AI Thinkers·AI researcher, founder of Ndea
François Chollet
Intelligence is not skill — it is the efficiency with which a system acquires new skills given minimal experience. LLMs are sophisticated memorisation engines: they store compressed representations of training data and retrieve them at inference time. They are powerful but not intelligent in any meaningful sense. His ARC-AGI benchmark makes this distinction empirical: it presents novel visual reasoning tasks that cannot be solved by prior memorisation. Humans solve 100% of them; frontier AI systems score below 5% on the harder ARC-AGI-2. His bet at Ndea is on deep learning-guided program synthesis as the path to genuine generalisation.
Key positions
LLMs are System 1, not System 2 — fast pattern-recognition, not deliberate reasoning.
LLMs are extremely good at retrieving and interpolating over training distribution. When a problem falls genuinely outside that distribution — requiring on-the-fly construction of a novel solution from first principles — they fail. Program synthesis is the natural substrate for System 2 thinking; LLMs can usefully guide that search without being sufficient on their own.
Chollet on X (@fchollet), June 2024 (x.com/fchollet/status/1801070604923347238); Dwarkesh Patel podcast, June 2024
ARC-AGI exposes that AI benchmark success is skill retrieval, not general reasoning.
ARC-AGI-1 (2019) and ARC-AGI-2 (March 2025) present visual grid transformation puzzles where each task is genuinely novel. ARC-AGI-2 was calibrated with 400 human participants to confirm humans find tasks straightforward. Frontier models ace every standard benchmark but score below 5% on ARC-AGI-2. This is a concrete, reproducible data point that cannot be dismissed as benchmark design failure.
ARC-AGI-2 paper, arxiv.org/abs/2505.11831 (May 2025); arcprize.org/blog/arc-prize-2025-results-analysis
Scaling LLMs has set back progress toward AGI by 5–10 years by creating an intellectual monoculture.
"LLMs have sucked the oxygen out of the room." Enormous investment in LLMs since 2020 has monopolised talent, compute, and attention that might otherwise have gone toward hybrid approaches or program synthesis. His stronger claim: OpenAI specifically set back AGI progress by 5–10 years by making scaling look like the only path.
Dwarkesh Patel podcast, June 2024 (dwarkesh.com/p/francois-chollet)
Intelligence is formally defined as skill-acquisition efficiency — not performance on fixed tasks.
"On the Measure of Intelligence" (2019) defines intelligence as "a measure of its skill-acquisition efficiency over a scope of tasks, with respect to priors, experience, and generalization difficulty." A lookup table can score 100% on any finite benchmark. The question is whether a system can acquire new skills efficiently from minimal experience — and on that measure, LLMs show diminishing returns as novelty increases.
"On the Measure of Intelligence," Chollet, 2019 (arxiv.org/abs/1911.01547)
OpenAI's o3 was a genuine but expensive breakthrough — progress, not AGI.
When o3 scored 75.7% on ARC-AGI-1 in December 2024, Chollet called it "a genuine breakthrough, marking a qualitative shift in AI capabilities." He simultaneously noted the high compute cost per task. ARC-AGI-2 (March 2025) showed the approach did not transfer: top scores remained below 5%, confirming that o3's ARC-AGI-1 performance was partly a function of that benchmark's limitations.
Chollet on X, December 2024 (x.com/fchollet/status/1870169764762710376); arcprize.org/blog/oai-o3-pub-breakthrough
The path to AGI is deep learning-guided program synthesis, not larger LLMs.
Programs are explicit, compositional, and generalisable: a program that solves a class of problems solves all instances, regardless of whether they appeared in training data. "I believe that program synthesis will solve reasoning. And I believe that deep learning will solve program synthesis" — through guided search, not direct end-to-end generation. He compares program synthesis's current state to where deep learning was in 2012.
Chollet on X (@fchollet), June 2024 (x.com/fchollet/status/1803096195684012371); TechCrunch on Ndea, Jan 2025
In their own words
“LLMs cannot plan and reason on their own — they're pattern-matchers. But that doesn't mean they can't be useful for reasoning. They work best in conjunction with systems that can actually plan or reason.”
“You can always achieve skill by memorization — by effectively just storing a lookup table. Intelligence is the efficiency with which you're going to make sense of new things, of new tasks that you've never seen before.”
“I see LLMs as more of an off-ramp on the path to AGI actually. All these new resources are actually going to LLMs instead of everything else they could be going to.”
“People have been rewriting history and saying that 'everyone has always believed that LLMs alone wouldn't be AGI.' No, throughout most of 2023 the mainstream Bay Area belief was that LLMs were [the path].”
Tensions with other thinkers
Where this view genuinely conflicts with others in the field.
Altman treats continued scaling as the near-certain path; Chollet says it sets AGI back by 5–10 years.
Both reject LLMs as the path to AGI, but disagree on mechanism: world models (LeCun) vs. program synthesis (Chollet).
Goodhart's Law applied to AI: once a benchmark becomes a target, it stops measuring intelligence. He has built ARC-AGI specifically to resist this.