AI summary by TwinMind
Melanie Mitchell: Concepts, Analogies, Common Sense & Future of AI | Lex Fridman Podcast #61
Lex Fridman
1h 52m 40s
Introduction and Background
- Melanie Mitchell: Professor of Computer Science at Portland State University, External Professor at Santa Fe Institute
- Background in adaptive complex systems, genetic algorithms, copycat cognitive architecture
- Doctoral work with Douglas Hofstadter and John Holland
- Recent book: "Artificial Intelligence: A Guide for Thinking Humans"
Terminology and Definitions in AI
- Problems with term "artificial intelligence" - means different things to different people
- John McCarthy coined the term to differentiate from cybernetics, later regretted it
- Herbert Simon suggested "complex information processing"
- Alternative terms discussed: cognitive systems, intelligent systems
- Distinction between narrow AI vs artificial general intelligence (AGI)
- Strong AI vs weak AI (John Searle's distinction)
- Moving goalposts problem: once machines accomplish tasks, we redefine intelligence
Predictions and the History of AI
- AI field consistently terrible at predicting future progress
- Marvin Minsky's famous computer vision "summer project" example
- Underestimated complexity of tasks that seem effortless to humans (vision, common sense)
- Mitchell's prediction: human-level AI more than 100 years away
- Quote: "100 Nobel Prizes away" from human-level intelligence
- Current approaches have fundamental limits, need breakthroughs beyond scaling
Current State and Approaches
- Diverse viewpoints in AI community:
- Singularity/transhumanism group: exponential progress, AGI in 30 years
- Deep learning scaling believers: current approaches will go "all the way"
- Hybrid approaches (Gary Marcus): combine deep learning with symbolic AI
- Causality advocates: deep learning lacks causal reasoning
- Cognitive science approach: developmental learning, intuitive physics/psychology
Deep Learning Analysis
- Mitchell surprised by deep learning success but sees fundamental limitations
- Atari game example: system didn't learn concepts like "paddle" or "ball"
- Moving paddle 3 pixels broke the system - shows lack of conceptual understanding
- Current networks treat all parts of input equally, lack dynamic feedback
- Missing: memory, mental models, human-interpretable concepts
- Question whether pure learning from data can achieve human-level intelligence
Analogy Making and Copycat
- Core thesis: "Without concepts there can be no thought, without analogies there can be no concepts"
- Copycat program (30+ years old): makes analogies in letter string domain
- Active perception approach: agents dynamically explore and build understanding
- Workspace architecture: blackboard system where agents post discoveries
- Example: ABC → ABD, what does IJK become?
- Analogy making fundamental to all cognition - every concept recognition is analogy
- How to form and fluidly use concepts = most important open problem in AI
Concepts and Mental Models
- Concepts are fundamental units of thought embedded in conceptual spaces
- Mental models enable simulation and prediction
- All reasoning potentially reducible to analogy making
- Enormous network of interconnected concepts underlies common sense
- Challenge: most knowledge invisible to us, not in Wikipedia
- CYC project limitations: wrong approach to encode common sense logically
Embodied Intelligence
- Growing belief that intelligence requires embodiment
- Need both ability to interact with world AND body to preserve
- Intelligence inseparable from emotions, self-preservation, social aspects
- 70 years of AI tried to separate thinking from "other stuff"
- Emotion not obstacle to intelligence but integral part
Autonomous Vehicles Case Study
- Illustrates difficulty of real-world AI applications
- Long tail problem: infinite edge cases not in training data
- Current systems overly conservative, brake frequently
- Two approaches: LiDAR-based (Waymo) vs vision-only (Tesla)
- Will likely require geofenced areas with infrastructure support
- Full autonomy requires common sense reasoning about human behavior
Complexity Science and Santa Fe Institute
- Complex systems: emergence from simple interactions at scale
- Reductionism limitations: can't understand intelligence from individual neurons
- Santa Fe Institute founded 1984 by Los Alamos scientists
- Interdisciplinary approach to big questions
- Cellular automata example: simple rules → unlimited complexity
- Educational programs: summer schools, online courses, public lectures
Existential Risk and Super Intelligence
- Critique of Bostrom/Russell superintelligence concerns
- Argument: superintelligence can't be intelligent in one dimension while lacking others
- Orthogonality hypothesis doesn't make sense
- More pressing existential threats: climate change, nuclear weapons, pandemics
- Timeline too distant (>100 years) to prioritize over immediate concerns
- Intelligence is holistic - can't separate rationality from values/emotions
Testing Intelligence
- Turing Test still best measure of intelligence
- Kurzweil-Kapoor bet: machine will pass Turing Test by 2029
- Real Turing Test requires deep common sense understanding
- Natural language conversation reveals conceptual understanding
Key Insights and Future Directions
- Need understanding of human cognition to build AI
- Dynamic perception with feedback loops essential
- Mental models and conceptual spaces fundamental
- Embodied, social, emotional aspects integral to intelligence
- Current scaling approaches insufficient for general intelligence
- Multiple breakthroughs needed across cognitive science, neuroscience, AI
