Knowledge catalogue
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