Isn't this violating the first rule of AI safety: do not let an AI change its code?