I am really wondering, how they tackle the long trail problem of unknown situations. Sure they are monitor them constantly but I am asking precisely, how to handle situations that occur ridiculously infrequently: do they sample them 1000x in sim? Or do they use a different model and track uncertainty in the model constantly?
How to get another 0 between the dot and the percent sign in situations where the model makes the wrong choice with a significant probability?
There was an article a while back that talked about their simulated world model. It has the ability to do things like floods, elephants on the road, etc.
