François Chollet

AI researcher, creator of the Keras deep learning library, and the field’s best-known critic of the idea that scaling language models leads to AGI. In On the Measure of Intelligence (2019) he defined intelligence as how efficiently a system learns new skills, not how many skills it has, and built the ARC benchmark to test it: puzzles that are easy for people but were very hard for AI. He co-founded the ARC Prize in 2024 and left Google to found Ndea, an AGI research lab, in 2025.

Timelines

• 2024-06, Dwarkesh Podcast: language models are “an off-ramp on the path to AGI”, and the focus on scaling them has set back progress toward AGI by years.
• 2025: AGI in about five years: “A year ago, I would have said ten years-ish,” and “now I think it’s probably, you know, five-ish.” He credits new methods that let models adapt at test time, not scaling alone.

He is the clearest example of a skeptic whose timeline shortened without his giving up the critique.

See AGI timelines.