How Many Miles Would It Take to Demonstrate Autonomous Vehicle Reliability?

The findings were more pessimistic than I expected, such as how even testing existing fleets for "hundreds of years" won't provide enough evidence that they meet safety standards.

There is some potential for "accelerated learning", right? Intelligence/lessons/data experienced by AVs can be distributed and shared relatively instantaneously.

Also, I wonder if there will be an inflection point in which autonomous driving becomes popular enough, and human driving deprecated enough, that most of our infrastructure will be overhauled to be safety-optimized for AVs?

It's not a question of miles, it's a question of scenarios. A car that takes on the entire interstate highways system but suffers no "close-calls" is practically untested, IMO. Also, that car never "experienced" rush hour in crazy places like downtown Cairo or the old&narrow streets of older Italy.