> What are wile e coyote attacks?

https://www.thedrive.com/news/tesla-autopilot-fails-wile-e-c...

> If Tesla FSD is better than the average driver using it, isn't it still a net win, even if it might crash in different scenarios than a human?

That was the analysis when the industry was in its infancy. I think a lot more work has to go into that argument for people to accept it now that the driverless car industry has been operating for a decade+, it's not really clear that this pans out.

For example, today you can look at a car and predict how it's going to behave because you have a good model for how people drive. But let's say in the future driverless cars are much "safer" on paper than human drivers, but they behave very differently from them such that it's hard for people to predict their behaviors.

Now you've created a highly dynamic system where you don't have a good model for the all the actors because some of them behave one way but others behave a completely different way. Does this increase the overall safety of the system or decrease it, despite the new actors being statistically safer than the current ones?

I don't think you can with great confidence say what's going to happen just by looking at crash rates and comparing to the current system. You're going to change the whole system by introducing large numbers of actors who "crash in different scenarios than a human"