Averting the convergent instrumental strategy of self-improvement — LessWrong