JEPA Explained: Yann LeCun's Answer to LLMs for Understanding the World

Show a model half of a photo and a generative model will paint every missing pixel. JEPA, the Joint Embedding Predictive Architecture proposed by Yann LeCun in 2022, does something smarter: it predicts what the hidden part means, in the compact numbers a network uses to describe an image. Inside: how JEPA works (hidden blocks, a context encoder, a target encoder, a predictor, and a loss measured in representation space), the collapse trap and the moving-average trick that avoids it, Meta's I-JEPA (632M parameters, 16 GPUs, under 72 hours), V-JEPA for video, and V-JEPA 2 controlling robot arms