Keras inventor Chollet charts a new direction for AI: a Q&A

Abstraction: Chollet critiques deep learning limits and proposes ARC generalization benchmark

Key points:

• Chollet (Google scientist, creator of Keras in 2015) argues deep learning performs interpolation over training data distribution, not extrapolation — "intelligence is not curve-fitting"
• Published "On the Measure of Intelligence" (2019) proposing a new formal definition of intelligence focused on skill-acquisition efficiency rather than task performance
• Abstraction and Reasoning Corpus (ARC): a benchmark of visual tasks solvable by humans with no training but "basically zero" performance by existing ML systems; relies on "core knowledge priors" (objectness, basic physics)
• Deep learning models are "brittle, extremely data-hungry, and do not generalize beyond their training data distribution" — an inescapable consequence of their architecture
• Criticizes multi-million dollar RL projects (e.g., AlphaStar, game-playing agents) as scientifically unproductive gimmicks generating press rather than new knowledge
• Chollet envisions a future "Keras for neuro-symbolic program synthesis" — a framework packaging general intelligence principles, but acknowledges this is still far away

Connections: Francois Chollet · Keras · Generalization · Deep Learning · AI Benchmarks

Source: https://www.zdnet.com/article/keras-creator-chollets-new-direction-for-ai-a-q-a/