“A comprehensive, "once-and-for-all" solution to the AI alignment problem would require four components: a formal theory for alignment, a process to elicit human values, techniques for training aligned AI, and formal verification tools.”