Created Keras and ARC, and examines how to distinguish learned task performance from the ability to solve unfamiliar problems.
Why read François Chollet?
Read Chollet when a benchmark score is being used to make a broad claim about intelligence. In his sole-authored 2019 paper, he proposes measuring how efficiently a system acquires new skills while accounting for prior knowledge and experience. ARC turns that question into small grid puzzles whose rule must be inferred from examples. It makes the evaluation question visible before the formal definitions become necessary.
The ARC Prize 2024 report, coauthored with the competition team, adds practical reasons to inspect the testing setup. Its two leaderboards use different compute budgets and conditions, so their scores are not directly interchangeable. Start with a grid task, then compare the leaderboard rules. This reading order helps you ask which dataset, adaptation method and resource budget a reported result actually describes.
Start with the original
Selected work
Research paper ·
On the Measure of Intelligence
His own paper introducing ARC and a proposed definition of intelligence. Start with section III.1's grid example, then the distinction between acquired skill and learning a new task.
Research report ·
ARC Prize 2024: Technical Report
Coauthored report. Sections 1 and 2 explain the puzzle and testing conditions; section 3 compares program search with adapting a model at test time. Results describe the 2024 competition.
Projects & roles
Community rating
Sources & further reading
- His current personal site links @fchollet and identifies Ndea, ARC Prize, Keras and the selected papers.
- Current X metadata names François Chollet and lists Ndea, ARC Prize, Keras and ARC-AGI.
- The company's primary team page confirms the co-founder relationship.
- Read for the proposed definition, ARC design and its acknowledged weaknesses.
- Read for leaderboard conditions, compute budgets and the 2024 methods comparison.