Hacker News Reader

I think every discussion like this should start with each commenter counting verbally to 3,700. That's the number of people killed on roadways worldwide every single day. [1]

And that's daily. It's only every few months that we see a story like this, and this one had no fatalities.

It's totally fair to criticize Tesla, Autopilot, and/or FSD, but these discussions always implicitly accept that the status quo is acceptable, and whether or not this technology should be allowed is debatable. The status quo is an unimaginable amount of human death and injury that we accept only because it's been normalized.

Delaying self-driving cars by one year comes at a cost of up to 1.35 million human lives [1]. That doesn't excuse failures like these, but the context should be understood.

[1] https://www.cdc.gov/injury/features/global-road-safety/index...

> Delaying self-driving cars by one year comes at a cost of up to 1.35 million human lives [1]. That doesn't excuse failures like these, but the context should be understood.

Citation needed. There is nothing so far proving that self driving cars currently have less crashes then humans in comparable conditions. And they disengage in difficult conditions when most crashes happen.

Dont just start with assumption "self driving is currently safer" and then go on blame people for not trusting it.

> Dont just start with assumption "self driving is currently safer" and then go on blame people for not trusting it.

My claim doesn't depend on that assumption. My claim depends on it being the case someday.

If the US tightens regulations on ADAS today, and that results in the day self-driving cars fully replace human-driven ones being 11 years away instead of 10, it has been delayed one year, at a cost of up to 1.35 million human lives.

Delaying self driving cars does not come at the cost of human lives, if those cars are not safer then human driver. If they are less safe with current tech, delaying then saves lives.

If they tighten regulation, they are saving lives. Also, mandating companies to put in more testing would force companies to test. Without that, they are economically motivated to make big claims, but to to not invest into reproducible quality.

The current tech is irrelevant. It is not safer than a human driver, but that's not relevant.

If regulation delays the day in the future that it is safer than human drivers and that human drivers are replaced, then it is costing lives. And it would.

The only way it wouldn't is if the regulation saves more lives in the meantime than would have been saved by having self-driving cars ubiquitous sooner, but that's tough to argue when we're already losing 3,700 per day.

The excess people killed by self driving cars matter too. Self driving is more dangerous until it is not and it is actual task of regulators to delay it until then.

And if regulators won't act, it will never be safe, because it is cheaper to produce unsafe products. It earns more to lie about safety of a product, unless you have functional regulations.

So if an automaker can make a system that is capable of handling any driving task safely but occasionally ran itself off the road on purpose for no reason, as long as it’s a net positive in lives saved it should be allowed?

I don’t agree with that assessment.

It needs to make fewer mistakes than humans at all times, not just most of the time. Great with a disaster here and there is terrible.

I see where you're going, but how do you actually justify your choices?

A: occasional random suicide (100 deaths/year), no other deaths B: no random suicide, 10k deaths per year from normal driving

You would be saying you choose B?

It's not really about the exact numbers here - just trying to get on the same page - would you really weigh your uncomfortableness with "random deaths" highly enough that you would accept increased actual deaths?

Another possibility is that the random death situation really is a signal of "actually hard-to-detect badness". If so, it should come out in statistics and we should easily be able to come to agreement in that case.

But, if the choice really is between accepting "weird mistakes+lower deaths" or "status quo+many deaths" I feel pretty uncomfortable accepting the latter knowing the real-world consequences.

This is a human comfort thing, not a reason thing. As someone else in this discussion said if computers do something that looks stupid to people, they won’t trust it.

What if the investigation here showed the Tesla was 100% correct to do what it did?

What it did is so odd I wouldn’t want to trust it. Maybe the problem is it needs to fail in a more human-like way?

If the death numbers get low enough, I think it will likely speak for itself and people will trust it. But it can’t be a 30% reduction in deaths. It needs to be like 95%.

Enough that it feels safer instead of questionable/dangerous. Even if it was actually safer the entire time.

I heard an interesting philosophical discussion many years back, where it was stated that we are morally obligated to outlaw non-self-driving cars as soon as it is proven that self-driving cars kill fewer people. It is not so easy to argue against that, because if you do, you rate your own right to choose higher than another persons right to live.

Firstly, human error in at least 7 out of the 8 cars caused this. FSD didn't take control away from the human, it encountered something outside of it's limits and returned full control to the human.

Secondly, that's what planes are. A lot of new plane safety comes from very automated, very complicated systems.

In some situations these planes fail in a way that leaves the pilot in a more dangerous plane than one with no electronics at all.

Why do we tolerate this? Because it doesn't matter how poorly something handles failure, what dominates is the base rate of the failure occuring.

If you ban these extra-automated driver aids, and people drive Teslas less, more people die. Same as if it causes a delay in developing real FSD.

I would rather live next door to an RBMK-1000, than drive a car 10 years older than my current one.

I caution you against changing automation policy based on an accident caused by human error.

So you categorize the 737MAX accidents as human errors too and thus they should not affect policy?

That's a very different situation. Namely, I don't own Boeing stock.

My thought was that the cars following needed to have adaptive cruise control engaged and perhaps then the accidents would be avoided.

> FSD didn't take control away from the human, it encountered something outside of it's limits and returned full control to the human.

We don’t know if that’s what happened.

The problem with point 1 is that it's just not good at convincing people. If you have a situation where automation fails but a human would have done better, then it doesn't matter unless the statistics are overwhelmingly in your favor, which I doubt they are in this case.

This is a discussion that comes up a lot in medical technology, and if I had to guess why this form of rhetoric fails is not just because it's easier to empathize with a human, but also because when the failure case seems "simple", the issue seems a lot more like an oversight and systemic issue. That in turn makes the probability of failure probably look much higher than it really is, while also implying further undiscovered oversights.

That also kinda addresses point 2, in that, specific failure case or not, it implies a weak system with obvious oversights. That definitely doesn't help the political case for complete FSD approval.

I'm generally skeptical of this utilitarian, rationalist form of rhetoric, if only because it's overly optimistic about the ability to overcome issues in some amorphous future. Sure, a future with full FSD is probably a net good, and we can even say it's probably in within our lifetimes, but the claim that future mitigated harms outweighs all current harms of live-testing FSD won't win enough people over, and drowns out other possible policies like advocating for adapting our infrastructure to support FSD rather than having cars attempt to read signs and signals designed for humans.

I agree. People take obvious failures as a meta-sign of "badness" and throw out the system and statistical information which we have which completely proves what we should do.

That said, at some point FSD could have been banned, it wasn't, and we are living in a world where it's likely saving lives on net. The only problem is if naive arguments to ban it are accepted despite them not actually challenging its net safety gains.

Other people have said it before, but the issue is that the Tesla crashes we are seeing are ones that a human would trivially avoid.

Most people, especially non technical people, aren’t going to care about numbers that say it’s safer or better on net when the car still occasionally does things that humans view as crazy or stupid.

I'm interested in understanding why you would want every subdivision of "type of crash" to strictly improve, rather than being content that the combined crash rate goes down?

i.e. if Tesla causes 100 new crashes of type A but saves 200 of type B, isn't that a gain? Even if type A crashes are one which humans don't normally make.

A way in which I agree with you is if type B crashes are more avoidable by conscientious drivers, and you feel that your personal base rate of B is therefore already low, and that introducing new type A risk is actually NOT an all-in improvement for you.

If you were hiring a human chauffeur would you insist they be better than your current driver in all conditions (rain, night-time, snow, off-road)? Or would you more likely ask that they be minimally competent in all likely conditions, and that his driving style in aggregate, for your current driving profile, be an improvement?

Insisting on not replacing a driver until every subcategory of driving type is strictly improved seems too stringent.

You still have an out here by saying that you're willing to accept more net risk due to "meta" factors like "driver not meeting minimal competence factors". But it would still be more honest in that case to fully admit that this reasoning requires knowingly accepting likely real-world increased fatalities.

And as an aside, for humans, "minimal competence" testing is reasonable - but at large scales & in statistical realms, such metrics value in protection from false claims are of less value since we already have millions of miles of real-world performance tests.

Is there any evidence that you can point to that Teslas are saving more crashes than they are causing, compared to some other vehicle with basic collision avoidance features?

I've googled and found some very partisan sites. It seems hard to tell what's going on.

My overall analysis is: Tesla has some very serious enemies in other automakers, lawyers, etc. And yet, there are no clear studies or statistics saying they're worse, and Tesla claims they are safer. This isn't a very good analysis and I'd love to have a clearer more well-defined group for statistical comparison, but it at least seems plausible to me.

> i.e. if Tesla causes 100 new crashes of type A but saves 200 of type B, isn't that a gain? Even if type A crashes are one which humans don't normally make

By that logic proffeshional racing drivers should be allowed to drive drunk and stoned

You and I actually agree completely. I was attempting to steelman an argument for why a lot of people remain hostile to the concept of self driving all together.

I guess people have no business being on the roads then, because all of the actual crashes in that video were vehicles driven by humans. If they had been following the rules of the road with proper spacing between vehicles, they wouldn't have gotten in this collision after all.

I think you're overestimating human drivers frankly. Collisions like this one happen between humans every day, multiple times per day across North America, you just don't hear about it because normal human error is at fault.

There was a human driving the car.

That a Tesla crashes in ways a human wouldn't doesn't seem true. From what I've seen, even the most common accident type for a Tesla occurs at a lower rate than non-Tesla cars.

Do you have more info that's more specific about the types of crashes?

The problem here is a lot of the worst crashes are caused by humans behaving crazy and stupid. It's not a good criticism.

> It'd be perfectly logical to accept road testing FSD, even if it's significantly worse in some areas, as long as this is not exploitable, AND the net gain from allowing it is still overall positive

So are all 20 manufacturers going to get the right to endanger my life just so they can train their self-driving? Or is this spesific to Tesla?

Do I and my family get some compensation for this danger, or in case we are killed? Do we get a share of the profit this self driving car will generate for Tesla ?

What if Tesla never produces a good self deiving car, goes bancrupt, then I died for nothing?

It hinges on whether you think it's net safe/unsafe today. if unsafe, I'd support not allowing it, of course. If on net safer, BUT with a different distribution of harm, I think it's quite hard to make a decision. That is what most discussion is about.