Knowledge catalogue
Francois Chollet shared a blog post analyzing failure modes of frontier AI models, specifically examining performance on the ARC (Abstraction and Reasoning Corpus) AGI benchmark with models including
RL is a bit of a double edged sword: in known territory performance increases, but in unknown territory the model tends to hallucinate that it is performing a completely different task it was trained
The latest crop of models remains below 1% on ARC-AGI-3 -- for now. Where will the scores be by the end of the year? GPT-5.5 & Opus 4.7 on ARC-AGI-3 - GPT-5.5: 0.43% - Opus 4.7: 0.18% We found 3 failu
AI automates tasks, not jobs, and when a task gets cheaper, demand for the job grows. AI cannot automate jobs end-to-end because it lacks autonomy and cannot operate without supervision. There is stil
I built a TPU-native medical Q&A fine-tuning pipeline using Gemma 3 with Keras and JAX The project fine-tunes Gemma-3 on medical dialogue data from ChatDoctor and evaluates it on MedMCQA, with a large
ARC Prize 2026: ARC-AGI-3 now has H100s Thank you to our partners at @kaggle for powering the ARC-AGI-3 competition with H100s Participants now have a pool of accelerators (subject to submission limit
Keras Kinetic has a new alpha release: v0.0.2! Including a new docs website: http://kinetic.readthedocs.io Kinetic is my favorite new release from the Keras team: a super simple Modal-like API to run
Kinetic version 0.0.2 release notes are available on GitHub, detailing updates and features for this Keras-related project. The announcement was shared by François Chollet, Keras creator, on X (former
Describe Seattle: I want to live near the sea and I want to live near the forest and I want to live near the mountains and I want to live near a café and I want to live near a library and I want to li
François Chollet criticizes anonymous AI influencer accounts on social media, warning that following their content can lead to diminished critical thinking and poor judgment. The post cautions against
No, the top score if you didn't account for action efficiency would be 100%, achievable with 20 lines of Python. All you need is to brute-force the state space. Please stop spreading complete disinfor
Francois Chollet comments on Twitter's bi-weekly payout system, noting that while it may be effective for engagement farming, the financial incentives are likely substantial. The post appears to criti
To all the clowns saying that Seattle is always rainy and never sees the sun: Seattle has *less* precipitation than NYC, Boston, Atlanta, Miami, New Orleans, etc... in fact it rains less in Seattle th
Percy Liang delivered a keynote address at ICLR 2026, a major machine learning conference. The talk was announced and promoted by Francois Chollet on social media. The specific content of Liang's keyn
GPT-5.5 achieved state-of-the-art performance on the ARC-AGI-2 benchmark, with scores ranging from 85.0% on maximum difficulty tasks to 33% on low difficulty tasks. The model demonstrated consistent i
Judging AGI by how well it can mimic us is a category error, because mimicry isn't intelligence and isn't general. We should judge AGI by how well it learns to do things we didn't teach it (including
Harmattan is a Paris-based military drone company that produces more than quadcopters, according to a post by AI researcher François Chollet. The company's product line extends beyond standard quadrot
One of the most jarring things about current AI is its lack of introspection ability and metacognition. It doesn't know what it doesn't know, how it knows, or how it could find out. It's a one-way sys
The very first quadcopter prototype was the Breguet-Richet Gyroplane No. 1, created in 1907 by Breguet Aviation in France. The first consumer quadcopter drone was the Parrot AR.Drone, released in 2010
Human biological limits, like our tiny working memory and shallow calculation depth, are actually a feature. They force us to abstract, compress, intuit. If we had infinite resources, we would never h
In a capitalist economy, absolute levels of value creation and productivity take a backseat to relative competitive advantage. That's why jobs still exist despite 150 years of intense automation. As l
In a hypothetical scenario where all tasks could be automated at zero cost, François Chollet argues that everyone would work in interpersonal roles, suggesting that human value would shift entirely to
Exactly. The number of iterations you do is far more important than the amount of 'thinking' you do. It doesn't matter how intelligent or knowledgeable you are. Reality will always push back. The worl
Human cognitive friction has long been acting as a regularizer for a lot of digital infrastructure. It made software APIs less terrible and codebases less complex. Now LLM disintermediation is causing
I'd say cognitive friction is even more than just a regularizer: it's an incentive to find the right interface abstractions, and in turn good abstractions are what enables compounding over time. Piles
There's no doubt that the world can consume tokens as fast as they're produced, even in the most maximalist infrastructure buildup scenarios imaginable. That's not the question. The question is whethe
Unlimited demand for a product or service does not guarantee business viability or economic sustainability, as other critical factors such as cost structure, scalability, profitability, and resource c
Constraints are the catalyst of invention. An infinite search space leads to paralysis. The most creative inventions happen when you are forced to solve a problem within appropriately narrow constrain
The quality of thinking acts as a multiplier on progress made during iterative work, but iteration frequency itself is the primary driver of success. This insight from AI researcher François Chollet e
Francois Chollet compares programming language philosophies between PHP and Go, likely discussing design trade-offs, developer experience, or use cases that differentiate the two languages. The compar
Francois Chollet discusses a hiring signal for identifying high-quality deep learning candidates based on their choice of machine learning frameworks, suggesting PyTorch proficiency (or lack thereof)
You cannot think your way to a perfect design. Only building and testing, over many iterations, can reveal the flaws in your mental model and provide the feedback you need to create the best design po
This post discusses how software development, like other engineering disciplines, requires substantially more effort and code when building scalable solutions compared to quick prototypes or one-off i
There's a broadly held misconception in AI that methods that scale well are simple methods -- even, that simple methods usually scale. This is completely wrong. Pretty much none of the truly simple me
ARC-AGI-3 is a new benchmark iteration designed so that any intelligent human applying genuine effort should achieve greater than 90% accuracy, maintaining the benchmark's core principle that tasks mu
ARC-AGI-3 has the lowest human bar of any AI benchmark out there. Almost all benchmarks require specialized knowledge that make them inaccessible to 99%+ of humans (like, say SWE-Bench). ARC-AGI-3 is
Francois Chollet makes the observation that achieving a perfect score on certain AI benchmark games or tasks only requires surpassing the median performance of an unfiltered random population sample,
ARC-AGI-3 Human Baseline Dataset Today we're open-sourcing the ARC-AGI-3 Human Baseline. This is the most exhaustive human testing study in the ARC-AGI series Every environment was solved by at least
In order to make sure ARC-AGI-3 games were solvable by humans we tested over 450 people If a game was too hard we either revised it or tossed it completely For maximum transparency we just open source
Simply retrieving a reasoning trace looks a lot like human reasoning, until it's time to navigate uncharted territory. If you memorized all reasoning traces of humans from 10,000 BC, you could automat
Francois Chollet argues that memorization and knowledge function as a cache for previously completed cognitive work, allowing it to be reused efficiently rather than recomputed. The purpose of storing
Francois Chollet, AI researcher and creator of Keras, argues that good design is fundamentally an act of compression — distilling an enormous number of implementation decisions, constraints, and consi
The power of JAX Introducing gyaradax 🐉: A JAX solver for local flux-tube gyrokinetics with custom CUDA kernels for acceleration. This entire code was vibecoded by @ggalletti_ and me in a month. Valid
The reason symmetry is so important in physics is because symmetry is a highly effective compression operator. If a system is invariant under some symmetry, you only need to explain one axis of it. Sc
Science needs a way to process models that are only 'mostly correct' in terms of their predictions, but are very compressive (high ratio between predictive power and model complexity). They are likely
Simplicity is a very strong signal of model quality. Galileo's heliocentric model was directionally correct, but its predictive power was quite bad compared to the much older Ptolemaic model, because
We should view the history of physics as a long-running program synthesis task. Kepler and Newton were searching the space of possible symbolic models to find the simplest one that would best satisfy
The new model from Meta is already looking like a disappointment: overoptimized for public benchmark numbers at the detriment of everything else. Knowing how to evaluate models in a way that correlate
Join the ARC Prize team -- help us build ARC-AGI-4 and ARC-AGI-5 Platform Engineer - Benchmark Lead ARC Prize Foundation is hiring a senior engineer to build our benchmark platform * Expand ARC-AGI-3