There's a positive feedback loop that will happen: As they get more cars on to the road, the log more real world miles, which makes the cars better, which gets more cars onto the road which log more miles...
The real challenge here is extracting valuable information - if you just dump the sensor data to disk you will consume the world's storage capacity with a handful of cars.
Not sure how much that is worth. Tesla collects a huge amount of data but that doesn't help as much as some had hoped. You need to interpret all that data which is expensive and complicated.
1. Storing large amounts of logs data is challenging and expensive for anyone. So even if you can collect data, good luck storing it all. Think about what a self driving car has to store: images, audio, 3d point clouds, all organized by time, ideally labeled in some manner.
2. Unorganized data is not useful. For example, part of the autonomous vehicle problem is proving that the car is safe for <condition parameters>. You need simulation to ensure model changes don't incur regressions. I've heard rumors that Tesla didn't have a sim team as recently as 1.5 years ago -- a lot of people argue "you can't sim everything" but most mature players now have realized that sim is vital.
If your data is not organized and accessible, or you can't afford to store the data, your development velocity is going to be really bad and/or your safety characteristics will not be provable.
One argument I've heard is "well with each new model launch why not just run the model in ghost mode for all live cars on the road, and see if it causes any issues" -- many life-threatening situations are rare combinations of factors. If you don't record them, you can't count on them to happen again during your regression test period. And again, even if it worked, the time it takes you to run a regression test is now a lot longer than a team that has sim that can run O(hours-days).
I wonder how valuable this is?