According to Elon, "This time, it’s right" and Dojo is the solution for the March of 9s. Dojo v1.0 will arrive in "maybe a year or so".
I'm optimistic enough to believe this...
If the platform is right, there is no significant need for more expensive software development, it is mostly just more real world examples, better training and smarter Neural Nets.
If Dojo partially generates training data with limited human intervention, we have a feedback loop where the system continually improves and learns from it's mistakes, those mistakes source better training data and that new training data improves the next round of training..
I think "operation vacation" is an aspiration to statement of goals, rather than an expected outcome. The real aim is maximum improvement in the NN with minimal human effort and expense.
People who doubt this are thinking it can't be automated, or it can't be done in the NN layer, or driving requires human logic.
We drive with an element of logic and memory a NN has similar properties...
Tesla a vague about the mapping aspect, but I think mapping exists in FSD and it is similar to our memory, FSD knows to expect a Stop sign at a particular location, and where the road lanes were last time, and vision confirms or denies that expectation.
So we need to identify something that FSD can't solve, something Elon and the smart people working for him haven't thought of.
IMO things like snow storms , obscured cameras, broken cameras or radar interference are more likely to be issues than a simple lack of training... I tend to think Elon and co have thought of my list, and just about everyone's list.