New Lex Fridman Insight: Douglas Lenat: Cyc and the Quest to Solve Common Sense Reasoning in AI
Sent June 11, 2026
Key Insights
- Psyche's common sense AI project began in 1984, requiring tens of millions of knowledge pieces, far more than the initial estimate of one million.
- The shift from global to local consistency in Psyche's knowledge base was crucial to handle real-world inconsistencies.
- AI systems must minimize error rates in critical applications, as even a 1% error rate can be unacceptable.
- Psyche's programming in Lisp is significantly faster, up to 50,000 times, than modern languages.
- AI doesn't need a physical body to be considered intelligent, but must understand body-related concepts.
How the conversation moved
Lex framed the discussion around the ambitious goal of the Psyche project, initiated in 1984, to encode common sense knowledge into AI systems. Doug Lenat explained the initial underestimation of the task, where experts believed a million pieces of knowledge would suffice, only to later realize that tens of millions were needed. This revelation highlighted the complexity of common sense and the challenges in making AI systems less brittle and more human-like in understanding.
Lenat detailed the shift from a model of global consistency to one of local consistency within Psyche's knowledge base. This change was necessary to accommodate the inherent contradictions in real-world knowledge. He emphasized the importance of efficient algorithms to retrieve information quickly, a critical factor in AI's ability to make real-time decisions. The conversation also touched on the historical context, including the influence of the Japanese Fifth Generation Computing Effort on U.S. research policies.
Despite the depth of the discussion, Lex did not challenge the underlying assumption that encoding common sense is the best approach to AI understanding. A potential counterargument could be the exploration of alternative methods, such as machine learning models that learn from vast data without explicit common sense encoding. Lenat also noted the dual nature of technology, like the internet, which can both enhance and diminish global understanding, but this point wasn't deeply contested.
The conversation concluded with a look at the future of AI, where Lenat expressed optimism about AI's potential to augment human intelligence and challenge existing paradigms. He highlighted the role of AGI in identifying blind spots in human understanding and emphasized the importance of pursuing long-term innovative ideas. The discussion underscored the need for AI to understand concepts related to human experience, even if it doesn't possess a physical form.
Surprising moments
In-depth
Common Sense Knowledge in AI
- Psyche began in 1984 to encode common sense knowledge, initially underestimated at one million pieces.
- The project requires tens of millions of knowledge pieces, highlighting the complexity of common sense.
- Psyche shifted from global to local consistency to manage real-world knowledge contradictions.
AI and Human Intelligence
- AI can enhance human decision-making and critical thinking.
- Error minimization in AI is crucial, especially in fields like medicine.
- AI doesn't need a physical body but must understand body-related concepts.
Programming and Efficiency
- Lisp's use in Psyche offers up to 50,000 times faster development than modern languages.
- Efficient programming accelerates AI development and innovation.
Notable Quotes
The programs didn't have what we would call common sense. They didn't have general world knowledge.
Still open
- Lex asked how many pieces of common sense knowledge are necessary for AI to avoid brittleness, highlighting uncertainty in this estimation.