Predicting in the Dark Room of Embeddings: Joint-Embedding Predictive Architectures, Latent-Space Geometry and World Models: A Companion Volume to Euclid in the Retrieval Engine

Joint-Embedding Predictive Architectures (JEPAs) learn by predicting the representation of a hidden part of an input from the representation of a visible part, rather than reconstructing pixels or contrasting positive and negative pairs. This paper, a companion to the author's Euclid in the Retrieval Engine, gives a fresh, self-contained account of the JEPA family (I-JEPA for images, V-JEPA and V-JEPA 2 for video and action-conditioned planning, MC-JEPA for motion and content, and LeJEPA's provable regulariser), read through the energy-based view of learning. It sets out the three paradigms of self-supervised learning side by side, explains why predicting in latent space invites representational collapse, and compares the mechanisms that prevent it: stop-gradient with an exponential-moving-average target, variance–covariance regularisation, and distributional matching to an isotropic Gaussian. The paper then connects JEPA's latent geometry to the themes of the earlier volume. Normalisation, cosine and Euclidean distance, distance concentration and isotropy are not afterthoughts for a JEPA: they decide whether collapse is visible, whether its latent energy is meaningful, and whether its embeddings can serve visual retrieval. Three small computed experiments, written for this paper, show that a scale-invariant rank measure can miss total collapse, that centring and normalisation do not cure distance concentration in high dimensions, and that planning by minimising an L1 latent energy converges quickly in a toy world model. The paper closes with implications for visual retrieval-augmented generation, open questions and a calibrated verdict. Companion volume to Euclid in the Retrieval Engine (KSCCN series).