Except that it's a statistic that's abused to hell and then some. For instance, frequently companies use total number of deaths per mile driven, without any consideration of what kind of miles they were, and you know, that's kind of important. Like, driving at night in certain places you are far more likely to be killed by a drunk driver, but that risk is simply extremely low when doing a low speed heavy traffic commute in the city. If your entire commute is being stuck in 10mph max traffic, how exactly can anyone argue that your driving risks include being hit by a drunk driver at 2am or a 120mph motorway crash? It's like using the statistics for "animal caused deaths" which include venomous snakes and stingrays to estimate the risk of playing with a hamster. Yes, technically it's also an animal but you're using the wrong stat for your argument.
And Tesla has in the past lumped together all vehicle deaths for marketing purposes - so, motorcycle deaths. Hopefully I don't need to explain how that one is misleading and possibly fraudulent.
Like the article said - in these kinds of scenarios that Tesla is showing everyone, you are at an extremely low risk. Or at least, if the system is still asking you to keep full attention, then really, you've gained absolutely nothing. A modern car with front based collision avoidance radar(so all cars sold for a few years now) will avoid any collision as well as a Tesla would, without using teraflops of computing power to navigate extremely mundane driving situations.
Again, Tesla tech might be saving lives elsewhere, I'm prepared to accept that argument. But in the scenarios they are showing? Nope, don't believe that in the slightest.
https://www.forbes.com/sites/bradtempleton/2020/10/28/new-te...
“If we reduce the estimate ratio of accidents/mile on freeway and non-free to just 2.3:1 (as a number from another source suggests) the chart shows Autopilot slightly below non-Autopilot until Q1 2020, then being around 15-20% better in 2020”
So, he’s also estimating it may have been slightly worse before 2020 and significantly better in 2020, yet he’s happy to post the 0.9 number based on what exactly?
I think that is what self-driving should be compared to, not the average person.
It's "applying for a job" as a professional driver, not as a simulated drunk, texting, shaving, eating, feet-on-dashboard average American driver.
Not too high a bar.
How can we ensure that the increase in the amount of driving (riding) that drunks do, doesn't more than cancel out the reduction in the rate of accidents?
What if, the self-driving being available to non-drunks, they use it because it's fun and futuristic, and that increases the amount of accidents even more?
We are well past that point. At even just 0.08 BAC a driver is 7 times as likely to cause a fatal accident and it gets dramatically worse as BAC increases. Self driving are close to the norm right now, even if they where say 20% more dangerous than an average alert driver it would take a massive increase in miles driven by drunk people for things to end up worse.
Anyway, the penalties are currently the same with and without self driving, hell sleeping in your car while drunk is illegal. As long as the laws stay the same things should stay about the same while saving 10’s of thousands of lives.
But estimates of the proportion of trips under the influence of alcohol that I was able to find were around 0.03%. (111 million out of 411 billion)
So I don't see a logical reason to rule out their driving increasing by ten times, or 100 times, or more.
And I don't see what the justification would be for assuming a self-driving system is within 20% of an arbitrary standard. Because drunk drivers are so bad, as you point out, there's a lot of space between "as bad as a drunk" and "as good as an average non drunk".
I think it's completely plausible that a system could be ten times or a hundred times worse than an average human and it not be obvious. That's because an average human has one fatal accident in approximately 100 lifetimes (in the US). Or, for another reference point, drivers had fatalities about 7 times more often per mile, in 1950 than now.
Without saying I know exactly what numbers to use, I'm pretty certain that betting on things matching to within 20% is unjustified without, well, a specific logical justification.
However, regarding penalties being the same, this seems like a lack of incentives that would mean neither lives saved nor lost.
Also, I don't think saving tens of thousands of lives per year is a possibility if we're talking about alcohol in the US. I believe there are only on the order of 10,000 fatalities per year.
I note that millions of people (again, in the US) seem to think 10,000 covid-19 fatalities in only a few weeks is not worth even wearing a mask to stop.
Some of the things I read while writing this:
https://en.wikipedia.org/wiki/Motor_vehicle_fatality_rate_in...
https://www.cdc.gov/transportationsafety/impaired_driving/im...
People don’t drink that much to really push the numbers all that high. 100x as much would mean people spent 3% of there time drunk while making trips. That would take drastic changes in behavior as most people generally either don’t drink or drink at home. Considering the average drunk has a much higher than .08 BAC I just don’t see this as a significant possibility.
Also, 20% was used as twice as bad as the pessimistic article about Tesla’s Autopilot posted in this comment thread.
The thing is if self driving systems where drastically worse we would have significant real world evidence of it. Basically the better the system the more data you need to collect. If they killed someone every 1,000,000 miles well there would be piles of bodies from these systems. Arguably only Tesla’s system has gotten anywhere near enough data to start doing some comparison and the error bars are still quite high. I have no problem saying near human performance on average but specific numbers depend on various assumptions. Some estimates say better, others worse that’s close enough to say it’s much better than a drunk driving.
Take Google’s completely autonomous system, it’s only at 20 million miles. So sure it’s got zero fatalities but a group of humans could easily clock 10 times that without any families.
If they killed someone every 10,000,000 miles, they would still be "drastically worse" than the average driver - virtually an order of magnitude.
But these are made up numbers. What data is there to work with? It seems like disengagement numbers from testing are on the order of 10-20K miles.
This article says that "everyone hates" these statistics and people claim they are easy to game:
https://www.theverge.com/2020/2/26/21142685/california-dmv-s...
But that doesn't matter, in my opinion, for our discussion, because even if they are gamed to the hilt, they provide a relevant upper bound of what is possible for self-driving in the best possible environments.
I don't know if you can see it my way, but I think switching to a safety driver, aka disengagements, are what we need to be comparing to fatal accidents, if the context is transporting people incapable of monitoring, let alone driving.
I am willing to estimate the worst habitual drunk drivers may well cause a fatal accident roughly every 100K miles. That's really bad!
But I seriously think that it would be foolhardy to plan on self-driving vehicles helping, when they punt about ten times as often as the worst drunks.
Sure, there is some uncertainty, not every disengagement has to be a fatality, but as a matter of public policy, nobody should be betting thousands of lives on that just because they would like it to be true.
Comparing fatalities of a self driving system combined with a safety driver, with fatalities of drunk drivers alone, is just not valid.
Unless you evaluate the drunks with a safety driver at dual controls as well.
Something I can't judge, and you will probably have an opinion on, is how much "low hanging fruit" remains with current approaches to self-driving. If we were to agree that it needs to be ten or a hundred times better to be better than the worst drunks, can that be done without a radical revolution in the technology, through incremental refinement? I'm skeptical, I think we probably are at least past the proverbial easy 80%, but don't have specific reasoning for that.
That’s no longer the case. Waymo is over 30k miles per disengagement, however it’s also a meaningless number.
The point of disengagement’s isn’t a proxy for accidents it’s to avoid cars simply shutting down in traffic and make it really clear the car is confused. Waymo is operating driverless taxis with as I said over 20 million miles often with an empty car and they’re using a call center to handle edge cases. People remotely pilot the car at very low speeds when needed, but the car must come to a safe stop on it’s own.
Having someone handle an edge case every 30k miles hours for is cheap enough to be bundled into a new car’s price. Worse case they pull off the road and call a tow truck, but more normally they can get the car to a place or can safely drive. Put a drunk driver in a car with that behavior and you don’t get accidents every 30k miles.
Saying something is meaningless doesn't make it so. Often things don't mean what people would like them to mean, but every observation means something.
If Waymo is operating in particularly easy environments, then it's plausible that it's not as good as a 30K mile interval would imply.
But that number definitely implies an upper bound, because if they could go longer, they would.
Do you have a source for there being no safety driver physically in the car and the car always finding a safe state on its own?
There’s been a fair amount of coverage for example:
https://www.wbur.org/hereandnow/2021/01/04/waymo-driverless-...
The safe state thing is inferred based on the list of crashes so far. Clearly it’s not going to always work perfectly to the end of time, but it’s good enough that their testing it without a safety driver in the car which says quite a bit.