It feels intellectually negligent to me that models like those used by Openpilot have no auditable concept of what a car, or a sign, or a person is.

Just where lines are and when a car should accelerate or break. The rest of the latent state is "based on pixels."