Defining Corrigible and Useful Goals — LessWrong