“Clean” vs. “messy” goal-directedness (Section 2.2.3 of “Scheming AIs”) — LessWrong