François Chollet Clarifies LLM Stance: Test-Time Compute Breaks Base Model Bottlenecks
fchollet · x · 2026-08-02
François Chollet clarified his evolving stance on Large Language Models (LLMs), stating that his past criticisms of base LLMs do not apply to systems utilizing test-time compute (TTA). He compared the difference to steam trains versus modern electric bullet trains—similar in appearance, but fundamentally different in underlying principles.
Chollet noted that he updated his views in December 2024, no longer believing that the LLM tech platform (enhanced with TTA) would stall. He emphasized that if we look strictly at base models without TTA, they still perform poorly on the 2019 ARC 1 benchmark, despite roughly a 100,000x increase in compute scaling. The paradigm shift to test-time compute, rather than simply scaling single-pass static inference, is what enables today's state-of-the-art systems to achieve advanced reasoning.
More from AGI Musings
- AI writes the CV, AI screens the CV: hiring is two robots talking — brandon_galang · 2026-10-16
- Is superintelligence a bigger crypto threat than quantum? Experts split — examachine · 2026-10-16
- 'AI fakes intelligence the way a crane fakes strength' — pushback on the 'AI is a scam' take — LucaAmb · 2026-10-16
- Same debate, with chart: can harnesses sustain AI progress past the plateau? — Otherwise-Report1848 · 2026-10-16
- Don't misread Musk: human-machine merger is a forecast of what's coming, not a plan — danfaggella · 2026-10-16
- MIT's Buehler: AI as 'a new language of science,' three years on — ProfBuehlerMIT · 2026-10-16