François Chollet: Deep Learning Will Eventually Be Replaced by Discrete Program Search

fchollet · x · 2026-08-02

François Chollet points out that to address deep learning's limitations, the AI field has widely adopted the first patch: active inference at test time. While this brings meaningful progress, it remains an incremental improvement rather than the ultimate solution.

He emphasizes that long-term, AI will inevitably shift to the second approach: overhauling the training mechanism itself. This requires moving away from deep learning (curve fitting) and SGD altogether, embracing discrete program search to truly tackle complex reasoning.

Related event: François Chollet: Test-Time Compute is Key as Pure Parameter Scaling Hits Limits(4 posts)→

Original post →

More from AGI Musings

AGI Musings channel →