Mitchell, Melanie

professor of computer science at Portland State University; external professor at the Santa Fe Institute; author of "Artificial Intelligence: A Guide for Thinking Humans"

Melanie Mitchell: Concepts, Analogies, Common Sense & Future of AI

Lex Fridman Podcast · Dec 28, 2019

• Melanie Mitchell
• analogy making
• Copycat cognitive architecture
• common sense reasoning
• deep learning limitations
• embodied cognition
• AI existential risk
• orthogonality thesis
• complexity science
• Santa Fe Institute

Core ideas

Mitchell opens by questioning the term "artificial intelligence" itself — noting John McCarthy coined it partly to distinguish the field from cybernetics and later regretted it, while alternatives like Herbert Simon's "complex information processing" are equally vague — and traces a recurring historical pattern where a machine achieving some task (chess, and later Go) forces a retroactive redefinition of what "real" intelligence must mean, since the achieved task no longer counts. She estimates human-level or superhuman general intelligence is "more than a hundred years" away, quoting a colleague's framing that it is "a hundred Nobel Prizes away," and argues current deep learning approaches, however impressive, face fundamental limits rooted in supervised learning's dependence on huge labeled datasets and its lack of causal or common-sense understanding — while candidly noting she has been repeatedly surprised by how far "brute force" data-driven approaches have gotten despite bearing no resemblance to how human minds actually work.

The intellectual core of the conversation is Mitchell's claim — inherited from her PhD advisor Douglas Hofstadter — that analogy-making, not any narrower reasoning technique, is the foundational mechanism of thought: "without concepts there can be no thought, and without analogies there can be no concepts." She describes her own contribution, the Copycat program (built starting in 1984 in an idealized "letter-string" domain, e.g., inferring what "IJK" becomes if "ABC" becomes "ABD"), as an attempt to model perception itself as fluid, active analogy-making rather than passive feed-forward processing — using an agent-based "workspace" (inspired by old blackboard systems) where competing agents dynamically decide what to look at next based on what they've already noticed, with perception and analogy-making treated as inseparable. She generalizes this to ordinary human cognition: recognizing a person, walking a dog, or even conducting "essentially the same" conversation twice are all, in her view, acts of analogy against internal mental models — models she believes function as generative simulations that produce expectations and actively guide where attention goes, not passive templates compared after the fact.

Mitchell argues the central open problem in AI is "how to form and fluidly use concepts," and that most of what humans rely on for common sense is invisible precisely because it's automatic and unconscious (citing Marvin Minsky's famously over-optimistic 1966 summer vision project as an early failure to appreciate this). She is skeptical that large hand-built or scraped knowledge bases (ConceptNet, Cyc) can capture this "invisible" layer of intuitive physics, psychology, and "intuitive metaphysics" (the built-in assumption that objects and causality exist at all), arguing Douglas Lenat's decades-long Cyc project failed because symbolic encoding is fundamentally the wrong representational approach — while stressing that nobody currently knows what representation would work instead, only that something beyond scaling current architectures is likely needed. She illustrates the concept-transfer problem using DeepMind's Atari-playing agent, which mastered Breakout via reinforcement learning but failed completely when researchers moved the paddle a few pixels — evidence, in her reading, that the system had learned surface statistical regularities rather than the actual concepts of "paddle," "ball," or "tunneling," and a caution against anthropomorphizing what such systems have really internalized. She increasingly leans toward embodiment being necessary for genuine understanding (going further than her former advisor's more purely "grounding"-focused view) and connects this to human intelligence's deep entanglement with emotion, self-preservation, and social motivation — arguing these aren't obstacles to rational thought but likely load-bearing components of it, since predicting others' behavior seems to require a "theory of mind" built from simulating one's own motivations and feelings.

On autonomous vehicles, Mitchell argues driving is a paradigm case of the "long tail" problem: systems handle normal situations well but struggle with the effectively infinite space of edge cases that require genuine common sense (a version of the Atari "paddle moved" problem) rather than more training data on typical scenarios — predicting that practical progress will come less from achieving full human-like driving intelligence and more from redefining "autonomous" to mean heavily instrumented, constrained environments rather than the popular sci-fi vision of a car that handles literally anything. On AI existential risk, Mitchell explains the argument of her New York Times op-ed responding to Stuart Russell: she rejects Nick Bostrom's "orthogonality thesis" (that a system's intelligence and its goals/values can vary independently, as in his paperclip-maximizer thought experiment) as incoherent, arguing that a system intelligent enough to solve climate change end-to-end would necessarily also be intelligent enough to recognize humans don't want to be exterminated in the process — her broader claim being that intelligence isn't cleanly separable into independent dimensions (raw optimization power vs. values vs. common sense). She agrees more pressing existential risks (nuclear weapons, climate change, pandemics) deserve priority over speculative superintelligence risk, which she thinks is centuries rather than decades away, while stressing she isn't criticizing researchers like Russell or Bostrom for taking the topic seriously, just their specific framing of what "super intelligence" would actually look like.

Turning to complexity science more broadly (the subject of an earlier book and her ongoing affiliation with the Santa Fe Institute, founded in 1984 by physicists and chemists frustrated with disciplinary silos), Mitchell describes the field's central puzzle as understanding emergence — how simple, well-understood low-level rules (a single neuron, a cellular automaton's transition rule) generate complex, often unpredictable higher-level phenomena that resist explanation by reductionism (illustrated by the Human Genome Project's failure to yield straightforward disease explanations, since it's gene interactions and networks, not individual genes, that matter). She recalls cellular automata as a personally humbling and awe-inspiring research area precisely because such minimal rule-sets can produce unlimited apparent complexity, and credits Hofstadter's methodological lesson — radically idealizing a problem down to its essential core while preserving what actually matters (as Copycat did for analogy-making) — as the throughline of her career, comparing this to AI's historical cycle of abandoning then returning to idealized "micro-world" problems like the blocks world.

Excerpts