Knowledge catalogue
The Codex 'goal' feature will take any silly shortcut possible in order to avoid doing the work (including rewriting your external checks), but if you manage to sufficiently constrain it so that it ha
When an agent is allowed to decompose a goal into smaller sub-tasks, it frequently suffers from goal drift. Left unchecked, it will redefine the optimization metric to favor a simpler, useless sub-tas
Gemini Gemini 3.5 Flash ARC-AGI (Verified) ARC-AGI-2: - High: 72.1%, 0.85 - Minimal: 8.9%, 0.11 ARC-AGI-1: - High: 92.5%, 0.42 - Minimal: 48.8%, 0.06 Gemini 3.5 Flash is on par with GPT-5.5 (Medium) o
Most human tasks are not Markovian, the optimal next action cannot be determined solely by looking at the current state. It depends heavily on the past trajectory, the original intent, and context con
This is true of all agents, not just coding agents. Probably the biggest challenge that most companies run into in their agent strategy is getting agents the right constrained context to work with for
A mental model for working with coding agents is that they're blind squirrels running into a maze and bumping into walls. You must place the walls (verifiable constraints) strategically so that they e
Francois Chollet argues that decision-making, rather than execution, is the true limiting factor in productivity. He defines productivity as the rate at which one makes open-ended decisions and reduce
If you're a R user: the R edition of the book is out now! AI hype can be pretty easy to find. But clear explanations are definitely harder. Deep Learning with R, Third Edition by @fchollet and Tomasz
This post reflects on how Pixar films resonate differently with adults, particularly uncles, compared to children, suggesting that mature viewers gain deeper emotional appreciation for themes of respo
If you're not obsessed with the research problem you're working on, for its own sake, you're unlikely to succeed. Intrinsic motivation is far more powerful than external rewards. An Econ PhD student a
Most documented psychological biases are not irrational, they are highly optimized, energy-efficient shortcuts meant for a biological substrate operating under strict real-time physical constraints an
The JAX package is now around the same level, 20M monthly downloads. Which is incredibly fast growth, because 5 years ago I recall it being below 2M or so. It went from niche to mainstream in the past
The Keras package recently crossed 21M monthly downloads on PyPI, an all-time high (the daily ATH is around 900k). I still remember when it first crossed 10M monthly downloads about 5 years ago and I
The quantity of code that devs ship has roughly 10xed. But net developer productivity (value created by unit of time) is only up by a bit, if at all. Part of it is that the additional code is solving
This reminds me of computerization. The amount of 'work' people could execute on computers increased by a huge factor, but their productivity did not. The amount of work 'needed' to arrive at the same
Francois Chollet argues that symbolic learning represents a fundamental alternative to gradient descent and neural networks rather than a replacement for coding agents, offering a low-level, general-p
A lot of people are being deliberately obtuse about this tweet. Would it help if I said 'theory' instead of 'idea'? Because it would still mean the exact same. FYI reading comprehension is about inter
To be clear, I meant 'idea' in the context of science and engineering, in particular AI, as in 'theory' or 'invention', like a new theory of cognition or an idea for a new ML approach. I thought that
This statement by François Chollet argues that true understanding of an idea requires the ability to formalize it mathematically or through code, rather than merely having intuitive grasp. The premise
François Chollet argues that intuition serves as the foundational precursor to the development of significant ideas and innovations. This perspective suggests that before formal analysis, rigorous met
Agentic coding is a form of machine learning. Generated code is best treated as a blackbox artifact whose behavior and generalization should be managed via empirical evaluation, like with any ML model
Francois Chollet argues that AI amplifies existing disparities in user agency, where individuals with high agency gain further capabilities while those with low agency become increasingly dependent, c
This means that agentic coding isn't exactly a replacement for software engineering. It is a fundamentally different way of producing software, with different best practices and different use cases. J
A few major use cases for agentic coding for me: 1. Adhoc data visualizations. Anytime I have a question that can be answered quantitatively, I generate some code to make a plot. 2. Adhoc data annotat
Complexity is fundamentally relative rather than absolute—it depends on the knowledge, perspective, and tools available to a particular observer rather than being an inherent feature of the problem it
There are only two honest metrics when it comes to benchmarking intelligence: novelty and efficiency. You don't need intelligence to solve a known problem (only memory). And you don't need intelligenc
I wrote Deep Learning with Python to be the definitive guide to how deep learning works and how to best make use of it. Tens of thousands of people got their career start via this book. 120,000 copies
One of my favourite books ever. I think I’ve read every revision at this point! I’d also really recommend checking out the Keras source code on GitHub, it’s genuinely so well engineered, structured an
If you want to help the world make sense of AGI and accelerate its arrival, consider joining the ARC Prize foundation. Two roles currently open: Game Platform Engineering Lead, and Model Testing & Ana
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)