“The goal of AI alignment is to build a model that is as good as humans at alignment research and is sufficiently aligned to be trusted with that research.”