The post guides readers through the full construction of Joint Embedding Predictive Architectures (JEPAs), a self‑supervised method that learns image representations by predicting latent codes of masked patches. Elon Litman recreates the Annotated Transformer approach, ending with a working training loop for I‑JEPA, the image instantiation first detailed in 2023.

JEPA’s core design separates two encoders, a predictor, and a distance function. Context patches are encoded into a shared latent space, and the predictor outputs a guess for the target Kuala. Loss minimises the distance between the predicted and true latent vectors, forcing the encoders to capture semantic structure rather than pixel‑level noise. This representation‑space prediction sidesteps the high‑entropy artifacts that plague pixel‑level reconstruction.

The article extends the idea to video with V‑JEPA (2024) and V‑JEPA 2 (2025), and showcases Le JEPA (2025), which replaces engineering heuristics with a distributional regulariser. All variants maintain the same encoder‑predictor framework but adjust modalities and latent handling.

Implementation notes deliberately omit production‑grade tricks such as Flash Attention, gradient checkpointing, mixed‑precision, and large‑batch strategies, keeping the focus on mathematical clarity while acknowledging that these optimisations are essential for scalable training.