Hacker News Reader

The problem is it's called "full self driving" and it runs red lights.

Just like the rest of the drivers out there you mean. Just think logically for a second. If they ran red lights all the time there would be nonstop press about just that and people returning the cars. Theres not though, which is enough evidence for you to conclude these are edge cases. Plenty of drivers are drunk and or high too, maybe autopilot prevents those drivers from killing others

We evolved to intuit other humans intentions and potential actions. Not so with robuts, which makes public trust much more difficult despite the statistics. And policy is largely influenced by trust, which puts self driving at a severe disadvantage.

You think you can out intuit a drunk driver? That’s some serious hubris.

I happened upon a swerving drunk-driving police car once. A tesla would have continued on, following the rules of the road, trying to pass the swerving drunk-driving police car, likely getting in an accident with it. I was smarter, and I stayed the fuck away, far far back from it and changed my course to avoid it.

Yours is a bit of a strawman, considering drunk driving is illegal. So the appropriate comparison is to unregulated (illegal) buggy software operation. Do I feel more comfortable intuiting the intention of a drunk driver compared to buggy software? Yes. Similarly, as the other poster said, if I see a weaving car I tend to stay away from them because I can infer additional erratic behavior.

That’s also why people tend to shy away from people with mental health issues. We don’t have a good theory of mind to intuit their behavior.

Its not a strawman. Having a conversation can also distract you. Being elderly can be a risk. Being tired can be a huge risk. Yet none of these things are illegal. I could have just as easily used one of these examples instead of drunk driving and my point would stand against your criticism.

Fact is, humans in general are imperfect operators. The ai driver only has to be an alright driver, not a perfect one, to route around the long tail of drivers that cause most all the fatalities.

If those are the stronger examples, then you should have went with them. It’s more inline with the HN guidelines than taking the weaker interpretation.

I think you missed my point. Because software is more opaque, it has a much higher threshold before the public feels comfortable with it. My claim is it will have to be an outstanding driver, not just an “alright” one before autonomous driving is given the reins en masse. In addition, I don’t think we know much about the true distribution of risk, so claims about the long-tail are undefined and somewhat meaningless. We don’t have a codified “edge” that defines what you call edge cases. Both software and people are imperfect. Given the opaqueness of software, I still maintain people are more comfortable with human drivers due to the evolved theory of mind. Do you think more people would prefer their non-seatbelted toddler to be in a an average autonomous vehicle by themselves or with a human driver give the current state of the art?

But more to my point, humans are also irrational so statistical arguments don’t translate well to policy. Just look at how many people (and professionals) trade unleveraged stocks when an index fund is a better statistical bet. Your point hinges on humans being modeled as rational actors and my point is that is a bad assumption. It’s a sociological problem as much as an engineering one.

> it runs red lights

Fixing that would require "full self stopping". Coming soon[1].

[1] ... for some value of "soon", that is.

> I am convinced, though, that using FSD judiciously will save far more lives than removing it will.

This is a statement of faith.

IMO cars are dangerous, whether they have a level headed experienced driver or an all seeing all knowing AI, sometimes a car pulls out from a blind drive or a tire fails and swerves into oncoming traffic. It's not clear to me why people think having a computer drive 95% of the time so they don't have to will make them better able to avoid accidents in exceptional cases.

In the case of pilots, even tho a computer can handle all manner of weather, pilots are still required to put in manual flying time so they actually know how the craft handles under their control, so that they can be ready to take over in case of exceptional events. I think we will see the same regulation in cars eventually, requiring drivers to have minimum hours per year in order to continue piloting craft, even if AI is good enough to do it most of the time.

In the case of the blind driveway situation, there's a safe, marked speed for roads with blind driveways on them. A driver (human or AI) going that speed can avoid a car pulling out from it. The assumption for higher safety is that the impatient emotional human that's late for work who's husband just left them is speeding excessively, while the AI isn't.

agreed, but in that case isn't the solution to simply require that all cars are manufactured with speed limiters that comply with the database of speed limits on a per road basis? if the goal is safety.

That's not out of the question. The Nissan GT-R in Japan is geofenced to only allow excessive speeding on racetracks.

But Tesla isn’t the only game in town, and eg Waymo seems to have a far better safety record. They’re doing “engineering” as it should be done, not “move fast and break people”, which is fine for websites but not great on the road.

That’s similar to how I feel about LLM’s. Amazing as an input to a system but you need real engineering guardrails around them.

Have the safety statistics been standardized, though? I vaguely remember articles about Waymo doing their own after-action reports and sanitizing their accident data to only keep those accidents that they felt a human would have avoided. This creates all sorts of data quality problems related to subjectivity and bad incentives. We wouldn't accept throwing out drunk-driving stats under the guise that a sober driver wouldn't have made that mistake. This says nothing of the differences in environments that can be used to game safety stats.

Not sure, and they’re dealing with fewer issues than Tesla which self drives all over the place. But their stance is more conservative, and Musk has more than a whiff of cowboy, over claiming etc. Less compatible with safety. I was a fan but less so over time. That’s my bias stated :)

You're expressing your opinions in this context, because Tesla didn't reveal the data that would let us tell whether its FSD system is objectively safer than an average human driver in the same driving conditions. You're leading a discussion that distracts from the core issue that Tesla is unwilling to release data that would enable independent researchers evaluate its FSD systems. In other words, you're doing marketing for Tesla for free.

> I am actually not able to adequately express how small a percentage of total highway deaths 51 people is

This is some kind of logical fallacy, a false equivalence or maybe a red herring. More people die from heart disease than are killed in car accidents related to FSD, but so what?

> I am convinced, though, that using FSD judiciously will save far more lives than removing it will.

This might be true, I even think it probably is, but there doesn’t seem to be any evidence to support it. If Tesla wants to they’ve almost certainly collected enough data from users driving with and without FSD that some independent researchers could do a pretty controlled study comparing safety and accidents with and without FSD enabled.

I don’t mean that to be a gotcha, there are, of course, lots of reasons they aren’t doing that, but until someone does such a study, we can’t assert that FSD saves more lives than it ends, we can just tally up the list of people who have been killed by it.