AI coordination needs clear wins — LessWrong