V-JEPA can acquire a level of physical common sense by learning to predict at the representation level, as demonstrated by the system being 'surprised' (high prediction error) when shown impossible physical scenarios like objects floating or changing shape.