This highlights a major risk: how well/poorly Tesla’s AI method generalizes across driving scenarios for human or better safety (eg, US vs Australia etc, or even region to region within a country).
We really don’t know.
There is no evidence that it does generalize well, but they haven’t yet offered a service accumulating millions of miles.
So there is no viable service yet existing that can be generalized.
The recent Reuters article on intensive efforts by Tesla to train their AI driver in Austin was not a good sign for the model’s ability to generalize, but that was just a little tidbit of information.
Neural networks are flexible models, so they generalize in that sense, but that can come with a cost of requiring a lot of data (and then of course compute, like HW3, HW4, HW5, etc).
And then do you need data that is hyper-specific to each city and weird case?