François Chollet: The engineer who made intelligence a measurable question

François Chollet was born in 1989 and trained as an engineer at ENSTA Paris, graduating in 2012. In March 2015 he released Keras, a high-level Python library that made deep neural networks legible to people who were not specialists. Google hired him shortly after, and Keras became the default front door to TensorFlow and, later, a multi-framework library in its own right.

His research output at Google was substantial. The Xception architecture, published in 2016 and presented at CVPR in 2017, became one of the most cited papers in that conference's history. His textbook Deep Learning with Python, first published in 2017 and now in a third edition, sold well over a hundred thousand copies. By any conventional measure, Chollet was a beneficiary and an author of the deep learning boom.

That position makes his later skepticism harder to dismiss. He was not a critic standing outside the technique. He was one of the people who industrialized it.

In November 2019 Chollet published On the Measure of Intelligence, a paper that opened with a provocation: the field did not have an agreed definition of the thing it claimed to be building. Surveying a century of attempts across psychology and AI, he argued that benchmarking systems by skill at specific tasks, from board games to video games, fails because skill can be bought with unlimited training data and engineered priors, masking the system's actual generalization power.

His alternative definition, formalized with tools from algorithmic information theory, was intelligence as skill-acquisition efficiency: the ability to turn limited experience into competence at tasks the system's designers never anticipated. To make the definition operational, he introduced the Abstraction and Reasoning Corpus, ARC, a set of visual puzzles trivial for most humans and brutal for machines, built on priors close to innate human ones.

The paper was as much a disciplinary intervention as a technical one. It told the field that its feedback signal was miscalibrated, and it offered a better one.

ARC behaved the way Chollet predicted a real intelligence test should. A 2020 Kaggle competition topped out at 21 percent on the private test set. Subsequent contests in 2022 and 2023 reached only around 30 percent, while ordinary benchmarks fell to scale one after another. In 2024 Chollet and Zapier co-founder Mike Knoop launched ARC Prize, a million-dollar competition to push open progress on the benchmark. The grand prize went unclaimed.

Then, in December 2024, OpenAI announced that a preview of its o3 reasoning model had scored 75.7 percent on ARC's semi-private evaluation at a compute limit of roughly ten thousand dollars, and 87.5 percent in a high-compute configuration estimated at thousands of dollars per task. Chollet's response was characteristic: the result was real progress, and it also proved his point. Humans solve ARC tasks for the cost of a sandwich. A system that matches them by spending orders of magnitude more compute per task is demonstrating skill bought with resources, not efficient skill acquisition.

In 2025 the ARC Prize Foundation released ARC-AGI-2, a harder version designed to restore the gap between human ease and machine difficulty, and confirmed that cost per task belongs on the leaderboard next to accuracy.

In November 2024, after more than nine years, Chollet left Google. Weeks later he announced Ndea, a research lab co-founded with Knoop, dedicated to building AGI through deep learning-guided program synthesis: systems that assemble short symbolic programs to solve novel problems from a handful of examples, rather than fitting ever larger parametric models. In parallel, the ARC Prize was expanded into a nonprofit foundation intended to serve as an open north star for AGI measurement.

His public position has evolved in a way that surprises people who filed him as a doomer about scaling. In 2026 interviews he described AGI progress as accelerating, suggested a timeline in the vicinity of 2030, and said AI progress is now effectively unstoppable. What has not changed is the architectural claim: the current LLM stack, in his view, automates domains with verifiable rewards like code, but it is an inefficient substrate for general intelligence and will eventually be replaced by something closer to optimal.

This is the unusual shape of his wager. He thinks AGI is coming, thinks the dominant paradigm is the wrong way to get there, and has organized his career around both claims at once.

The case for Chollet's foresight is strong. His 2019 critique anticipated the memorization-versus-reasoning controversy that consumed the LLM era. ARC exposed, in a way no saturated benchmark could, the difference between retrieving patterns and generating solutions. And his insistence on measuring cost per task, once an eccentricity, is now standard practice in serious model evaluation.

The contested parts are equally clear. Reasoning models with test-time compute blurred the line he drew, converting scale into partial novel-problem competence faster than he expected, and his own timeline estimates have moved accordingly. Skeptics of program synthesis note that the approach has been promising for decades without a breakthrough. The honest summary is that his definition of intelligence has survived every round so far, while his predictions about which architecture will satisfy it remain unproven. He is now spending his own career finding out.