5 ms·

> Humans drive correctly 99.99999% of the time. We can't can't even keep a computer up and running that well, never mind driving perfectly. This comparison doe

> Humans drive correctly 99.99999% of the time. We can't can't even keep a computer up and running that well, never mind driving perfectly. This comparison doesn't make any sense. You're comparing _continuous_ operation with the aggregate of the average human's bouts of driving. Even robotaxis aren't going to run constantly. A sensible analogue would be something like aircraft software, which I'd guess does have near the reliability of human drivers (and can last much longer continuously than humans do).

ars 5y ago
> You're comparing _continuous_ operation with the aggregate of the average human's bouts of driving. I don't understand what you mean. I'm comparing actual time behind the wheel - hours of correct driving vs hours (well minutes) of incorrect driving. Do the math yourself, take total hours behind the wheel and divide by the accident rate, like I did. You'll get the same results. Depending on your assumptions anywhere from 99.999 to 99.99999% correct driving by humans. > Even robotaxis aren't going to run constantly. Again, I'm not comparing constantly, I'm comparing actual hours behind the wheel.
- iso1210 5y ago
  84 billion hours of driving[0] for 35,485 deaths[1]. That's one death per 2.4 million hours, so that's an indicator of the magnitude that's needed before even talking about it. That number also means no situations where the computer gives up and hands control back to the driver. If autonomous cars can never hand back (at most stopping legally, safely and out of traffic) and cause 1 death per 10 million hours, that would seem about right. [0] https://www.volpe.dot.gov/news/how-much-time-do-americans-spend-behind-wheel https://www.volpe.dot.gov/news/how-much-time-do-americans-sp... [1] https://en.wikipedia.org/wiki/Motor_vehicle_fatality_rate_in_U.S._by_year https://en.wikipedia.org/wiki/Motor_vehicle_fatality_rate_in...
- wutbrodo 5y ago
  Is this not a reference to uptime (as in servers)? > We can't can't even keep a computer up and running that well That's the comparison that I quoted in my comment and that I was responding to. You're comparing continuous uptime to humans' disconnected (and brief) bouts of driving.
bumby 5y ago
>A sensible analogue would be something like aircraft software Does automobile software development have the level of oversight that aircraft development has in terms of regulation and approval for use? I know the NHTSA approves safety with crash tests (and are currently working on AV standards) but I'm not sure it's an equivalent to what the FAA currently does. Not to say they won't be equivalent in the future, but I don't know how well the analogue holds up if they don't have equivalent levels of quality oversight currently.
- wutbrodo 5y ago
  > Does automobile software development have the level of oversight that aircraft development has in terms of regulation and approval for use? Nothing in my comment implied this, and I don't disagree that it doesn't. The question is immaterial to the specific topic of discussion: the comparison of server uptime to human-driver reliability as an indicator of the feasibility of functional AV software.
  - bumby 5y ago
    The tone of your response seems to imply I’ve was attacking your position, but that wasn’t my intent. I am legitimately asking how much parity there is regarding oversight regulation between flight and autonomous software, which would help make your point clearer since it has a direct relationship to reliability. However, I do think it’s relevant to the point. If you are making an analogy about reliability but one is legally forced to demonstrate reliability while the other is not, does that comparison really make sense with that additional context? While I've worked in the aerospace software domain, I don't know to what standard AV is held which is why I was asking.
    - wutbrodo 5y ago
      > seems to imply I’ve was attacking your position, but that wasn’t my intent. Not quite attacking, but disagreeing with an implied assumption that I didn't make (ie, that I was arguing that AVs were ready for primetime instead of narrowly disputing the flawed reliability comparison in the parent comment). > I am legitimately asking how much parity there is regarding oversight regulation between flight and autonomous software, which would help make your point clearer since it has a direct relationship to reliability The reason I figured it was rhetorical is that I thought it was fairly universally known that there's no comparable regulatory apparatus for AVs (yet?). But you're mistaken that this is relevant to my point: the parent comment was discussing the feasibility of AVs by saying "we can't build computers with the uptime of human drivers", and I was pointing out that he had this backwards. > If you are making an analogy about reliability but one is legally forced to demonstrate reliability while the other is not, does that comparison really make sense with that additional context? I didn't say AVs were as reliable as aircraft software, nor that they would definitely get to that pt (perhaps they don't need to, as safe stops are much easier on the ground). I was disputing the claim that computer system uptime can't be made as reliable as humans in the current day, using aircraft systems as an in-production example.

ars 5y ago

> You're comparing _continuous_ operation with the aggregate of the average human's bouts of driving. I don't understand what you mean. I'm comparing actual time behind the wheel - hours of correct driving vs hours (well minutes) of incorrect driving. Do the math yourself, take total hours behind the wheel and divide by the accident rate, like I did. You'll get the same results. Depending on your assumptions anywhere from 99.999 to 99.99999% correct driving by humans. > Even robotaxis aren't going to run constantly. Again, I'm not comparing constantly, I'm comparing actual hours behind the wheel.

iso1210 5y ago
84 billion hours of driving[0] for 35,485 deaths[1]. That's one death per 2.4 million hours, so that's an indicator of the magnitude that's needed before even talking about it. That number also means no situations where the computer gives up and hands control back to the driver. If autonomous cars can never hand back (at most stopping legally, safely and out of traffic) and cause 1 death per 10 million hours, that would seem about right. [0] https://www.volpe.dot.gov/news/how-much-time-do-americans-spend-behind-wheel https://www.volpe.dot.gov/news/how-much-time-do-americans-sp... [1] https://en.wikipedia.org/wiki/Motor_vehicle_fatality_rate_in_U.S._by_year https://en.wikipedia.org/wiki/Motor_vehicle_fatality_rate_in...
wutbrodo 5y ago
Is this not a reference to uptime (as in servers)? > We can't can't even keep a computer up and running that well That's the comparison that I quoted in my comment and that I was responding to. You're comparing continuous uptime to humans' disconnected (and brief) bouts of driving.

bumby 5y ago

>A sensible analogue would be something like aircraft software Does automobile software development have the level of oversight that aircraft development has in terms of regulation and approval for use? I know the NHTSA approves safety with crash tests (and are currently working on AV standards) but I'm not sure it's an equivalent to what the FAA currently does. Not to say they won't be equivalent in the future, but I don't know how well the analogue holds up if they don't have equivalent levels of quality oversight currently.

wutbrodo 5y ago
> Does automobile software development have the level of oversight that aircraft development has in terms of regulation and approval for use? Nothing in my comment implied this, and I don't disagree that it doesn't. The question is immaterial to the specific topic of discussion: the comparison of server uptime to human-driver reliability as an indicator of the feasibility of functional AV software.
- bumby 5y ago
  The tone of your response seems to imply I’ve was attacking your position, but that wasn’t my intent. I am legitimately asking how much parity there is regarding oversight regulation between flight and autonomous software, which would help make your point clearer since it has a direct relationship to reliability. However, I do think it’s relevant to the point. If you are making an analogy about reliability but one is legally forced to demonstrate reliability while the other is not, does that comparison really make sense with that additional context? While I've worked in the aerospace software domain, I don't know to what standard AV is held which is why I was asking.
  - wutbrodo 5y ago
    > seems to imply I’ve was attacking your position, but that wasn’t my intent. Not quite attacking, but disagreeing with an implied assumption that I didn't make (ie, that I was arguing that AVs were ready for primetime instead of narrowly disputing the flawed reliability comparison in the parent comment). > I am legitimately asking how much parity there is regarding oversight regulation between flight and autonomous software, which would help make your point clearer since it has a direct relationship to reliability The reason I figured it was rhetorical is that I thought it was fairly universally known that there's no comparable regulatory apparatus for AVs (yet?). But you're mistaken that this is relevant to my point: the parent comment was discussing the feasibility of AVs by saying "we can't build computers with the uptime of human drivers", and I was pointing out that he had this backwards. > If you are making an analogy about reliability but one is legally forced to demonstrate reliability while the other is not, does that comparison really make sense with that additional context? I didn't say AVs were as reliable as aircraft software, nor that they would definitely get to that pt (perhaps they don't need to, as safe stops are much easier on the ground). I was disputing the claim that computer system uptime can't be made as reliable as humans in the current day, using aircraft systems as an in-production example.