Keras Creator Admits Underestimating LLMs, Points to Test-Time Compute as Key

fchollet · x · 2026-08-07

François Chollet, creator of Keras, admitted that he underestimated the long-term importance of LMs during 2023 and early 2024. He noted that his perspective shifted completely in late 2024 following the test-time compute (TTC) breakthrough of the o3 model.

Chollet clarified that he initially did not see LLMs as a viable base for systems capable of fluid intelligence. However, he emphasized that the early 2023 narrative—claiming that simply scaling up base LLMs would solve AGI—has not panned out. Despite being scaled up significantly, current base LLMs still struggle with simple math and tests like ARC 1. He argues that TTC and external harnesses are the actual critical factors, and the TTC breakthrough is what unlocked their potential.

Original post →

More from AGI Musings

AGI Musings channel →