Amazon research find Video-Action Models fail at physical tasks despite making accurate visual predictions · Digg