@francois-chollet · 53 positions · 2 changes of mind · On ppll ↗
Creator of the Keras deep-learning library and the ARC-AGI benchmark.
François Chollet did not write this page.
We collected these quotes from things they published elsewhere, and every quote links to where it was said. They have no account here and have not endorsed this site. Quotes are word for word; the short line under each one is our own restatement, not their wording.
Their own site.
Is this you? Claim it or ask us to remove it.Or tell us what is wrong here.
Each column is a stretch of time in which the same views came up; a new column starts when one appears or goes quiet. Quiet means nothing on it in our sources for that stretch, which is not the same as a change of mind; Dropped means they said they no longer hold it; tap the card for which claim and when. Revised means the view carries on but one of its claims was dropped. Each row is one view, in the same place all the way along. The number is how many claims in that stretch; we hold 5 in 2019, 0 in 2020, 0 in 2021, 0 in 2022, 0 in 2023, 4 in 2024, 5 in 2025, 46 in 2026.
Intelligence is how efficiently a mind acquires new skills, not how skilled it is at any one task
Our label. 6 claims in François Chollet's own words.
Their wordsIf intelligence lies in the process of acquiring skills, then there is no task X such that skill at X demonstrates intelligence, unless X is actually a meta-task involving skill-acquisition across a broad range of tasks.
François CholletCreator of the Keras deep-learning library and the ARC-AGI benchmarkWe argue that solely measuring skill at any given task falls short of measuring intelligence, because skill is heavily modulated by prior knowledge and experience: unlimited priors or unlimited training data allow experimenters to "buy" arbitrary levels of skills for a system, in a way that masks the system's own generalization power.↗On the Measure of Intelligencearxiv.org · 5 Nov 2019All korrents from this piece
Their wordsWe argue that solely measuring skill at any given task falls short of measuring intelligence, because skill is heavily modulated by prior knowledge and experience: unlimited priors or unlimited training data allow experimenters to "buy" arbitrary levels of skills for a system, in a way that masks the system's own generalization power.
Their wordsToday, it is increasingly apparent that both of these views of the nature of human intelligence - either a collection of special-purpose programs or a general-purpose Tabula Rasa - are likely incorrect
Their wordsWe argue that ARC can be used to measure a human-like form of general fluid intelligence and that it enables fair general intelligence comparisons between AI systems and humans.
François CholletCreator of the Keras deep-learning library and the ARC-AGI benchmarkTo make deliberate progress towards more intelligent and more human-like artificial systems, we need to be following an appropriate feedback signal: we need to be able to define and evaluate intelligence in a way that enables comparisons between two systems, as well as comparisons with humans.↗On the Measure of Intelligencearxiv.org · 5 Nov 2019All korrents from this piece
Their wordsTo make deliberate progress towards more intelligent and more human-like artificial systems, we need to be following an appropriate feedback signal: we need to be able to define and evaluate intelligence in a way that enables comparisons between two systems, as well as comparisons with humans.
Their wordsPassing ARC-AGI does not equate to achieving AGI, and, as a matter of fact, I don't think o3 is AGI yet. o3 still fails on some very easy tasks, indicating fundamental differences with human intelligence.
Their wordsWe contend that games provide particularly rich and controlled environments uniquely well-suited for systematically evaluating rapid model adaptation and the process of world model induction.
Their wordsWe expect that building and evaluating AI systems capable of this kind of rapid world model induction will be critical for achieving robust, general AI capable of functioning effectively in the complex and fast-changing real world, and especially in human worlds - the environments that human beings have evolved in, created, and are continually changing and re-creating.
Their wordsHowever, current understanding and evaluation of world models in artificial intelligence (AI) remains narrow, often focusing on static representations learned from training on massive corpora of data, instead of the efficiency and efficacy in learning these representations through interaction and exploration within a novel environment.
Their wordsWe argue that this profound adaptability is fundamentally linked to the efficient construction and refinement of internal representations of the environment, commonly referred to as world models, and we refer to this adaptation mechanism as world model induction.
François CholletCreator of the Keras deep-learning library and the ARC-AGI benchmarko3's improvement over the GPT series proves that architecture is everything. You couldn't throw more compute at GPT-4 and get these results. Simply scaling up the things we were doing from 2019 to 2023 -- take the same architecture, train a bigger version on more data -- is not enough.↗OpenAI o3 Breakthrough High Score on ARC-AGI-Pubarcprize.org · 20 Dec 2024
Their wordso3's improvement over the GPT series proves that architecture is everything. You couldn't throw more compute at GPT-4 and get these results. Simply scaling up the things we were doing from 2019 to 2023 -- take the same architecture, train a bigger version on more data -- is not enough.
François CholletCreator of the Keras deep-learning library and the ARC-AGI benchmarkI believe current techniques are 4-6 orders of magnitude away from optimality in terms of data efficiency and test-time compute efficiency. But far future AI will be near-optimal.↗@fchollet on Xx.com · 7 Aug 2026
Their wordsI believe current techniques are 4-6 orders of magnitude away from optimality in terms of data efficiency and test-time compute efficiency. But far future AI will be near-optimal.
François CholletCreator of the Keras deep-learning library and the ARC-AGI benchmarkHowever, looking ahead, I still do not believe that future AI (say, in 15 years) will be based on the LLM stack. I believe it will necessarily have to move closer to its optimal, final form -- symbolic learning. Obviously this is a risky and contrarian belief -- the safe bet would be LRMs. But let's see.↗@fchollet on Xx.com · 7 Aug 2026
Their wordsHowever, looking ahead, I still do not believe that future AI (say, in 15 years) will be based on the LLM stack. I believe it will necessarily have to move closer to its optimal, final form -- symbolic learning. Obviously this is a risky and contrarian belief -- the safe bet would be LRMs. But let's see.
François CholletCreator of the Keras deep-learning library and the ARC-AGI benchmarkIf we want actual autonomous robotics we have to engineer causality and discrete reasoning back into the loop.↗@fchollet on Xx.com · 14 Oct 2026
Their wordsIf we want actual autonomous robotics we have to engineer causality and discrete reasoning back into the loop.
François CholletCreator of the Keras deep-learning library and the ARC-AGI benchmarkHow AI shifted from a transductive paradigm (token-by-token answer completion) to an inductive paradigm (on-the-fly reasoning chain synthesis), in one chart. This is the big story of 2025 and 2026.↗@fchollet on Xx.com · 10 Oct 2026
Their wordsHow AI shifted from a transductive paradigm (token-by-token answer completion) to an inductive paradigm (on-the-fly reasoning chain synthesis), in one chart. This is the big story of 2025 and 2026.
François CholletCreator of the Keras deep-learning library and the ARC-AGI benchmarkSeems like test time scaling has gained a 3rd axis: latent space reasoning iterations in looped transformers.↗@fchollet on Xx.com · 5 Sept 2026
Their wordsSeems like test time scaling has gained a 3rd axis: latent space reasoning iterations in looped transformers.
François CholletCreator of the Keras deep-learning library and the ARC-AGI benchmarkIn verifiable domains, model capability scaling should remain unbounded. Models will simply keep improving by "absorbing more and more of the computational universe", which is infinite by construction.↗@fchollet on Xx.com · 27 Aug 2026
Their wordsIn verifiable domains, model capability scaling should remain unbounded. Models will simply keep improving by "absorbing more and more of the computational universe", which is infinite by construction.
François CholletCreator of the Keras deep-learning library and the ARC-AGI benchmarkSo I would not "declare AGI" until we have AI that is, at last, capable of invention -- conceptual breakthroughs, novel insights, new real-world technology, etc.↗@fchollet on Xx.com · 7 Sept 2026
Their wordsSo I would not "declare AGI" until we have AI that is, at last, capable of invention -- conceptual breakthroughs, novel insights, new real-world technology, etc.
François CholletCreator of the Keras deep-learning library and the ARC-AGI benchmarkThe narrative around why building AGI was a good idea has always been about invention -- "it will cure cancer", "it will solve fusion energy", etc.↗@fchollet on Xx.com · 7 Sept 2026
Their wordsThe narrative around why building AGI was a good idea has always been about invention -- "it will cure cancer", "it will solve fusion energy", etc.
François CholletCreator of the Keras deep-learning library and the ARC-AGI benchmarkIncredible watering down -- the Singularity is now redefined to mean "the rate of new firm creation has increased somewhat" Vernor Vinge described the Singularity as an event horizon past which everything (e.g. what happens tomorrow) becomes entirely unimaginable and unpredictable to human understanding↗@fchollet on Xx.com · 19 Aug 2026
Their wordsIncredible watering down -- the Singularity is now redefined to mean "the rate of new firm creation has increased somewhat" Vernor Vinge described the Singularity as an event horizon past which everything (e.g. what happens tomorrow) becomes entirely unimaginable and unpredictable to human understanding
François CholletCreator of the Keras deep-learning library and the ARC-AGI benchmarkThe only way to avoid extreme power concentration is to ensure we have multiple independent providers of frontier AI models, including open-source options.↗@fchollet on Xx.com · 13 Sept 2026
Their wordsThe only way to avoid extreme power concentration is to ensure we have multiple independent providers of frontier AI models, including open-source options.
François CholletCreator of the Keras deep-learning library and the ARC-AGI benchmarkThe ideal outcome for science isn't a couple of labs using proprietary models and millions of dollars to crack trophy problems, it's to have millions of AI-aided human scientists who follow their curiosity and become way more ambitious↗@fchollet on Xx.com · 14 Oct 2026
Their wordsThe ideal outcome for science isn't a couple of labs using proprietary models and millions of dollars to crack trophy problems, it's to have millions of AI-aided human scientists who follow their curiosity and become way more ambitious
François CholletCreator of the Keras deep-learning library and the ARC-AGI benchmarkDo not confuse competent information processing and sentience, they are wholly unrelated.↗@fchollet on Xx.com · 3 Oct 2026
Their wordsDo not confuse competent information processing and sentience, they are wholly unrelated.
François CholletCreator of the Keras deep-learning library and the ARC-AGI benchmarkThe notion that current AI models are sentient and can suffer, combined with the foolish idea that suffering can be mathematically quantified and weighted between humans and non-humans, could lead us down an incredibly dark and dystopian path.↗@fchollet on Xx.com · 3 Oct 2026
Their wordsThe notion that current AI models are sentient and can suffer, combined with the foolish idea that suffering can be mathematically quantified and weighted between humans and non-humans, could lead us down an incredibly dark and dystopian path.
François CholletCreator of the Keras deep-learning library and the ARC-AGI benchmarkHumanity has not yet created a conscious computer program, and there are no signs we are close to doing so.↗@fchollet on Xx.com · 4 Oct 2026
Their wordsHumanity has not yet created a conscious computer program, and there are no signs we are close to doing so.
François CholletCreator of the Keras deep-learning library and the ARC-AGI benchmarkLikewise a static input-output program that has *none* of the properties we associate with conscious beings (e.g. information integration, interoception, temporal binding, embodiment, etc.) has no more reason to be presumed conscious than a rock has to be presumed alive.↗@fchollet on Xx.com · 4 Oct 2026
Their wordsLikewise a static input-output program that has *none* of the properties we associate with conscious beings (e.g. information integration, interoception, temporal binding, embodiment, etc.) has no more reason to be presumed conscious than a rock has to be presumed alive.
Noah SmithEconomics writer; author of the Noahpinion newsletterHowever, this is the most convincing argument against machine consciousness that I have ever come across.↗@Noahpinion on Xx.com · 3 Oct 2026
Their wordsHowever, this is the most convincing argument against machine consciousness that I have ever come across.
François CholletCreator of the Keras deep-learning library and the ARC-AGI benchmarkcreating artificial consciousness should *not* be a goal of AI research. Understanding consciousness, sure, synthesizing it, no.↗@fchollet on Xx.com · 4 Oct 2026
Their wordscreating artificial consciousness should *not* be a goal of AI research. Understanding consciousness, sure, synthesizing it, no.
François CholletCreator of the Keras deep-learning library and the ARC-AGI benchmarkFor this reason, intelligence is practically never the bottleneck to solving impactful problems.↗@fchollet on Xx.com · 14 Oct 2026
Their wordsFor this reason, intelligence is practically never the bottleneck to solving impactful problems.
François CholletCreator of the Keras deep-learning library and the ARC-AGI benchmarkAutomating one aspect of a system creates its own friction and moves the bottlenecks elsewhere.↗@fchollet on Xx.com · 17 Oct 2026
Their wordsAutomating one aspect of a system creates its own friction and moves the bottlenecks elsewhere.
François CholletCreator of the Keras deep-learning library and the ARC-AGI benchmarkIf an agent saves an engineer 2 hours to implement a feature but collectively costs the team 3 hours in alignment and review, the ROI is negative.↗@fchollet on Xx.com · 13 Oct 2026
Their wordsIf an agent saves an engineer 2 hours to implement a feature but collectively costs the team 3 hours in alignment and review, the ROI is negative.
François CholletCreator of the Keras deep-learning library and the ARC-AGI benchmarkTo achieve sustained exponential progress, the AI industry will need to achieve matching exponential profits -- it will need to close the resource loop.↗@fchollet on Xx.com · 8 Oct 2026
Their wordsTo achieve sustained exponential progress, the AI industry will need to achieve matching exponential profits -- it will need to close the resource loop.
François CholletCreator of the Keras deep-learning library and the ARC-AGI benchmarkBut super-exponential injection of *external* resources is not sustainable in the long-term.↗@fchollet on Xx.com · 8 Oct 2026
Their wordsBut super-exponential injection of *external* resources is not sustainable in the long-term.
François CholletCreator of the Keras deep-learning library and the ARC-AGI benchmarkThis is why science as a system requires exponential resources (input) to produce linear impact (output).↗@fchollet on Xx.com · 9 Oct 2026
Their wordsThis is why science as a system requires exponential resources (input) to produce linear impact (output).