I've observed all the things you did (in fact, all of them on Sunday). But I don't do any of them, and you probably wouldn't either if you ever used a car.
I'm comparing self-driving cars to an average Human, who truly does drive acceptably. Over 99% of people drive acceptably. The 2 dudes on I-5 without headlights at dusk are a problem, but there are Hundreds of vehicles in sight driving very safely.
Yeah we agree?
The point is human drivers aren't a proportions problem. It's that any unsafe driver is bad.
Adding more safe drivers doesn't lower the risk of unsafe drivers, even if they're a smaller proportion of drivers.
Of course the US has dug itself in a hole where many places require being able to drive to get around, putting unsafe drivers in a bit of a bind...
Stopping dead on a road with nobody in front of you because your AI detected a hazard might be objectively safe and conservative driving, but it is unpredictable and a driverless car that behaves this way is going to cause problems.
Sure, but less problems than if it just kept on going and hit something. While rare regular cars have unpredictable things happen and many of them are a lot worse than a car coming to a controlled stop on the road. As a driver you need to be able to handle that. When there is a catastrophic failure of the vehicle the other vehicles need to safely react.
Big is different, and scaling this kind of thing to majority of vehicles on the road can cause all sorts of interesting problems.
I’m not saying it’s inevitable; sure, we can and most likely will improve things on the long road to majority AV. This simply occurred to me and I hadn’t seen it articulated before so thought I’d share.
That said, I don't know if the average human drives acceptably, and I'm very skeptical of a "99%" estimate, but I will say that I would prefer a fleet that consistently drove at the 40%ile of human quality vs. the current human fleet that follows the normal distribution.
If self-driving cars reach 40%ile quality, and enable us to stop the 20% worst drivers from driving (or average drivers when operating at 20%ile due to being drunk or tired), that'd be a huge win for safety.
Right now it's socially unacceptable to increase the difficulty of getting your license, or to take it away for poor performance, because cars are a necessity or near-necessity in most places. If we could provide self-driving vehicles as a fallback, perhaps we could make it easier to permanently lose your license than—checks local laws—repeatedly killing people by driving while intoxicated[0].
[0] https://www.idrivesafely.com/dmv/indiana/laws/dui-and-dwi-la...
But I also recognize that many people, having the suspended license, must now choose between never leaving home or driving with a suspended license.
I feel like better public transit, more than Self-driving cars (which will be expensive to actually own) is the answer here.
We need to lower the cost to an individual of losing driving privilege, so that we can lower the cost to society from bad drivers.
Autonomous vehicles don't get drunk or sleepy, don't speed, don't leave the scene of an accident...
The bar for AI should be way higher than just "outperform mediocre humans".
https://arstechnica.com/cars/2023/08/how-cities-can-stem-the...
Requiring cellphones to disable themselves in cars might be even more effective.
That means the net danger of traffic as a whole goes down.
Setting an unreasonably high bar - like "drive with perfect record" - will simply prevent that happening, and we're stuck with humans. Perfection is a north star, not an attainable goal. As long as we get net positive, the world is a tiny bit better. And then you make it a tiny bit better again. And again.
Why do you think the bar for AI should be higher than "better than what we have"?
Right now, any incident involving a autonomous car is going to immediately be blamed on the autonomous car and not any other party, even if the fault lies entirely on the dickhead running the red light.
It won't be enough for autonomous cars to merely not cause crashes, people will expect them to actively avoid them, even in cases where a human would have failed.
Is it? Seems like heavy enforcement could pay for itself in terms of fines.
But the better question is why would you need to have overpaid humans drive around in cars to catch speeding violations? In most of Europe you almost never get stopped by cops for traffic violations. Most speeding violations are caught by stationary radar traps, a machine can run the plates and send you the ticket by mail, including a pretty picture of the driver of the car if you want to dispute who drove it. Sometimes there are similar traps activated by passing a red light. It'll even tell you how the long the light had been red when you passed it.
In the case of headlights, I disagree. Many other countries have laws that require always on headlights, or at least daytime running lights (DRLs) in some cases.
For example, I work in the auto industry and the cars that we distribute in America have different light programming than the cars we send next door to Canada. It's stupid. We should at least require DRLs in America and then we'd be on par with countries like Poland, Russia, Serbia, Slovakia, Slovenia, etc. who recognize it's safer when every moving vehicle is lit up. There's no day-to-day enforcement needed, the car turns on the lights automatically every time the vehicle is started.
Also bikes absolutely should have lights and are required by law to have them in my country (and to have them on at night, dusk or in reduced visibility conditions). E-bikes and e-scooters also have lights which are typically always-on regardless of lighting conditions.
Pedestrians don't normally have lights (although I've seen joggers wear hi-vis vests or head lamps, especially during early morning or late night runs) but pedestrians also don't move as rapidly and suddenly as vehicles. Plus in urban environments drivers are expected to look out for pedestrians, especially in narrower roads. "Child peeking out between parked cars" is a scenario covered several times during driver ed and the practical exam includes one emergency break without prior warning (tho in practice learners can make an educated guess about when it might be coming up, e.g. being led to a low-traffic street and instructed to "pick up speed a little").
Not to mention, I said "at least" DRLs because DRLs are a watered down version of what the European studies show as safer, which is headlights on all the time. In the US, the brightness of DRLs is limited, so the studies are shit to begin with because they're not even using the right variables to find the safest solution.
I find it remarkable that we haven't done so with cars as well. It seems like a win-win-win? Automakers (and service centers) get more money replacing more headlights, the public gets safer roads, and politicians get an easy win.
baseline in this case means minimum to be considered acceptable. We have been allowing humans to drive vehicles for as long as there have been vehicles.
Why would we not use that as the baseline? Do autonomous vehicles need to be absolutely perfect before they're allowed on our roads? or just be as good as humans, with the possibility of being better?
Otherwise "it's better than humans" is in fact a win.
When I had Hodgkins, I found out there is no baseline data for untreated patients. Scientific medicine was invented after the first cures for Hodgkins, so by the time anybody knew "We should measure properly what happens if you do nothing so we know whether we're improving on that outcome" it would be unethical because you're giving people a death sentence in order to do science, which, you know, we outlawed that because of the Nazis.
We can assume not treating Hodgkins is extremely bad but we don't have scientific data, because all the untreated people in a civilised country self-selected - they said "No" having been told that means they'll almost certainly die, and right soon if they say "No". That's not a normal sample, that's... crazy people maybe? Very religious people? Not normal. But we don't have actual baseline numbers.
However, although we don't have a baseline, we do run trials, they compare against what we're currently doing, called the Gold Standard, in my case ABVD. When I was treated, trials were investigating Stanford V, the fifth attempt by a team at Stanford to improve on the existing treatment regime. It was similarly effective at curing Hodgkins, but the side effects were, if anything, worse, so, no dice.
For cars that "Gold Standard" in the US is basically every adult except the very poor and the blind. If an AI can do better than the Gold Standard, even if it's not as good as you'd like, it's an improvement.
Indeed and the most individual form of transport is walking and most cities in the world are built around walking. Only the US lays its cities out to require cars. It never really laid its cities for horses and carts (with the exception of Salt Lake City.) Before the car, cities were being built around the streetcar, and before that around walking. It's not like people lived in suburbs and went around in horse and buggy lol.
I don't know where this comes from, other than the usual "US bad". Go to Johannesburg and live in a regular estate without a car.
I'm not familiar with South Africa, but Australia also copied a lot of US auto-centric code regulations, so have no doubt that they did as well. But they very much aren't the norm. Most of Asia (outside of Middle Eastern cities and even here it's a mixed bag with Oman being different than the rest) and Europe are built around pedestrians.
If you think I'm just angry at the US due to some internet lefty baggage, then I don't know what to say. Japan shows that cars and pedestrians can coexist fine when you don't cater to just the car.
Also relevant is when Anthony Levandowski at Google / Waymo caused a freeway incident through unsafe testing: https://www.businessinsider.com/anthony-levandowski-google-s...
While at the time the DMV did not require reporting, later when Levandowski was on trial Google / Waymo hid the video of the incident when it’s existence was finally reported on.
Imagine driving by a playground and, instead of a dog, you see a ball roll across the street in front of your car, but you can already intuit that it is rolling fast enough that it won't be an obstacle in your lane. A human driver can intuit from the context (playground + ball) there is a chance a child may come running after that ball, from between those same cars.
While self-driving tech may be faster at 'seeing and classifying'* objects, is it able to predict from a larger context? I haven't seen much to indicate it can. I think that's where a lot of the edge cases are going to come from. Driving is more than just 'see object and make a decision'.
*Because most of the tech is relatively opaque to the public, this performance seems to be generally taken on faith. But what the Uber accident in AZ showed is that (with that system, at that time, at least) the self-driving tech really wasn't very good at classification in real-time.
Rebuilding just a mile of a side street would cost $10 million. If it is a major road the cost goes up, and if you need to put in bridges the costs is very high. A useful change to a single intersection can costs more than a million dollars.
We have the ability to do more testing and training. However may drivers would hate the fact that they can't pass a license tests - enough to vote out anyone who tries to pass such things and get the law changed.
Long term running great public transit is the cheapest answer, but anything less than great won't move the needle, so your minimum costs are in the mulit-billion dollar range for tiny cities - cost goes up greatly from there.
This claim gets made a lot but, while I'm not sure how you would test this, I'm not really convinced that an expensive driving school and more rigorous driving test moves the needle much relative to just a lot of hours on the road and perhaps becoming more mature with age. My recollection is that my driving almost certainly approved with a year or two of driving after I got my license; you're not going to replicate that amount of driving with training/testing.
Commercial (not private) pilots would be a good example. They regularly take refresher courses to get updated on the latest laws. They have regular simulator training so they can practice situations harder than anything they will ever encounter. They have to provide proof they are in good health. (I'm not what is legally required, or what something all airlines require but isn't legally needed).
I've seen driving simulators - not video games, but actual driving simulators. They are very good, we can require everyone spend a couple hours per year in one to show they know how to drive. We can setup situations where the other car does something stupid. We can setup situations where you will crash and have to choose the least bad crash. We can setup bikes and pedestrians to ensure that you drive safe around them. We can setup bad weather, ice on the roads. We can setup mechanical car failures. These are things that rarely happen and so you need refreshers to ensure you do the right thing should they happen to you.
A rounding error?
The point is human drivers aren't a proportions problem. It's that any unsafe driver is bad.
Adding more safe drivers doesn't lower the risk of unsafe drivers, even if they're a smaller proportion of drivers.
It is not hard to imagine that self driving cars are normally better except for some situation. If we can know those situations we can also talk about hybrid where we don't allow humans to drive where they are bad, and don't allow cars to drive where they are bad. However someone with real numbers needs to talk.
Every single accident involving a self driving car in California must be reported to the DMV. In addition, miles driven and disengagements are reported. All that information is available today online (e.g. accidents [1]).
> nobody is talking
No one talks data about the biggest player Waymo because "human driver causes an accident" is not news.
In September, there were three crashes involving a Waymo car. On September 3rd and again on the 11th, a driver hit a parked Waymo car[1]. On September 30th, a car rear-ended a Waymo car stopped at a red light[1].
[1] https://www.dmv.ca.gov/portal/vehicle-industry-services/auto...
Not just the kind where they try to make a yellow that they clearly weren’t gonna make, but also they stop at a red light, wait to see if there’s cross traffic, and then intentionally run the red light.
Did the bike lanes have a "no parking in bike lane" sign? Because (and I learned this by getting the question wrong on my CA license test when moving here) it's not illegal to park in bike lanes in CA.
> That seems like a comically bad baseline.
Now consider a version of autonomous cars that would kill 20k.
The argument to switch to those cars, to save 10k lives a year, has to use 30k as a baseline.
If we don't use 30k as a baseline, then we end up saying "20k is too many, stick with the status quo", and then more people die.
Also I expect that a single autonomous taxi can replace more than one private car.
From my perspective as a person who lives in San Francisco and also bikes a LOT: human driver also do not perform acceptably.
I counted 14 cars without headlights on after sunset on a 2 mile bike ride yesterday.
I counted 5 cars parked in either a bike lane or active traffic.
2 cars took blind rights (no stops) onto Polk from various alleyways.
This was a good day, FWIW.
Cruises have obvious problems, but my hope is we don't use humans as a baseline, because they're not very good either.