Melanie Mitchell: Professor of computer science at Portland State University and external professor at the Santa Fe Institute. A former student of Douglas Hofstadter and John Holland, she created the copycat analogy-making architecture and wrote 'Artificial Intelligence: A Guide for Thinking Humans.'
What this episode covers
Melanie Mitchell joins Lex Fridman to question what intelligence really is and whether today's AI is anywhere close to it. She argues that concepts and analogy-making are the core of cognition, drawing on her copycat project built with Douglas Hofstadter. Mitchell is skeptical that brute-force deep learning can reach human-level intelligence, pointing to missing common sense, embodiment, and innate structure. She discusses autonomous driving as a case where the open-ended 'long tail' demands common sense, and critiques the orthogonality view of superintelligence held by Bostrom and Russell. She closes on complexity science, the Santa Fe Institute, and her pride in copycat.
The guest's own work
Artificial Intelligence: A Guide for Thinking Humans
Melanie Mitchell
“she has contributed a lot of important ideas to the field of AI including her recent book simply called artificial intelligence a guide for thinking humans”— Lex Fridman
Analogy-Making as Perception
Melanie Mitchell
“I have a book called analogy making as perception which is a version of my PhD thesis on it”— guest
Also referenced (named, not recommended)
Surfaces and Essences
Douglas Hofstadter and Emmanuel Sander
“he has a really nice quote in in the book by by himself and Emmanuel Sanders called surfaces and essences I don't know if you've seen that book”— guest
Human Compatible
Stuart Russell
“a earlier New York Times op-ed by Stewart Russell which was summarizing his book called human compatible”— guest
Gödel, Escher, Bach
Douglas Hofstadter
“douglas hofstadter author of get olestra bach was your PhD adviser he mentioned a couple times and collaborator”— Lex Fridman
Fluid Concepts and Creative Analogies
Douglas Hofstadter
“there's a Douglas Hofstadter's book called fluid concepts and creative analogies talks in great detail about copycat”— guest
Breakout (Atari)
Atari
“it learned to play the game break out thousand percent better than humans okay that was one of the results and it was great”— guest
Big reveals from this episode
- Mitchell bets we cannot create real intelligence without first understanding our own minds.
- She predicts human-level AI is more than a hundred years away, citing 'a hundred Nobel Prizes' worth of discoveries.
- Frames how to form and fluidly use concepts as the single most important open problem in AI.
- Uses DeepMind's Atari Breakout agent to show systems can be superhuman yet never learn the concept of a 'paddle' or 'ball.'
- Predicts we won't have true fully-autonomous self-driving 'for a very long time' without engineered common sense.
- Calls Stuart Russell's climate-AI-kills-humans scenario incoherent, rejecting the idea intelligence is separable along one dimension.
- Reveals the provocative op-ed title was chosen by New York Times editors, not her.
Worth remembering
- John McCarthy coined 'artificial intelligence' to distinguish it from cybernetics and later regretted the name.
- Graduating in 1990 during an AI winter, Mitchell was advised to write 'intelligent systems' on her CV instead of AI.
- Marvin Minsky famously assigned computer vision as a summer project to undergrads, underestimating its difficulty.
- Hofstadter found a music-generation program terrifying because beautiful music felt core to being human.
- Mitchell has become an embodiment adherent, doubting intelligence can be learned without a body.
- Copycat, started in 1984, makes analogies in an idealized world of letter strings (ABC to ABD, what does IJK become?).
- Kurzweil and Kapoor have a bet that a machine will pass a rigorous Turing test by 2029.
- The Santa Fe Institute, founded in 1984 by Los Alamos scientists, has lived 'on the edge of chaos' with no significant endowment.
- The reductionist promise of the human genome failed because diseases arise from networks of interactions, not isolated genes.