Meta's V-JEPA AI Learns Physics Like a Baby, Achieves 98% Accuracy on Reality Tests

Meta's V-JEPA AI learns physics from videos like a human baby, achieving 98% accuracy on reality tests by discovering object permanence and showing measurable surprise at impossible events, though its memory currently lasts only seconds.

• Meta's V-JEPA AI system learns physics concepts like object permanence from ordinary videos without built-in assumptions about how the world works, achieving nearly 98% accuracy on physical plausibility tests
• The model uses latent representations instead of individual pixels to focus on essential details, allowing it to ignore irrelevant information like moving leaves while tracking important elements like traffic lights
• V-JEPA demonstrates measurable 'surprise' when shown physically impossible events, similar to infant reactions, but currently has limited memory spanning only a few seconds like a goldfish