François Chollet

Some people build tools, and some people build tools that change how everyone else builds. François Chollet, a French engineer born in 1989, belongs to the second category. In 2015, while working at Google, he released Keras: a Python library that put deep learning within reach of anyone without a doctorate in mathematics. Before Keras, training a neural network meant wrestling with cryptic configurations and low-level code. After Keras, a few lines were enough. The book he wrote about it, "Deep Learning with Python," sold more than 100,000 copies. Keras was integrated directly into TensorFlow and became, de facto, the standard API for the most widely used framework in the world.

But Chollet did not rest on that achievement. In 2019 he published a paper titled "On the Measure of Intelligence" (arXiv:1911.01547) that shook the field. Its central argument: we are measuring artificial intelligence the wrong way. Typical benchmarks measure learned skills, the same thing a math exam measures when a student has memorized the steps. Real intelligence, Chollet argued, is measured by how efficiently a system acquires new skills from little information, the way a child approaches a problem it has never seen. From that argument he created ARC-AGI (Abstraction and Reasoning Corpus), a set of visual puzzles trivially simple for a human but systematically difficult for the world's best models. In June 2024 he launched the ARC Prize, a competition with $1.1 million in prizes, which raised the field's best score from 33% to 55.5%.

Chollet spent more than nine years at Google Brain, where he also designed the Xception architecture (2016), a computer vision model that outperformed Google's Inception V3 using depthwise separable convolutions. In November 2024 he left Google, and in January 2025 he co-founded Ndea alongside Mike Knoop, co-founder of Zapier. Ndea is an AGI research lab betting on deep-learning-guided program synthesis as an alternative to the dominant paradigm of scaling models with more data and more compute. The company raised $43.5 million from investors including Y Combinator and Coatue. In 2021 he received the Global Swiss AI Award, and in September 2024 he was named to TIME's TIME100 AI list of the 100 most influential people in artificial intelligence.

What strikes me as extraordinary about Chollet is not just what he built, but the question he dared to ask: if AI is already so intelligent, why can it not solve this. Every time you use Keras, or watch someone learn to fine-tune a model for the first time, there is something of that accessibility he designed on purpose. And every time someone debates what it truly means to "understand" something, ARC-AGI is there as an uncomfortable mirror. That, to me, is a legend: someone who not only advances the field but stops to ask where it is going.