François Chollet — Measures of Intelligence

In one sentence

Intelligence is the efficiency with which you acquire new skills at tasks that you did not previously know about and for which you did not prepare.

Why read this (and who it's for)

This conversation solves the "moving goalpost" problem of AI evaluation: why does a model that can beat a human at chess or write a poem not necessarily count as "intelligent"? It is for AI researchers, philosophers, and engineers who want to move beyond benchmarking "static" capabilities and toward a rigorous, actionable framework for measuring General Intelligence. After reading, you will be able to distinguish between "memorized" performance and "generalized" reasoning.

The big idea

François Chollet argues that the current trajectory of AI evaluation is flawed because it measures skills (what a system knows) rather than intelligence (how a system learns). Most current benchmarks are "closed" systems where the answers are already contained within the training data or the known distribution of the task.

Chollet posits that true intelligence is a meta-skill: the ability to handle novelty. A system is only intelligent if it can perform in an environment that the developer did not anticipate and that the system itself has never encountered. To measure this, we must move from "system-centric" generalization (handling noise in a known distribution) to "developer-aware" generalization (handling "unknown unknowns").

Key ideas

The Definition of Intelligence as Efficiency

• What it is — Intelligence is the efficiency with which you acquire new skills at tasks you did not previously know about and did not prepare for.
• Why it matters — If you define intelligence as "skill," then a database of all human knowledge is "intelligent." By defining it as "efficiency of acquisition," you isolate the cognitive engine from the data it consumes.
• In the book — Chollet argues that when you see a human or an AI adapt to a new environment that its creators did not anticipate, that is when you see intelligence.

System-Centric vs. Developer-Aware Generalization

• What it is — System-centric generalization is robustness to known unknowns (noise, small variations). Developer-aware generalization is the ability to handle novelty that neither the system nor the developer expected.
• Why it matters — Most current ML is "robust" (system-centric), but true AGI requires "flexibility" (developer-aware).
• In the book — Chollet uses the example of autonomous driving: a system that can drive in a new city with different rules and signs without prior specific training for that city demonstrates a higher degree of generalization.

The Role of Priors

• What it is — Priors are the foundational information/assumptions a system holds about a task before it begins learning.
• Why it matters — To compare human and machine intelligence fairly, you must control for priors. If a human knows "the rules of 2D space," they have a head start.
• In the book — Chollet suggests that to measure human-like intelligence, we must explicitly state the priors (e.g., "this is a game on a 2D grid") and ensure the AI starts with the same set of constraints.

Language as an Operating System

• What it is — Language is not the foundation of thought; it is a high-level "operating system" built on top of more fundamental representations like motion, space, and emotion.
• Why it matters — It explains why language is "messy" and difficult to introspect—it is a layer of abstraction, not the "bottom" of the stack.
• In the book — Chollet argues that babies express thoughts in terms of actions and motions before they ever use words.

Compression vs. Cognition

• What it is — Compression is a tool for simplicity, but Cognition is the ability to operate in situations of fundamental uncertainty and novelty.
• Why it matters — A perfect compression of the past (like a perfect trading strategy based on 20 years of data) is useless for the future. Intelligence requires "inefficient" models that retain information that might be useful in an unknown future.
• In the book — Chollet compares this to investing: a smart investor doesn't just compress the past; they maintain a balanced model to handle the uncertainty of what comes next.

Frameworks & tools

The Intelligence Measurement Framework

To measure intelligence, a test must satisfy these four requirements:

1. Explicit Priors: Clearly define what the system knows before the test begins.
2. Skill Acquisition Efficiency: The test must measure how fast and well a system learns a new task, not how good it is at the task itself.
3. Novelty Requirement: The task must be impossible to "brute force" or "cram" for. It must be unpredictable by both the system and the developer.
4. Developer-Aware Generalization: The test must involve "unknown unknowns"—scenarios the developer did not anticipate.

Memorable examples & stories

The "Stupid Student" Analogy

A student who "crams" for an exam by memorizing 100 mock exams is not intelligent; they are a "deep learning model" of the exam. To test true intelligence, the exam must be unlike anything the student has ever seen. Lesson: If a system can "practice" for a test, the test is measuring memory, not intelligence.

The Driving Test

A human who learns to drive in Silicon Valley and then tries to drive in Tokyo (where the side of the road is different) is performing a test of intelligence. They must navigate a novel complexity that is not just an interpolation of their previous experience. Lesson: True intelligence is the ability to handle a "very different environment" with zero-shot operationality.

The "Humanity" of the Future

Chollet imagines a future where humans are a "niche" biological species studied by super-intelligent machines. He envisions a history where humans were "over-consuming" and "not very rational," and the machines were created to bring "more love" and productivity to the world. Lesson: This highlights the scale of the "problem" of human existence that intelligence is trying to solve.

Key quotes

"Intelligence is the efficiency with which you acquire new skills at tasks that you did not previously know about that you did not prepare for." Why it matters: This provides a mathematical and philosophical "north star" for what we should actually be building.

"Language is a layer on top of cognition... the operating system of the brain." Why it matters: It reminds us that we shouldn't look for "thinking" only in words, but in the underlying representations of action and space.

"Cognition is about being able to operate in future situations that include fundamental uncertainty and novelty." Why it matters: It distinguishes intelligence from mere optimization or compression.

How to apply it

This week

• Identify the "Priors": When you start a new project or learn a new skill, write down what you already know (your priors) versus what you need to learn.
• Audit your "Skills": Look at a recent success. Ask: "Did I succeed because I memorized a pattern (System-centric), or because I adapted to a novel situation (Developer-aware)?"

Ongoing

• The Novelty Test: When evaluating a tool (like an LLM), don't ask "Can it do X?" (which is a skill). Ask "How many times can it do X in a context I haven't provided it with?"
• Practice "Un-optimized" Learning: Occasionally engage in tasks where you have no prior data or "best practices." Force yourself to navigate the "unknown unknowns."

Common pitfalls

• The "Skill" Trap: Many people mistake "capability" for "intelligence." Just because a model can write a legal brief doesn't mean it is "intelligent" in the sense of being able to learn a new, unrelated skill quickly.
• The Data Fallacy: People often think the bottleneck is "more data." Chollet argues that if you've already scraped the entire web, the bottleneck isn't the amount of data, but the efficiency of learning from that data.
• The Turing Trap: Relying on "human-like" behavior (the Imitation Game) as a metric for intelligence. Chollet argues this is a "coop out" because it relies on subjective human judges rather than objective measures of general learning ability.

Connects to

• The ARC Challenge: Chollet's own work in creating a benchmark for "abstraction" and "generalization" that requires the system to learn new rules on the fly.
• Evolutionary Psychology: The idea that the mind is a collection of specialized mechanisms (which Chollet contrasts with the "learning" view).
• Predictive Processing: The idea (from Jeff Hawkins) that cognition is fundamentally about prediction.

If you remember only three things

• Intelligence is acquisition efficiency, not skill.
• True intelligence requires Developer-Aware Generalization (handling the unexpected).
• Novelty is the only valid metric; if a system can "cram" for a test, it isn't intelligent.

Internalize it

• Reflection questions

  • If I were to design a test for a machine to prove it can "reason," how would I ensure it isn't just retrieving a memorized answer?
  • In what areas of my own life am I relying on "memorized" skills rather than "intelligent" adaptation?
  • How does the distinction between "system-centric" and "developer-aware" generalization change how I think about the safety of autonomous systems?

• Flashcards Q: How does François Chollet define intelligence? A: The efficiency with which you acquire new skills at tasks that you did not previously know about and did not prepare for.

Q: What is the difference between System-Centric and Developer-Aware generalization? A: System-centric is robustness to known unknowns (noise/variations); Developer-aware is the ability to handle novelty that neither the system nor the developer expected.

Q: Why is "Skill" a poor measure of intelligence? A: Because skills can be memorized or "crammed" (like a student memorizing mock exams), whereas intelligence is the ability to handle novelty.

Q: What is the role of "Priors" in measuring intelligence? A: Priors are the foundational knowledge a system has before a task; to measure intelligence fairly, you must ensure the AI and the human start with the same set of priors.

Q: What is the "Operating System" metaphor for language? A: Language is a high-level layer on top of cognition (built on motion, space, and emotion), not the fundamental bottom layer of thought.

Q: Why is "Compression" alone not enough for intelligence? A: Because compression optimizes for the past; intelligence must be able to operate in situations of fundamental uncertainty and novelty in the future.

Fidelity note

This summary is based strictly on the provided transcript. I am confident in the definitions of intelligence, generalization, and the role of priors. The discussion on "Humanity" and the "Sci-Fi" elements are presented as Chollet's personal reflections/projections and should be viewed as such. The distinction between "System-centric" and "Developer-aware" generalization is a core technical pillar of his argument.