Model progress at spitting out unhallucinated facts is slowing down hard. Model progress at solving hard math challenges/programming tasks doesn't seem to be slowing down that I can tell.