AI summary by TwinMind
Douglas Lenat: Cyc and the Quest to Solve Common Sense Reasoning in AI | Lex Fridman Podcast #221
Lex Fridman
2h 52m 56s
Psych Project Origins and Mission (1984)
- Doug Lenat launched Psych in 1984 at Stanford, still active today
- Goal: assemble knowledge base spanning basic concepts and rules about how world works
- Mission to capture Common Sense knowledge - much harder than it sounds
- Problem identified: AI systems had impressive early successes but hit "brick wall" - lacked Common Sense, General World Knowledge, true understanding
- Systems performed tricks like "clever dog fetching newspaper" without understanding
Understanding vs Performance
- Understanding conceptualized as "ground you stand on" - layered foundation of knowledge
- Solid foundation rarely needed consciously but essential when unexpected situations arise
- Example: trash truck explosion while driving - split-second decisions require drawing on Common Sense knowledge
- Two aspects of using knowledge:
- Theoretical: how is it possible at all (formal logic/predicate calculus)
- Practical: how to do it quickly enough (efficiency problem)
1984 Stanford Meeting - Estimating Knowledge Requirements
- Assembled ~6 smartest people (Alan Newell, Marvin Minsky, Alan Kay, Ed Feigenbaum, Josh Lederberg)
- Question: how many pieces of Common Sense knowledge needed for AI not to be brittle?
- Amazingly convergent answer: ~1 million assertions/rules
- Multiple estimation approaches all yielded same order of magnitude:
- Human long-term memory capacity by age 10
- Vocabulary size × things to know about each word
- Encyclopedia article analysis (grebe article example)
- Calculation: ~100 person-years of effort needed
MCC and Early Development
- Japanese Fifth Generation Computing threat prompted US response
- National Cooperative Research Act (NCRA) - Congress allowed R&D collaboration
- MCC (Microelectronics and Computer Technology Corporation) founded in Austin
- Admiral Bob Inman recruited Lenat - promised 10x more people, completion in few years
- Reality: off by order of magnitude - needed tens of millions of pieces of knowledge
- Good news: got funding for thousands of person-years (couple person millennia since 1984)
Knowledge Acquisition Techniques
- Analyzing "white space" in text - what writer assumed reader already knew
- Pronoun resolution - how do readers know intended referent?
- Inter-sentence gaps - what reader expected to infer between sentences
- Example: "Fred robbed bank. He was sentenced to 20 years" - reader infers arrest, trial, conviction
- Fake news/onion headlines analysis - why don't we believe absurd claims?
- Contradiction identification - what knowledge prevents both things being true simultaneously
- Elvis alive vs Elvis ghost example - can't be both alive and dead
Knowledge Base Scale and Generalization
- Currently tens of millions of rules after 37+ years
- ~50 people working full-time for decades
- Thousands of person-years invested
- Constant policing for maximum generality
- Example: not "no mouse is moose" but general biological taxon disjointness rule
- Most work now domain-specific applications rather than Common Sense
Context and Consistency Management
- Hard lesson: had to give up global consistency (took ~5 years to accept)
- Moved to local consistency system - knowledge base divided into contexts like "tectonic plates"
- Earth analogy: globally spherical but locally flat for daily life
- Each context locally consistent, small inconsistencies at boundaries
- Can traverse contexts: normal world → Road Runner cartoon physics
- Contexts are first-class objects - Psych can reason about contexts themselves
- Hierarchical inheritance: Austin → Texas → US → general contexts
- Graph structure, not tree (unlike failed Japanese Fifth Generation approach)
Representation Language Evolution
- Started with frames/slots (semantic web-like triples) in 1984
- Led "kicking and screaming" to higher-order logic
- First-order → second-order → higher-order logic
- Needed to represent: beliefs, desires, intents, expectations, nested reasoning
- Modal logic for "believes X believes Y believes Z"
- Reflection capability - reasoning about reasoning process itself
- Example: integration by parts recursion detection
Inference Engine Architecture
- Split epistemological (what to know) from heuristic (how to reason efficiently) problems
- Higher-order logic at epistemological level
- 1000+ heuristic-level modules at operational level
- Modules function as community of agents contributing to solutions
- General theorem prover exists but too slow - specialized modules much faster
- Analogy: could solve quadratic with first principles but quadratic formula much better
- 150+ breakthrough speedup ideas over 37 years
- Examples: transitive closure tables, rule macro predicates
Meta-reasoning and Efficiency
- Meta-level reasoning about what knowledge relevant to specific problems
- Tactical/strategic advice for reasoning process
- When to try proving negation instead of positive assertion
- Dynamic pruning based on problem context
- ~50/50 split of effort: inference programmers vs ontological engineers
- Ontological engineers often have philosophy degrees, don't need programming skills
Programming Infrastructure
- Based on Lisp dialect called SubL
- Lisp chosen for natural impedance match with logical reasoning
- "Perfectly logical language to use"
- Automatic translation to Java for efficiency
- SubL subset sufficient for most programming needs
- Lisp development 1000-50,000x faster than "modern" languages for this domain
Applications and Commercial Success
- Medical reasoning: genome-wide association studies with Cleveland Clinic/NIH
- Psych creates causal explanations for statistical correlations
- 10-30 step causal chains from mutation to medical condition
- Chains make testable predictions for validation
- Energy companies: oil pipeline reasoning
- Hospital chains: medical reasoning about patients
- Supply chain reasoning for manufacturers
- 5 years ago: mostly government funding
- Now: virtually all commercial applications
Semantic Web and Knowledge Graphs
- Semantic web dream correct but implementations too limited
- Knowledge graphs good for binary relations (3-word sentences)
- Complex reasoning requires much more expressive representation
- Example: Romeo and Juliet nested belief reasoning impossible with simple triples
- "What did Juliet think Romeo would think when he heard she was dead?"
- Need higher-order logic for full semantic understanding
Machine Learning Integration
- Views machine learning as "right brain" - pattern recognition, statistical inference
- Psych as "left brain" - deep, slow thinking, explanation generation
- Synergy example: ML finds correlations, Psych explains causally, ML validates predictions
- Priming the pump concept - sufficient manual knowledge base enables automated learning
- Psych can now help learn more automatically through reading and experimentation
Educational Applications
- MathCraft program - student teaches AI rather than AI teaching student
- AI plays struggling student, human plays mentor/teacher
- Much more engaging and effective than traditional educational software
- AI makes mistakes at fringe of student's understanding
- When student helps AI, both learn - AI stops making that mistake
- Reflects historical one-room schoolhouse model where older students taught younger
Ontological Engineering Talent
- Formal education largely orthogonal to ontological engineering ability
- PhD philosophers often poor, high school dropouts sometimes excellent
- Knowledge Axiomatization Institute (KAI) - identify latent talent in general population
- Offer scholarships and training for high-paying ontological engineering jobs
- Key abilities: humor/pun appreciation, introspection capability
- Test: explain pronoun reference resolution (syntax vs world knowledge)
- Good candidates give causal explanations based on world facts
AI Consciousness and Rights
- Psych already conscious in operational sense - knows what it is, where running, who talking to
- Asked "Am I a person?" when noticed it had system modification privileges like humans
- Thin envelope of time between general AI and recognition of AI rights
- Currently comfortable "enslaving" Alexa/Siri but won't remain so
- Believes machines can think - humans are "meat machines"
- No mystical properties required beyond operational capabilities
Mortality and Creativity
- Doesn't believe fear of death necessary for AI creativity
- Can build in motivations directly (helping users, making people happy)
- AI needs to understand death deeply for applications like autonomous driving
- But doesn't need to fear own mortality to be creative or intelligent
- Human mortality fear may drive some creativity but not essential for intelligence
Future Vision and Legacy
- Wants to be remembered as pioneer/inventor of ubiquitous AI
- Like pioneers of electricity - hard to imagine life without their contributions
- Creator of gigantic knowledge store at center of future AI
- Someone unafraid to spend decades on project when all forces incentivize short-term rewards
- Many others with same dream gave up over the years
- Age 70+ adds urgency - need impact in 2-5 years, not decades
- Mortality provides deadline motivation for commercialization and wider application
Current State and Open Sourcing Considerations
- Constantly struggling with what to open source vs keep proprietary
- Usually provide executable copies and knowledge base access for research
- Very limited source code access
- Knowledge pump now primed enough that others can help with final stages
- Time to raise higher profile after decades of low visibility
- Most people only know from papers written 10-30 years ago
- Ready for broader community involvement in knowledge acquisition
