[SEQ RERUN] Leaky Generalizations — LessWrong