Chollet: Advanced agentic coding is essentially machine learning, with ML's classic failure modes
fchollet · x · 2026-11-17
François Chollet argues that sufficiently advanced agentic coding is essentially machine learning: the engineer sets the optimization goal and constraints (the spec and its tests), and coding agents iterate until the goal is reached. The resulting codebase is a blackbox artifact deployed without inspecting its internal logic — like ignoring individual neural network weights. He predicts classic ML problems will soon plague agentic coding: overfitting to the spec, non-generalizing Clever Hans shortcuts, data leakage, concept drift. He also asks what will be the "Keras of agentic coding" — the right high-level abstractions for steering codebase "training" with minimal cognitive overhead.
More from AGI Musings
- Alison Gopnik to speak on intelligence and consciousness in AI and children at Berkeley — AlisonGopnik · 2026-11-17
- Microsoft's iWSP: Claude writes transformer weights directly, 166 params solve algorithmic tasks — dair_ai · 2026-11-17
- Fleuret pushes back on anti-physicalist argument: mechanisms can have feelings — francoisfleuret · 2026-11-17
- On intellectual consumption: academics debate representations of arguments, not arguments — clarejtbirch · 2026-11-17
- 'Formalizing All Math' went from absurd to consensus in 8 months; CFSG audit found 96 paper errors — AlexKontorovich · 2026-11-17
- Learning to Code in 2026: boot.dev Crunches the Job Data — theobeers · 2026-11-17