Hacker News Reader

Comment by MontyCarloHall | Hacker News Reader

>human drivers, who aren't actually all that great.

A large fraction of human drivers are actually all that great. The majority of accidents/deaths are caused by a minority of terrible drivers, or good drivers who found themselves in terrible but rare circumstances. The majority of drivers drive hundreds of thousands of miles without any accidents that were their fault, or even any accidents at all.

In other words, it's probably easy to beat the mean human driver, which is greatly dragged down by a minority of terrible drivers. It's probably very difficult to beat the median human driver, and near impossible to beat the top 20% of human drivers.

View on HN

I don't think it's easy to beat the mean human driver and to demonstrate with solid data that you've done so.

In 2019 in California, there were 1.06 deaths per 100 million vehicle miles traveled. Any self-driving automobile technology that doesn't have at least 1 billion vehicle miles of data is in no position to claim that it is safer than human drivers and less likely to kill people.

Self driving cars don't make the same kinds of mistakes as human drivers do, but they make different kinds of mistakes. Some of these can be fatal.

I'd be curious to see a source for your number

Not really answering your question, but CDC says 28% of all traffic-related deaths in 2016 involved alcohol. Excluding these would immediately improve the mean performance.

I see why LMGTFY had its day in the sun: you can literally paste the first sentence of parent’s post into DDG, and the first link answers your question. Hell, the preview answers your question, you don’t even need to click it.

the writer has the responsibility give a source, not the reader

* The 2019 Mileage Death Rate (MDR) – fatalities per 100 million miles traveled – is 1.06.

https://www.ots.ca.gov/ots-and-traffic-safety/score-card/

California Numbers:

* 3,606 traffic fatalities in 2019.

* 1,066 Alcohol-impaired driving fatalities (fatalities in crashes involving a driver or motorcycle rider with a blood alcohol concentration, or BAC, of 0.08 or higher) in 2019.

* 620 Unrestrained passenger vehicle occupant fatalities in all seating positions in 2019.

* 164 Teen motor vehicle fatalities (age 16-19) in 2019.

* 972 Pedestrian fatalities in 2019.

* 133 Bicycle fatalities in 2019.

assuming the above (alcohol-impaired, unrestrained passenger, teens, pedestrian, bicycle) are all all poor-driver related that leaves 651 traffic fatalities.

>I don't think it's easy to beat the mean human driver and to demonstrate with solid data that you've done so.

Agreed. I should have written "relatively easy."

> Any self-driving automobile technology that doesn't have at least 1 billion vehicle miles of data is in no position to claim that it is safer than human drivers and less likely to kill people.

The circumstances under which those miles are driven (e.g. road type, location, weather, time of day, etc.) also have to be consistent with circumstances under which humans are driving. 10 billion autonomous vehicle miles driven only on highways in broad daylight is a worthless point of comparison, whereas 500 million miles driven across a variety of conditions representative of the full human driving population is worth a lot more.

> but they make different kinds of mistakes.

This is key, there's expectation and some wiggle room that as a human driver, humans will fuck up predictably and experienced drivers know how to avoid getting into incidents when this happens (usually).

Self-driving cars are weird to drive around. They will absolutely stop in situations where no human would think to stop. I think about this as a motorcycle rider, what if I'm committed to cornering on a corner I can't see around and the software decides on a self-driving car that it should just stop in the middle of the road after the apex? A human driver could do this too but many will know that this is a dangerous place to stop and try to put the car on the shoulder or minimize the amount of time it's stuck there.

I don't know if this is something we need to tolerate a temporary increased incident rate on as people get used to them being on the road, or if we need to make the software drive more like humans (with the assumption that means potentially making the behavior act sloppier than it can handle so that increased software reaction rate doesn't cause humans with slow reaction rate to slam into them)

> I don't think it's easy to beat the mean human driver and to demonstrate with solid data that you've done so.

It is. The mean is dragged down by alcoholics who drive drunk every single day. If you never drive drunk, that gives you a significant advantage.

We were talking about self-driving cars.

The mean is given by the number I posted, about 1 death per 100 million miles traveled. That number includes drunk drivers, distracted drivers trying to text, everything.

The point is that "1 death per 100 million miles traveled" is the mean average, but most drivers do better than the mean. Mean, median, and mode are not the same and the mean crash rate is not relevant to most drivers.

Not sure about that. I'd say from the people that are very close to me (friends & family), I wouldn't want to be a passenger with half of them. AI is /so much better/, can't wait for it to be mainstream. And it's not just about the AI driving, it's about the AI reacting 100X faster and having eyes all around the car to avoid accidents before they could even happen.

I wonder what humans will actually do better than AI in 50 years. I have a personal theory but I'm a bit off topic here

Do you have a source for that?

I don't know where the op got the statistic from. But this claims 1.02 per 100 million. https://worldpopulationreview.com/state-rankings/fatal-car-a...

Exactly right. Furthermore, most risky behavior is a choice. Crashes aren't random "acts of god" that strike anybody with equal likelihood. If you choose not to drive drunk, drive in bad weather, or for many hours without rest, then you can greatly improve your odds and almost certainly reduce your risk below the average. In these discussions I often see far too much fatalism; "Everybody thinks they're above average but half of you aren't". ignores both the fact that crashes aren't distributed like that, and the fact that the riskiest behavior is a choice. The mean number of miles driven drunk is greater than zero, but the number of miles I drive drunk is zero.

Reply on news.ycombinator.com