Also don't forget that while each of us is somewhat unique in terms of driving, the cars are not, so a bug in a car's software can amplify issues even further.
There are issues also how cars will behave when they will have to interact with other self driving cars. Those problems are already difficult when writing distributed applications and with cars we don't have luxury of just letting things crash together until things improve.
Only when that's deemed acceptable.
http://www.fastcompany.com/28121/they-write-right-stuff
> But how much work the software does is not what makes it remarkable. What makes it remarkable is how well the software works. This software never crashes. It never needs to be re-booted. This software is bug-free. It is perfect, as perfect as human beings have achieved. Consider these stats : the last three versions of the program — each 420,000 lines long-had just one error each. The last 11 versions of this software had a total of 17 errors. Commercial programs of equivalent complexity would have 5,000 errors.
You say that so glibly. Have you ever done the math? I did, I posted it here:
https://news.ycombinator.com/item?id=12012166
Summary: The computer would have to be better than 99.99999% perfect to save any lives.
That's a hypothetical. You say autonomous cars are "poised" to save many lives -not even avoid many deaths, actively "save".
That's very strong language to use based on an "if" that nobody yet knows the answer to with any certainty.
That is, unless you are privy to some information that the rest of us are not? Are you?
What is it that you base your assertion on? Do you base it on anything or is it just a pretty turn of phrase you thought sounded cool?
There are already videos of assisted-driving cars avoiding accidents which likely would have been fatal. We already know they can potentially save lives. It's something we already accept for flying. Autopilot in planes has resulted in deaths of crew and passengers. So we have precedent in at least one mode of transportation.
Drivers get T-boned or rear end other people all the time. Those are two scenarios where driver-assist could have applied the brakes and actively saving lives for drivers inattentiveness. You'll also find no shortage of Liveleak clips of people being killed by drivers who failed to see them and brake in time. Many being scenarios that Google's self-driving cars have the technology to avoid.
I'm not concerend about assisted driving. I'm concerned about fully autonomous cars being handed complete control in real-world situations.
I'm not just "concerned" about them- they scare the hell out of me.
>> Many being scenarios that Google's self-driving cars have the technology to avoid.
What Google says it programs its cars to do, what it really programs its cars to do and what its cars can really do are all separate things.
The problem is that the current technology level is nowhere near advanced enough to allow fully autonomous vehicles to operate safely. There is a huge number of situations that those cars aren't programmed for, that they can't be programmed for, because those situations are completely unforseeable.
Machine learning has a huge problem with data sparsness. You may train a learner in petabytes and petabytes of data collected from the real world and still miss the vast majority of events that may occur.
That is why, like I say elsewhere, machine learning-based AI makes utterly ludicrous mistakes that humans would never do, even in difficult situations were they can't be expected to perform with 0% error. I've used a few metaphors- here's another one: a human would never mistake a truck for a cucumber. A machine learning algorithm, might.
And what's worse, there's no way to prevent this sort of mistake, or even correct it, because most of the time the models built by such algorithms are simply too complex to be processed by humans in the way a hand-crafted system would be (and goddess knows how hard those can be to process).
Autopilot in a plane doesn't have to deal with obstacles besides the odd mountain here and there.
So I'd say that in many (most?) ways it's easier and safer to autopilot a plane than a car, and as you say, even planes with autopilot have failed.
Planes most certainly do have to deal with wildlife (e.g. birds) appearing suddenly in their path regularly.
In other words, the plane autopilot doesn't "deal" with wildlife at all unless you count running into it. Which is apparently what Tesla autopilot does as well.
https://en.wikipedia.org/wiki/List_of_motor_vehicle_deaths_i...
From which I quote:
The number of deaths – and deaths relative to the total population – have declined over the last two decades. From 1979 to 2005, the number of deaths per year decreased 14.97% while the number of deaths per capita decreased by 35.46%.
So at least the rate of fatalities is decreasing. I won't claim I know whether that's because humans get better at driving, carse get safer, traffic regulations get safer, or any other factor, but the point is roads are getting safer already and without any autonomous vehicles on them.
[Original bit:] Again- based on what? Why do you say that? What is even your measure of humans getting "better"? Better at what? Better drivers? Better cognitive systems? Better what?
I said no such thing, you are putting words in my mouth.
What I said is that the baseline for traffic deaths is quite high, and if (or when) autonomous cars are even a little better than humans, then less lives will be lost.
I was replying to someone else and you jumped in. You can hold the accusations of putting stuff in places, thank you.
>> What I said is that the baseline for traffic deaths is quite high, and if (or when) autonomous cars are even a little better than humans, then less lives will be lost.
And I said that autonomous cars being better than humans is just a hypothetical.
I'll add that we have no evidence that it's the case. Computers are better at humans at tasks that require fast and accurate retrieval, or computation, but they'll not just magically become better than humans at driving. Somebody has to program them to do that.
So who is going to do that? Do you know how to do that? Do you know anyone who knows how to do that?
The theory is that we'll put enough (real or virtual) cars in enough training situations that they'll learn to drive on their own, but that approach has serious limitations, not least the fact that in the real world there may be an infinite number of situations that a learner will never encounter during training. In any case, with any other cognitive task that we train computers to carry out they end up making completely ridiculous mistakes, which in the case of driving will cost lives.
Until we have a way to develop systems that understand their surroundings, autonomous cars being better drivers than humans is just a dangerous fantasy.
Bad compared to pretty much every other form of transportation that humans employ.
Human drivers are very bad. We kill over 32,000 people a year (and that's a 60 year low, just a few years ago it was over 40,000).
So if self driving cars are just a little better than humans, they will save many lives. If they are a lot better than us, they will save many more.