AIs should also refuse to work on capabilities research — LessWrong