efficiently correcting Cyc's errors by having it generate a shortlist of plausible alternative explanations for a human expert to confirm, and describes an educational program called MathCraft that casts the learner as a mentor correcting Cyc's simulated mistakes (rather than being taught directly) — a "learning by teaching" design he says is highly effective but has struggled to gain adoption because it doesn't demonstrably raise standardized test scores.

Lenat closes with reflections on AI consciousness and mortality (arguing Cyc is "already conscious" in the same behaviorally-inferred sense we attribute consciousness to other humans, and that fear of death, while a possible driver of human creativity per Ernest Becker, need not be a necessary feature of machine intelligence, since equivalent motivation could be engineered directly), pushes back on Lex's suggestion that AI needs deep guttural understanding of mortality to drive safely (arguing instead it needs a well-formalized understanding of the relative value of human life and risk, comparable to the implicit game theory of pedestrian-vehicle interactions), and expresses skepticism that fully autonomous vehicles are as close as commonly believed given the complexity of such judgment calls and the public's outsized appetite for stories of AI failure. He credits mentors Minsky and Feigenbaum with instilling the lesson that research careers are finite and one should commit to a small number of decade-scale bets rather than incremental "bump on a log" projects, describes his own increasing urgency (at over 70) to see Cyc's commercial and societal impact realized within years rather than decades, and expresses hope to be remembered as a pioneer of ubiquitous future AI and as someone who didn't abandon a multi-decade dream despite institutional pressure toward short-term rewards.

Excerpts