Comment by snowwrestler | Hacker News Reader

Comment by snowwrestler | Hacker News Reader

Parent Full thread

On HN I frequently see comments to the effect that self-driving cars will be commercially available in a decade or two, and that they will be safer than human-driven cars. Personally, I question whether either is true, and I strongly doubt that both can be true at once.

Humans have a lot of shortcomings--emotions, sleepiness, distraction, the effects of drugs and alcohol--but humans have some amazing strengths too. We've got cognitive abilities in recognizing patterns and making decisions that were honed by millions of years of evolution, and directly trained by, on average, decades of actually driving cars. I think it's easy for technology-minded folks to overestimate the impact of the weaknesses, and underestimate the deep complexity of the strengths.

We barely have fully-autonomous self-flying planes today, and I don't know of any research showing that they are safer than human-piloted planes. And flying is a far less complicated domain to automate than driving.

There are some cases where we want amazing drivers, for example ambulances. But for an ordinary day to day commute, people shouldn't have to be at the top of their game. I think your point boils down to variance. Human drivers have a high variance, we all can be very good or very bad somewhat randomly.

Your point about research is really interesting. I think a large majority of cars actually following the speed limit, looking before changing lanes, and using turn signals will have a positive effect on overall safety, even if they aren't anywhere near as good at dealing with extraordinary situations. It would be nice to know what % of "good drivers" is required to minimize those extraordinary situations, if any at all.

We've all seen that one car weaving in and out of traffic. Perhaps they do indeed make us safer because they force us to pay attention. But they do seem to create a whole bunch of extraordinary situations for dozens or hundreds of drivers to deal with. It only takes the one weak driver (i'm assuming weaker than self driving here) to panic and then the freeway is a parking lot for an hour. That research would be interesting.

Ambulances may kill more people than they save by driving at high speed http://usatoday30.usatoday.com/news/nation/2002/03/21/usat-a...

It seems like you'd still want to hire for the best drivers, rather than trying to automate away ambulance driving. Coping with snow, rain and ice is something we want emergency services to deal with. A school bus on the other hand, meh, just don't run them in scary weather.

For a school bus you still want an adult on the bus to be responsible for the kids not killing and eating each other. So you aren't going to have any real savings there.

Well, it might be better to have an adult that can actively prevent the kids from "killing and eating each other" instead of the current state where one person has to contend with traffic, possibly inclement weather, and dozens of loud, obnoxious (and possibly cannibalistic) children.

I think it's easy for technology-minded folks to overestimate the impact of the weaknesses, and underestimate the deep complexity of the strengths.

I'm also really astounded when I'm reading the answers about self driving cars here on HN.

If you have ever worked on algorithms that are mostly based on heuristics, then you know all of these ugly edge cases, and this uncertain feeling, that you most likely don't have covered all edge cases.

I think the risk here is that you're aiming the bar for 100% perfection (very high) when the accurate bar to compare it against is human drivers (far from 100%) So self driving cars just need to be better than your average human driver. That's a much lower bar from the parking lot fender bender to the prevalence of speed in fatalities etc.

EDIT: Changed to "100% perfection" from 0% to remove any confusion.

They don't even need to be better than "the average" human. They need to be better than some subset of humans we let drive. Replace those humans with computers, and iterate up from there.

I could see within not too long people with DUIs being restricted from self-driving. The elderly are another demographic, but there are social forces working both ways there. (AARP will reject any calls to restrict driving rights of seniors, but self-driving cars will be awesome for some seniors who just can't cut it.)

Logically, it makes sense that self-driving cars just need to be safer than humans, but the political and consumer decisions that shape the self-driving car industry will not be based on logic. Realistically, self-driving cars will need to be damn near perfect before regulations and consumers get behind them.

I know what you're saying. I think Tesla is a great example. When they had car fires it was front page news. And the first deaths that happen in self driving cars will also be front page news. But just like those fires had no impact on Tesla sales or interest in electric vehicles the same thing will happen here is my guess.

Why do you think that human drivers are "far from 100%"?

Anecdotally, I can only recall a handful of car crashes involving myself or any friends or family. I can't recall any injuries or deaths due to car crashes among any friends or family. Is it a regular occurence in your circles?

U.S. government data I could find quickly reported that the death rate per 100 million vehicle miles travelled in 2010 was 1, which is a per-mile success rate of 99.999999%. The injury rate was 75, which drops the success rate to 99.999925%.

edit: speling

A person will be in a crash every 15 years or so. At 10,000 miles per year, that's about 150,000 miles per crash.

Six months ago Google said they've logged 700,000 miles, with 0 accidents.

The failure mode for the self-driving car isn't that you crash into a tree. It's that there are newspapers blowing across the road and it says "nope, not going anywhere" and you sit and wait.

The linked article details several failures that would be real accidents on an actual street.

Google's cars are just one, very carefully controlled, experiment in self-driving cars.

How much higher than 0% though? Human still have a high number of accidents, but that's because we drive a lot. The 0% error for human driving is very, very, very, very low.

No Humans really aren't that good at driving cars, 33,561 people were killed just in 2012 in the US[1].

[1] https://en.wikipedia.org/wiki/List_of_motor_vehicle_deaths_i...

...Which, as this [1] commenter points out, works out to a per-mile success rate of 99.999925%.

(Or, alternatively, works out to on average one injury or death per 1.33 million miles traveled.)

Or, to put that number in perspective, there are 435,000 deaths per year caused by tobacco in the US - 13 times that number. [2]

[1] https://news.ycombinator.com/item?id=8700342

[2] https://en.wikipedia.org/wiki/List_of_preventable_causes_of_...

Those coments are mostly extrapolating results gotten in a place with a nice weather, and every inch of the road mapped.

Reply on news.ycombinator.com