Hacker News Reader

Comment by dekhn | Hacker News Reader

That's not a rocket blasting off, it's a rocket failing to reach orbit by barely clearing the launch platform.

More importantly: that chart isn't normalized by miles driven. As the companies have been ramping up, you'd expect (under a random model) that incidents would go up.

It could also be explained by greater knowledge of SDCs, increased news coverage, and more awareness of how to report incidents.

More importantly, I wasn't considering the chart, I was considering what the source of the quote said:

Julia Friedlander, SFMTA’s senior manager of automated driving policy, told state regulators in late June that driverless taxi incidents began “skyrocketing” this year. Though city leaders suspect it coincides with a rise in driverless activity, Friedlander said the city can’t make definitive conclusions because it doesn’t have detailed data.

View on HN

You can normalize the report by miles driven, but if the number of miles driven by those services increases rapidly then you can expect the number of incidents to increase just as rapidly, and that is what appears to be happening. There is a level at which this becomes untenable and that level is a direct function of the number of miles driven.

So unless you see some kind of cap to the number of miles driven (which given the ambition to scale up is not something I would subscribe to) I believe this is an early indication of something that may well develop into an actual problem in a relatively short amount of time.

I think this is not taking the analysis far enough. You need to compare incidents per passenger mile, with humans vs. AV.

If AVs are safer, and you increase the AV miles driven, presumably you mostly displace human miles driven (e.g. Uber). In this possible word you would see the missing graph of human-driven harm go down, but you’d still see the graph from the OP go up. Indeed in this scenario the more the AV mischief goes up, the better, as that would be implying that there are fewer humans on the road. (Obviously it’s not only displacement but I think that’s the first-order effect.)

Without quantifying the level of human-related mischief, I don’t think it’s wise to draw strong conclusions.

Different kinds of incidents compared to TFA.

That’s the point, you want to compare different kinds of harm so you don’t pick the worse world.

Per mile incident rate is a measure only the companies care about. Why wouldn't the city care about an increase in the total number of traffic incidents? Especially when the self-driving cars are not displacing human drivers, so at the end of the day the roads are getting less safe.

>Especially when the self-driving cars are not displacing human drivers, so at the end of the day the roads are getting less safe.

what makes you think that? "Driverless taxis" imply they're being used to transport passengers. A robotaxi taking a fare means that there isn't a human taxi taking the same fare, so human drivers are essentially being displaced.

Robotaxis are cheaper than human taxis (in the long run), don't need to take breaks, etc. There's no reason to assume that taxi rides would remain flat if cost falls and availability increases.

If you care about absolute numbers, then you need to compare it against human-caused incidents too. If there's 100,000 human-caused incidents then AVs going from 3 -> 300 isn't even a blip.

>Per mile incident rate is a measure only the companies care about

This is a very frustrating sentence to read. If the per mile incident rate of driverless cars is 1 and the per mile incident rate for normal cars is 2. What happens when you do a 1:1 replacement of all normal cars with driveless cars? The number of incidences halves. i.e the total number of traffic incidences halves.

You assume that a driverless mile replaces a human driven mile. I'm not sure that's a reasonable assumption. (I'd assume people would drive more miles if they don't need to worry about parking or paying for a human driver)

Well its just an example to clarify that what is needed is some context for the incidence rate (i.e something to normalise for a fair comparison).

If you have alternate information, it's better to offer it than to describe what it could look like hypothetically.

Reply on news.ycombinator.com