To me safety is mostly about the risk of injury and death to me. And mostly about death.
Humans die from driving only once every 100 million miles on average.
So until there is a comparison at this scale, to me it’s a very incomplete picture of safety to say that you are outperforming humans.
If driving my own car means colliding 100x as often but I expect to die 2x less often, I would consider self driving to be much much much more unsafe. There is really no way for me to understand the risk of fatalities from a sample of only 7 million miles, since either humans or a system with 2x the fatality rate of humans would both be expected to have 0 fatalities at this scale.
Given we are at 7m miles, hopefully this comparison is coming soon and I will be much more convinced.
I guess to me it seems like common sense that a system that has substantially fewer crashes also has substantially fewer deaths. Maybe we can't make definitive statements about the expected number of deaths yet, but I think the most reasonable best guess with the information we have is that waymo deaths will be much lower.
The alternative requires a scenario where waymo is especially likely to get into fatal accidents while being very good at avoiding non-fatal ones, right? Seems far-fetched.
I would not make the same inference because we know that ML systems generally struggle with robustness and the long tail, while humans tend to be exhibit much more flexibility to adapt to distribution shifts or unusual situations. For a concrete example of this in computer vision, see work such as the Imagenet—C dataset where simple distribution shifts generally tank ML models but do not impact human performance.
But regardless, the claim here isn’t “under some assumptions that some people (and not others) find reasonable, we can extrapolate and predict that self driving cars will be found to be safer”
It’s “self driving cars are safer”, which there isn’t enough evidence to claim yet.
This, plus there's a control aspect. People fear flying largely because they aren't in control, even though it's technical safer than driving. But driving...you're in control.
Humans die from driving only once every 100 million miles on average.
So until there is a comparison at this scale, to me it’s a very incomplete picture of safety to say that you are outperforming humans.
If driving my own car means colliding 100x as often but I expect to die 2x less often, I would consider self driving to be much much much more unsafe. There is really no way for me to understand the risk of fatalities from a sample of only 7 million miles, since either humans or a system with 2x the fatality rate of humans would both be expected to have 0 fatalities at this scale.
Given we are at 7m miles, hopefully this comparison is coming soon and I will be much more convinced.