Eventually accidents are going to happen and the author's projection from human statistics and a simple relative safety multiplier is not a bad rough model, but there are reasons to believe that the distribution of FSD accidents could be much more heavily weighted towards minor events than human accidents are. The failure modes for FSD are going to be different from the way that humans make mistakes and that's going to change the character of accidents. Human accidents are dominated by attention failures. Whether due to illness, impairment, distraction, or just the human limitation of only being able to see in one direction at a time, human accidents involve a lot of events where the driver(s) didn’t see the accident coming, or didn’t see it until very late. In those accidents the impact velocity is sometimes very high. FSD will have fewer of these failures and as a result will be able to take action sooner and reduce impact velocities further than the average human accident. FSD also has faster reflexes and will employ better brake modulation than the average human. Additionally FSD is less likely to engage in high speed driving than more adventurous human drivers. All of these things will tend to reduce the average impact velocity when FSD is driving. Because accident severity follows impact energy, and impact energy goes up with the square of impact velocity, even moderate reductions in impact velocity can have a big impact on the severity of an accident and we should see that in the accident statistics. There’s an irreducible floor on accident rates. Even perfect driving on the part of one vehicle isn’t perfect protection on roads that other drivers share. This floor is probably something like 10x reduction relative to the human rate of accidents and supervised FSD, currently at about 7x reduction, might already be getting into that territory. But the nature of the accident matters to severity and the different characteristics of FSD as a driver give hope that the rate of minor incidents won’t be linearly predictive of the rate of serious incidents. We might know soon if this is true. Once we have data on 5M unsupervised miles we can set a 95% confidence rate for minor incidents. At 50M we’ll be able to say something about serious incidents. There are going to be accidents. And it’s possible that FSD ends up being “only” 10x better than human drivers. This will save a lot of lives and avoid enormous misery and waste. But in a world where the knives are coming out for anyone who threatens incumbent business models *any* accident will be an opportunity for opponents to smear FSD. Still, the jury isn’t out on the possibility that FSD ends up being so dramatically safer than human drivers that opponents have to seriously demean themselves.