← Glossary Model
JEPA
Meta's world model architecture, which predicts an abstract summary of what comes next rather than the pixels.
JEPA stands for Joint Embedding Predictive Architecture, an approach championed by Meta’s Yann LeCun. The core idea is a reaction against video generators: instead of predicting exactly what the next frame looks like, pixel by pixel, JEPA predicts a compressed abstract description of it.
The argument is that most pixels are irrelevant detail. If you want a system to understand that a dropped glass will fall and break, you do not need it to paint every shard correctly, you need it to get the situation right. Whether that abstraction beats generation is one of the live disagreements in AI research.
Mentioned in