Predictors that don't try to manipulate you(?) — LessWrong