Joint Embedding Predictive Architecture
Joint Embedding Predictive Architecture, or JEPA, is the representation-space predictive approach discussed by Yann LeCun in [src-102]. The practical idea is to predict useful abstract representations of future observations instead of reconstructing every pixel or detail.
Key points
- JEPA is positioned as a non-generative architecture for learning predictive representations [src-102].
- The architecture predicts in embedding space, which lets it ignore unpredictable or irrelevant low-level detail [src-102].
- LeCun connects JEPA-style systems to World Models because a useful agent needs to anticipate what could happen after an action [src-102].
- He distinguishes this from LLM-style next-token prediction: text and code are discrete symbolic domains, while physical observations are continuous and high-dimensional [src-102].
- The source also mentions anti-collapse mechanisms such as VICReg and SIGReg as part of making representation learning stable [src-102].
Related entities
Related concepts
Source references
- [src-102] Vivatech / Yann LeCun – "Beyond Language Models: Building AI that Understands the World" (2026-06-17)
Recommended next
Keep reading from this thread
From 491 indexed pages and articles.
- Wiki concept Advanced Machine Intelligence Labs Or AMI Labs, is Yann LeCun's post-Meta AI company represented in [src-102] as a longer-term research and product Related by joint
- Wiki concept Yann LeCun An AI researcher represented in this wiki by his VivaTech interview on why useful LLMs still fall short of human-level intelligence Related by joint
- Insight Recommendation Systems in Production How recommendation systems become production decisioning systems through signals, ranking, constraints, feedback loops, and experimentation Readers have engaged with this next