Learning the smooth prior — LessWrong