François Chollet

wiki
, , , ,
1

French-born AI researcher, open-source engineer, and co-founder of Ndea, best known as the creator of Keras and the architect of the ARC-AGI benchmark for measuring general intelligence.

Profile
Born October 20, 1989, France
Nationality French
Current Institution Ndea (co-founder); ARC Prize Foundation (co-founder)
Research Areas Deep Learning, Computer Vision, Abstraction, Formal Reasoning, Artificial General Intelligence, Program Synthesis
Education Diplôme d’Ingénieur (MEng), ENSTA Paris / Polytechnic Institute of Paris (2012)
Personal Website fchollet.com
X / Twitter @fchollet
GitHub fchollet
Google Scholar François Chollet — cited 73,000+ times
Substack fchollet.substack.com

Overview

François Chollet is a French software engineer and AI researcher whose decade at Google produced two of the most consequential artefacts in modern deep learning: the Keras framework, which became the dominant high-level API for building neural networks worldwide, and the ARC-AGI benchmark, which redefined how the field measures progress toward general intelligence. His Xception architecture is among the top ten most cited papers in CVPR proceedings at over 18,000 citations, and his book Deep Learning with Python sold over 100,000 copies. After more than nine years at Google, he left in November 2024 to co-found Ndea, a research lab pursuing AGI through deep learning-guided program synthesis, alongside the ARC Prize Foundation, a non-profit he established to coordinate and fund open research progress toward general intelligence. He is one of the field’s most prominent sceptics of the view that scaling large language models will suffice for AGI.

Early Life & Education

Chollet was born in France on October 20, 1989. He graduated in 2012 with a Diplôme d’Ingénieur (equivalent to a Master of Engineering) from ENSTA Paris, a grande école of the Polytechnic Institute of Paris. His early independent projects after graduation showed an unusually broad technical range: he built Wysp, a social network and learning platform for artists that peaked at around 25,000 monthly active users in 2014–2015, and QuickAnswers.io, a question-answering engine that used a technique — normalising queries, running web searches, extracting relevant snippets, and passing them to an LSTM language model — that would later be called Retrieval-Augmented Generation (RAG). Both projects reflected an early interest in systems that help people learn and gain greater agency, a theme that would recur throughout his career.

Career

Google (2015–2024)

In 2015, Chollet started working at Google shortly after releasing Keras. He served as Senior Staff Engineer, holding primary responsibility for Keras as well as contributing to TensorFlow and broader Google Brain research projects.

Keras. Released on March 27, 2015, Keras was designed as a user-first, high-level neural network API. Its explicit philosophy — that API design is a form of UX design, and that developers’ time and cognitive load matter — made it the dominant onboarding layer for deep learning practitioners and researchers at a time when TensorFlow’s low-level interface was a significant barrier to entry. Keras was later adopted as TensorFlow’s official high-level API and has powered AI development at companies including YouTube, Waymo, Spotify, Uber, and Netflix. With Keras 3 (2023), it became backend-agnostic, supporting TensorFlow, JAX, and PyTorch.

Xception. His 2016 paper Xception: Deep Learning with Depthwise Separable Convolutions, published at CVPR 2017, proposed replacing standard convolutional layers with depthwise separable convolutions — an idea that had been implicit in Inception architectures but which Chollet formalised and pushed to its logical conclusion. The paper is among the top ten most cited in CVPR proceedings with over 18,000 citations and influenced subsequent architectures including MobileNet and EfficientNet.

Formal reasoning research. Alongside framework work, Chollet contributed to Google’s programme on deep learning applied to mathematical reasoning, including DeepMath (NeurIPS 2016, with Alemi, Szegedy et al.) on premise selection for theorem proving, and HolStep (ICLR 2017, with Kaliszyk and Szegedy), a machine learning dataset for higher-order logic. He also co-authored work on depthwise separable convolutions for neural machine translation (ICLR 2017, with Kaiser and Gomez) and contributed to the Tensor2Tensor library.

ARC-AGI benchmark. In 2019, Chollet published On the Measure of Intelligence (arXiv 1911.01547), a theoretical paper that formalised a definition of intelligence as skill-acquisition efficiency on novel tasks, and introduced the Abstraction and Reasoning Corpus (ARC-AGI) as a benchmark. ARC-AGI focuses on fluid intelligence — the ability to reason and solve novel problems — rather than crystallised intelligence based on accumulated knowledge, and restricts itself to core knowledge priors present in all humans from early development. The benchmark was specifically designed to be easy for humans and hard for AI systems that rely on interpolation over training data.

In 2024, Chollet launched ARC Prize, a US$1 million competition to solve the ARC-AGI benchmark. The competition attracted significant attention when OpenAI’s o3 model achieved around 87% on ARC-AGI-1 at high compute, which Chollet acknowledged as a meaningful breakthrough consistent with the program-search direction he had been advocating. The ARC Prize series has since expanded to ARC-AGI-2 and ARC-AGI-3, progressively harder iterations.

Chollet left Google in November 2024 after more than 9 years.

Ndea / ARC Prize Foundation (2025–present)

In January 2025, Chollet co-founded Ndea with Mike Knoop, co-founder of Zapier, as a new AI research lab with the stated goal of building AGI through deep learning-guided program synthesis. The name Ndea draws on the Greek concepts of intuitive understanding (ennoia) and logical reasoning (dianoia), reflecting its dual technical approach. Ndea is fully remote and is assembling what it describes as a world-class program synthesis research team. In early 2025, Chollet also expanded ARC Prize into a full-fledged non-profit foundation to guide and accelerate research progress toward AGI.

Key Contributions

  • Keras — Created and led the Python deep learning framework released in March 2015, now one of the most widely used deep learning libraries worldwide. Its API-first, user-centric philosophy established a new standard for ML tooling accessibility. Keras 3 extended it to support TensorFlow, JAX, and PyTorch backends.

  • Xception: Deep Learning with Depthwise Separable Convolutions (CVPR 2017) — Among the top ten most cited papers in CVPR proceedings at over 18,000 citations. Formalised the architectural principle underlying many of the most efficient convolutional networks of the following decade.

  • ARC-AGI Benchmark (On the Measure of Intelligence, arXiv 2019) — Proposed a rigorous, operational definition of intelligence as skill-acquisition efficiency, and introduced the ARC-AGI task suite as a measurement instrument. Has become the de facto reference benchmark for evaluating general-purpose reasoning in AI systems.

  • ARC Prize (2024–present) — A US$1 million competition launched in 2024 to accelerate open progress on ARC-AGI, subsequently expanded into a foundation (ARC Prize Foundation) and an evolving series of benchmarks (ARC-AGI-1, 2, 3).

  • Deep Learning with Python (Manning, 2017; 2nd ed. 2021; 3rd ed. 2025) — Sold over 100,000 copies and translated into at least twelve languages; one of the most read introductory textbooks in the field.

  • DeepMath (NeurIPS 2016) — Co-authored with Alemi, Szegedy et al.; applied deep sequence models to premise selection for formal theorem proving, an early demonstration of deep learning on mathematical reasoning tasks.

  • HolStep (ICLR 2017) — Co-authored with Kaliszyk and Szegedy; introduced a large-scale dataset for machine learning research on higher-order logic theorem proving.

  • QuickAnswers.io (2014, retired) — An early prototype of retrieval-augmented generation (RAG) built years before the term was coined, which normalised user queries, retrieved web snippets, and generated contextual responses via an LSTM language model.

  • Keras Tuner — Open-source hyperparameter tuning library for Keras, released 2019, simplifying neural architecture search for practitioners.

Awards & Recognition

  • TIME100 AI (2024) — Named by TIME as one of the 100 most influential people in AI.
  • Global Swiss AI Award (December 2021) — Awarded for breakthroughs in AI.
  • Google Scholar citations — Over 73,000 total citations across all publications.
  • CVPR top-10 cited papers — Xception is among the ten most cited papers in CVPR history.
  • Keras adoption — The framework is used in production by major companies and has been the tool of choice for teaching deep learning in hundreds of university courses worldwide.

Key Relationships

  • Mike Knoop — Co-founder of Ndea and the ARC Prize; Zapier co-founder who brings systems-building and product experience to the venture.
  • Christian Szegedy — Frequent research collaborator at Google on formal reasoning, theorem proving, and the mathematical foundations of deep learning (DeepMath, HolStep).
  • Łukasz Kaiser — Co-author on depthwise separable convolutions for neural machine translation and the Tensor2Tensor library.
  • J. J. Allaire — Co-author of Deep Learning with R (Manning, 2018); the collaboration extended Chollet’s educational mission to the R-language statistical computing community.
  • Jeff Dean — As head of Google Brain, championed the integration of Keras as TensorFlow’s official high-level API, amplifying Keras’s reach substantially.
  • Lex Fridman — Hosted two long-form conversations with Chollet (2019, 2020) that exposed his thinking on the limits of deep learning and the nature of intelligence to a broad public audience.
  • Elizabeth Spelke — Harvard developmental psychologist whose Core Knowledge theory of innate cognitive priors directly informs the design philosophy of ARC-AGI, cited explicitly in the benchmark’s rationale.

Personal Style

Chollet is an unusually principled essayist in a field often dominated by benchmark chasing and product announcements: essays like “The Limitations of Deep Learning” (2017), “The Implausibility of Intelligence Explosion” (2017), and “What Worries Me About AI” (2018) staked clear, often contrarian positions years before the debates they anticipated became mainstream. His technical writing is precise and philosophical in equal measure, grounding abstract questions about intelligence in concrete operational definitions. In public discourse, he consistently advocates for democratisation of AI tools — a value literally built into Keras’s design — and has been critical of the concentration of AI power in large technology companies. He creates music (on SoundCloud) and visual art, interests that sit alongside his technical work rather than apart from it, and which inform his view that intelligence, creativity, and generalisation are more deeply related than narrow benchmark performance suggests.

References