A joint-embedding predictive architecture (JEPA) is a model that learns by predicting the representation of a future or masked state — not by reconstructing raw pixels. Where a generative model asks "what does this look like?", a JEPA asks "what representation will this produce?" That distinction changes everything about how the model learns to understand scenes.
Darwin builds an internal representation of a video scene — a compact embedding that captures the semantic state: what is happening, how objects relate, what is likely to happen next. It then predicts forward in that representation space, giving it semantic understanding without a language model anywhere in the perception loop. Language only enters at the query layer, when a user or agent asks a question in plain English. This is why Darwin's outputs are deterministic: the perception pipeline is pure computation, not sampling.
Further reading: JEPA deep dive — Joint-embedding predictive architectures for video.
The diagram-led JEPA 101 explainer is coming during our August launch.