I think we're pretty much out of "spend more compute on even longer training runs" atp.