Object permanence (just from images/video) seems like a particularly hard problem for a super-smart prediction engine. Is it the old thing, or a new thing?