Isn't the problem that without exploration we won't actually get new goals e.g. can AI come up with new math conjectures instead of "just" solving existing ones.