They may by safer but as long as there are still humans driving I believe they are not capable of managing perplexing states that humans have no problems dealing with. I’d not rush the transition to autonomous vehicles. I know there are vested interests to hasten this transition but that’s because it’s all at our cost…
They can't do this everywhere or under all conditions, and they still make plenty of mistakes. Still, if a Cruise car can navigate through a crowd of drunks who walking in the street, they're certainly capable of managing complex situations.
good catch. The "double parked" car was even using the indicator.
still impressive how the car handled the overall situation. I don't think I would have done much better. (except for maybe that single situation you pointed out)
I agree, but not with the second half of your statement regarding perplexing states that humans have no problems dealing with.
Humans have a lot of problems dealing with those states! It's why they're so difficult; funny enough a big reason why car insurance is a must but health insurance isn't. It's also why I believe we should aggressively push for driverless ecosystems, as fewer humans actually making decisions on the road = fewer potential problems due to the perplex human states.
Of course we are still a long ways from that goal, but it's hard to imagine a future where human drivers are still a thing. Makes more sense for the transportation network to be exclusively driverless and all following the same exact ruleset precisely. Again, thinking long term here not the next 20 years.
State Patrol Capt. Bryan Niewind said both the pick-up and the semi were southbound on Highway 281 just after 6:00 p.m. when the pick-up slowed to take a left turn at the junction with 75th St. SE. The semi, which was traveling behind the pick-up, pulled into the northbound lane to pass the the pick-up and hit the drivers side of the pick-up as it turned.
There is no such thing as a left turn on a limited access freeway/motorway.
You're describing an older state road or US Highway that has intersections, which has all the risks that come with intersections amplified by higher speed limits.
I didn't describe anything, I was merely refuting the claim "reason being you can’t really get t-boned on the driver side on a highway".
I quoted a news article that specifically described a highway where a vehicle was t-boned on the driver side and exactly how it occurred. You can add additional parameters such as "new highways" to narrow the original claim and refute my argument, I really can't defend against that. I really shouldn't need to.
It's rare but with enough traffic volume eventually most things happen, eg:
on a highway with no physical divider a car executing a poor overtaking procedure can pull in sharply after passing and violently "pit manoeuvre" the vehicle just ahead of it, causing that vehicle to spin 270 degrees and present the drivers side to oncoming traffic.
Rare physics aside, a highway t-bone by an oncoming truck on a turning car is enough to cause fatalities and destroy a car regardless of which side presents.
That said, in Australia country roads are far more dangerous than city streets:
It found the per capita road death rate for regional Australians in 2022 was 10.6 deaths per 100,000 people, while the corresponding rate for urban Australians was 2.24 deaths per 100,000 people.
Beware of statistics - the quote I gave above re: Australia is 100% true - but Australia is overwhelmingly an urban nation clinging to the fringes of the continent.
There are absolutely (in raw numbers) more urban automotive fatalities than rural in Australia but proportionally more (per 100,000 drivers) rural deaths from driving.
ie. Country driving is riskier.
Interestingly your link appears (to my eye) to be confusing - the first graph is described as proportional but it's not "proportional" in the sense of being "normalised", it's proportional in the sense of showing a percent division of the total absolute number; as more and more people live in urban areas, more and more people die in urban areas (thus obfuscating the relative dangers).
The second graph gets closer, showing the rural US areas still 'winning' the death stakes in terms of risk per miles travelled.
The statistics show that "[m]ore than 50 percent of the combined total of fatal and injury crashes occur at or near intersections", at least in the US.
Until they're driving everywhere humans drive, I don't see how it's possible to make any such comparison. Is Waymo taking skiers on I-70 in Colorado on a powder day? Heck, how about we get central Boston[1]? It's rather telling that the only areas of Boston were tiny slices of the very spacious, well-marked, and recently developed expansions of the city.
It's been 10 years, and the real driverless car scene has been unchanged in that time. The Silicon Valley Season 1 jokes about them have yet to age out out of relevance.
Until then we're extrapolating safety by comparing the car equivalent of train tracks to, well, people driving cars.
Exactly right. Try driving past a mob of kangaroos at the side of the road. There's a definite pace that works and they'll often break straight in front of you with little warning.
Until self driving cars have been deployed internationally in a huge range of driving environments I'll take my chances behind the wheel.
> Until they're driving everywhere humans drive, I don't see how it's possible to make any such comparison.
You raise a good point but surely it's at least possible to compare the safety records of driverless cars vs. human operated cars within the domain that driverless cars typically operate. One way to do that would be to only examine all trips, human and driverless, that occur entirely within a bounded area that driverless cars are known to operate. Another option would be to look at the total safety record of a city like SF in the years prior to widespread driverless car usage and in the years after common adoption and see if there is a broad reduction in accidents generally.
Apropos of this, the other day I was thinking about the adoption of mobile devices. Originally the unadorned word "phone" referred to a wired home/office telephone, and if you needed to refer to a mobile device, you'd say, call me on my mobile/cellphone. Over the course of a generation it's now that if you say "phone," the assumption is that you are talking about a mobile device, and the exceptions become either landline or office phone. It makes me wonder how long, if ever, it will be until "car" almost universally refers to driverless vehicles, and you'll need to say "manual car," "meatmobile," or something similar to those to refer to a human-driven vehicle.
I'm certain the best self-driving cars are better than the worst human drivers. I've personally been in the car with older relatives who terrified me.
I doubt they're better than the best humans today, and that's OK. The cars are getting better with time. My uncle isn't. We could improve road safety right now by swapping his keys for a Waymo.
Just got accepted into the Waymo One program today after being waitlisted forever :D I'm excited to be able to offer a first-hand opinion on the quality of the ride here soon.
Has anyone seen an analysis of driving safety statistics if you avoid:
* Drunk driving
* Drugged driving
* Driving very late at night
* Sleep deprived driving
* Driving with a vehicle with known safety issues or no safety inspection
* Driving without a license
* Driving while young, dumb and reckless
* Drag racing or other similar behaviour
* Motorbiking
* Driving while holding a cellphone
In other words, if you're not doing stupid things, what are your odds of death, an accident etc.
Because it would actually an important question for an individual choosing whether to use a self driving car rather than a manual car. You want the car to be safer than you, rather than safer than the average.
I don't just want it to be safer than me, I want it to be shockingly safer. If self driving cars are safer than an average driver that's great. But the cars need to be 10x as good as an average driver and 2x as good as a 99th percentile driver before the objections start to fade.
Safety isn’t the only consideration. Fully autonomous vehicles that let you close your eyes and not pay attention are worth a slight decrease in safety, just as most people consider driving faster than the posted speed limit to be worthwhile.
Avoiding all of those things makes you a safer driver, but you have to remember the most important thing on the road, you will be hit by some other driver, and they could easily be doing any or all of those things.
Which itself can be drilled down to drunk motorbiking, first timers on 200hp bikes, idiots without helmets etc.
An average motorcyclist dies 14x as often per km driven as car drivers, however it seems to just attract idiots.
That said, if drunk/distracted/too old drivers switch to self-driving cars that could be a positive change for everyone earlier than general population.
The above look to me like things of life that you would have preferred to avoid, but you'll end up doing anyway.
Driving a kid/spouse to an hospital in the middle of the night while you're also under your flu medications is a thing that happens. You could call a doctor/cab/ambulance but more often than not it's either not an affordable option, or you're too tired to have thought about it.
> * young, dumb and reckless
Could be a guaranteed Nobel price if you find the cure to this.
Honestly, that's kind of the best reason to want self driving cars. They're never drunk, texting, hitting their kids, watching Instagram videos, etc. Upgrade all of those drivers to "boring and competent" and viola, you've saved 10,000 lives a year or more.
I thought that the whole point of self driving cars is it eliminates every single one of the problems you just listed. Humans get tired, drunk, distracted, ignore warnings about low pressure in the tires etc. A computer doesn’t worry about any of these things. It just complies or it doesn’t.
It’s kind of like the whole rust vs C++ debate. Rust proponents say that Rust as a language just eliminates a whole class of errors(memory errors). They’re just nonexistent anymore. Sure there are new problems, but they’re hopefully less commonplace. Your argument would be like a C++ programmer saying, “can we see some statistics on how many security vulnerabilities show up in code written by a modern C++ developer using smart pointers and RAII? That way they can make more informed decisions over whether rust is necessary.”
I think it would be interesting to see some alternative statistics, but I’m looking forward to the day nobody gets killed by an irresponsible drunk driver because cars don’t allow people to drive at all. Plus, once all the humans are off the road, I presume new technology can be made to reduce traffic patterns and other cool things.
If you implement stuff with AI for either school or a company, I'd get used to this kind of goalpost moving.
AIs currently beat "average" humans at just about every task, if done by itself. Multimodal models will do more, but they're for next year. But reading, writing, translating, talking, listening, controlling robots, designing proteins, modeling chemical reactions, driving, ... so currently it is fancy to compare AIs to absolute top human performance, while ignoring the problems even top performing humans have (such as needing to sleep, not knowing every language, cost, ...)
I'm very excited about self driving cars. But I think this is a real question to be answered. The 80:20 rule applies in much of life, and it's probable that a small minority of drivers cause the majority of accidents and deaths.
It's quite possible that replacing all drivers with self driving cars would both:
* Improve road safety dramatically
* Worsen road safety for the people not in the accident-prone minority
I don't this is a frivolous question. My post got a lot of replies, but no one came forward with any data on the distribution of accidents.
The 80:20 “rule” is interesting because it’s not a rule, it’s an observation. The Pareto Principle isn’t a law for a reason, because it’s observing the natural result of categorizing data. This medium article was surprisingly good[0] and makes a good point: even if you eliminate those 20% of drivers you won’t see an 80% decrease in accidents because the system as a whole fluctuates daily.
I don’t think anyone came forward with any data in the replies because this distribution is impossible to come up with. Like another commenter pointed out, we all account for those distracted drivers, or tired drivers, or “accident prone” drivers at some point in our lives (possibly many points in our lives). Because of this it’s impossible to remove the 20% of the poorest drivers and have a dramatic increase in road safety.
The only way to fix this system would be to remove all the factors that you listed completely. The easiest way to remove those would be to stop having people drive cars. We won’t know if it’s safer or not until we get there, and it will still have its own 80:20 distributions, just based on different factors (and hopefully to a smaller degree).
That's a good point, but I purposefully didn't included "distracted driver" in the list. I ought to have put "extremely sleep deprived driver".
I personally have never driven drunk, extremely sleep deprived, holding a cellphone, or any of the rest. Apart from driving while young.
It's actually pretty easy to avoid the list. I guess to clarify my question you could ask: do self driving cars have the kinds of accidents people avoiding that list would never have?
* Driving after having an emotional argument or any other serious emotional stress
Those are all things that we're all probably going to do at the low points in our lives.
We conveniently edit those out of our perception of ourselves, though, because we had excuses for them (possibly quite good ones, but in reality those all count towards how good or bad of a driver you are).
> You want the car to be safer than you, rather than safer than the average.
You probably aren't as good of a driver as you think you are.
And half of you out there are just chronic tailgaters who incorrectly believe that you have to do that in order for traffic to flow faster, or just believe that people "cutting you off" will significantly increase the time it takes to get to your destination. If you don't keep about a semi-trailers worth of distance between you and the car in front of you, then an AI probably would be a safer driver.
[ Or just the passive-aggressive shithead that wasn't going to slow down for me in the crosswalk the other night ]
It's or course true that if you exclude all dangerous human driver scenarios, you're left with something unrealistically safe.
But it's not entirely an apples to apples comparison. People who want to drive recklessly will not switch on self driving. Self driving cars are pretty new simply because it's a new technology, how will they compare against a banged up 30 yo car when they are themselves 30 yo? They may in fact fare worse, because there is more reliance on fragile components like cameras or lidar. How will they deal with the things like dodgy brakes or the one gear that doesn't work? These are things humans can mostly adapt to.
The other point, not raised by GP, is that often self driving only works on fairly safe roads anyway, whereas humans drive everywhere, skewing the stats.
But I guess broadly I agree, it is possible that on average, self driving is already safer, simply because it eliminates the variance. Best driver is better than self driving, but average - not.
> Has anyone seen an analysis of driving safety statistics if you avoid:
What exactly is your algorithm to avoid that stuff? You realize that there are other people on the roads, right? This is like arguing we don't need to bother buying defibrilators because 70% of heart disease is preventable with diet.
I'm still waiting for one that advertises that it can drive a rural road in the Midwest after sunset in the winter with poor cell reception as it goes through valleys.
All of the demonstrations I've seen so far have been restricted to well mapped city roads where there's no snow or climate originated navigational hazards worse than fog.
While getting it working for Seattle to San Diego, Phoenix, Texas and maybe New York City... I want to see it in December in Minneapolis, Chicago, and Denver.
The existing 80% solution of the major metro areas in favorable climates may be good, but the flip side of Pareto is that the remaining 20% of time takes up the next 80% of the allocated budget.
I've actually driven FSD Beta on rural roads (though not at night, since they are generally unsafe at night IMO) and it performs well, with a few hiccups. It dodges cars, slows as it approaches hill crests, and halts when it can't see terrain.
The commute home for me in December is "at night". Sunset on December 1st will be at 4:25 pm. For that matter, I'd also be leaving for work before sunrise.
Driverless cars need to be able to operate at all hours, in all seasons, and in all climates and conditions.
If it is considered to be "not safe to operate at night" then the article's title of "Driverless cars may already be safer than human drivers" is working out of cherrypicked conditions in cherrypicked locales.
San Francisco and Phoenix (the places cited in the article) are not representative of the conditions for a significant part of the United States.
Would you rather take a human taxi from Minneapolis airport to Lake City at 6pm on December 1st or a driverless one?
SF is a very complex city with lots of narrow and weird streets and random obstacles. Phoenix is more representative of most newer, griddy, car friendly cities. Real rural stuff isn't that far away. The one thing they're missing is snow and ice, but I don't know, you can't do everything at once.
> I'm still waiting for one that advertises that it can drive a rural road in the Midwest after sunset in the winter with poor cell reception as it goes through valleys.
First, this is a fallacy: "products have no value unless they meet the performance of their competition in every possible use case" is simple not true and never has been.
And it's wrong anyway. FSD on my Model Y does great on unlit night time roads in the mountain west. So unless you're positing something special about the "Midwest" I don't see the issue here.
And in fact it's wrong in silly ways. That's not the hard part! Following the pavement is the easy bit, and the car does it more or less perfectly. Where it has trouble is in urban traffic. It's timid[1], it makes poor long-term lane planning choices (i.e. if there's a long line waiting for a left turn it will try to "go around" in the right lane and then get stuck), it will occasionally refuse to pass construction cones if it can't model out what's supposed to happen, it sometimes will slow down to "avoid" cars that are technically on collision courses but will "obviously" stop (e.g. cars on the opposite side of the road pulling into middle turn lanes -- the Tesla hates that).
There's clearly work to do. But it's not the work you think needs to happen, and in particular it's not in areas that impact "safety" in the way you're imagining.
Frankly at this point I'm more comfortable on the roads, especially on the highways, with FSD doing the microdriving for me than I am driving myself. The contention of this article matches my experience more or less perfectly.
[1] A recent regression is that it really doesn't like oncoming motorcycles and will slow down inexplicably for them where it's just fine with the cars in the same lane.
There is a lot of value in handling the roads in the costal and southern metropolitan areas - a significant portion of the pollution lives there and its the 80% use case.
I'm waiting for a car that is able to handle https://www.kttc.com/2022/11/29/current-minnesota-iowa-road-... without having cell phone connectivity in night time conditions because that more closely resembles my evening commute for a season of the year.
Driverless cars may be good for commuting in cities... and highway miles, and fair weather trips (I'd be curious to see someone go driverless all the way to a ski resort in the winter when you need chains on the tires to go on I-80).
It may be good for all the driving needs for much of the population. But it's not categorically shown to be safer than human drivers (which is what the article is about) outside of the carefully crafted and restricted areas.
I can't take a driverless car to go see my parents for Christmas (rural road, midwest, no cell phone coverage for 25% of the route)... or for them to come visit me - and that's where the driverless car would have a great deal of value if it was safer than me driving there.
> I'm waiting for a car that is able to handle https://www.kttc.com/2022/11/29/current-minnesota-iowa-road-... without having cell phone connectivity in night time conditions because that more closely resembles my evening commute for a season of the year.
If you're going to be in the car, why? I mean, what if someone offered you a car that would drive you around except in comically cherry picked conditions of unplowed roads during active snowstorms? Drive yourself if there's snow on the ground[1].
This just seems like a specious and silly argument about vehicle autonomy to me.
[1] Or, in fact, don't. Because that shit's dangerous and you know it, no matter how much you protest that the upper midwest is cold. Stay off the roads in snowstorms and let the emergency crews clear the roads, it's their job.
It's not the utility that I have issue with. It's the claim "driverless cars may already be safer than human drivers" because it leaves out the criteria.
I'll accept "driverless cars may already be safer than human drivers when driving in urban conditions in San Francisco or Phoenix."
No problem.
The unqualified statement however - I haven't seen anything saying (or even attempting to claim) that it is safer than humans when driving in Minneapolis at 5pm on December 1st, 2023.
I'm still waiting for the unqualified statement to be attempted to even see if it holds up.
> I'm still waiting for one that advertises that it can drive a rural road in the Midwest after sunset in the winter with poor cell reception as it goes through valleys.
When the automobile was first introduced it couldn't drive far from gas stations, it needed oil changes at garages that didn't exist yet and you couldn't drive to a far away town that didn't have a gas station. It got better.
I could tell that same story about laptops, or airplanes or boats or washing machines or even the electrical and water supply to your house.
All products get better over time, but even now there are hundreds of millions of people that live in places that simply don't need all the stuff you mentioned, so for them a self-driving vehicle is already extremely useful.
It's even possible the product will be "finished" and your use case will never be satisfied, because it isn't deemed important enough to include. For example, I recently drove a vehicle 1,050 miles on the world's most remote road [1]. 99.99999% of all vehicles sold in the entire world can't make that drive because they can't carry enough fuel. Just because I wanted to do it with a vehicle, that doesn't mean auto manufactures are going to suddenly add duel fuel tanks to all vehicles.
(FWIW, the Land Cruiser 70 series comes from the factory with dual 90 liter diesel tanks, and can drive the road end to end no problem directly form the factory. So ONE manufacturer decided it was worth satisfying that hyper specific use case, but obviously the others do not)
Yes, you have a use case that the product doesn't do yet.
self driving cars are only good for city built around cars.
I was in Marseille - France - last week and a driverless car would not drive more than 100 yards there before being stuck
I do believe the same but don't think the comparison has any meaning. Any car in the traffic has to second-guess how other cars would behave, no matter whether the car has a human driver or not. If the overall traffic is actually overspeeding and trying not to be detected (can often happen when the speed limit is substantually lower), it is indeed safer to overspeed as well even though it would be illegal. Human drivers can---and will gladly---take that risk; driverless cars have less freedom to do so.
If driverless cars have learned so much that it can simulate the human behavior, you can't really expect that they would be much safer. So driverless cars have to behave differently from other cars for the safety reason, and that behavior would affect human drivers. Humans will eventually adapt, but not until a substantial portion of road traffic is driverless, and meanwhile human drivers will be more vulnerable. I think roads with no human drivers will be eventually needed for safety much earlier than the driverless driving becomes a norm.
Driverless cars will NEVER ever be safer than human drivers in the big picture. But it doesn't matter because safety is just the sales pitch, the real goal is of course control.
I live in Austin TX. Cruise cars circle our neighborhood ever night in fleets at low speeds, repeating the same basic traffic patterns over and over and over. Statistics about distance and time traveled are flawed, because the vast majority of autonomous driving time is spent in essentially closed-course conditions.
I read the article, and it seemed like the driverless car companies were reporting every accident no matter how minor (as they should). I don't know what it's like in the USA but where I live in Australia I know people don't report accidents that need only low costs repairs. It's just not worth the excess insurance payment + the increased insurance premiums.
I have no idea what that ratio of unreported accidents are, but my wild guess is it would be over 50%. I can't remember when my wife last had an insurance claim, but the number of chunks has in the side walls of her tyres (and the gouges on the rims) tell their own story. When I was younger, I recall boy racing and hitting some gravel. The car ended up in the bush but miraculously avoided hitting anything too solid.
If indeed the driver less cars are reporting every incident and they are using insurance claims or worse police reports to measure human diving performance, it ain't a fair comparison.
66 comments
[ 6.2 ms ] story [ 125 ms ] threadI don't understand how you can believe that if you watch a few videos of self-driving cars navigating complex situations. Here's a good one: https://twitter.com/kvogt/status/1641123102858919953
They can't do this everywhere or under all conditions, and they still make plenty of mistakes. Still, if a Cruise car can navigate through a crowd of drunks who walking in the street, they're certainly capable of managing complex situations.
It goes around the car in the oncoming lane and forces an oncoming car to stop
still impressive how the car handled the overall situation. I don't think I would have done much better. (except for maybe that single situation you pointed out)
Humans have a lot of problems dealing with those states! It's why they're so difficult; funny enough a big reason why car insurance is a must but health insurance isn't. It's also why I believe we should aggressively push for driverless ecosystems, as fewer humans actually making decisions on the road = fewer potential problems due to the perplex human states.
Of course we are still a long ways from that goal, but it's hard to imagine a future where human drivers are still a thing. Makes more sense for the transportation network to be exclusively driverless and all following the same exact ruleset precisely. Again, thinking long term here not the next 20 years.
But probably deadlier. Avoiding highways keeps crash severity down quite a bit.
reason being you can’t really get t-boned on the driver side on a highway
https://kfgo.com/2023/08/29/one-injured-in-cattle-semipick-u... Just two days ago.
You're describing an older state road or US Highway that has intersections, which has all the risks that come with intersections amplified by higher speed limits.
I quoted a news article that specifically described a highway where a vehicle was t-boned on the driver side and exactly how it occurred. You can add additional parameters such as "new highways" to narrow the original claim and refute my argument, I really can't defend against that. I really shouldn't need to.
on a highway with no physical divider a car executing a poor overtaking procedure can pull in sharply after passing and violently "pit manoeuvre" the vehicle just ahead of it, causing that vehicle to spin 270 degrees and present the drivers side to oncoming traffic.
Rare physics aside, a highway t-bone by an oncoming truck on a turning car is enough to cause fatalities and destroy a car regardless of which side presents.
That said, in Australia country roads are far more dangerous than city streets:
which is likely also true in the US.https://www.iihs.org/topics/fatality-statistics/detail/urban...
There are absolutely (in raw numbers) more urban automotive fatalities than rural in Australia but proportionally more (per 100,000 drivers) rural deaths from driving.
ie. Country driving is riskier.
Interestingly your link appears (to my eye) to be confusing - the first graph is described as proportional but it's not "proportional" in the sense of being "normalised", it's proportional in the sense of showing a percent division of the total absolute number; as more and more people live in urban areas, more and more people die in urban areas (thus obfuscating the relative dangers).
The second graph gets closer, showing the rural US areas still 'winning' the death stakes in terms of risk per miles travelled.
https://highways.dot.gov/research/research-programs/safety/i...
It's been 10 years, and the real driverless car scene has been unchanged in that time. The Silicon Valley Season 1 jokes about them have yet to age out out of relevance.
Until then we're extrapolating safety by comparing the car equivalent of train tracks to, well, people driving cars.
1. https://www.bostonglobe.com/2023/08/29/business/san-francisc...
Until self driving cars have been deployed internationally in a huge range of driving environments I'll take my chances behind the wheel.
You raise a good point but surely it's at least possible to compare the safety records of driverless cars vs. human operated cars within the domain that driverless cars typically operate. One way to do that would be to only examine all trips, human and driverless, that occur entirely within a bounded area that driverless cars are known to operate. Another option would be to look at the total safety record of a city like SF in the years prior to widespread driverless car usage and in the years after common adoption and see if there is a broad reduction in accidents generally.
Apropos of this, the other day I was thinking about the adoption of mobile devices. Originally the unadorned word "phone" referred to a wired home/office telephone, and if you needed to refer to a mobile device, you'd say, call me on my mobile/cellphone. Over the course of a generation it's now that if you say "phone," the assumption is that you are talking about a mobile device, and the exceptions become either landline or office phone. It makes me wonder how long, if ever, it will be until "car" almost universally refers to driverless vehicles, and you'll need to say "manual car," "meatmobile," or something similar to those to refer to a human-driven vehicle.
I doubt they're better than the best humans today, and that's OK. The cars are getting better with time. My uncle isn't. We could improve road safety right now by swapping his keys for a Waymo.
* Drunk driving
* Drugged driving
* Driving very late at night
* Sleep deprived driving
* Driving with a vehicle with known safety issues or no safety inspection
* Driving without a license
* Driving while young, dumb and reckless
* Drag racing or other similar behaviour
* Motorbiking
* Driving while holding a cellphone
In other words, if you're not doing stupid things, what are your odds of death, an accident etc.
Because it would actually an important question for an individual choosing whether to use a self driving car rather than a manual car. You want the car to be safer than you, rather than safer than the average.
Which itself can be drilled down to drunk motorbiking, first timers on 200hp bikes, idiots without helmets etc. An average motorcyclist dies 14x as often per km driven as car drivers, however it seems to just attract idiots.
That said, if drunk/distracted/too old drivers switch to self-driving cars that could be a positive change for everyone earlier than general population.
> * Driving very late at night
> * Sleep deprived driving
The above look to me like things of life that you would have preferred to avoid, but you'll end up doing anyway.
Driving a kid/spouse to an hospital in the middle of the night while you're also under your flu medications is a thing that happens. You could call a doctor/cab/ambulance but more often than not it's either not an affordable option, or you're too tired to have thought about it.
> * young, dumb and reckless
Could be a guaranteed Nobel price if you find the cure to this.
It’s kind of like the whole rust vs C++ debate. Rust proponents say that Rust as a language just eliminates a whole class of errors(memory errors). They’re just nonexistent anymore. Sure there are new problems, but they’re hopefully less commonplace. Your argument would be like a C++ programmer saying, “can we see some statistics on how many security vulnerabilities show up in code written by a modern C++ developer using smart pointers and RAII? That way they can make more informed decisions over whether rust is necessary.”
I think it would be interesting to see some alternative statistics, but I’m looking forward to the day nobody gets killed by an irresponsible drunk driver because cars don’t allow people to drive at all. Plus, once all the humans are off the road, I presume new technology can be made to reduce traffic patterns and other cool things.
AIs currently beat "average" humans at just about every task, if done by itself. Multimodal models will do more, but they're for next year. But reading, writing, translating, talking, listening, controlling robots, designing proteins, modeling chemical reactions, driving, ... so currently it is fancy to compare AIs to absolute top human performance, while ignoring the problems even top performing humans have (such as needing to sleep, not knowing every language, cost, ...)
It's quite possible that replacing all drivers with self driving cars would both:
* Improve road safety dramatically
* Worsen road safety for the people not in the accident-prone minority
I don't this is a frivolous question. My post got a lot of replies, but no one came forward with any data on the distribution of accidents.
An insurer might have good insight here.
I don’t think anyone came forward with any data in the replies because this distribution is impossible to come up with. Like another commenter pointed out, we all account for those distracted drivers, or tired drivers, or “accident prone” drivers at some point in our lives (possibly many points in our lives). Because of this it’s impossible to remove the 20% of the poorest drivers and have a dramatic increase in road safety.
The only way to fix this system would be to remove all the factors that you listed completely. The easiest way to remove those would be to stop having people drive cars. We won’t know if it’s safer or not until we get there, and it will still have its own 80:20 distributions, just based on different factors (and hopefully to a smaller degree).
[0]: https://medium.com/pinstriped/when-the-pareto-principle-goes...
I personally have never driven drunk, extremely sleep deprived, holding a cellphone, or any of the rest. Apart from driving while young.
It's actually pretty easy to avoid the list. I guess to clarify my question you could ask: do self driving cars have the kinds of accidents people avoiding that list would never have?
> * Sleep deprived driving
* Driving after having an emotional argument or any other serious emotional stress
Those are all things that we're all probably going to do at the low points in our lives.
We conveniently edit those out of our perception of ourselves, though, because we had excuses for them (possibly quite good ones, but in reality those all count towards how good or bad of a driver you are).
> You want the car to be safer than you, rather than safer than the average.
You probably aren't as good of a driver as you think you are.
And half of you out there are just chronic tailgaters who incorrectly believe that you have to do that in order for traffic to flow faster, or just believe that people "cutting you off" will significantly increase the time it takes to get to your destination. If you don't keep about a semi-trailers worth of distance between you and the car in front of you, then an AI probably would be a safer driver.
[ Or just the passive-aggressive shithead that wasn't going to slow down for me in the crosswalk the other night ]
There's a real "No True Scotsman" problem here. Or like the legal fiction of the "Rational Person". We have the "Average Safe Driver".
After you've removed all these drivers:
Is anyone actually left?But it's not entirely an apples to apples comparison. People who want to drive recklessly will not switch on self driving. Self driving cars are pretty new simply because it's a new technology, how will they compare against a banged up 30 yo car when they are themselves 30 yo? They may in fact fare worse, because there is more reliance on fragile components like cameras or lidar. How will they deal with the things like dodgy brakes or the one gear that doesn't work? These are things humans can mostly adapt to.
The other point, not raised by GP, is that often self driving only works on fairly safe roads anyway, whereas humans drive everywhere, skewing the stats.
But I guess broadly I agree, it is possible that on average, self driving is already safer, simply because it eliminates the variance. Best driver is better than self driving, but average - not.
What exactly is your algorithm to avoid that stuff? You realize that there are other people on the roads, right? This is like arguing we don't need to bother buying defibrilators because 70% of heart disease is preventable with diet.
All of the demonstrations I've seen so far have been restricted to well mapped city roads where there's no snow or climate originated navigational hazards worse than fog.
While getting it working for Seattle to San Diego, Phoenix, Texas and maybe New York City... I want to see it in December in Minneapolis, Chicago, and Denver.
The existing 80% solution of the major metro areas in favorable climates may be good, but the flip side of Pareto is that the remaining 20% of time takes up the next 80% of the allocated budget.
Driverless cars need to be able to operate at all hours, in all seasons, and in all climates and conditions.
If it is considered to be "not safe to operate at night" then the article's title of "Driverless cars may already be safer than human drivers" is working out of cherrypicked conditions in cherrypicked locales.
San Francisco and Phoenix (the places cited in the article) are not representative of the conditions for a significant part of the United States.
Would you rather take a human taxi from Minneapolis airport to Lake City at 6pm on December 1st or a driverless one?
You don't need to begin by solving the universe.
First, this is a fallacy: "products have no value unless they meet the performance of their competition in every possible use case" is simple not true and never has been.
And it's wrong anyway. FSD on my Model Y does great on unlit night time roads in the mountain west. So unless you're positing something special about the "Midwest" I don't see the issue here.
And in fact it's wrong in silly ways. That's not the hard part! Following the pavement is the easy bit, and the car does it more or less perfectly. Where it has trouble is in urban traffic. It's timid[1], it makes poor long-term lane planning choices (i.e. if there's a long line waiting for a left turn it will try to "go around" in the right lane and then get stuck), it will occasionally refuse to pass construction cones if it can't model out what's supposed to happen, it sometimes will slow down to "avoid" cars that are technically on collision courses but will "obviously" stop (e.g. cars on the opposite side of the road pulling into middle turn lanes -- the Tesla hates that).
There's clearly work to do. But it's not the work you think needs to happen, and in particular it's not in areas that impact "safety" in the way you're imagining.
Frankly at this point I'm more comfortable on the roads, especially on the highways, with FSD doing the microdriving for me than I am driving myself. The contention of this article matches my experience more or less perfectly.
[1] A recent regression is that it really doesn't like oncoming motorcycles and will slow down inexplicably for them where it's just fine with the cars in the same lane.
I'm waiting for a car that is able to handle https://www.kttc.com/2022/11/29/current-minnesota-iowa-road-... without having cell phone connectivity in night time conditions because that more closely resembles my evening commute for a season of the year.
Driverless cars may be good for commuting in cities... and highway miles, and fair weather trips (I'd be curious to see someone go driverless all the way to a ski resort in the winter when you need chains on the tires to go on I-80).
It may be good for all the driving needs for much of the population. But it's not categorically shown to be safer than human drivers (which is what the article is about) outside of the carefully crafted and restricted areas.
I can't take a driverless car to go see my parents for Christmas (rural road, midwest, no cell phone coverage for 25% of the route)... or for them to come visit me - and that's where the driverless car would have a great deal of value if it was safer than me driving there.
If you're going to be in the car, why? I mean, what if someone offered you a car that would drive you around except in comically cherry picked conditions of unplowed roads during active snowstorms? Drive yourself if there's snow on the ground[1].
This just seems like a specious and silly argument about vehicle autonomy to me.
[1] Or, in fact, don't. Because that shit's dangerous and you know it, no matter how much you protest that the upper midwest is cold. Stay off the roads in snowstorms and let the emergency crews clear the roads, it's their job.
I'll accept "driverless cars may already be safer than human drivers when driving in urban conditions in San Francisco or Phoenix."
No problem.
The unqualified statement however - I haven't seen anything saying (or even attempting to claim) that it is safer than humans when driving in Minneapolis at 5pm on December 1st, 2023.
I'm still waiting for the unqualified statement to be attempted to even see if it holds up.
When the automobile was first introduced it couldn't drive far from gas stations, it needed oil changes at garages that didn't exist yet and you couldn't drive to a far away town that didn't have a gas station. It got better.
I could tell that same story about laptops, or airplanes or boats or washing machines or even the electrical and water supply to your house.
All products get better over time, but even now there are hundreds of millions of people that live in places that simply don't need all the stuff you mentioned, so for them a self-driving vehicle is already extremely useful.
It's even possible the product will be "finished" and your use case will never be satisfied, because it isn't deemed important enough to include. For example, I recently drove a vehicle 1,050 miles on the world's most remote road [1]. 99.99999% of all vehicles sold in the entire world can't make that drive because they can't carry enough fuel. Just because I wanted to do it with a vehicle, that doesn't mean auto manufactures are going to suddenly add duel fuel tanks to all vehicles.
(FWIW, the Land Cruiser 70 series comes from the factory with dual 90 liter diesel tanks, and can drive the road end to end no problem directly form the factory. So ONE manufacturer decided it was worth satisfying that hyper specific use case, but obviously the others do not)
Yes, you have a use case that the product doesn't do yet.
Yes, it likely will in the future.
Yes, you will have to wait.
[1] https://www.youtube.com/watch?v=7UR6agkRbT4
> it couldn't drive far from gas stations
You could buy gasoline at a pharmacist.
> it needed oil changes at garages that didn't exist yet
You were expected to to it yourself. This 1909 manual describes how for a Model T: https://archive.org/details/ford-price-list-model-t-1909/pag...
I suspect local bike mechanics and blacksmiths could also do this work.
Given your recent drive, you might like to read about the first drive across the US, in 1903 - https://en.wikipedia.org/wiki/Horatio_Nelson_Jackson .
[1] https://youtube.com/watch?v=d6tgmGqXysM
If driverless cars have learned so much that it can simulate the human behavior, you can't really expect that they would be much safer. So driverless cars have to behave differently from other cars for the safety reason, and that behavior would affect human drivers. Humans will eventually adapt, but not until a substantial portion of road traffic is driverless, and meanwhile human drivers will be more vulnerable. I think roads with no human drivers will be eventually needed for safety much earlier than the driverless driving becomes a norm.
This is my anecdotal take.
Arthur: The Ravenous Bugblatter Beast of Traal? Is it safe?
Ford: Oh Yes! It’s perfectly safe - it’s just us who are in trouble.
- Hitchhiker's Guide to the Galaxy, Fit the Sixth
I have no idea what that ratio of unreported accidents are, but my wild guess is it would be over 50%. I can't remember when my wife last had an insurance claim, but the number of chunks has in the side walls of her tyres (and the gouges on the rims) tell their own story. When I was younger, I recall boy racing and hitting some gravel. The car ended up in the bush but miraculously avoided hitting anything too solid.
If indeed the driver less cars are reporting every incident and they are using insurance claims or worse police reports to measure human diving performance, it ain't a fair comparison.