Towards more cooperative AI safety strategies — LessWrong