Hacker News Reader

Comment by overgard | Hacker News Reader

I don't think you can just look at the statistics though. What I mean is that, the main issue is that the category of mistakes they make are going to be inherently different from the category of mistakes that humans make. IE, humans are more likely to make emotional mistakes, but self driving is probably more likely to make perceptual mistakes or fail to respond to novel situations. I think the "fail to respond to novel situations" is probably the biggest blocker for people to trust them enough for widespread deployment. I'd love to see one of these things drive on a poorly mapped dirt road up in the mountains, for instance (I'm not saying it can't -- I just mean I would personally love to see how it can adapt to that kind of situation)

View on HN

> I don't think you can just look at the statistics though. What I mean is that, the main issue is that the category of mistakes they make are going to be inherently different from the category of mistakes that humans make

Would you rather have a X% chance of a loved one dying from a human-category mistake? Or a 0.5X% chance of a loved one dying from a ML-category mistake? I would take the latter in a heartbeat. Why waste time optimizing for "categories" when you can optimize for your loved ones not dying

While I agree with you, not everyone does.

You and I are thinking with a purely logical thought process, but most people don’t reason that way.

A lot of people want to be able to hold someone accountable, and that is more important than the pure probabilities.

They want to be able to blame a person for making a mistake.

That is not the way human beings typically evaluate risks. 'Acceptable risks' are more about psychology and not a statistical calculation. A self driving car that is twice as safe as the human average would be considered an unacceptable risk.

Add something that is new/unknown + lack of control, and you will have to clear a very high threshold to be considered 'acceptably safe'. My guess is a minimum of a 0.1x.

I don't think that human beings factor the rate of injury or death in at all, because human beings are notoriously bad with expected values from extreme conditions.

Once people get used to taking robotaxis, they won't care. The bar is "safe enough for very few people to have a bad experience." If you take 10 robotaxi rides in a row and they all go flawlessly your subjective interpretation of the risk will be that there isn't any.

I feel confident saying this because if you die in the US to anything other than old age, it was almost certainly a car crash. Despite this, most people don't think twice before getting in a car.

"Trust" isn't just about "% of a chance I die". It's like "can this system actually function well in the environment I'm in". For instance, I live in the midwest. Our roads are AWFUL compared to most coastal areas, the weather can get really extreme (-20 in the winters), there are things like invisible black ice which is a huge hazard, people drive giant brodozers, etc. If I'm in a taxi or an uber or something, I trust that the driver can adapt to these conditions. I do not trust that a Waymo can. I'm not saying that's a "never" thing, but it is a big deal.

> humans are more likely to make emotional mistakes, but self driving is probably more likely to make perceptual mistakes or fail to respond to novel situations

Humans fail chaotically. AVs, at least Waymos, fail safely. In urban environments, there is a case to be made within a few years for banning (or making it massively cumbersome) to manually pilot a private vehicle.

That's an odd claim. Human failures are uncorrelated, but the Waymos all running the same software must be expected to fail in the same way, in the situations in which they fail. Which is why they often make the news; one Waymo stuck soon sees many of its brethren join it.

Don't humans fail in the same way too, broadly speaking? Wouldn't it depend on how you categorize failure?

If a Waymo failed because it missed a stop sign, and then another Waymo failed because it was confused about some strange markings on the road, is that not failing "the same way" in the same sense that a person failing to yield or driving drunk?

I think they're saying if one waymo misses a stop sign every waymo will miss the same stop sign

Failing in ways that confuse the crap out of all the humans around (including pedestrians and cyclists) and demand their time and attention and thereby detract from the safety of the overall system is not "failing safely".

Nobody would make excuses for a rookie driver coming to a stop in a busy interchange. Only someone with the moral character of a career politician would attempt to construe that as anything other than detracting from overall safety. But a computer does it and it's fine?

Personally, I would fight hard against that kind of outcome. While I don't mind self driving cars on the road if they drive safely, being able to own and operate a vehicle is a huge deal for personal freedom. Part of why I refuse to own a Tesla is because I don't want a vehicle that's subordinate to big tech above my own wishes. (I even still drive a stick shift and my car is essentially air gapped)

Statistics is still the most reliable indicator, because you can’t really tell if a situation is “novel”. How novel is novel? I doubt even Waymo knows for sure.

People will trust them by familiarity, not by any objective measures. Their trust grow as they see them more often, without bad accidents.

My understanding is that they only operate in areas they’ve had approved and pre-mapped out for them though. Practically I don’t think the scenario you’ve described would ever happen, but I too would like to see it (I suspect they wouldn’t handle it well).

I live East of DTLA and twice I’ve seen people driving them around, presumably to map those areas out. It’s pretty annoying that I can’t use them for my late night commutes yet, but I’m eagerly looking forward to when I can!

IDK, I've been in the car with someone when they just up and up drove off the road because they failed to notice that there was a corner - and that's a sober human. Another time, someone else, they rounded a corner, saw a tractor in the other lane, and were so surprised that they hit a tree. My mother has totalled six cars - she insists it was someone else's fault every time, but the statistics are somewhat damning. My wife's driving has given me haemmorhoids, I clench so fucking hard. Blind corners at twice the speed limit, looks over her shoulder to talk to passengers, and I constantly find myself buying wing mirrors, fenders, headlights, body panels, hubcaps, and anything else she can break off on traffic, rocks, or street furniture. Well, I did until someone finally took her license away from her for a year.

So... not so sure. Every AV I've ever been in has driven incredibly cautiously compared to a human driver.

Reply on news.ycombinator.com