I wonder how easy is it for Tesla to 1) find the exact problem and 2) fix it (provided it is software based). Does it mean that they first have to figure out what some "black box" neural network(s) are getting wrong and then retraining (on simulations), etc?

What if city X decides to repaint all its buses and that confuses the object detection system? How log would it take to get the (hot)fix out of the door...