Self-Driving Cars

To reference what I said earlier, when something does go wrong and it should be easy to prove that it's misuse and actually the technology is fine if used correctly...

It kind of needs to be spread by something other than:

1. Elon Musk alone
2. Random social media fans
3. People arguing on small media forums.

And the National Transportation Saftey Board isnt the transportation professionals?

Because there the ones that published the data on accidents, not Tesla.

That part you would think, but a lie spreads on social media faster then the truth, especialy when it targets acceptable targets.

See the amount of people that think electric cars start on fire better then zippo lighters.

Edited by Imca on Mar 29th 2024 at 11:51:37 PM

To clarify, it's a little of both. The NTSB and NHTSA collect accident data, but most of that reporting is voluntary; they have no way to compel it from automakers. Tesla is somewhat unique in that its active vehicle telemetry allows it to provide substantially more comprehensive data than anyone else.

In other words, the stats are probably distorted a bit by the accuracy of Tesla's reporting, not the other way around.

Tesla also publishes its own crash and fire data on its safety page, although in aggregate.

Edited by Fighteer on Mar 29th 2024 at 10:57:54 AM

The irony is that I'm pretty sure I'm the only person here who has had some first-hand experience with Tesla's FSD. And I will admit it didn't cause me any grief during the times I was driven in the car.

This is more difficult to do than it seems. The numbers that most people like to cite are the ones provided by Tesla, but they seem to count "an accident" as the airbag firing, while the industry in general typically goes off of police reports or insurance claims. Indeed, if you look at insurance numbers, Tesla actually has the highest accident rate of any major car brand.

Another problem is that comparing "Autopilot active" vs "all non-Tesla driving" is that Autopilot is something like 90%+ highway driving — which has lower accident rates (per mile) even in completely non-automated driving. So you're comparing Tesla's accident rate under the easiest driving conditions to everyone else's accident rate under all driving conditions, which is not apples to apples. There are additional statistical correlations that should be controlled for to make a proper analysis, as well, but generally are not — for example, as a more expensive vehicle, Tesla owners are more likely to be owned by people who are older and richer than average. Older and richer drivers are also among the demographic groups with the lowest accident rates. So if we're trying to determine the role of the vehicle in accident rates, we need to control for that, but this is basically never done. Here's a source that goes into the details of some of these discrepancies.

The tldr is that best guess we can make based on the available data is that Tesla accident rates are roughly equivalent or slightly higher than the rest of the industry. The most disappointing thing is that Tesla very definitely has the data to make apples-to-apples comparison (they just need to sort it the same way the rest of the industry does), but they do not make it public. The second source I posed goes into this as well.

In any case, all of this is actually besides the point I was making earlier: that by doing the Full Self Driving beta with regular Tesla owners on public roads, Tesla is privatizing benefit (the data they gather) and socializing risk (the chance of an accident). This behavior affects not just Tesla owners (who can choose for themselves whether to use the beta service after considering the risks and benefits), but everyone on the road. Even if is statically safer (which, as mentioned above, is a dubious claim), Tesla does not have the right to make that decision for the public at large. But that's exactly what they've done.

As an aside, Fighteer, you continue to be dismissive (calling my arguments a "performative emotional response" and "fearmongering") and I just want to be on the record as saying that I do not appreciate it. I'm trying to engage the topic in good faith and from a neutral perspective (in the sense that I'm willing to be convinced, not that I don't have an existing opinion). Do me the favor of at least keeping the contempt from showing in your replies.

Edited by NativeJovian on Mar 29th 2024 at 2:23:54 PM

There's a lot of criticism of Electrek's analysis of that LendingTree report in the comments, which I don't have time to dig into. It is unclear if Tesla drivers are self-reporting accidents at a higher rate, for one thing.

Anyway, since there's no data showing whether these insurance claims are related to Autopilot or FSD use, there's no way to usefully determine the value of the system. You're broadly correct that Tesla only reports crashes in which the airbags deployed, and this could skew the data.

Further, there's no breakdown of the fault in said accidents. If someone hits a Tesla, that's still an accident involving the Tesla.

The lack of accurate crash reporting is a serious issue that goes well beyond the scope of this thread.

Edited by Fighteer on Mar 29th 2024 at 2:24:33 PM

I find it frustrating that you repeat "the data is the only thing that matters here" several times, and then when I post data you say "well I don't have time for that".

Well, Tesla drivers are presumably using Autopilot and FSD, so at minimum we can say that the real-world use of Autopilot and FSD do not make Teslas safer overall. We can't say whether that's because Autopilot and FSD are extremely safe but manually-driven Teslas are extremely unsafe, manually-driven Teslas are extremely safe but Autopilot and FSD are extremely unsafe, or some point in between those two extremes. But we can definitely say that, for whatever reason, having Autopilot and FSD available objectively does not make Teslas less accident-prone.

I see no reason to believe why this wouldn't be a random distribution. Why would people be hitting Teslas at higher rates than other brands?

The data I posted is just as accurate as anything else in the industry, where is this coming from?

The problem is that the statistics presented by LendingTree don't analyze the causes of crashes or incidents, and therefore there's no way to tell if Autopilot or FSD were involved in any of them. I agree that Tesla's safety statistics may be somewhat misleading when it comes to Autopilot, since it's mainly used on highways and highways have lower crash rates in general, but there's a lack of normalization.

Subaru, which makes a lot of vehicles that have built-in ADAS (EyeSight), is third on the list of accident rates. It seems counterintuitive that the use of ADAS would lead to more crashes, but again there is no way to tell how frequently Subaru drivers use EyeSight, nor whether its use was involved in any of the reported accidents.

LendingTree's data covers insurance claims, but the NHTSA data cited by Tesla includes only crashes where a police report was filed and a vehicle was towed. LT would naturally report higher raw collision numbers since not all accidents result in a tow.

The data are simply too messy to use them to support any particular conclusions.

X thread from Ashok Elluswamy, AI director at Tesla. It is in response to a video clip from an FSD 12.3 user whose vehicle perfectly navigated offroad through a marked construction zone. (The original poster of the video noted that FSD made that maneuver accurately without a lead vehicle earlier, off-camera.)

Elluswamy is claiming that FSD is capable of contextual intelligence: in other words, real-world problem solving rather than following rote instructions. It seems to be getting really good. Still looking for counterexamples or cases of it screwing up, but not many of those are making it to my feed.

Incidentally, it is officially no longer in beta. FSD is now called "Supervised" in the marketing, a label that is repeated multiple times throughout the instructions and user acknowledgments.

Edited by Fighteer on Apr 3rd 2024 at 3:37:27 PM

Full Self-Driving (Supervised) is very much a better name, I’d rather the “Supervised” went first but it provides much clearer context as to the use case.

It doesn’t seem beside the point to me at all. If the question is “does Tesla’s FSD beta make people less safe”, then whether FSD cars are more likely to get in accidents than other cars is precisely the point. And even from your analysis of the stats that adjusts for FSD generally being used in lower-risk situations than the average driver, it doesn’t sound like there’s a clear, demonstrable difference:

So if that’s the case - if they’re basically the same as the rest of cars on the road - how are they making you less safe?

You could make a much stronger statistical case than drivers in their teens and early 20s were “endangering you without your consent” by being on the roads at all, given their higher rates of accidents, and I think most people would easily recognize that as an unreasonable characterization. You could likewise say that the very existence of trucks and SUV is “endangering you without your consent”. It’s…not how consent works. A person’s choice is to drive or not, taking into account the existence and behaviour of other drivers. If a thing presents a clear and demonstrable risk (like drunk driving), then people push for laws to ban it; but even from what you’ve cited, the evidence on FSD doesn’t seem to come close to that threshold.

(All of this is separate from the question of whether Tesla should be calling something that isn’t full-self-driving “Full-Self Driving. They shouldn’t. I feel that’s easily answered but not germane to the policy question of what forms of driver assistance and self-driving should be allowed on the roads.)

Edited by Galadriel on Apr 3rd 2024 at 4:29:23 AM

"does Tesla's FSD beta make people less safe" is not the point I was making. You literally quoted me saying that it wasn't the point so that you could say "it was the point, actually". I don't know what else to say except to repeat myself that no, that is not the point I was making, even if it's the point you want to make.

You're conflating two different ideas. Teenagers may well be more statistically likely to cause an accident, but that's not the nature of my complaint. People should be allowed to use public roads because "public roads that no one is allowed to use" is a self-defeating concept. Tesla using public roads as a beta testing facility for their self-driving software is a categorically different thing. The benefit of "other people get to use roads" is "I also get to use the roads", which seems like a fair trade to me. The benefit of "Tesla gets to test their software on public roads" is "Telsa gets to develop a product that they sell for a profit".

Why should Tesla be allowed to put the public in danger (however slight) for their own benefit? No one else developing self-driving software feel the need to have untrained consumers test it on public roads. That's the point I'm making about privatizing benefit and socializing risk.

really, it should be "Supervised Self-Driving" as that's descriptive, accurate, and doesn't sound near so silly. But at least they found the parentheses to try and downplay it with.

Yeah, it does rise the question what “full” refers to.

Edited by Smeagol17 on Apr 4th 2024 at 2:56:37 PM

The name is aspirational. It describes what the product is intended to do when it is mature. The people complaining about the name will look fairly silly if/when it gets certified for L4/5 autonomy.

If.

'Aspirationally' naming widely-available consumer products, especially if the name doesn't represent what that product actually does in the here and now, is at best dumb and at worst purposely dishonest. Even dangerous, in this case—we all know full well people don't read EULAs or fine print for shit.

Edited by Chortleous on Apr 4th 2024 at 7:29:05 AM

Especially since there's literally nothing stopping Telsa from just changing the name when it's fully operational.

Like they just did.

There is no compelling reason to name a thing misleadingly, and "it's aspirational" as a defense is moronic.

It's the sort of thing that trading standards (at least in the UK and EU) may have Strong Views on, but it's also a thing many businesses do.

Is it shoddy, misleading marketing? Maybe. But branding based on the day one 'minimum viable product' doesn't age well, and companies always prefer to sell you the dream.

Integrity in marketing and branding is rarer than it should be. And when it is there, it's often because a regulator intervened.

Why would anyone look silly for pointing out they had a misleading name for years if and when it gets certification? It just vindicates that they were being overoptimistic for years of providing the service. <_>

It's not misleading if it labels itself as developmental. No reasonable person would read "Full Self-Driving Beta" and believe it's a mature, finished product. That is literally what "beta" means.

There are more reasonable grounds to go after "Autopilot" as a brand name, an argument we've had to death already and that several courts have agreed with. I still don't agree, but it's not worth fighting over.

What I care about is whether it works, not what it's called, and I refuse to apologize for that position. Maybe if it was called "Miracle Perfect Driver 9000 That Totally Works Always" there would be some valid concerns.

Edited by Fighteer on Apr 4th 2024 at 11:03:29 AM

I'd argue that calling it a beta this entire time is also wrong, because I don't think anyone could call "Self-driving system that needs to hand over to a human sometimes" a feature complete release just looking for bugs.

Are you seriously going to nitpick over the meaning of "beta" now? You can't win that one.

Well, it was either not Full Self Driving or it wasn't feature complete despite its name and was therefore an alpha release. So yes, I'm going to be nitpicky, because one way or the other it was misnamed. It is misleading to give something a name that should mean 'finished except for bug testing' when it still needs work to even do its primary role. If I use beta software, and occasionally I've had to (i.e. the whole remote development frontend/client that I use sometimes is beta software), then I expect that it does everything that it's for. Not perfectly, and more buggy than normal software, but I wouldn't expect it to suddenly give up and go "Nope, not doing anything now, back to you".