Amazing. I can practically smell that owl it looks so darned owl-like.
From the article it doesn’t seem as though photorealism per se was a goal in training; was that just emergent from human preferences, or did it take some specific dataset construction mojo?