Learning the prior — LessWrong