Abstracting The Hardness of Alignment: Unbounded Atomic Optimization — LessWrong