#120 – François Chollet: Measures of Intelligence

At a Glance

The guest is François Chollet, a Google AI researcher and creator of the Keras framework. This episode centers on his 2019 paper On the Measure of Intelligence, exploring how to define and measure general intelligence in machines and humans. Chollet’s core argument is that intelligence should not be defined as “skill” (what you know or can do) but rather as “efficiency”—the speed and efficiency with which a system acquires new skills when confronted with a novel task.

Intelligence Is Efficiency, Not Skill

François Chollet argues that the essence of intelligence is “the efficiency of adapting to new environments and acquiring new skills,” not the skills themselves. He points out that the current AI field commonly conflates “the process of intelligence” with “the product of that process (skill).” A program that can play chess merely records the output of human intelligence; it is worthless when faced with a task outside chess, and therefore is not intelligent. True intelligence lies in the ability to create new skills.

• Historical context: Chollet traces two opposing views in AI research. One is the “evolutionary psychology” view, which treats the mind as a set of static, specialized programs (e.g., early AI and Minsky’s work). The other is the “blank slate” view, which treats the mind as a blank information sponge that absorbs knowledge from experience (e.g., the current dominant deep learning paradigm). Chollet believes neither view is complete.
• Mechanisms and data: He cites Einstein’s quote, “The measure of intelligence is the ability to change,” to support his view, but defines it more precisely as efficiency. He offers a core analogy: “Don’t confuse a road-building company with a specific road. A road can only take you from A to B, but a road-building company can build a road from anywhere to anywhere.” Intelligence is the “road-building company,” and skills are the “roads.”
• Reasoning and falsifiability: If this definition holds, then the benchmark for evaluating an AI system should not be its performance on a specific task, but rather its learning efficiency when faced with completely new, unforeseen tasks. This view can be falsified by observing a system’s performance in entirely novel situations that cannot be handled through pattern matching or memorization.

Critique of GPT-3: Approaching the “Unforeseeable Future”

Chollet is highly skeptical of GPT-3, calling it a “plausible text generator” rather than true intelligence. He acknowledges that GPT-3’s scale yields impressive capabilities, but explicitly states that its core flaw is its inability to perform genuine reasoning or generalization, especially when confronted with entirely novel situations.

• Mechanism breakdown: Chollet believes GPT-3’s success stems from its massive parameter count (175 billion) and training data (nearly the entire internet), making it a “huge associative memory” system. He notes that the text generated by GPT-3 is constrained only by “plausibility,” not by factuality or self-consistency, making it easy to induce contradictory or factually incorrect statements.
• Data and limitations: He makes a key argument: GPT-3’s bottleneck is not compute, but training data. Because the model has already consumed the entire internet, future improvement via scaling is limited. He predicts that GPT-3’s successors will produce increasingly “context-aware” text, but the fundamental flaw will not be resolved by simply scaling up, “because its only purpose is plausibility, and there is no other constraint.”
• Reasoning and uncertainty: Chollet argues that to achieve true intelligence, future AI systems will need to combine this “pool of intuitive knowledge from unsupervised learning” with “explicit, programmable reasoning programs.” He judges that simply scaling up model size without changing the architecture will not achieve genuine generalization.

On Autonomous Driving: L5 Is Not a Test of Intelligence

Chollet argues that Level 5 autonomous driving itself does not constitute a test of general intelligence, because the problem can be solved through “infinite data” or “manual engineering” without requiring true intelligence. He emphasizes that a true test of intelligence is not about dealing with a known “long tail distribution,” but about facing a completely unfamiliar environment that requires immediate adaptation.

• Data chain and competitive landscape: He mentions a Google paper showing that 30 million different road scenarios were still insufficient to train a reliable driving model. In contrast, humans, building on existing knowledge, need only 20–30 hours of driving practice to learn to drive. This data contrast highlights the huge gap in generalization efficiency between AI and humans.
• Mechanism breakdown: He distinguishes two approaches: “pure end-to-end deep learning,” which requires “near-infinite” training data to cover all edge cases and is extremely inefficient; and “the engineering approach,” which uses deep learning for perception combined with explicit environmental models and reasoning. He believes the latter is more efficient, and even if L5 is achieved, it does not mean the system possesses general intelligence.
• Reasoning: To test the true intelligence of an L5 driving system, one should see whether it can be “transferred” to learn other tasks, such as “learning to fly a commercial aircraft.” The expert must observe how much human intervention and training data the system needs when mastering a new task—this is what measures its “intelligence.”

Ultimate Test: The ARC Challenge and “Developer-Unaware” Generalization

Chollet introduces the concept of “developer-unaware generalization” and designs the ARC Challenge as a test for measuring true intelligence: the system must be able to handle entirely new situations that even its creator could not foresee. This goes beyond “robustness” (dealing with known unknowns) and “flexibility” (dealing with unknown unknowns within the same task domain) in machine learning, aiming for “extreme generalization” (dealing with unknown unknowns across arbitrary task domains).

• Mechanism and design: The design philosophy of ARC is to “make prior knowledge explicit.” It is based on innate “core knowledge priors” that humans are born with, such as “objectness,” “agentness,” “basic geometry and topology,” and “numerosity.” All tasks are built on these foundations, using no language or external knowledge, ensuring that what is tested is “fluid intelligence” rather than “crystallized intelligence.”
• Data and current status: The ARC test set is kept secret to prevent developers from achieving high scores through “hacking” or “memorization.” In the Kaggle competition, the best AI system could solve only about 20% of the test set, while humans achieve nearly 100%. This huge gap proves that current AI has a fundamental deficiency in “extreme generalization.”
• Reasoning and falsifiability: Chollet argues that when an AI system reaches human-level performance on ARC, it will mean we have developed a general fluid intelligence close to human level. He acknowledges that the system may not behave like a human, but would possess similar generalization ability. He also points out that any system that scores high on ARC proves the intelligence of its “process,” not the rigidity of its “output.”

Mentioned Positions

Position	Guest's Stance	Key Data
GPT-3 (OpenAI)	Risk Warning (skeptical of its true intelligence, considers it pattern matching)	175 billion parameters; training data is “nearly the entire internet”
Tesla Autopilot	Risk Warning (believes L5 can be achieved without intelligence, skeptical of pure end-to-end approach)	Humans need 20–30 hours of training; Google paper shows 30 million scenarios still insufficient

Judgments Worth Remembering

1. Intelligence Is Efficiency, Not Skill (Chollet): The key measure of intelligence is the efficiency with which a system acquires new skills when faced with entirely novel, unforeseen tasks, not how well it performs on a specific task.

2. GPT-3 Is a “Plausible Text Generator” (Chollet): Its output is constrained only by “plausibility,” not by factuality or self-consistency, making it very easy to induce erroneous or contradictory answers. Its core flaw cannot be resolved by simply scaling up.

3. GPT-3’s Bottleneck Is Data, Not Compute (Chollet): Because its training data is already “the entire internet,” future improvement from merely adding more data is very limited. The model is more like a “huge associative memory” than a system that understands.

4. The “Road-Building Company” Analogy (Chollet): Do not confuse “the process of intelligence” with “the product of that process.” A program that can play chess is a “road,” while the agent that can “learn to play chess” is the “road-building company.” The former is not intelligent; the latter is.

5. “Developer-Unaware Generalization” (Chollet): A true test of intelligence requires the system to handle new tasks or situations that even its creator could not foresee—this is the key dividing line for measuring AI capability.

6. Human Intelligence Is Not Universal (Chollet): Although human cognitive ability is highly generalizable, it is still constrained by the specific environment of “the human condition.” Humans cannot “fly” and cannot survive at the bottom of the ocean; intelligence is not omnipotent.

7. Cognition Is Not Compression (Chollet): Although equating intelligence with compression (e.g., the Hutter Prize) is a tempting idea, the two are fundamentally different. Cognition requires “hedging” against future uncertainty, retaining some information that seems useless now but may be useful later—this contradicts compression’s goal of seeking the “most concise” representation.

8. The ARC Challenge Reveals the Fatal Flaw of Current AI (Chollet): On the ARC test, human performance is close to 100%, while the best AI system solves only about 20%. This huge gap shows that current AI fundamentally fails on abstract reasoning tasks requiring “extreme generalization.”