Even then there'll be a distribution shift effect they need to account for. Not everyone is allowed into the FSD program, and the fact that they're self-selecti

Even then there'll be a distribution shift effect they need to account for. Not everyone is allowed into the FSD program, and the fact that they're self-selecting to use FSD additionally biases the roads and conditions where they choose to do so. Moreover, how you count is still very relevant if all the hairy events happen when a human operator takes control -- the car screwed up, but the resulting crash gets blamed on the operator who couldn't react to the ensuing shitshow in time, and even if total crashes/fatalities are lower when counting those events you still have the self-selection problem of perhaps those FSD operators in those conditions being better at recovery from car-induced errors than the general population in arbitrary conditions would be. Like...maybe it's fine. But even at 0 fatal crashes in 5 billion miles they won't have a strong argument that a more widely deployed FSD would have similar success without a lot more data to explain how the thing they measured directly connects to the thing they'll want to claim it represents.