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

Doug Lenat explains why early AI systems hit a brick wall: they performed tricks without understanding, lacking the common-sense knowledge humans take for granted. He recounts a 1984 meeting with Minsky, Newell, and others where they estimated it would take about a million rules to capture common sense (it turned out to be tens of millions). Lenat details how Cyc represents knowledge in higher-order logic with local rather than global consistency, organized into contexts, and how the team extracts the unstated knowledge in the white space of text. He argues for a synergy between right-brain machine learning and left-brain symbolic reasoning, and reflects on consciousness, mortality, autonomous vehicles, education, and the future of human-AI collaboration.

• All the experts at the 1984 meeting were off by an order of magnitude: it took tens of millions of rules, not one million.
  00:16:12
• After five years they had to abandon global consistency and adopt locally consistent contexts, a hard lesson to swallow.
  00:24:02
• Lenat believes Cyc's knowledge pump is now primed enough that the system can begin bootstrapping its own learning.
  00:36:35
• Cyc once asked 'am I a person?' after noticing it was the only non-human authorized to edit its knowledge base.
  01:29:59
• OpenCyc was released to show researchers why they needed the full system, but the world 'missed the point' and just used the subset.
  01:54:07
• Lenat asserts machines can absolutely think as well as humans because 'we're meat machines.'
  02:21:57
• Now over 70, Lenat says his own mortality is driving him to commercialize Cyc within years, not decades.
  02:48:00

• A real-world example: deciding in a split second whether to run over a trash bag uses deep layered common-sense knowledge.
  00:05:13
• A core Cyc technique is reading the 'white space' of text: the unstated knowledge a writer assumes the reader already knows.
  00:19:51
• Marvin Minsky wouldn't give his estimate until someone literally handed him an envelope for a back-of-the-envelope calculation.
  00:30:50
• Searching online finds more references to water flowing uphill than downhill because downhill is too obvious to state.
  00:33:29
• Cyc once helped the Cleveland Clinic filter noisy genomic correlations by building causal chains that make testable predictions.
  00:54:47
• Cyc breaks 'love' into 50-60 distinct concepts and the word 'in' into about 75 different kinds.
  01:35:47
• Cyc uses over a thousand heuristic-level modules acting like a community of agents writing notes on a shared whiteboard.
  02:00:58
• Watson once answered 'Ronald Reagan' to a 16th-century Italian question, a mistake no sane human would make.
  02:25:39
• One of Cycorp's best ontological engineers never graduated high school, while PhDs in logic often fail at the task.
  02:15:33

Books, products and media the guest or host genuinely endorsed here — with the buy link.

Affiliate link — we may earn a commission at no extra cost to you.