Hinton vs. LeCun reignites: did reasoning models vindicate his autoregression doubts?

OnlyBath9046 · reddit · 2026-09-23

The old disagreement between the two AI godfathers resurfaced as inference-time search and latent reasoning gained ground. One side argues these developments vindicate LeCun's claim that pure autoregressive Transformer LLMs were never enough; the other counters that today's reasoning systems are still Transformer-based and LeCun is moving the goalposts after years of calling them an off-ramp. In his detailed response, LeCun argues human-like reasoning must involve search in continuous representation space rather than token space — connecting to a growing latent-reasoning research landscape: continuous-thought LMs, compressed non-linguistic tokens, recurrent-depth models, task-trained recursive solvers, and Pathway's BDH-CQ, whose ARC-AGI results make it one of the most empirically demonstrated teams in latent reasoning. The open question: at what point does an LLM augmented with search, recurrence, or latent computation stop being an LLM in itself?

Original post →

Hinton vs. LeCun reignites: did reasoning models vindicate his autoregression doubts?

More from AGI Musings