7 ms·

>Cruise and Waymo say city officials have mischaracterized their safety track records. Their driverless taxis, the companies say, have lower collision rates tha

>Cruise and Waymo say city officials have mischaracterized their safety track records. Their driverless taxis, the companies say, have lower collision rates than human drivers and public transit. This is comparing the mean driverless incident rate to the mean human driver incident rate. This is disingenuous, since a small minority of human drivers cause the vast majority of incidents, thereby severely inflating the mean human incident rate. Today's self-driving cars may be safer than the mean human driver, but I would wager they are far from the median human driver, and absolutely nowhere close to the top 10% of human drivers. It may be possible (but very difficult) to beat the median human with dramatic improvements to current self-driving algorithms, but beating the top 10% of human drivers will require AGI.

minwcnt5 3y ago
> I would wager they are far from the median human driver, and absolutely nowhere close to the top 10% of human drivers. I think you would likely lose that wager. In the information Waymo has published about their safety methodology, they benchmark themselves against an always-attentive driver (this already rules out most accidents), which they still outperform by a large margin. Even a driver who is never distracted doesn't have constant 360 degree vision or near-instant reaction times.
scarmig 3y ago
Mean is the relevant metric, though. If it's only the bottom 10% of human drivers who cause all crashes and driverless cars are as safe as the 20p human driver, you can replace all human drivers with autonomous ones and eliminate all crashes. That's to say nothing of the fact that driverless cars will improve from their current level, and the fact that companies operating them can be held liable for crashes and forced to account for their actions in a way that the worst human drivers cannot (which provides a strong incentive for them to improve).
- cozzyd 3y ago
  you can more easily prevent the bottom 10% from driving by more rigorous driving standards / more stringent revocation of licenses.
  - scarmig 3y ago
    About a month ago a human driver ran into me while I was riding a bike and sped away (luckily I was unharmed, aside from a scrape or two). How exactly would stricter driving tests or revocation of licenses have prevented that? If a Waymo had done that, on the other hand, it'd have stopped and I'd be getting a nice big check from Google.
    - cozzyd 3y ago
      traffic enforcement cameras (e.g. automatic tickets for speeding, running red lights, hitting cyclists etc.) would probably help, but for some reason motorists are against that. I guess the nice thing about robocars is that they can be effectively regulated not to drive at unsafe speeds (though probably people would be mad about that too...).
      - tiltowait 3y ago
        Red light cameras don't seem to increase safety: https://www.scientificamerican.com/article/red-light-cameras-may-not-make-streets-safer/ https://www.scientificamerican.com/article/red-light-cameras... Motorists hate them because they're a money grab from local municipalities.
        cozzyd 3y ago
        Rear end crashes are much more survivable than t-bones though, and this study doesn't consider pedestrian injuries. To me it sounds like the cameras should also be fining people who are stopping too quickly, as it implies they were going at an unsafe speed for the condition. Anyway, relatively few people run red lights. Speed cameras would likely have a much higher safety return.
        Thlom 3y ago
        Unless you move them around drivers will just slow down at the camera and then commence speeding. What works is strategically placed cameras that measure average speed over a distance. We have these a few places in Norway and most people stick to the speed limit or a bit below between the cameras.
johnfn 3y ago
Hmm, I think that, in fact, this comparison is disingenuous! If you want to reduce net collisions, you want to compare to the mean, not median. In that case, once you've produced a self-driving car that's better than the mean, you're saving lives. You don't need to create a self-driving car which is better than every driver, ever.
- scarmig 3y ago
  It doesn't even need to be at the mean. It just needs to drive more safely than the worst drivers who cause fatal crashes, and you're saving lives.
  - dublinben 3y ago
    SDC will only move the average upwards if they're taking less safe drivers off the road. If a 40% safe SDC replaces a 50% safe human driver, then you've shifted the average downwards.
valine 3y ago
Why does it require AGI? That's a bold, unsubstantiated claim. Let me make my own bold claim: Simple common sense reasoning is very close to being solved with our latest LLMs. You would need AGI if you wanted to replace the team developing your self-driving car. A decent LLM, with vision input that's been fine-tuned on driving scenarios, will be more than sufficient for Level 5 driving.
- reso 3y ago
  You need AGI for good self-driving because driving requires predicting the actions of other human drivers at an extremely high level. This is second-nature for humans so we barely notice that we are doing it, but it is extraordinarily difficult for non-humans.
  - yazaddaruvala 3y ago
    This is overcome with super human sensing and reaction time and better visual angles. i.e. radar based deceleration to avoid accidents already helps many humans avoid collision. It'll help the robot too. The rest of driving is relatively simple, methodological, and slow.
    - reso 3y ago
      That is a conjecture which does not represent the current state of the technology.
    - lanstin 3y ago
      But also ambiguous and from time to time requiring judgment. Should I let that dumb ass driver go next or pull around them? I agree it’s insane not to allow the automated driving use more sensors than humans have. I wish I had vision that can cut thru rain and glare.
  - valine 3y ago
    No you don’t. I can feed images of crazy driving scenarios into LLaVA and get reasonable responses. That’s a general purpose LLM with $500 worth of fine tuning running locally on my PC. You should look into what can be done with the current state of the art LLMs. Your intuition for what’s possible is out of date. If I can do that with open source LLaMA variants, I can only imagine what’s possible if you have an actual annotated dataset of driving scenarios. Imagine a LLaMA model thats been fine tuned for lane selection, AEB, etc.
prpl 3y ago
P75 would be nice
bluGill 3y ago
I care about the overall mean though. Do you know what we call the bottom 10% drivers? Drivers! So long as overall they are better, then get humans off the road, sure the top 10% are worse off, but overall I'm better off as I will never have to deal with that bottom 10%.
speedgoose 3y ago
Thankfully driving safer than the mean or median taxi driver seems very achievable.
paxys 3y ago
They also conveniently ignore everything except collisions. For example number of traffic jams caused because your car got confused and decided to stop in the middle of the street.
bhuga 3y ago
I hadn't heard this statistic: > since a small minority of human drivers cause the vast majority of incidents It's tough to find anything relevant on DDG or google, they come up with information about racial disparities in traffic stops...can you link anything to read about this?
abeppu 3y ago
I think a key disconnect is that the 'incidents' being complained of most are not collisions, but they can still be disruptive, and indeed dangerous. Suppose one were to just install brightly painted immobile bollards on streets, and insist they were "driving" just very very slowly. They wouldn't hit anyone. They wouldn't kill anyone. They would piss everyone off. This disconnect is repeatedly part of where this conversation gets tripped up. > “Cruise’s safety record is publicly reported and includes having driven millions of miles in an extremely complex urban environment with zero life-threatening injuries or fatalities,” Cruise spokesperson Hannah Lindow told The Chronicle. > The city’s transportation agencies documented several incidents where driverless cars disrupted Muni service. During the night of Sept. 23, five Cruise cars blocked traffic lanes on Mission Street in Bernal Heights, stalling a Muni bus for 45 minutes. On at least three different occasions, Cruise cars stopped on Muni light-rail tracks, halting service. https://www.sfchronicle.com/projects/2023/self-driving-cars/ https://www.sfchronicle.com/projects/2023/self-driving-cars/ Ok great, they didn't kill or injure people which is nice, but they were _disruptive_ in a way which a human driver would not have been. And critically, we can't _have_ a fact-based conversation around those non-collision incidents because these companies aren't even required to report them: > But officials said it’s been difficult to assess their effectiveness because companies aren’t required to report unplanned stop incidents — some of which have been captured on social media — when they happen. > San Francisco officials want the state to require that companies report incidents when they happen. I've certainly seen cases where self-driving cars behaved in a way that I would consider reckless if a human was at the wheel, but were not collisions, did not result in injury, and I expect will not form part of any statistics reported to any government body ... but I think that's too low a bar.
- reaperducer 3y ago
  I've certainly seen cases where self-driving cars behaved in a way that I would consider reckless if a human was at the wheel, but were not collisions, did not result in injury, and I expect will not form part of any statistics reported to any government body A few weeks ago I saw one blow a red light. Nobody hurt or killed, so it wasn't reported to any responsible agency. It might not have even been tallied by the company, if the car thought there wasn't a problem. But its action was clearly unsafe. It's hard to get a grip on the problem when the data is so faulty.
gruez 3y ago
>This is comparing the mean driverless incident rate to the mean human driver incident rate. This is disingenuous, since a small minority of human drivers cause the vast majority of incidents, thereby severely inflating the mean human incident rate. Why is this disingenuous? Unless you're also for advocating for banning those top 10% of drivers from driving, you can't just cherry pick only the good drivers.
- mtsr 3y ago
  Banning those drivers does make sense, though. Repeat offenders (speeding, dangerous driving, etc) should just have their license revoked, right?
- okdood64 3y ago
  > unless you're also for advocating for banning those top 10% of drivers from driving Yes maybe they should be at least suspended from driving for some period of time. The problem that won't happen because we don't catch them all.
mdorazio 3y ago
> a small minority of human drivers cause the vast majority of incidents Citation needed.
lamontcg 3y ago
> beating the top 10% of human drivers will require AGI. even if it is true, this doesn't matter at all. if they can beat the median driver and you replaced all drivers with driverless vehicles, they would eliminate whatever bottom percentile driver is responsible for most of the accidents and the world would be vastly safer. by your own assertion, the 10% best drivers don't matter because they already don't cause accidents.
dontreact 3y ago
I personally don’t care that much the collision rate. I care about the fatality or injury rate (deaths or injuries per miles driven) I don’t mind picking an Uber driver if it reduces my chances of getting hurt, even if we are slightly more likely to experience a fender bender.
civilized 3y ago
> a small minority of human drivers cause the vast majority of incidents, thereby severely inflating the mean human incident rate. Could you cite a source for this claim? I have tried many different searches and can't find any support for it.

minwcnt5 3y ago

> I would wager they are far from the median human driver, and absolutely nowhere close to the top 10% of human drivers. I think you would likely lose that wager. In the information Waymo has published about their safety methodology, they benchmark themselves against an always-attentive driver (this already rules out most accidents), which they still outperform by a large margin. Even a driver who is never distracted doesn't have constant 360 degree vision or near-instant reaction times.

scarmig 3y ago

Mean is the relevant metric, though. If it's only the bottom 10% of human drivers who cause all crashes and driverless cars are as safe as the 20p human driver, you can replace all human drivers with autonomous ones and eliminate all crashes. That's to say nothing of the fact that driverless cars will improve from their current level, and the fact that companies operating them can be held liable for crashes and forced to account for their actions in a way that the worst human drivers cannot (which provides a strong incentive for them to improve).

cozzyd 3y ago
you can more easily prevent the bottom 10% from driving by more rigorous driving standards / more stringent revocation of licenses.
- scarmig 3y ago
  About a month ago a human driver ran into me while I was riding a bike and sped away (luckily I was unharmed, aside from a scrape or two). How exactly would stricter driving tests or revocation of licenses have prevented that? If a Waymo had done that, on the other hand, it'd have stopped and I'd be getting a nice big check from Google.
  - cozzyd 3y ago
    traffic enforcement cameras (e.g. automatic tickets for speeding, running red lights, hitting cyclists etc.) would probably help, but for some reason motorists are against that. I guess the nice thing about robocars is that they can be effectively regulated not to drive at unsafe speeds (though probably people would be mad about that too...).
    - tiltowait 3y ago
      Red light cameras don't seem to increase safety: https://www.scientificamerican.com/article/red-light-cameras-may-not-make-streets-safer/ https://www.scientificamerican.com/article/red-light-cameras... Motorists hate them because they're a money grab from local municipalities.
      - cozzyd 3y ago
        Rear end crashes are much more survivable than t-bones though, and this study doesn't consider pedestrian injuries. To me it sounds like the cameras should also be fining people who are stopping too quickly, as it implies they were going at an unsafe speed for the condition. Anyway, relatively few people run red lights. Speed cameras would likely have a much higher safety return.
        Thlom 3y ago
        Unless you move them around drivers will just slow down at the camera and then commence speeding. What works is strategically placed cameras that measure average speed over a distance. We have these a few places in Norway and most people stick to the speed limit or a bit below between the cameras.

johnfn 3y ago

Hmm, I think that, in fact, this comparison is disingenuous! If you want to reduce net collisions, you want to compare to the mean, not median. In that case, once you've produced a self-driving car that's better than the mean, you're saving lives. You don't need to create a self-driving car which is better than every driver, ever.

scarmig 3y ago
It doesn't even need to be at the mean. It just needs to drive more safely than the worst drivers who cause fatal crashes, and you're saving lives.
- dublinben 3y ago
  SDC will only move the average upwards if they're taking less safe drivers off the road. If a 40% safe SDC replaces a 50% safe human driver, then you've shifted the average downwards.

valine 3y ago

Why does it require AGI? That's a bold, unsubstantiated claim. Let me make my own bold claim: Simple common sense reasoning is very close to being solved with our latest LLMs. You would need AGI if you wanted to replace the team developing your self-driving car. A decent LLM, with vision input that's been fine-tuned on driving scenarios, will be more than sufficient for Level 5 driving.

reso 3y ago
You need AGI for good self-driving because driving requires predicting the actions of other human drivers at an extremely high level. This is second-nature for humans so we barely notice that we are doing it, but it is extraordinarily difficult for non-humans.
- yazaddaruvala 3y ago
  This is overcome with super human sensing and reaction time and better visual angles. i.e. radar based deceleration to avoid accidents already helps many humans avoid collision. It'll help the robot too. The rest of driving is relatively simple, methodological, and slow.
  - reso 3y ago
    That is a conjecture which does not represent the current state of the technology.
  - lanstin 3y ago
    But also ambiguous and from time to time requiring judgment. Should I let that dumb ass driver go next or pull around them? I agree it’s insane not to allow the automated driving use more sensors than humans have. I wish I had vision that can cut thru rain and glare.
- valine 3y ago
  No you don’t. I can feed images of crazy driving scenarios into LLaVA and get reasonable responses. That’s a general purpose LLM with $500 worth of fine tuning running locally on my PC. You should look into what can be done with the current state of the art LLMs. Your intuition for what’s possible is out of date. If I can do that with open source LLaMA variants, I can only imagine what’s possible if you have an actual annotated dataset of driving scenarios. Imagine a LLaMA model thats been fine tuned for lane selection, AEB, etc.

prpl 3y ago

P75 would be nice

bluGill 3y ago

I care about the overall mean though. Do you know what we call the bottom 10% drivers? Drivers! So long as overall they are better, then get humans off the road, sure the top 10% are worse off, but overall I'm better off as I will never have to deal with that bottom 10%.

speedgoose 3y ago

Thankfully driving safer than the mean or median taxi driver seems very achievable.

paxys 3y ago

They also conveniently ignore everything except collisions. For example number of traffic jams caused because your car got confused and decided to stop in the middle of the street.

bhuga 3y ago

I hadn't heard this statistic: > since a small minority of human drivers cause the vast majority of incidents It's tough to find anything relevant on DDG or google, they come up with information about racial disparities in traffic stops...can you link anything to read about this?

abeppu 3y ago

I think a key disconnect is that the 'incidents' being complained of most are not collisions, but they can still be disruptive, and indeed dangerous. Suppose one were to just install brightly painted immobile bollards on streets, and insist they were "driving" just very very slowly. They wouldn't hit anyone. They wouldn't kill anyone. They would piss everyone off. This disconnect is repeatedly part of where this conversation gets tripped up. > “Cruise’s safety record is publicly reported and includes having driven millions of miles in an extremely complex urban environment with zero life-threatening injuries or fatalities,” Cruise spokesperson Hannah Lindow told The Chronicle. > The city’s transportation agencies documented several incidents where driverless cars disrupted Muni service. During the night of Sept. 23, five Cruise cars blocked traffic lanes on Mission Street in Bernal Heights, stalling a Muni bus for 45 minutes. On at least three different occasions, Cruise cars stopped on Muni light-rail tracks, halting service. https://www.sfchronicle.com/projects/2023/self-driving-cars/ https://www.sfchronicle.com/projects/2023/self-driving-cars/ Ok great, they didn't kill or injure people which is nice, but they were _disruptive_ in a way which a human driver would not have been. And critically, we can't _have_ a fact-based conversation around those non-collision incidents because these companies aren't even required to report them: > But officials said it’s been difficult to assess their effectiveness because companies aren’t required to report unplanned stop incidents — some of which have been captured on social media — when they happen. > San Francisco officials want the state to require that companies report incidents when they happen. I've certainly seen cases where self-driving cars behaved in a way that I would consider reckless if a human was at the wheel, but were not collisions, did not result in injury, and I expect will not form part of any statistics reported to any government body ... but I think that's too low a bar.

reaperducer 3y ago
I've certainly seen cases where self-driving cars behaved in a way that I would consider reckless if a human was at the wheel, but were not collisions, did not result in injury, and I expect will not form part of any statistics reported to any government body A few weeks ago I saw one blow a red light. Nobody hurt or killed, so it wasn't reported to any responsible agency. It might not have even been tallied by the company, if the car thought there wasn't a problem. But its action was clearly unsafe. It's hard to get a grip on the problem when the data is so faulty.

gruez 3y ago

>This is comparing the mean driverless incident rate to the mean human driver incident rate. This is disingenuous, since a small minority of human drivers cause the vast majority of incidents, thereby severely inflating the mean human incident rate. Why is this disingenuous? Unless you're also for advocating for banning those top 10% of drivers from driving, you can't just cherry pick only the good drivers.

mtsr 3y ago
Banning those drivers does make sense, though. Repeat offenders (speeding, dangerous driving, etc) should just have their license revoked, right?
okdood64 3y ago
> unless you're also for advocating for banning those top 10% of drivers from driving Yes maybe they should be at least suspended from driving for some period of time. The problem that won't happen because we don't catch them all.

mdorazio 3y ago

> a small minority of human drivers cause the vast majority of incidents Citation needed.

lamontcg 3y ago

> beating the top 10% of human drivers will require AGI. even if it is true, this doesn't matter at all. if they can beat the median driver and you replaced all drivers with driverless vehicles, they would eliminate whatever bottom percentile driver is responsible for most of the accidents and the world would be vastly safer. by your own assertion, the 10% best drivers don't matter because they already don't cause accidents.

dontreact 3y ago

I personally don’t care that much the collision rate. I care about the fatality or injury rate (deaths or injuries per miles driven) I don’t mind picking an Uber driver if it reduces my chances of getting hurt, even if we are slightly more likely to experience a fender bender.

civilized 3y ago

> a small minority of human drivers cause the vast majority of incidents, thereby severely inflating the mean human incident rate. Could you cite a source for this claim? I have tried many different searches and can't find any support for it.