“Initial automated alignment research AIs will likely have legible thought processes that can be understood by reading their output, without needing complex mechanistic interpretability techniques.”