Yann LeCun · NYU · Meta · 2022-06-27
LeCun's position paper is the technical scaffolding underneath every JEPA, V-JEPA and world-model release that has followed. He argues that autoregressive token prediction is a dead end for genuine machine intelligence and proposes a hierarchy of joint-embedding predictive architectures, configured by a world model and trained on observation rather than text.
It is long, and parts have aged faster than others, but no other single document gives a learner the precise vocabulary — energy-based models, JEPA, configurator, intrinsic cost — that the LeCun camp uses to push back on the LLM-only roadmap. Read it as the manifesto it is, not as a survey.
> JEPA is not generative — it captures the dependencies between x and y without explicitly generating predictions of y.