Robustness to Scaling Down: More Important Than I Thought — LessWrong