Subsystem Alignment — LessWrong