LeCun clashes over whether modern AI is still autoregressive token prediction

john__allard · x · 2026-11-23

LeCun insists his old slide still holds, claiming modern AI systems are no longer autoregressive predictors of discrete tokens. johnallard pushes back: tool calls and closed loops were expected (ReAct predates the slide), and reasoning is just autoregressively sampled tokens under a different training signal — so either the slide only covered open-loop LLMs nobody planned to deploy, or it was wrong.

Related event: LeCun Reiterates Modern AI Is No Longer a Discrete Token Autoregressive Predictor, Faces Pushback(3 posts)→

Original post →

More from Models

Models channel →