Chollet's Shift: Endorsing Test-Time Compute over Scaling

Episode 1 · François Chollet: Test-Time Compute is Key as Pure Parameter Scaling Hits Limits (2026-08-02, 4 posts)

François Chollet stated that while pure parameter scaling has hit a bottleneck, test-time compute and post-training scaling have broken through these limitations. He emphasized that these approaches still have a 10x to 100x room for improvement, with program search being the ultimate path to AGI.

• François Chollet: Deep Learning Will Eventually Be Replaced by Discrete Program Search — fchollet · 2026-08-02
• fchollet: Pure Parameter Scaling Hit a Wall; Test-Time Compute Is Necessary — fchollet · 2026-08-02
• François Chollet Clarifies LLM Stance: Test-Time Compute Breaks Base Model Bottlenecks — fchollet · 2026-08-02
• Keras Creator François Chollet: Test-Time Compute Scaling Has 10-100x Left to Deliver — fchollet · 2026-08-02

Episode 2 · Keras Creator Admits Underestimating LLMs, o3 Test-Time Compute as Turning Point (2026-08-07, 5 posts)

Keras creator François Chollet posted on social media to reflect and admit that he underestimated the long-term importance of large language models (LLMs) from 2023 to early 2024. He explicitly stated that the breakthrough in test-time compute achieved by the o3 model in late 2024 was the key turning point that completely changed his view. He also clarified that some of his earlier judgments still hold: base LLMs will hit capability ceilings, scaling laws alone are not the only path to AGI, and he predicts LLMs will eventually be superseded by methods like symbolic learning.

Confirmed

• Chollet confirmed that he underestimated the potential of LLMs as a foundation for building fluent intelligent systems from 2023 to early 2024.
• The key turning point was the breakthrough in test-time compute by the o3 model in December 2024.
• He clarified that some of his earlier judgments still hold: base LLMs will hit capability ceilings, and scaling laws alone are not the only path to AGI.
• He believes LLMs will eventually be replaced by methods like symbolic learning.

Why it matters

As a highly influential researcher in AI, Chollet's change of view signals growing recognition of the test-time compute approach, which could influence future directions in AI system architecture.

• Keras Creator Admits Underestimating LLMs, Points to Test-Time Compute as Key — fchollet · 2026-08-07
• Keras Creator Reassesses LLMs: Underestimated Long-Term Value, But Scaling Isn't AGI — fchollet · 2026-08-07
• Keras Creator Admits Underestimating LLMs, Cites o3 Breakthrough as Turning Point — khademinori · 2026-08-07
• François Chollet Admits Underestimating LLMs: Test-Time Compute Was the Turning Point — yacineMTB · 2026-08-07
• Keras Founder François Chollet: Future AI Will Shift from LLMs to Symbolic Learning — burny_tech · 2026-08-08