All that shows is, that Google is as happy spouting bullshit about safety as they are at bullshitting about privacy. Seriously, the longer I think about this, t

All that shows is, that Google is as happy spouting bullshit about safety as they are at bullshitting about privacy. Seriously, the longer I think about this, the worse it gets. Thing is the structure of the data for traffic accidents is, that we have a lot of data were we don't care much and very little about the accidents we care about a lot. We care more about fatalities than severe injuries, more about severe injuries than about light injuries and more about bodily harm than about fender benders. However, the number of events is precisely the opposite, which makes it easy to just swamp the signal we care about with events we don't care about as much. > For example, Waymo’s driverless cars have traveled roughly 4 million miles, but traffic-related fatalities are routinely measured per 100 million miles. Across America's roadways last year, there were 1.37 fatalities for every 100 million miles traveled, totaling 42,795 lives lost in vehicle traffic crashes That is, the expected number of fatalities at this point of data taking assuming Waymo's self driving cars are comparable to humans is 0. If we assume that rate is ten times the human rate, then the expected number of fatalities observed is still 0. If we assume that they kill people at a 100 times the rate of human drivers, then there is still a decent chance, something slightly worse than 50/50, that we didn't observe a fatality yet. > “It’s interesting that we were able to show statistical significance with so little data,” Victor said. “That has to do with two things: one is that we have very robust baselines from Swiss Re, and the other is that we have incredible safety performance.” Now this claim is probably true, in the sense of conflating all accidents with accidents we care about. And actually I think that there is a type of accident, low speed collisions when one tries to squeeze past a badly parked car, that is probably a very large part of all insurance claims where computers just should be really good. Computers have a engineered very precise world model and distance sensors all around and don't rely on human intuition guided by human impatience to just hope that squeezing past that will just fit.