Joint Embedding Predictive Architecture

Joint embedding predictive architecture, or JEPA, is presented in 133. 对谢赛宁的7小时马拉松访谈：世界模型、逃出硅谷、AMI Labs、两次拒绝Ilya、杨立昆、李飞飞和42 as more than a single self-supervised algorithm. Xie Saining says he moved through questioning JEPA, understanding JEPA, and “becoming JEPA”, meaning he came to see it as a broader cognitive architecture for prediction, planning, and world understanding.

Source View

The source connects JEPA to Yann LeCun’s World Models route. The important shift is from predicting raw pixels or tokens toward predicting abstract representations that are useful for action, reasoning, memory, and controllability.

Connections

• Yann LeCun, Xie Saining, and AMI Labs — people and company context.
• World Models — broader technical destination.
• Representation Learning and Self-Supervised Learning — representation and objective foundations.
• Multimodal Intelligence and Embodied AI — domains where predictive abstract representation may matter.
• Causal World Models — adjacent route that emphasizes causal variables and structures rather than only predictive embeddings.