Extrapolating values without outsourcing — LessWrong