3 ms·

Which claim needs a citation? Mine or X3874s?

by zaroth 3y ago

glitchcrab 3y ago
Both claims should be backed up.
- zaroth 3y ago
  Tesla does provide crash statistics. Those stats used to be criticized on the basis that AutoPilot was driving predominantly highway miles and it’s safety record was being compared to overall driving crash statistics which have much more non-highway driving. Now Tesla breaks out AutoPilot and FSD statistics and shows that FSD driving, noting that those miles are predominantly non-highway, still has 5x fewer crashes per million miles than average. As you can see, the goalposts moved however. Now the “standard” for convincing stats includes controlling for road surfaces, weather patterns, and phase of the moon. https://twitter.com/WholeMarsBlog/status/1650601088981307392?s=20 https://twitter.com/WholeMarsBlog/status/1650601088981307392...
  - FireBeyond 3y ago
    > Now the “standard” for convincing stats includes controlling for road surfaces, weather patterns, and phase of the moon. Much as you'd like to imply there is goalpost shifting, these are quite valid concerns, and not facetious. Road surfaces: poor paint, marking, potholes, heavy use of crack filler - all things which will cause FSD to disengage as it can't determine where the road actually is. Weather: FSD will disengage in sufficiently heavy rain/show/fog/hail. It will also be compromised by bright sunlight and road reflections hitting the "vision only" camera system. > Now Tesla breaks out AutoPilot and FSD statistics and shows that FSD driving, noting that those miles are predominantly non-highway, still has 5x fewer crashes per million miles than average. All of those factors above happen on all types of roads. By absolutely no coincidence, many of these factors result in increased collisions. FSD gets turned off in these conditions, so "conveniently" it doesn't suffer the collision adjustment as a result. Human drivers have no such ability, and nor do NHTSA statistics say "Well, these crashes happened but they weren't under optimal and level playing field conditions so, well, they don't count". Rephrased: to be actually comparable and worthy of anything beyond Tesla "marketing", crash stats should be looking at "similar constraints" between FSD and humans, i.e. if FSD can't or won't engage on a given stretch of road, or in given weather conditions, then the human crashes that happened on that same stretch of road or weather conditions shouldn't count for comparison, because you have no way of saying whether FSD would have driven better, because _it wouldn't even engage_. Breaking it down to "highway vs non-highway" is certainly _easier_ but reductive to the point of near-meaninglessness.
    - mardifoufs 3y ago
      The stats account for a 5s delay between deactivation and the accident. Meaning that crashes that happens in a 5s window after deactivation count as an FSD crash. Imo 5s is usually a lot in a crash situation.
    - zaroth 3y ago
      I think you may be confused with how FSD works. FSD does not just disengage. For any of those reasons. If the FSD stack crashes, or if it is completely unable to track its position on the road, you will get a forced disengagement - a loud alarm with a big flashing red “Take Control Now” alert. That doesn’t happen if the lines get hard to see. That doesn’t happen if snow starts to fall. That doesn’t happen if the sun is at a bad angle. Forced disengagements of FSD are very rare. The human driver of course can engage or disengage FSD at any time. The system will want a clear view of the lanes at the moment in time that FSD is first engaged, after which it will stay on until the driver decides to turn it off. You can engage FSD at night, in the rain/snow, etc. But I can see how you might assume there are dangerous conditions that humans are compelled to drive where it’s unlikely FSD would be used. And let’s even presume without evidence that is true. However the premise is flawed. The vast majority of traffic accidents are not due to driving conditions, but rather are due to driver conditions. Drivers in accidents are usually either drunk, fatigued, or distracted. Drivers in accidents are likely to be inexperienced, or overly aggressive. This is why the gap is so large (5x safer is an absolutely massive gap). Not because humans can’t be better drivers, but mostly because humans choose not to be better drivers because, frankly, humans make lots of really bad choices.

Which claim needs a citation? Mine or X3874s?

by zaroth 3y ago

Which claim needs a citation? Mine or X3874s?

glitchcrab 3y ago

Both claims should be backed up.

zaroth 3y ago
Tesla does provide crash statistics. Those stats used to be criticized on the basis that AutoPilot was driving predominantly highway miles and it’s safety record was being compared to overall driving crash statistics which have much more non-highway driving. Now Tesla breaks out AutoPilot and FSD statistics and shows that FSD driving, noting that those miles are predominantly non-highway, still has 5x fewer crashes per million miles than average. As you can see, the goalposts moved however. Now the “standard” for convincing stats includes controlling for road surfaces, weather patterns, and phase of the moon. https://twitter.com/WholeMarsBlog/status/1650601088981307392?s=20 https://twitter.com/WholeMarsBlog/status/1650601088981307392...
- FireBeyond 3y ago
  > Now the “standard” for convincing stats includes controlling for road surfaces, weather patterns, and phase of the moon. Much as you'd like to imply there is goalpost shifting, these are quite valid concerns, and not facetious. Road surfaces: poor paint, marking, potholes, heavy use of crack filler - all things which will cause FSD to disengage as it can't determine where the road actually is. Weather: FSD will disengage in sufficiently heavy rain/show/fog/hail. It will also be compromised by bright sunlight and road reflections hitting the "vision only" camera system. > Now Tesla breaks out AutoPilot and FSD statistics and shows that FSD driving, noting that those miles are predominantly non-highway, still has 5x fewer crashes per million miles than average. All of those factors above happen on all types of roads. By absolutely no coincidence, many of these factors result in increased collisions. FSD gets turned off in these conditions, so "conveniently" it doesn't suffer the collision adjustment as a result. Human drivers have no such ability, and nor do NHTSA statistics say "Well, these crashes happened but they weren't under optimal and level playing field conditions so, well, they don't count". Rephrased: to be actually comparable and worthy of anything beyond Tesla "marketing", crash stats should be looking at "similar constraints" between FSD and humans, i.e. if FSD can't or won't engage on a given stretch of road, or in given weather conditions, then the human crashes that happened on that same stretch of road or weather conditions shouldn't count for comparison, because you have no way of saying whether FSD would have driven better, because _it wouldn't even engage_. Breaking it down to "highway vs non-highway" is certainly _easier_ but reductive to the point of near-meaninglessness.
  - mardifoufs 3y ago
    The stats account for a 5s delay between deactivation and the accident. Meaning that crashes that happens in a 5s window after deactivation count as an FSD crash. Imo 5s is usually a lot in a crash situation.
  - zaroth 3y ago
    I think you may be confused with how FSD works. FSD does not just disengage. For any of those reasons. If the FSD stack crashes, or if it is completely unable to track its position on the road, you will get a forced disengagement - a loud alarm with a big flashing red “Take Control Now” alert. That doesn’t happen if the lines get hard to see. That doesn’t happen if snow starts to fall. That doesn’t happen if the sun is at a bad angle. Forced disengagements of FSD are very rare. The human driver of course can engage or disengage FSD at any time. The system will want a clear view of the lanes at the moment in time that FSD is first engaged, after which it will stay on until the driver decides to turn it off. You can engage FSD at night, in the rain/snow, etc. But I can see how you might assume there are dangerous conditions that humans are compelled to drive where it’s unlikely FSD would be used. And let’s even presume without evidence that is true. However the premise is flawed. The vast majority of traffic accidents are not due to driving conditions, but rather are due to driver conditions. Drivers in accidents are usually either drunk, fatigued, or distracted. Drivers in accidents are likely to be inexperienced, or overly aggressive. This is why the gap is so large (5x safer is an absolutely massive gap). Not because humans can’t be better drivers, but mostly because humans choose not to be better drivers because, frankly, humans make lots of really bad choices.