Or buggy. We're already running plenty of buggy code, but at least the errors can usually be found. If an AI overflows on some intermediate calculation sometimes because of implementation error I'm almost sure nobody will find that.