François Chollet

Created Keras as a high-level, user-friendly API for building neural networks, explicitly designed to be accessible to non-specialists. His philosophy: the most important factor in AI progress is the number of people who can run experiments. Keras accelerated deep learning adoption by making it approachable.

Created the Abstraction and Reasoning Corpus (ARC) — a set of visual pattern completion tasks trivially easy for humans but consistently difficult for state-of-the-art AI systems. His thesis: ARC exposes the gap between 'crystallized' skill (LLMs excel here) and 'fluid intelligence' (generalizing to novel situations with few examples).

In 'On the Measure of Intelligence' (2019), Chollet proposed defining intelligence as the ability to acquire new skills efficiently given limited prior knowledge and data. By this measure, LLMs score poorly: they require massive data to learn each skill and fail to generalize across slight variations. True intelligence requires learning programs, not patterns.

Argues that LLMs are essentially compressed databases of human skills, retrieved and remixed on demand. This is useful but not intelligence. Intelligence is what happens when you face a situation your training distribution doesn't cover. LLMs fail systematically in those situations — which is exactly what ARC-AGI is designed to test.

After OpenAI's o3 model showed dramatic ARC-AGI improvements in late 2024, Chollet updated his view: test-time compute (letting models 'think longer' via chain-of-thought and search) may bridge the gap between skill retrieval and reasoning. He moved ARC-AGI 2 goalposts: o3 solved v1, but new harder tasks remain.