Learning with catastrophes — LessWrong