Just as with wealth the average is drastically skewed by outliers. I don't recall precise numbers off the top of my head but there are plenty of people who have commuted daily for multiple decades and have never been in a collision. I myself have only ever hit inanimate objects at low speeds (the irony) and have never come anywhere near totaling a vehicle; my seatbelts and airbags have yet to actually do anything. Freight drivers regularly achieve absurd mileage figures without any notable incidents.
As I stated earlier I agree with the broader point you were trying to make. I like what they're doing. It's just important to be clear about what human skill actually looks like in this case - a multimodal distribution that's highly biased by category.
Yeah, I agree with you too. Per IIHS, the fatality rate per 100,000 people ranged from 4.9 in Massachusetts to 24.9 in Mississippi, so clearly there's a huge variance even with "US population".
The other person's comment was "we won't ever have self-driving cars" because they aren't good enough: but something like Waymo already is, particularly for the population. If we waved a wand and replaced everyone's car with a Waymo, accident rates would fall, at a population level and at a per-mile driven level.
It's even tough to see that a Waymo would be more dangerous for a good driver: they too have never been the cause of a serious accident and have certainly driven more miles across the fleet than any human driver. All 4 serious injury accidents and both fatalities were essentially "other driver at fault, hit Waymo".
This isn't meant to glaze Waymo, but point out that self-driving cars in certain environments are "solved". They're expensive, proprietary, aren't suitable for trucking or deployment to cold climates (yet?); but self-driving that is safer than people-driving is already here. To your point: human skill in driving is variable: Waymo won't replace Verstappen right now, but just like the AGI argument with LLMs, they're already "smarter" than the average person in certain domains.
As I stated earlier I agree with the broader point you were trying to make. I like what they're doing. It's just important to be clear about what human skill actually looks like in this case - a multimodal distribution that's highly biased by category.