Slept on the autonomous car issue. Here is what I think Tesla is doing:
In google car and other autonomous cars, the human is the proctor, and the software is the driver. The human reports exceptions to what he considers normal driving behavior.
I think Tesla has flipped this relationship. In the newer sensor equipped cars, the software is proctoring the human driver, and reporting exceptions. To do this proctoring, the software has to know where the human is going in the car. This is done by tagging repetitive trips (to and from work). The software "drives along" without actually controlling the car.
The second type of exceptions that the autonomous car software could report are abrupt maneuvers when the software and driver disagrees: 1) The software calling for an abrupt maneuver to avoid a hazard that the human driver doesn't take. 2) The driver making what seems to be an emergency action that was not flagged as such by the software. e.g. slamming on the brakes.
With this approach, Tesla would be adding the equivalent of a hundred or more autonomous test cars per week. There are major advantages of this approach as the software gets good. The poor human in the google car sits for hours before something interesting happens to report to the development team.
Musk seems to suggest that Tesla has an advantage in autonomous cars due to the very high miles driven. Where are these test cars, if he is not referring to regular production vehicles?