“Jan Leike's favored approach to AI alignment research is to build a system that can perform alignment research better than humans.”