francois chollet

High-quality datasets and RL environments are crafted products requiring domain expertise (doctors, lawyers, traders), not commodity zip files. This creates defensible vertical data companies analogous to apps on a phone platform.

Current LLM scaling hits a wall on fluid intelligence benchmarks (ARC); reasoning models and RL post-training only automate verifiable domains. True AGI needs human-level sample efficiency via symbolic program synthesis, not bigger parametric curves.

Coding agents succeeded because code provides formal verification (unit tests), enabling massive RL post-training data generation. This paradigm will extend to math and other verifiable domains, but not to fuzzy tasks like essay writing where reward signals are noisy.

Trillions in compute poured into one architecture (transformers + gradient descent) creates fragility. Chollet argues the same resources applied to alternative paradigms (genetic algorithms, state space models, symbolic search) would yield breakthroughs, and that recursive self-improvement without human bottlenecks is the key criterion.

Open source AI tools win by lowering onboarding friction (simple API, educational docs) and converting enthusiastic users into core maintainers. This compounding community flywheel turned Keras into a Google-supported standard.