4 ms·

>Seeing as the cars have been on the road in Arizona for a while, and this is the first we are hearing about a pedestrian collision, it seems the algorithm is p

by hnaccy 9y ago

>Seeing as the cars have been on the road in Arizona for a while, and this is the first we are hearing about a pedestrian collision, it seems the algorithm is pretty safe. One fatality in 4 million miles is safe? Humans are 20 times better than that.

invisible 9y ago
I wonder what the margin of error is on that one in 4 million rate since the sample size is one. Also of note: > Arizona has experienced a surge of pedestrian fatalities recently, with more than 10 in a single week of March in Phoenix alone. The state has the highest rate of pedestrian fatalities in the United States.
- ajb 9y ago
  The sample size is not one. The sample size is 4 million. The 5% and 95% quantiles of the posterior are roughly one in 11 million and one in 800,000.
w0utert 9y ago
>> One fatality in 4 million miles is safe? Humans are 20 times better than that. And this is considering the fatality rate for all human drivers, including those that are drunk, playing on their smartphone, sleepy, speeding, street racing, etc. I would not be surprised if those cases account for a very sizable majority of fatal car accidents. Now imagine you would reduce those numbers by relatively simple measures that don't involve AI. Let the car enforce speed limits, mandate smartphone companies to have a 'car mode' that cannot be shut off, mandate attention tracking capabilities for cars, dead-zone indicators, automatic emergency braking, etc. Maybe even mandate an alcohol lock, or at the very least a black box that records driving behavior so it can be analyzed in case of a crash, with sever penalties after the fact if the black box shows signs of drunk driving, etc. The list of things you can do to improve road safety without requiring AI cars is endless, yet billions are invested in trying to fully automate driving. This just shows that self-driving cars are not about safety, but purely about business opportunities.
- ACow_Adonis 9y ago
  And conversely, those 4 million miles will (presumably) be highly biased to safe/above-standard driving conditions. As you've basically said, the human stats include all the very worst drivers in the full range of driving conditions, whereas the auto-cars not only have the very best driving conditions, but likely also have the statistical benefit of a likely number of cases where the human actually stepped in and upped the auto-car score even further (since we can't account for the auto-car deaths/fatalities that humans have successfully stepped in and stopped).
  - w0utert 9y ago
    Yes, so far the statistics we have about the safety of AI cars are mostly meaningless to say anything useful about how they would scale if applied universally. Another thing that bothers me is how often the argument 'human drivers are terrible, far worse than computers for situations A, B, C, etc' is used to suggest that AI systems will improve road safety. I will be the first to admit that many people are terrible drivers, they make mistakes, don't pay enough attention, don't obey traffic laws, have bad reaction times, etc, no argument there. But so far I have never seen any statistics about how many of the accidents caused by these kinds of human failures actually lead to fatalities. Taking the human out of the equation could in theory improve all kinds of car accidents, but how many would be just fender-benders as opposed to road fatalities? It's not like it is super-easy to accidentally cause a fatal road accident, and everyone knows one or more people who were involved in one. Who can prove that AI cars will not decrease the total number of accidents of any kind, but increase the kinds of accidents with fatalities, like this one with the Uber car? Maybe human drivers, with all their faults, would have avoided this particular accident by driving more defensively or using social cues to be more prepared for this woman crossing the street? I'm not trying to make an argument about the viability of AI cars that are safer than humans here by the way. I personally don't believe AI cars will ever be safer overall than humans in all possible situations that people drive in, but that's a different discussion. What bugs me is that suddenly it seems as if the whole world is willing to bend over backwards to interpret statistics about the presumed safety of AI cars in the most favorable way possible, and at the same time ignore all these relatively simple things that could actually improve safety right now, at much lower costs than developing fully autonomous cars.

by hnaccy 9y ago

invisible 9y ago

I wonder what the margin of error is on that one in 4 million rate since the sample size is one. Also of note: > Arizona has experienced a surge of pedestrian fatalities recently, with more than 10 in a single week of March in Phoenix alone. The state has the highest rate of pedestrian fatalities in the United States.

ajb 9y ago
The sample size is not one. The sample size is 4 million. The 5% and 95% quantiles of the posterior are roughly one in 11 million and one in 800,000.

w0utert 9y ago

>> One fatality in 4 million miles is safe? Humans are 20 times better than that. And this is considering the fatality rate for all human drivers, including those that are drunk, playing on their smartphone, sleepy, speeding, street racing, etc. I would not be surprised if those cases account for a very sizable majority of fatal car accidents. Now imagine you would reduce those numbers by relatively simple measures that don't involve AI. Let the car enforce speed limits, mandate smartphone companies to have a 'car mode' that cannot be shut off, mandate attention tracking capabilities for cars, dead-zone indicators, automatic emergency braking, etc. Maybe even mandate an alcohol lock, or at the very least a black box that records driving behavior so it can be analyzed in case of a crash, with sever penalties after the fact if the black box shows signs of drunk driving, etc. The list of things you can do to improve road safety without requiring AI cars is endless, yet billions are invested in trying to fully automate driving. This just shows that self-driving cars are not about safety, but purely about business opportunities.

ACow_Adonis 9y ago
And conversely, those 4 million miles will (presumably) be highly biased to safe/above-standard driving conditions. As you've basically said, the human stats include all the very worst drivers in the full range of driving conditions, whereas the auto-cars not only have the very best driving conditions, but likely also have the statistical benefit of a likely number of cases where the human actually stepped in and upped the auto-car score even further (since we can't account for the auto-car deaths/fatalities that humans have successfully stepped in and stopped).
- w0utert 9y ago
  Yes, so far the statistics we have about the safety of AI cars are mostly meaningless to say anything useful about how they would scale if applied universally. Another thing that bothers me is how often the argument 'human drivers are terrible, far worse than computers for situations A, B, C, etc' is used to suggest that AI systems will improve road safety. I will be the first to admit that many people are terrible drivers, they make mistakes, don't pay enough attention, don't obey traffic laws, have bad reaction times, etc, no argument there. But so far I have never seen any statistics about how many of the accidents caused by these kinds of human failures actually lead to fatalities. Taking the human out of the equation could in theory improve all kinds of car accidents, but how many would be just fender-benders as opposed to road fatalities? It's not like it is super-easy to accidentally cause a fatal road accident, and everyone knows one or more people who were involved in one. Who can prove that AI cars will not decrease the total number of accidents of any kind, but increase the kinds of accidents with fatalities, like this one with the Uber car? Maybe human drivers, with all their faults, would have avoided this particular accident by driving more defensively or using social cues to be more prepared for this woman crossing the street? I'm not trying to make an argument about the viability of AI cars that are safer than humans here by the way. I personally don't believe AI cars will ever be safer overall than humans in all possible situations that people drive in, but that's a different discussion. What bugs me is that suddenly it seems as if the whole world is willing to bend over backwards to interpret statistics about the presumed safety of AI cars in the most favorable way possible, and at the same time ignore all these relatively simple things that could actually improve safety right now, at much lower costs than developing fully autonomous cars.