François Chollet

Tiny Recursive Model (TRM) with only 7M parameters achieves 45% accuracy on ARC-AGI-1 and outperforms much larger LLMs like Deepseek R1 and o3-mini using less than 0.01% of their parameters.

“Created ARC-AGI benchmark in 2019”

“Perspectives on intelligence measurement in AI, though article content is inaccessible.”

“Creator of ARC benchmark commenting on Astra's step-function change in capability for interactive reasoning.”

“Researcher who published ARC-AGI benchmark and provided different score figures than those in ARC Prize's published table.”

“Provided early review of ARC-AGI-3 analysis and leads ARC Prize research on measuring progress toward AGI.”

“Recommended chapters 15-16 of Deep Learning with Python as best explanations of dot-product attention mechanism.”

“Chollet conjectured that intelligence may be subject to diminishing returns, proposing that AI gains may be asymptoting rather than continuing to accelerate exponentially.”

“Creator of the Keras deep-learning library and co-founder of Ndea, speaking on a panel about durable open source projects.”

“Created ARC-AGI benchmarks showing frontier models struggle with abstract reasoning tasks that untrained humans solve easily.”

“Well-known machine learning researcher who created the ARC-AGI benchmark to measure intelligence based on how efficiently a system learns new skills.”

“Created ARC-AGI-3, an interactive reasoning benchmark where humans can solve 100% of tasks but top frontier AI models score less than 1%.”