Say it turns out you're only mediocre at doing AI safety research. How much of a positive impact can you have then?