At the risk of invoking the ire of Google's marketing team further (I just got -6'd on this thread in less than two minutes. It happens actually as predictably as clockwork on any case where I point out Google isn't quite as amazing as it seems.)
So, I'm sitting here watching this, and nothing he's saying is particularly a solid logical argument. He's pointing out that a always-running self-driving car has to make more decisions or actions than a driver assistance device, but that's irrelevant: 99% of the time, the car's basically just moving forwards. And arguably, lane assist devices make decisions or corrections just as often as an autonomous car, same for automatic breaking. They constantly need to check the conditions on the road and decide whether or not to adjust. I'm not seeing any real reasoning driver assistance and fully autonomous are so shockingly different... short of the fact that this guy is marketing a Google product.
If anything, it seems Urmson is heavily pidgeon-holing what a "driver assistance system" is, so that he can discredit the way all of their competitors are developing the same types of technology. Nothing particularly special about what Google is tracking that couldn't be similarly developed as part of a driver assistance system.
In actuality, his final argument is the most ludicrous of them all... that gradually developing the technology through driver assistance is too slow. That this needs to be rushed out the door in the next couple years.
I didn't downvote your first comment, but in my opinion the reason it was downvoted was that it amounted to ad-hominem ("Don't listen to him because he's from Google.") rather than a reasoned argument like you've posted here. Sure, it's wise to take into account the source of an argument, but you can't reasonably stop there.
I do disagree with your first point, that an always-running self-driving car having to make more decisions is irrelevant because a car spends the majority of its time moving forward. I think the point is that a driver assistance system basically only handles that moving forward case. It may cover 99% of the time, but it only covers a small fraction of the decisions and actions the car has to take. It's the other 1% (probably higher than 1% actually, but it doesn't matter) that's the tricky part, and it is difficult to iterate from handling the special (if very common) case of basically just moving forward, to handling all the other situations a car can find itself in. You would quickly lose any easily-understood distinction between what situations the car is and is not able to handle, making it difficult for the human driver to know when their input will be required.
But we have other types of driver assistance for that as well. For instance, parking assist has to handle the complex maneuvering of a vehicle into a narrow space. Presumably while both avoiding hitting the cars near it which are parked, but also any cars which might come across your car while it is parking.
I could see driver assistance eventually move to a point where it's tracking and simulating just as much as a fully autonomous vehicle, but allowing the driver to remain in control until it believes a collision is imminent, or that the user is making an unsafe judgment call.
Things like the automatic emergency braking feature will eventually learn how to detect different types of objects as well. Say, to avoid slamming on the brakes because a plastic bag was pushed by the wind past a camera or sensor.
Presumably, these currently independent assistance features will converge over time, into your car simply being aware of it's surroundings and intervening to protect you. And then someday it'll just drive itself.
All of this can (and probably will) be done gradually, but it sounds like that isn't fast enough for Urmson's son to reach driving age.
