[SEQ RERUN] Measuring Optimization Power — LessWrong