LeCun argues that world models can replace knowledge accumulation with understanding

Yann LeCun used a post on X to draw a wider conclusion from his new paper with Randall Balestriero. He argued that large language models need large parameter counts because they store declarative knowledge, and that world models built on Joint Embedding Predictive Architectures (JEPAs) can replace that accumulation with understanding, working with less memory and less data.

He tied the claim to recent mathematical results on the geometry of representations under anti-collapse criteria, the conditions that stop a model mapping every input to the same point. The paper supplies the evidence for one part of this: a text JEPA of 109 million parameters aligned better with vision models than language models of up to 72 billion parameters. The broader claim, that such models replace accumulation with understanding, is LeCun’s own argument, and the alignment results do not by themselves settle it.