Yann LeCun

LeCun shares the 2018 Turing Award for the work that made deep learning workable at scale. The position he has held publicly for the past several years — that autoregressive next-token prediction is a dead end as a path toward anything resembling general reasoning — is one of the few senior dissents from current LLM consensus that is argued in technical architectural detail. Reading him is useful even when you suspect he's wrong, because the disagreement is specific: world models, joint-embedding predictive architectures, energy-based learning, all set against why he thinks the current generation of LLMs cannot get there from here.

Worth following when

you want a senior voice arguing against current LLM consensus in concrete architectural terms rather than in vague philosophical objection.

Topics

self-supervised learning; the limits of autoregressive generation; world-model and joint-embedding approaches to learning systems.

Key works

Turing Award–era foundational deep-learning contributions (shared 2018); ongoing JEPA / world-model line (2022 onward); public-facing critiques of the autoregressive LLM paradigm (2022, ongoing).