François Chollet

Open-source work

Chollet created Keras in 2015 and has led the project since then; Chollet also contributed to TensorFlow from 2015 to 2021, has led Keras Tuner since 2019 and has contributed to AutoKeras. KerasHub, the Keras library of pretrained models, offers presets for open-weight models that run on JAX, PyTorch or TensorFlow.¹²³

Earlier projects listed on Chollet's site include Wysp, a social network with drawing lessons for artists (2011–2014), and QuickAnswers.io (2014–2015), a question-answering engine that passed snippets from web search results to an LSTM language model to write its answers.¹

Research

Chollet's papers include "Xception - Deep Learning with Depthwise Separable Convolutions" (2016, listed with CVPR on Chollet's site) and co-authored work on neural theorem proving and machine translation such as DeepMath, HolStep and Tensor2Tensor. Chollet's primary research interest is understanding abstraction and building algorithms that can form abstractions on their own.¹⁴

In the 2019 paper "On the Measure of Intelligence", Chollet introduced the Abstraction and Reasoning Corpus, now known as ARC-AGI, as a benchmark for skill-acquisition efficiency.⁵⁶

ARC Prize and Ndea

Chollet hosted the first ARC-AGI competition on Kaggle in 2020 and ran the ARCathon competitions with Lab42 in 2022 and 2023. In 2024 Chollet and Mike Knoop created ARC Prize 2024, and in 2025 ARC Prize became a non-profit foundation; the foundation lists Chollet as co-founder, board member and creator of ARC-AGI. In 2026 ARC Prize introduced ARC-AGI-3, an interactive reasoning benchmark for agents.⁶⁷⁸

Chollet is also a co-founder of Ndea, which is building AI systems that combine intuitive pattern recognition with formal reasoning in a single architecture, aimed at scientific discovery.¹⁹

Work and contributions

• Keras and KerasHub

  Creator and project lead of Keras since 2015. KerasHub pairs Keras 3 model implementations with pretrained checkpoints, including Gemma, Llama, Mistral, Qwen, Phi-3, and SmolLM3 presets, for training and inference on JAX, PyTorch, or TensorFlow.¹²³

• ARC-AGI

  Co-founded ARC Prize, which runs the ARC-AGI benchmark.¹

• Deep Learning with Python Show, channel or publication

  Author of Deep Learning with Python (first edition 2017, second edition 2021), an introduction to deep learning with Python and Keras, and co-author with J. J. Allaire of Deep Learning with R (2018).¹

Timeline

1. 2015Created the Keras deep learning library and began contributing to TensorFlow.¹
2. Oct 7, 2016Released "Xception - Deep Learning with Depthwise Separable Convolutions" as sole author.⁴
3. 2017Published the first edition of the book Deep Learning with Python.¹
4. Nov 5, 2019Published "On the Measure of Intelligence", which introduced the ARC-AGI benchmark.⁵⁶
5. 2020Hosted the first ARC-AGI competition on Kaggle.⁶
6. 2021Published the second edition of Deep Learning with Python.¹
7. 2024Created the ARC Prize 2024 competition with Mike Knoop.⁶
8. Dec 5, 2024Co-authored the ARC Prize 2024 technical report.⁸
9. 2025ARC Prize became a non-profit foundation, with Chollet as co-founder and board member.⁶⁷
10. 2026ARC Prize introduced ARC-AGI-3, an interactive reasoning benchmark for agentic intelligence.⁶

Other documented roles

Roles the sources document that are not confirmed as current.

• Contributor (2015–2021), TensorFlow¹

Where they work

• ARC Prize Foundation 3 people in the wiki
• Ndea 1 person in the wiki

Related people

• Mike KnoopAlso at ARC Prize Foundation
• Greg KamradtAlso at ARC Prize Foundation
• Nils ReimersML frameworks and libraries · Evaluation and benchmarks
• Shinji WatanabeML frameworks and libraries · Evaluation and benchmarks
• Eric HorvitzEvaluation and benchmarks
• Hamel HusainEvaluation and benchmarks
• Mark CollierML frameworks and libraries
• João CarreiraEvaluation and benchmarks
• Matthew JohnsonML frameworks and libraries
• Shreya ShankarEvaluation and benchmarks