Yes, like teaching a driver who can't understand complex situations, who is fearless and without conscience. Besides that it's the same.

That's probably part of the calculus, but mainly it's to deliver at least something on an over-promised feature.

I also wonder how good the training examples are, if they're just using average drivers without any quality control. Like when I'm teaching someone to drive, I'm following traffic laws to the tee. But when I'm doing my day to day driving, it's mostly about optimizing time. There's probably a reason why Waymo hired and trained drivers to teach their algo in a standardized way.