But they don't, they just understand pixel relationships (right?)
You can model a lot of basic physics through observing 1,000,000 videos
You can model a lot of basic physics through observing 1,000,000 videos