Rohin Shah comments on Language models seem to be much better than humans at next-token prediction