Lex Fridman PodcastFrançois Chollet: Measures of Intelligence | Lex Fridman Podcast #120
CHAPTERS
- 0:00 – 5:10
Framing the problem: why measuring intelligence matters
Lex sets the stage for a rigorous, benchmark-driven conversation about François Chollet’s paper “On the Measure of Intelligence.” He contrasts narrow-ML progress with the AGI community’s philosophical leanings and argues for bridging the two with actionable tests.
- •Motivation for formal definitions and benchmarks for general intelligence
- •Tension between narrow AI benchmarks and broader AGI ambitions
- •Why measurement should guide research (not just describe it)
- •Overview of the episode’s focus on Chollet’s paper
- 5:10 – 7:58
Early influences: Piaget, Hawkins, and intelligence as development
Chollet describes formative ideas from Jean Piaget and Jeff Hawkins. The emphasis is on intelligence as something constructed in stages and cognition as hierarchical prediction and representation.
- •Piaget’s staged development view of intelligence
- •Hawkins’ hierarchy of temporal prediction modules
- •How early cognitive science metaphors shaped AI thinking
- •Deep learning as a practical instantiation of hierarchical ideas
- 7:58 – 14:05
Language, memory, and cognition: language as an operating system
The discussion moves from memory retrieval to language’s role in structuring thought. Chollet argues language is not the foundation of cognition, but an upper layer that lets us query, program, and retrieve memories deliberately.
- •Brain as sparse-access memory vs neural nets as distributed storage
- •Language as a tool to query and ‘program’ memory
- •Disagreement with Chomsky: language sits atop deeper cognition
- •Pre-linguistic thought in space, motion, and action
- 14:05 – 21:25
Thinking visually and using mind maps to structure ideas
Chollet and Lex compare mental representations—visual concept spaces versus writing on paper. Chollet explains mind maps as a graph-like externalization that reduces cognitive messiness and supports reorganization of ideas.
- •Visual thinking: navigating “concept space”
- •Mind maps vs bullet lists: graph/topology vs trees
- •Mind maps as low-syntax scaffolding before formal writing
- •Associative structure and reorganization as a thinking aid
- 21:25 – 23:42
From mind maps to deep learning: topology vs geometry in representations
A technical digression connects mind maps to representational form. Chollet contrasts topological structure (connectivity) with geometric vector spaces (distances), arguing deep learning is biased toward geometry due to differentiability.
- •Topology: connections without distance; geometry: distances/metrics
- •Deep learning’s reliance on vectors and differentiability
- •Why embedding discrete/topological structures into continuous spaces is common
- •Implications for how thoughts might be encoded vs learned embeddings
- 23:42 – 34:49
Defining intelligence: efficient skill acquisition under novelty
Chollet offers a crisp definition: intelligence is efficiency at acquiring new skills for tasks you weren’t prepared for. A major theme emerges: don’t confuse intelligence (process) with skill (artifact/output).
- •Intelligence as learning efficiency for unknown tasks
- •Adaptation, improvisation, and out-of-distribution generalization
- •Skill vs intelligence: program output vs the creator process
- •Analogies: road-building company vs a single road; Einstein quote on change
- 34:49 – 36:13
Purpose of the paper: clearing misconceptions and making measurement actionable
Chollet explains the paper’s goal: to correct how AI evaluates progress and to enable meaningful claims about AGI. He argues a measure should be explanatory and usable as feedback to build more intelligent systems.
- •Need for precise definitions before AGI extrapolations
- •Benchmarks should assess generality, not narrow task prowess
- •Actionable metrics: explanatory power, not binary labels
- •Measurement as a research steering signal
- 36:13 – 42:22
Two views of mind: hand-coded programs vs tabula rasa deep learning
Chollet surveys cognitive science’s historical split: evolutionary-psychology modular mechanisms versus blank-slate learning. He connects each to corresponding AI eras—symbolic logic systems versus modern deep learning metaphors.
- •Mind as static specialized mechanisms (evolutionary psychology)
- •Early AI’s logical programs + database-like knowledge
- •Tabula rasa and connectionism resurgence with deep learning
- •Risk of ‘intellectual laziness’ in assuming scaling alone yields AGI
- 42:22 – 53:07
GPT-3 and scaling: pattern matching, data bottlenecks, and missing constraints
The conversation turns to GPT-3’s few-shot behavior and whether it represents real task learning. Chollet argues scaling improves plausibility, but doesn’t fix core issues like factuality, consistency, and controllability—plus data becomes the bottleneck.
- •Few-shot performance may be task recognition vs learning
- •Scaling yields more plausible text, not grounded truth/consistency
- •Prompt sensitivity and adversarial phrasing expose lack of constraints
- •Training data (not compute) as the main limiter; quality beats noisy scale
- •Need for explicit programmability/reasoning over learned latent knowledge
- 53:07 – 56:43
Semantic web, web knowledge, and the limits of unsupervised text reasoning
Lex asks about the semantic web dream of structured machine-readable knowledge. Chollet is skeptical about incentives for manual structuring and argues models like GPT are a more realistic path—while still limited in genuine novelty handling.
- •Why semantic web likely won’t work at internet scale (incentives/cost)
- •Unsupervised learning as a pragmatic alternative for leveraging web knowledge
- •Illusion of reasoning from text patterns vs true novelty adaptation
- •Continuous learning and improvisation as requirements for intelligence
- 56:43 – 1:13:57
Autonomous driving as a generalization problem: L4 vs L5 and explicit world models
Driving becomes a concrete case study in long-tail generalization. Chollet suggests L4 is plausible but L5 (human-level across environments) is much harder, and argues pure end-to-end learning is inefficient versus hybrid systems with explicit models.
- •Long-tail edge cases demand flexibility beyond typical ML robustness
- •Human zero-shot adaptation example: driving in a new country/city
- •End-to-end deep learning could work with vast data, but is inefficient
- •Deep learning as perception/intuition; explicit models enable stronger generalization
- •Driving skill alone is not a direct measure of general intelligence
- 1:13:57 – 1:36:00
Psychometrics and human intelligence testing: validity, bias, and g-factor
Chollet introduces psychometrics as the study of measuring cognitive traits, emphasizing reliability and validity. He explains g-factor as a latent statistical variable and cautions it doesn’t imply universal general intelligence—using a physical fitness analogy.
- •Psychometrics principles: reliability, validity, standardization, bias control
- •Tests measure observable performance (skills) to infer latent ability
- •g-factor as statistical structure across correlated cognitive tasks
- •CHC theory hierarchy: g, broad abilities, narrow abilities
- •Fitness analogy: broad capability within constraints, not universality
- 1:36:00 – 1:38:12
Designing machine intelligence tests: controlling priors and preventing leakage
Chollet lays out requirements for AI intelligence tests: make priors explicit, focus on skill acquisition, and ensure tasks cannot be anticipated or brute-forced by developers. The key is to test developer-aware generalization, not overfitting to a known test set.
- •Control for priors (what’s known before learning) and experience (data)
- •Novel tasks must be unpredictable even to system creators
- •Avoid benchmarks that can be ‘solved’ by hand-coding test-specific rules
- •Intelligence tests should measure learning efficiency, not static competence
- 1:38:12 – 1:44:40
Human core knowledge priors: objectness, agency, space, and number
Drawing on Elisabeth Spelke’s core knowledge theory, Chollet enumerates foundational priors humans bring to problem solving. These priors underpin ARC’s design: minimal, domain-general cognitive building blocks rather than language or cultural facts.
- •Core priors: objectness/basic physics (coherence, occlusion)
- •Agentness: goal-directed behavior and theory-of-mind projections
- •Geometry/topology and navigation (grid/place cells)
- •Number sense: small counting and quantity comparison
- •Priors may extend to social reasoning; rotation may be hardwired (mental turntable)
- 1:44:40 – 1:59:11
ARC Challenge: format, goals, creation difficulty, and current progress
Chollet explains ARC as an IQ-test-like benchmark of abstract reasoning over colored grids, designed to require only core priors and to be easy for humans yet hard for machines. He discusses dataset creation challenges, the need for a private test set, Kaggle results, and the benchmark’s evolution.
- •ARC task structure: few I/O examples → infer transformation → produce output grid
- •Explicitly limited priors; no language or real-world object concepts
- •Task creation requires ‘unfakeable novelty’; humans are a limited novelty source
- •Private test set prevents engineers from hard-coding solutions
- •Kaggle validated benchmark resilience; SOTA ~20% vs humans near 100%
- •Future: crowdsourcing, refinement, and human studies (NYU collaboration)
- 1:59:11 – 2:09:55
Generalization spectrum: robustness, flexibility, and extreme generalization
Chollet formalizes types of generalization from standard ML robustness to long-tail flexibility and finally extreme generalization across domains. He argues AGI corresponds to extreme generalization at human-like efficiency, while noting humans are likely far from any theoretical upper bound.
- •System-centric vs developer-aware generalization
- •Robustness: known distribution, known unknowns
- •Flexibility: unknown unknowns within a domain (e.g., driving long tail)
- •Extreme generalization: rapid transfer across domains (cook → garden)
- •Humans generalize remarkably but are not universally general
- 2:09:55 – 2:20:44
Turing Test critique: subjectivity, incentives, and the value of interactivity
Chollet argues the Turing test outsources measurement to biased judges and incentivizes tricks over scientific progress. Lex pushes back on the importance of interactive adaptation; they converge on the idea that interactivity is valuable but difficult to scale and standardize.
- •Turing test as a ‘cop-out’: lacks reliability/standardization and clear definition
- •Anthropomorphism bias makes humans easy to fool (agentness projection)
- •Perverse incentives: mimicry and stage magic vs understanding
- •Interactivity can elicit adaptation; challenge is scalable objective scoring
- •Idea: attempt-based scoring to quantify efficiency in reducing uncertainty
- 2:20:44 – 2:27:40
Hutter Prize and compression: why cognition isn’t just compression
Chollet discusses compression-based intelligence measures and why he no longer equates cognition with compression/prediction. He argues intelligence must hedge against fundamental uncertainty and novelty, so models can’t be perfectly compressed descriptions of the past.
- •Compression of Wikipedia as a clean, elegant benchmark concept
- •Why ‘cognition = compression’ is flawed: future novelty breaks past-optimality
- •Brains use compression as a tool, not as the definition of intelligence
- •Need for diversity and exploration (childhood curiosity) vs optimal encoding
- •Analogy to investing: balanced portfolios vs best-fit-to-history strategies
- 2:27:40 – 2:34:20
Meaning of life: culture, ripples, and immortality through contribution
In a philosophical closing, Chollet frames humans as products of culture—made of inherited words and concepts. Meaning comes from the ripples our actions create in the cultural future, motivating kindness and constructive contribution.
- •Personal thoughts are built from cultural artifacts we didn’t invent
- •Contributing to culture as a path to a kind of immortality
- •Actions propagate forward as ripples—good or harmful
- •Ethical takeaway: choose the ripples you want to create
