Story perspectives

Meta's V-JEPA: AI Learns Surprise from Video Insights

12/7/2025

26 5

Summary

Story summary

Meta researchers developed Video Joint Embedding Predictive Architecture (V-JEPA) that learns about the world through videos.
V-JEPA shows a notion of "surprise" when information contradicts its learned knowledge, akin to infant object permanence.
Unlike traditional pixel-space models, which misinterpret irrelevant details, V-JEPA aims to better interpret complex scenes.
Cognitive scientist Micha Heilbron says the claims are plausible and the results are intriguing.

1 of 1

Story summary

Meta researchers developed Video Joint Embedding Predictive Architecture (V-JEPA) that learns about the world through videos.
V-JEPA shows a notion of "surprise" when information contradicts its learned knowledge, akin to infant object permanence.
Unlike traditional pixel-space models, which misinterpret irrelevant details, V-JEPA aims to better interpret complex scenes.
Cognitive scientist Micha Heilbron says the claims are plausible and the results are intriguing.