4 ms·

Do you need everything to work reliably 99.9999% of the time? Is that the reliability rating of human drivers?

Retric 4y ago
Depnds on how you count. If you consider every ~5 seconds you could have an accident, and a minor or worse accident happens once every 2,000 to 20,000 hours of driving that’s 99.99993% to 99.999993% reliability. On the other hand if you start talking per trio then we aren’t even close to that.
- fastball 4y ago
  Why in the world would you count in 5 second intervals? If the measure is accidents per hour you don't then divide it by 5s. Reliability with 1 accident per 2k hours would be 99.95%. Reliability for 1 accident per 20k hours is 99.995%.
  - Retric 4y ago
    Five seconds seems roughly how long it would take for really bad choices to cause an accident. In heavy traffic you can cause an accident by steering into a car in under a second. I am sure someone has had two independent accidents in under a minute. An hour on the other hand is completely arbitrary, why not a day or a minute?
    - sgtnoodle 4y ago
      Any fixed timespan chosen to analyze risk is arbitrary, and you can convert between timespans using probability math. I'd say it's typical to choose the average "mission" length, because that's the most relevant to the purpose. For a self driving car, an hour would make sense. For the first stage of a rocket, 5 minutes. You don't ride in a car appreciating every 5 seconds that you haven't crashed yet, do you? Maybe if it's a taxi ride in Los Angeles! For a system that isn't "mission" oriented, you would probably think more in terms of mean time between failure.
      - saltcured 4y ago
        You're trying to talk about system reliability as a whole, while the others are talking about reliability of individual decisions or event handling. It's like the difference between talking about all-causes mortality versus the reliability of a bullet proof vest. Some don't buy that the vest is more reliable when we are not being shot.
      - Retric 4y ago
        The average American only drives ~35 miles per day. 1 hour may represent the average time driven per day but it’s much longer than the average trip length. Which doesn’t mean it’s a bad point of comparison but distance seems like the more common metric. Failure intervals generally use multiple different metrics depending on context. You see statistics for fatalities per year or per passenger mile as both being relevant depending on the context. A life insurance company and a car insurance company would focus on different statistics. The natural comparison for AI seems to be each individual situation. With City and Highway driving being different enough to be considered separately. In that context ~5 second intervals seem fairly natural because the car who accurately identified everything 0.001 seconds ago is probably going to avoid maying a major mistake for a few seconds. In much the same way a human driver can blink or look at a reed view mirror without significant risk.
        maxerickson 4y ago
        Decision events probably correlate better with distance than time. Like your self driving car shouldn't be meaningfully more reliable if it slows down by 5 mph (because if it is the operating parameters are poorly chosen!).
        Retric 4y ago
        5MPH makes a dramatic difference in stopping distance so that alone is going to make a car noticeably safer. However, safety isn’t the only thing people care about or it would be illegal to sell cars with the downright ludicrous top speeds currently available.
spookie 4y ago
A single person might die, go to the hospital, or even go to jail depending on the accident. But if it's the car's software fault... Who gets blamed? Bear in mind we are trying to make this economically sustainable.
- fastball 4y ago
  Presumably the company making the car/software/hardware – whichever part is at fault. If one company is making all three (Tesla) it's pretty easy.
hansvm 4y ago
Even assuming one "event" per mile that needs to be handled satisfactorily (pothole, pedestrian, inconsistent embedded reflectors, chipped or graffitied sign, ....), 99.9999% is about the same reliability as the rate that human drivers crash in the US, and fatal crashes are around 100x less common. If, as in this comment thread, self-driving cars are being proposed as a solution to the deaths caused by human ineptitude then 99.9999% seems like an absolute bare minimum for reliability.
- fastball 4y ago
  Well in terms of fatalities, FSD actually has a 100% reliability rating, so presumably it's ok for it to be deployed, right?
  - hansvm 4y ago
    Not really: - All FSD miles put together are still only 1/3 of the average miles needed to generate a fatal crash using a human driver. - The vast majority of those miles are in ideal conditions. - A cursory glance indicates there have been 15 FSD-induced fatalities.
    - fastball 4y ago
      - Fair point, guess we will see in the next 60m miles, which by the sounds of it will be in the next few months. - Tesla's FSD in particular (being an entirely vision-based system) likely operates in less-than-ideal conditions much more often than competitors. - Where are you finding the "15 FSD-induced fatalities" stat?
      - hansvm 4y ago
        - Another 60m miles will be suggestive, but you need at least 2x-3x the ordinary fatality mile count to even have a 90% confidence that the results weren't a fluke. - Tesla's FSD might operate in less-than-ideal conditions more often than competitors, but I'd posit that the typical human driver operates in much less ideal conditions than FSD. Humans are explicitly instructed to not enable it in sketchy scenarios and to be fully attentive when it's in control, and performance in that environment is at best an upper bound on what you might expect if deployed more widely. - I skimmed the stats from https://www.tesladeaths.com https://www.tesladeaths.com and was hoping Cunningham's law would kick in if that was too far off base.

Depnds on how you count. If you consider every ~5 seconds you could have an accident, and a minor or worse accident happens once every 2,000 to 20,000 hours of driving that’s 99.99993% to 99.999993% reliability. On the other hand if you start talking per trio then we aren’t even close to that.

fastball 4y ago
Why in the world would you count in 5 second intervals? If the measure is accidents per hour you don't then divide it by 5s. Reliability with 1 accident per 2k hours would be 99.95%. Reliability for 1 accident per 20k hours is 99.995%.
- Retric 4y ago
  Five seconds seems roughly how long it would take for really bad choices to cause an accident. In heavy traffic you can cause an accident by steering into a car in under a second. I am sure someone has had two independent accidents in under a minute. An hour on the other hand is completely arbitrary, why not a day or a minute?
  - sgtnoodle 4y ago
    Any fixed timespan chosen to analyze risk is arbitrary, and you can convert between timespans using probability math. I'd say it's typical to choose the average "mission" length, because that's the most relevant to the purpose. For a self driving car, an hour would make sense. For the first stage of a rocket, 5 minutes. You don't ride in a car appreciating every 5 seconds that you haven't crashed yet, do you? Maybe if it's a taxi ride in Los Angeles! For a system that isn't "mission" oriented, you would probably think more in terms of mean time between failure.
    - saltcured 4y ago
      You're trying to talk about system reliability as a whole, while the others are talking about reliability of individual decisions or event handling. It's like the difference between talking about all-causes mortality versus the reliability of a bullet proof vest. Some don't buy that the vest is more reliable when we are not being shot.
    - Retric 4y ago
      The average American only drives ~35 miles per day. 1 hour may represent the average time driven per day but it’s much longer than the average trip length. Which doesn’t mean it’s a bad point of comparison but distance seems like the more common metric. Failure intervals generally use multiple different metrics depending on context. You see statistics for fatalities per year or per passenger mile as both being relevant depending on the context. A life insurance company and a car insurance company would focus on different statistics. The natural comparison for AI seems to be each individual situation. With City and Highway driving being different enough to be considered separately. In that context ~5 second intervals seem fairly natural because the car who accurately identified everything 0.001 seconds ago is probably going to avoid maying a major mistake for a few seconds. In much the same way a human driver can blink or look at a reed view mirror without significant risk.
      - maxerickson 4y ago
        Decision events probably correlate better with distance than time. Like your self driving car shouldn't be meaningfully more reliable if it slows down by 5 mph (because if it is the operating parameters are poorly chosen!).
        Retric 4y ago
        5MPH makes a dramatic difference in stopping distance so that alone is going to make a car noticeably safer. However, safety isn’t the only thing people care about or it would be illegal to sell cars with the downright ludicrous top speeds currently available.

spookie 4y ago

A single person might die, go to the hospital, or even go to jail depending on the accident. But if it's the car's software fault... Who gets blamed? Bear in mind we are trying to make this economically sustainable.

fastball 4y ago
Presumably the company making the car/software/hardware – whichever part is at fault. If one company is making all three (Tesla) it's pretty easy.

hansvm 4y ago

Even assuming one "event" per mile that needs to be handled satisfactorily (pothole, pedestrian, inconsistent embedded reflectors, chipped or graffitied sign, ....), 99.9999% is about the same reliability as the rate that human drivers crash in the US, and fatal crashes are around 100x less common. If, as in this comment thread, self-driving cars are being proposed as a solution to the deaths caused by human ineptitude then 99.9999% seems like an absolute bare minimum for reliability.

fastball 4y ago
Well in terms of fatalities, FSD actually has a 100% reliability rating, so presumably it's ok for it to be deployed, right?
- hansvm 4y ago
  Not really: - All FSD miles put together are still only 1/3 of the average miles needed to generate a fatal crash using a human driver. - The vast majority of those miles are in ideal conditions. - A cursory glance indicates there have been 15 FSD-induced fatalities.
  - fastball 4y ago
    - Fair point, guess we will see in the next 60m miles, which by the sounds of it will be in the next few months. - Tesla's FSD in particular (being an entirely vision-based system) likely operates in less-than-ideal conditions much more often than competitors. - Where are you finding the "15 FSD-induced fatalities" stat?
    - hansvm 4y ago
      - Another 60m miles will be suggestive, but you need at least 2x-3x the ordinary fatality mile count to even have a 90% confidence that the results weren't a fluke. - Tesla's FSD might operate in less-than-ideal conditions more often than competitors, but I'd posit that the typical human driver operates in much less ideal conditions than FSD. Humans are explicitly instructed to not enable it in sketchy scenarios and to be fully attentive when it's in control, and performance in that environment is at best an upper bound on what you might expect if deployed more widely. - I skimmed the stats from https://www.tesladeaths.com https://www.tesladeaths.com and was hoping Cunningham's law would kick in if that was too far off base.