From Pixels to Predictive Codes: What V-JEPA 2 Means for Embodied-Simulation Research
A technical reading of predictive video representations and their relevance to embodied simulation.
Predicting useful structure
This note examines V-JEPA 2 through the lens of embodied simulation. It focuses on the distinction between predicting compressed representations and reconstructing every detail of an image.
Richard considers how prediction-first systems might relate to SIML's research questions. The piece is a reading of public research and a discussion of possible connections, rather than a claim to have developed V-JEPA 2.
From reading to experiment
The later SIML FEPGate benchmark explores one concrete connection with a latent planner. Keeping the two records separate makes the progression clear: interpretation and proposed bridges first, then a reported implementation and comparison.
The original note contains the technical discussion and links for further reading.
SOURCES & FURTHER EXPLORATION