JEPA (joint embedding predictive architecture) is an architecture that predicts abstract representations of future states rather than reconstructing raw sensory details, eliminating the need to predict unpredictable information by focusing only on predictable, relevant information.