Francois Chollet

Software Engineer — Google

Profile

François Chollet is the French software engineer who made deep learning usable. In 2015 he released Keras — a high-level neural networks API that prioritized human readability over framework purity — and changed who gets to do AI. Before Keras, building a neural network meant wrestling with Theano or raw TensorFlow graphs. After Keras, you wrote a few lines of Python that read like pseudocode and got a working model. It became the default entry point for millions of practitioners and was later absorbed into TensorFlow as its official high-level interface. If you’ve ever typed model.fit(), you’ve used his work.

He spent more than nine years at Google as a software engineer before leaving in late 2024. His second act is louder than the first. In 2019 he published On the Measure of Intelligence, a paper that argued the field was measuring the wrong thing — that benchmark scores on tasks with unlimited training data test memorization, not intelligence. The paper introduced ARC-AGI, a set of visual reasoning puzzles that humans solve easily and that deep learning models, including the biggest LLMs, could not. For years it was a quiet embarrassment to the scaling-is-all-you-need crowd. In 2024 he and Zapier co-founder Mike Knoop turned it into the ARC Prize — a million-dollar public competition. In March 2026 they launched ARC-AGI-3, an interactive agent benchmark where the best frontier model (Gemini 3.1 Pro Preview) scores 0.37%.

In January 2025 he announced Ndea, a research lab he co-founded with Knoop to pursue AGI through deep learning-guided program synthesis — the bet being that pure scaling won’t get there and that systems need to compose programs, not just interpolate patterns. Whether he’s right is the central open question in AI today.

For developers learning AI, Chollet matters for two reasons. First: his third edition of Deep Learning with Python remains one of the cleanest, most honest introductions to the field, updated for Keras 3 (which now runs on TensorFlow, PyTorch, and JAX). Second: he is the most rigorous skeptic in a field drowning in hype. He’s been wrong before — he underestimated how far LLMs would go — but when Gary Marcus criticizes LLMs it sounds like a rant, and when Chollet does it sounds like an engineering spec.