François Chollet Defends Test-Time Training as 'Pure' Deep Learning vs Neurosymbolic Tradition

teortaxesTex · x · 2026-08-14

Keras founder François Chollet recently argued that test-time training is the only form of test-time adaptation that is "pure" deep learning. According to Chollet, it adapts in a continuous latent space rather than adapting in a discrete symbol space like neurosymbolic approaches.

This sparked a debate over foundational AI paradigms. A researcher pushed back, claiming that LLMs autoregressively generating discrete symbols was a victory for the neurosymbolic tradition all along. They argued this vindicates early symbolic AI pioneers like Newell & Simon, joking that "Hinton is destroyed" and "Minsky is proven right again."

Original post →

More from Fun

Fun channel →