Hacker News Reader

> I believe the number one cause of death for Americans under the age of 30 is traffic accidents.

Traffic accidents are not included in the "accidental death" category, and haven't been for as long as I've been looking. From page 2 of the full report[0]:

"For example, Malignant neoplasms of the trachea, bronchus and lung (lung cancer) and Motor vehicle accidents are not rankable causes of death (see “Procedures for ranking causes of death”), although they can be identified using the standard mortality tabulation lists. If these causes were included in the current rankings, lung cancer would be placed among the 10 leading causes of death with a rank of 6th, whereas Motor vehicle accidents would rank 13th."

It's not even close to being #1, for any age group. More people die from suicide every year than from car accidents. Please also keep in mind that 1 out of every 6 traffic fatalities in the United States was a pedestrian.

[0]: https://www.cdc.gov/nchs/data/nvsr/nvsr68/nvsr68_06-508.pdf

Good catch. However, that paragraph goes on to say "However, each of these is incorporated into broader rankable categories, namely, Malignant neoplasms and Accidents (unintentional injuries), respectively." So they are included in accidental deaths -- they're just not broken out.

Being #13 overall definitely doesn't preclude them from being #1 in younger age brackets. Most of the deaths from cancer and heart disease are the older age brackets ("Diseases of the heart" are 23% of overall deaths, but only 2-4% for each age rank among minors; ~6% in the 20s-30s). "Accidents (unintentional injuries)" is the #1 listed cause for every male and female age bracket until the 45–54 bracket (pp 19-22) (this precludes infants under 1). The question is what portion of those were traffic accidents.

If we assume traffic accidents at the same rate among all age groups, then we have 36,000/169,936[0] = ~20%. That drops it to around fifth place for most of the younger brackets. (Personally, I'd guess it's higher for at least the 15-19 age bracket, but that's just a guess at this point.)

So, okay, fine, traffic accidents is roughly sixth most common cause of death for kids in the US, and thirteenth for all ages. The original question was why we'd prefer that if the cost is the same. Even while my numbers were way off, it's still a big benefit.

> Please also keep in mind that 1 out of every 6 traffic fatalities in the United States was a pedestrian.

I'm assuming a significant portion of pedestrians traffic fatalities are drivers killing pedestrian, not pedestrians...I don't even know what to put as the alternative. Stepping in front of a bicycle?

My position is not that human-driven cars kill human drivers -- it's that human-driven cars kill lots of people. If drivers only ended up harming themselves, I'd be slightly less concerned about it.

[0] Just grabbing most recent number from https://en.wikipedia.org/wiki/Motor_vehicle_fatality_rate_in...

> A 16-year-old doesn't get to inherit the experience of the 70-year-old they're replacing. A new self-driving car does inherit the experience of older self-driving cars. They also learn from each other in a way humans can't.

> Human drivers also improve at a rate of O(1). Self-driving cars can share their experience perfectly, so they improve at a rate of O(n). I get a day of driving experience per day. Each one of a million self-driving cars on the road will acquire a million days of driving experience. They won't drive traffic accidents down to zero in a week, but in a few years, if we were to deploy them en masse, they could.

This is all very nice, but if true, why are self-driving cars still unable to drive in snow, avoid crashing into trucks pulled over on the side of the road, or navigate a parking lot without nearly getting t-boned?

The rate of improvement might be O(N) (I strongly doubt it), but the problem may lie in the constant multiplier...

> why are self-driving cars still unable to drive in snow

Have you seen people driving on the first snowfall of the year?

> or navigate a parking lot without nearly getting t-boned?

Driving in a parking lot is tough. Lots of pedestrians, lots of other cars, and way more people driving in reverse than you'd see in most scenarios.

> Have you seen people driving on the first snowfall of the year?

Most of those people don't have winter tires.

> Driving in a parking lot is tough.

So when is the O(N) rate of growth going to get SDVs to the point of an attentive 16 year old?

> Most of those people don't have winter tires.

You'd think that they'd drive extra cautiously in snow in that case.

> So when is the O(N) rate of growth going to get SDVs to the point of an attentive 16 year old?

Does it need to beat an attentive 16 year old or does it just need to beat a distracted early to mid 30s parent with a small child or two in the car? Self driving cars only need to be better than the average human to reduce overall accidents. Having said that, I'd probably be more likely to hop into a car driven by a random person than a self driving car that I know is only marginally better than an average human.

> I believe the number one cause of death for Americans under the age of 30 is traffic accidents.

And it's possible that the number one cause of shortened lifespan for older Americans is also cars, by virtue of encouraging (perhaps enforcing, if they've sufficiently influenced your town's urban planning) a sedentary lifestyle, and all the health risks associated with one. I also wouldn't be surprised if they contribute to the USA's relatively high rate of mental health issues, by contributing to social isolation.

Self-driving cars won't help with those bits. If anything, they'll intensify them.

> Self-driving cars won't help with those bits. If anything, they'll intensify them.

One common reason people give for not biking is that they're afraid someone driving a car will kill them. Self-driving cars should reduce that risk.

Of course, they also should make commuting by car cheaper and more relaxing/less of a waste of time, so you're likely correct, but there is that competing force that may increase cycling rates.

> Self-driving cars can share their experience perfectly,

I'm not a machine learning expert, but I'm not convinced this assertion is well grounded. If you have two ML systems, I don't know of a simple, general, reliable way to transfer the "knowledge" stored in one to the other without disrupting the "knowledge" already within the other.

If you mean just copying the video feed, sure, but that's only one input variable. There's no guarantee of a linear relationship between hours of video feed and driving quality. Humans seem to hit an asymptotic lower bound in the collision probability even as they accumulate experience, and while that might not be true for self-driving cars, it isn't guaranteed a priori.

You're right, it's not perfect, and it might be sub-linear improvement, but it can practice on its predecessors' input in a way a human can't. (There are plenty of scifi stories where they found a way to do this for humans as well, so it's certainly conceivable, but seems a long way off, whereas "feed all existing training data into the new model" is something we have now.)

You're pretty much guaranteed that self-driving technology "skills" will collectively improve over time, at some positive rate, whereas we don't expect "Gen Z" to be more skilled drivers than previous generations.

This is all entirely skill-focused. We're not guaranteed that self-driving cars will actually get safer if we don't have the proper incentives to favor safety over speed or cost. Similrly, despite the stagnancy of human driving skill, I anticipate safety will improve with human drivers for non-skill-related reasons -- specifically, driver assistance technology, and a shift towards just getting a Lyft if you're impaired.

One (probably not so simple) way could be to build up a shared database of input-decision-outcome records, so when a car is faced with some unknown object on a particular street, or its own risk assessment is high (e.g. heuristics- "I am planning to decelerate quickly"), it could search for similar input patterns (e.g. color, shape, gait) or similar decision plans in the database, and bias the car's own decision accordingly to avoid an undesirable outcome. Then that decision would get posted to the database as well, linked to its parent. Previous outcomes in the database could be assessed by humans in some kind of legal framework to help the cars determine undesirable vs desirable outcomes.

Wouldn't the safety of self-driving cars relative to human drivers also become moot if cars were no longer needed?

We are a LOT further away from cars no longer being needed than we are full self-driving cars.

Who's "we"? It seems to me that not all parts of the world are equally dependent on cars.

This is an odd comment considering how humans have lived hundreds of thousands of years without cars, and only a few decades with them. You’re saying that a solution that’s as old as time - and used by billions of people today - is not as realistic as a technology that hasn’t been developed yet?

We've had wheeled chariots and carts powered by animals for almost as long as we've had civilization (about 5000 years). Cars and trucks are simply more efficient (a defensible statement... https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5102965/ particularly with electrification), self-propelled versions of these.

So absolutely we will have cars and trucks well after we're able to develop driverless versions of them.

You're misusing big O notation

It’s pretty clear that the author isn’t using the technical definition of big O but rather is using it in a descriptive way. It’s not ambiguous in any way so what’s the harm?

I would say no harm, but rather confusing to people that do not know big O, and useless to those who do.

Can you explain how it's wrong? Obviously, normally you want things to scale more slowly rather than faster, but other than applying to a different thing than usual, it seems accurate to me -- human drivers improve at a constant rate regardless of how many there are, and self-driving cars improve linearly with their number.

A function being in O(f(x)) means the function grows no more than a (multiplicative) constant factor faster than f(x). So O(1) is a subset of O(n). O is analogous to <=.

The semantics that GGP comment meant to convey are better represented by saying that autonomous cars' driving experience grows with Ω(n) (analogous to >=) or Ɵ(n) (analogous to =).

How so? There, n means "the number of [this type of] drivers", thus the GP is saying that "self driving cars improve in a way proportional to the number of self driving cars [that are collecting and sharing data], while humans improve their driving skills in a way independent of the number of humans driving."

(I think that's slightly wrong -- humans are probably O(log n) because large populations of humans invest in heuristics and teaching methods that enhance driver improvement rate, and for SDCs, the "lessons learned" between units are probably somewhat redundant, so it might scale as something like sqrt(n). But I don't see a problem with the O(n) usage.)

Just yesterday while driving to the gym, I witnessed the person behind me at a red light, violently gas it and swerve out into oncoming traffic lanes, drive past all the cars waiting for the light, and blow straight through the red light.

He went on to blow another red light before turning off onto a side street.

This kind of shit would never happen with a self driving car. The bar for self driving cars is incredibly low.

Well, that kind of stuff would still happen since in that case the driver was deliberately driving recklessly.

In fact, people could start accounting on other people's self driving cars being cautions and deliberately betting on their safety features. e.g If most people were driving a Tesla I could more safety run red lights since their other cars will stop for me automatically.

Unless we are going to make self-driving ONLY cars that is.

> Unless we are going to make self-driving ONLY cars that is.

This is the dream.

EDIT: More specifically, I envision a world where to manually drive your car you have to past a MUCH more rigorous test than the current licensing process. Including background checks for infractions such as DUI, would ban / block you from obtaining a manual driving license.

You would of course retain your freedom to have a car that can drive itself.

Maybe your bar. A reasonable (minimal) bar is that they have to perform at the 50th percentile of drivers by safety. This guy sounds like the first percentile.

Depends what the curve is, doesn't it? If the worst 1% of drivers cause 90% of the accidents, then replacing all drivers with vehicles at the 2nd percentile would drop the accident rate.

> get to inherit the experience of the 70-year-old they're replacing

A lot of 70 year olds should not be driving.