> Self-driving is very easy to evaluate. I disagree. We can say that someone is driving safely in one given moment. A drunk driver could drive quite safely thr

> Self-driving is very easy to evaluate. I disagree. We can say that someone is driving safely in one given moment. A drunk driver could drive quite safely through a main road while we watch them, and then turn around a corner and run over someone the moment they leave our sight. We can only extrapolate the "goodness" of a driver by expanding the time we spend observing their driving. 2-3 hours of evaluation will be quite solid. It would be even more solid if we are inside the car with the driver, and if we have received training about how to evaluate a driver. And the result would be even more reliable if the evaluator is financially independent from the driver. And for safety, the car should have controls that allow the examinator to take over if needed. So yeah, I am describing a driving test. As far as I know we don't have those for self-driving cars. We should. They should be way more thorough than than the ones we give to a human, because machines lack the shared context that humans have with each other. Tests like driving offroad, driving while there's tiny (fake) animals around, behavior in a (staged) accident, etc. Things that we know a human will have enough context to deal with, but for a machine it's not a given. One advantage is that these tests could be done in parallel. You can load the same AI in 50 identical vehicles and have 50 examinators run tests on it at the same time.