François Chollet
Last reviewed
May 31, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v2 · 3,527 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 31, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v2 · 3,527 words
Add missing citations, update stale details, or suggest a clearer explanation.
François Chollet (born 20 October 1989) is a French software engineer and artificial intelligence researcher best known as the creator of the keras deep-learning library, the originator of the Abstraction and Reasoning Corpus (ARC) benchmark for measuring general intelligence, and a co-founder of the AI research lab ndea.[1][2] He spent close to a decade at google before leaving in November 2024 to launch a new company, and in January 2025 he announced Ndea, an AGI-focused research lab pursuing deep learning-guided program synthesis.[3][4]
Chollet is the author of the textbook Deep Learning with Python and the lead engineer behind multiple generations of Keras, including the multi-backend Keras 3 rewrite released in 2023.[5][6] In 2019 he published the influential essay "On the Measure of Intelligence", which proposed a formal definition of intelligence as "skill-acquisition efficiency" and introduced the ARC benchmark.[7] The benchmark became the basis of the arc agi competition series, which Chollet and mike knoop turned into the ARC Prize in June 2024 with a US$1 million Kaggle competition; the two later set up the ARC Prize Foundation as a nonprofit dedicated to AGI benchmarks.[8][9]
Chollet has become one of the most widely read AI commentators of the 2020s, both for his technical contributions and for his vocal skepticism that scaling large language models alone will produce artificial general intelligence. In September 2024 TIME named him one of the 100 most influential people in AI.[10]
| Born | 20 October 1989 (age 36) in France[1] |
| Nationality | French[1] |
| Education | Diplôme d'Ingénieur (Master of Engineering), ENSTA Paris, 2012[1] |
| Known for | keras, ARC / ARC-AGI benchmark, Deep Learning with Python, xception architecture[1][7][11] |
| Current roles | Co-founder, ndea (Jan 2025–); co-founder, ARC Prize Foundation (2024–); external lead of Keras open-source project[4][9][12] |
| Previous role | Senior Staff Engineer, google (2015 – November 2024)[1][3] |
| Notable papers | "Xception: Deep Learning with Depthwise Separable Convolutions" (CVPR 2017); "On the Measure of Intelligence" (arXiv:1911.01547, 2019)[7][11] |
| Awards | Global Swiss AI Award (2021); TIME 100 AI (2024)[1][10] |
Chollet was born on 20 October 1989 in France.[1] Public biographies do not identify his hometown.
He pursued his higher education at the École Nationale Supérieure de Techniques Avancées (ENSTA Paris), an engineering school that is part of the Institut Polytechnique de Paris, graduating with a Diplôme d'Ingénieur (equivalent to a Master of Engineering) in 2012.[1] On his personal site he summarises his early research interests as the application of machine learning to computer vision and natural language processing, along with a long-standing interest in understanding and democratising AI.[12]
Before joining google he worked on a series of independent software projects. Wysp, an online social network and learning community for visual artists that he created and ran from 2011 to 2014, peaked at around 25,000 monthly active users; QuickAnswers, a 2014–2015 question-answering and dialogue system, used a retrieval-augmented technique similar to what later became known as retrieval-augmented generation (RAG) but predated the wide use of large language models.[12] These projects established many of the design instincts, including an emphasis on usability, clean APIs and short feedback loops, that would later define Keras.
Chollet began developing Keras in early 2015 as a personal project, with the goal of making it easier to prototype deep neural networks in Python. The library was released as open source in March 2015, originally running on top of the Theano library; support for tensorflow was added later the same year after Google released that framework.[13][14]
Keras was designed around three principles that Chollet has frequently emphasised: user friendliness, modularity and extensibility. By providing a high-level API that hides much of the boilerplate of low-level tensor frameworks, it became one of the most widely adopted deep-learning libraries in the world; the Google Developers blog noted at the time of his departure in 2024 that Keras had more than two million users and powered systems including Waymo's autonomous-driving stack and major recommendation systems.[15]
In 2017, Keras was adopted by google as the official high-level API of tensorflow, and Chollet led the integration that produced the tf.keras namespace and, later, TensorFlow 2.0.[14] In an interview at the time of the TensorFlow 2.0 launch he described the goal of the integration as preserving the simple, layer-oriented programming model of Keras while exposing the production tooling and scalability of TensorFlow underneath, so that the same code could be used for research prototypes and large industrial deployments.[14]
Keras 2 followed, and in November 2023 Chollet and the Keras team released Keras 3, a full rewrite of the library as a multi-backend framework able to run identical models on top of JAX, TensorFlow and PyTorch.[6] The rewrite was a significant strategic shift: rather than coupling Keras to a single tensor backend, Keras 3 made the choice of backend a configuration option, so that users could write Keras code once and execute it on the framework most suited to their hardware or workflow.[6] Keras 3.8, released in 2025, added an inference-only OpenVINO backend.[6]
After Chollet left google in 2024, he confirmed that he would continue to lead the open-source Keras project from outside the company; Google said that the Keras team at Google would continue collaborating with him on its roadmap.[15][16]
Chollet joined Google in 2015 as a software engineer working on machine learning, shortly after the open-source release of Keras.[1] Over the following decade he rose to the title of Senior Staff Engineer, with Keras formally absorbed into Google's machine-learning stack as part of TensorFlow.[1] During his time at the company his work covered a range of areas, including computer vision and the application of deep learning to formal reasoning and abstraction.
Two pieces of research from this period stand out. The first is the xception architecture, published as "Xception: Deep Learning with Depthwise Separable Convolutions" at CVPR in 2017 (preprint arXiv:1610.02357, 2016). The paper reinterprets Inception modules as an intermediate step between standard convolutions and depthwise-separable convolutions, and uses this insight to design an architecture that outperformed Inception V3 on ImageNet and on an internal 350-million-image, 17,000-class classification task at the same parameter count.[11] Xception has been cited tens of thousands of times and remains a reference architecture for efficient convolutional networks.
The second is "On the Measure of Intelligence", posted to arXiv on 5 November 2019, which introduced both a formal definition of intelligence grounded in algorithmic information theory and the Abstraction and Reasoning Corpus (see below).[7] Although Chollet wrote the paper while at Google, the ARC benchmark itself was released as an independent open project.
Chollet announced his departure from Google on 13 November 2024 via the Google Developers blog and a personal post on X dated 14 November 2024.[15][17] In his X post he wrote that he was "leaving Google to go start a new company with a friend" and that he would remain "deeply involved with the Keras project from the outside".[17] At that point he had been at the company for close to ten years.
The departure attracted significant industry attention, partly because of Chollet's profile but also because it came amid a wave of high-profile exits from large AI labs.[3] The Google Developers blog post acknowledging the change emphasised that Keras would remain Google's recommended high-level deep-learning API and would continue to support JAX, TensorFlow and PyTorch under his external stewardship.[15]
For roughly two months Chollet did not reveal the name of the new venture. On 14 January 2025 he announced ndea, an AI research and science lab co-founded with mike knoop, the co-founder and former head of AI at Zapier.[4][18] In an accompanying post on X, Chollet wrote that the company would focus on "deep learning-guided program synthesis" and that it was "betting on a different path to build AI capable of true invention, adaptation, and innovation".[19]
Ndea positions itself as a "frontier AI lab" pursuing artificial general intelligence through a hybrid approach that combines deep learning with discrete program synthesis (the search for short programs that explain observed data) rather than relying primarily on scaling up large language models.[4][20] Chollet has argued that program synthesis is computationally intensive but generalises far more efficiently than gradient-based pattern matching, because a compact program can capture a rule that generalises perfectly from a handful of examples to an unbounded set of new inputs, whereas a neural network typically requires many more examples to interpolate the same regularity.[4][20] Overcoming the combinatorial cost of program search is, in Chollet's framing, the central technical problem on the path to agi, and the founding thesis of Ndea is that the right way to attack it is to use deep learning models as fast, approximate guides that bias program search toward plausible candidates.[4][20]
Knoop, who stepped back from Zapier's day-to-day operations to focus on the new venture while remaining on Zapier's board, has described the team's ambition as building "the world's top program synthesis team".[4] At launch, Ndea was hiring remote research roles and had not publicly disclosed details of any outside funding.[4]
Ndea operates separately from but is intellectually aligned with the ARC Prize Foundation, which Chollet and Knoop set up in 2024 to develop benchmarks for AGI; the foundation was announced as transitioning to a 501(c)(3) nonprofit in early 2025 with Greg Kamradt, a former Salesforce engineering director, as president.[9] Chollet has been clear that Ndea is a for-profit research company whose products will be built on top of the ideas it develops, while the foundation is intended as a neutral benchmarking organisation that any lab can engage with.[4][9]
Chollet's most influential piece of basic research is the 2019 essay "On the Measure of Intelligence" (arXiv:1911.01547).[7] Drawing on algorithmic information theory, the paper proposes that intelligence be defined as skill-acquisition efficiency: the rate at which a system can convert priors and experience into new skills over a broad scope of tasks. The essay argues that benchmarks built around fixed tasks measure only specific skills and reward memorisation rather than the ability to generalise to genuinely novel problems.[7]
The same paper introduces the Abstraction and Reasoning Corpus (ARC), a benchmark of visual grid puzzles in which a system is given a handful of input-output example grids and must infer the underlying transformation rule in order to solve a held-out test grid. The tasks are designed to rely only on a small set of human "core knowledge priors" (objectness, basic geometry, counting and elementary causality) and to require no specific outside knowledge.[7][21] Although the puzzles are visually simple, they are conceptually rich: solving them typically requires identifying compositional rules involving objects and relations, geometry, counting, visual instructions and logical operations, in a setting where each task differs substantially from every other, so memorising solutions is not a viable strategy.[7]
Chollet ran the first ARC competition on Kaggle in 2020; the winning team reached only about 20 percent accuracy, which Chollet later cited as the first empirical evidence that the ideas in "On the Measure of Intelligence" were directionally correct.[21] Subsequent competitions were organised in partnership with the Swiss nonprofit Lab42: ARCathon 2022, which attracted 118 teams from 47 countries, and ARCathon 2023, which expanded to more than 265 teams from 65 countries and saw the top score reach roughly 30 percent.[21]
On 11 June 2024, Chollet and mike knoop launched the ARC Prize 2024, a US$1 million competition on Kaggle structured around a $600,000 grand prize for the first solution to reach 85 percent accuracy on the private ARC-AGI evaluation set, alongside $50,000 in progress prizes tied to the Kaggle leaderboard and $75,000 in paper-submission awards.[8] At launch it was described as the largest prize pool ever offered on Kaggle.[9] The announcement was accompanied by a high-profile appearance on Dwarkesh Patel's podcast in which Chollet argued that current large language models alone would not scale to AGI and that program synthesis was a more promising direction.[22]
Although the grand prize was not claimed, the state of the art on the private set rose from 33 percent to 55.5 percent over the course of the year, propelled by deep learning-guided program synthesis and test-time training; results were published in a technical report at NeurIPS.[8][9] By the end of the 2024 cycle, more than 1,400 teams had submitted approximately 17,000 entries.[9]
After the 2024 cycle, Chollet and Knoop expanded the effort into the ARC Prize Foundation, a nonprofit charged with developing rigorous AGI benchmarks. Greg Kamradt, a former Salesforce engineering director, joined as president.[9] The foundation announced in early 2025 that it would operate as a 501(c)(3) and that its long-term goal was to establish ARC-AGI as an academic, policy and industry standard for measuring and guiding progress towards agi.[9]
On 24 March 2025, the foundation released arc agi 2, a successor benchmark that preserved the input-output grid format of the original ARC but introduced a more curated and substantially harder task set designed to provide a finer-grained signal at higher levels of fluid intelligence.[23] At release, every task had been solved by at least two non-expert human testers, while frontier AI systems struggled to make significant progress. The corresponding ARC Prize 2025 ran from 26 March to 3 November 2025, attracted 1,455 teams and more than 15,000 submissions, and saw the top team reach 24 percent on the private ARC-AGI-2 set at a cost of US$0.20 per task.[23][24] A technical paper describing the benchmark was posted to arXiv in May 2025.[25]
The foundation unveiled arc agi 3 on 25 March 2026 at Y Combinator's headquarters in San Francisco, with a launch event featuring a fireside conversation between Chollet and Sam Altman.[26] In contrast to the static grid-puzzle format of the previous corpora, ARC-AGI-3 is the first fully interactive benchmark in the series: it consists of hundreds of bespoke turn-based environments containing thousands of levels in which an agent must, without any instructions or stated goals, explore the environment, infer the rules and acquire goals through trial and observation.[26][27] At release, humans scored 100 percent on the environments while frontier AI agents scored below 1 percent.[26]
The accompanying ARC Prize 2026 competition is the largest in the series so far, offering more than US$2 million in total prizes across an ARC-AGI-3 agent-building track and a parallel ARC-AGI-2 track, with a US$700,000 grand prize on the agent track for the first eligible system to reach 100 percent on the private set.[26]
Chollet's principal book-length work is "Deep Learning with Python", published by Manning. The first edition appeared in 2017 and has been translated into more than ten languages; Manning describes the book as having sold over 100,000 copies across its editions.[28] The book is structured around Keras and aims to teach deep learning by building working models for image classification, text generation and sequence modelling, rather than by starting from mathematical first principles, which made it a widely adopted teaching text in university courses and industry training programs through the late 2010s and 2020s.[28][5] A substantially rewritten second edition was published in 2021, updated to reflect the integration of Keras with TensorFlow 2 and to expand the treatment of generative models and transformers, and a third edition co-authored with Matthew Watson is in preparation.[5][29]
He is also a co-author, with J.J. Allaire and Tomasz Kalinowski, of "Deep Learning with R", an R-language adaptation of the same material first published in 2018.[5][12]
Beyond his books, Chollet has been a prolific writer of long-form essays, most notably "On the Measure of Intelligence," and an active contributor to the Keras documentation, blog and developer community. He is also one of the most-followed AI researchers on X (formerly Twitter), where he posts both technical commentary and longer threads on the foundations of intelligence, generalisation and the limits of current AI systems.[5][12][22]
Chollet is widely associated with a distinctive set of public positions about the path to artificial general intelligence. The core of his argument is that skill is not the same as intelligence: a system that has been trained to solve a fixed set of tasks demonstrates skill on those tasks, but intelligence is the efficiency with which a system can pick up new skills given limited priors and experience.[7][22]
From this premise he has argued, most prominently in his June 2024 conversation with Dwarkesh Patel and in subsequent commentary, that scaling up large language models will not by itself produce AGI: LLMs improve on benchmarks largely because their training distributions expand to cover those benchmarks, not because they become better at adapting to genuinely novel problems.[22] He has used the persistent difficulty of ARC for state-of-the-art LLMs as a key piece of empirical support, observing that even very large models trained on enormous corpora struggle on the benchmark because each task is unfamiliar and cannot be solved by retrieving or interpolating a memorised pattern; the system has to synthesise a new solution program on the fly.[22]
In Chollet's view, the path forward is a hybrid of deep learning and discrete program synthesis, in which neural networks guide a search over compact programs that can compose to solve unfamiliar tasks, an approach that motivates the technical research agenda at ndea.[4][22] He has framed this position not as a dismissal of deep learning, of which he is a leading proponent and educator, but as an argument that deep learning needs to be complemented with mechanisms for explicit search and compositional generalisation.[4][22]
He has also been a vocal critic of conflating benchmark progress with progress toward general intelligence and has used ARC's continued resistance to even the largest frontier models as evidence that current systems lack a key component of human-like fluid intelligence.[22][26] At the ARC-AGI-3 launch in March 2026 he reiterated, in conversation with Sam Altman, that the very low frontier scores against essentially perfect human performance underscored how far current agents remained from genuine fluid reasoning.[26]