Self-driving cars are headed toward an AI roadblock • Hacker News Board

Re: Self-driving cars are headed toward an AI roadblock

#211

Post by notahacker » Thu, Jul 05, 2018, 11:52 PM UTC

the8472 wrote: post #207

notahacker wrote: post #188

Earlier quoted context omitted.

If cars which require no effort whatsoever to operate and cost significantly less than taxis are on the market, it's a safe bet that being slightly less dangerous per passenger mile will kill far more people than are currently killed on the roads due to more, longer car journeys. That's before we even get into the fact a car which is slightly less likely to cause a fatal accident than the average human driver on an a…

Of course if people make more journeys, i.e. become more mobile, you could argue that the increased number of deaths should be balanced against life-hours of drivers freed up while also improving quality of life of passengers at the same time, i.e. it might still be QALY-neutral or even positive. > than the average sober, careful and experienced human driver But ultimately we do accept the fatality rates caused by al…

True, there are QALY arguments for even unreliable SDVs (and also for other things like abolishing speed limits or allowing people to continue to operate older vehicles currently deemed unroadworthy if they can't afford better), but they're dependent on how you adjust for quality and at best orthogonal to standard safety arguments.

> But ultimately we do accept the fatility rates caused by all drivers and not just the good ones as a cost of having individual transportation. If we didn't only teetotalers would be allowed to operate cars to remove one risk factor. And we also accept the risk of inexperienced drivers because we have to teach them eventually, the same reasoning can be applied to autonomous cars (just substitute teaching with data gathering)

We don't accept the fatality rates caused by all drivers; we confiscate the licenses of or jail bad drivers on a regular basis, and we certainly don't permit drivers being taught to operate commercially. Driver accountability for accidents is part of the social contract that permits individual transportation despite its cost; an corollary of that is that the accident rate is higher than we believe it should be, and that some degree of negligence is involved in the current level of accidents which it would seem unreasonable to exonerate corporations from if they were to replicate them using software. And on a purely practical basis, if you estimate SDV safety based on narrowly beating the mean accident rate across all drivers in tests, and then early commercial adoption tends to be replacing experienced commercial drivers who drive the most miles with the fewest incidents first, you're certainly going to push that accident rate up in the short run even without additional miles driven.

Re: Self-driving cars are headed toward an AI roadblock

#212

Post by jfoutz » Thu, Jul 05, 2018, 11:56 PM UTC

mathiasben wrote: post #111

jfoutz wrote: post #30

Earlier quoted context omitted.

This is hard to parse. All software that can be represented as if and goto (and some sort of assignment i suppose)? when, while and do are just if and goto. alpha go is all just if and goto. I'm not sure what you're looking for or alluding to. Is there some other operation that isn't just if and goto? because i think any function you come up with, i can just make a big table of inputs that result in specific outputs.…

the idea of creating a pre-determined program which will be followed is limiting. creating AI will require methods we can't yet conceptualize.

Are you sure? evolution is pretty darn simple. just not trying the same solution twice is a pretty big speedup.

Re: Self-driving cars are headed toward an AI roadblock

#213

Post by the8472 » Fri, Jul 06, 2018, 12:45 AM UTC

notahacker wrote: post #211

the8472 wrote: post #207

Earlier quoted context omitted.

Of course if people make more journeys, i.e. become more mobile, you could argue that the increased number of deaths should be balanced against life-hours of drivers freed up while also improving quality of life of passengers at the same time, i.e. it might still be QALY-neutral or even positive. > than the average sober, careful and experienced human driver But ultimately we do accept the fatality rates caused by al…

True, there are QALY arguments for even unreliable SDVs (and also for other things like abolishing speed limits or allowing people to continue to operate older vehicles currently deemed unroadworthy if they can't afford better), but they're dependent on how you adjust for quality and at best orthogonal to standard safety arguments. > But ultimately we do accept the fatility rates caused by all drivers and not just th…

> And on a purely practical basis, if you estimate SDV safety based on narrowly beating the mean accident rate across all drivers in tests, and then early commercial adoption tends to be replacing experienced commercial drivers who drive the most miles with the fewest incidents first, you're certainly going to push that accident rate up in the short run even without additional miles driven.

Uber could be skewing statistics in similar ways.

Anyway, this is all handwavy speculation. What I'm trying to get at is that the overton window for AV behavior may be surprisingly large, especially if they exhibit their worst behavior in the learning years when the fleets are still small. For some metrics possibly even larger than that for humans.

Re: Self-driving cars are headed toward an AI roadblock

#214

Post by rohit2412 » Fri, Jul 06, 2018, 12:48 AM UTC

InclinedPlane wrote: post #186

ebikelaw wrote: post #11

Earlier quoted context omitted.

All journalists covering the self-driving car beat are basically required to ignore Waymo so they can continue writing “When will cars drive themselves? Experts say nobody knows” articles. If you ignore Waymo it’s a race; if you include Waymo it’s not. This article makes the further error of assuming “AI” is required for self-driving cars, rather than just a shitload of code.

Nobody has self-driving cars, not even Waymo. Every "self-driving car" has "disengagements" (a euphemism for "it stopped working") periodically every few thousand miles. That's simply unacceptable and needs to be sorted out. If a human driver was performing a panic stop for no reason or hitting a pedestrian or another car every 3 months or so we would take away their license. Waymo is doing better than almost everyon…

I repeat this again and again, and am always surprised that nobody else likes to raise the unfair comparisons.

I am willing to assume that waymo/Tesla etc have found the holy Grail, and made a driverless car safer than the average current human driver on the road. I'm not convinced that we should still allow these driverless cars on the road with no human backups. Not because of expectations or liability, but because the benchmark is wrong. We should compare them to a human assisted by the same driverless technology being used passively to prevent accidents. And it's not clear if a driverless car with no human backup is safer than a human assisted by the same technology.

Re: Self-driving cars are headed toward an AI roadblock

#215

Post by rohit2412 » Fri, Jul 06, 2018, 12:52 AM UTC

the8472 wrote: post #159

rifung wrote: post #149

Earlier quoted context omitted.

> So put them on the road. Let's give it a try. Maybe it works. Maybe it doesn't and people die and they sue. That's been part of driving since day one. Roll the dice. I'm sick of the endless debates. Well I'd certainly hope that there is a high level of confidence that they don't kill people BEFORE they're put on the road with other people.

All we need is that they don't kill more people than average human drivers do per passenger-mile. Assuming ride-sharing we could even let them be a little more deadly than human drivers and still save lives.

No, that's a wrong benchmark. We shouldn't compare driverless cars with the average human driver. We should compare them with the average human driver assisted by the best technology (which is say waymo's driverless software running passively to prevent accidents).

Re: Self-driving cars are headed toward an AI roadblock

#216

Post by CPLX » Fri, Jul 06, 2018, 1:01 AM UTC

jonknee wrote: post #178

CPLX wrote: post #158

Earlier quoted context omitted.

Why can't they just have normal cars for sale that also drive themselves to start with? Like I have a car and recline and read a book on a commute most of the time. If snow gets crazy or there's very strange construction or a flooded road or a flat tire we go back to manual driving. That's plausible and incremental. No cleaning problems, no homeless people living in them, no need for remote monitoring, a far simpler…

> That's plausible and incremental. No cleaning problems, no homeless people living in them, no need for remote monitoring, a far simpler liability and insurance situation, and so on. These are all solved problems, I mean when was the last time you saw a homeless person camped out in a Zip Car? BMW's ReachNow vehicles are all over Seattle and they are clean, insured, fueled/charged and exactly 0 people live in them.

I live in Brooklyn, where ReachNow launched with much excitement and pulled out within a year or so due to problems with maintenance, damage to cars, and difficulty with the technology.

Re: Self-driving cars are headed toward an AI roadblock

#217

Post by rohit2412 » Fri, Jul 06, 2018, 1:14 AM UTC

freerobby wrote: post #137

Zhenya wrote: post #125

Earlier quoted context omitted.

But the other companies aren't imitators even if what you're saying holds true. They have products that did, before or at the same time, similar functions. Did they actually copy Tesla once the functionality was released on the Model S? The answer is 'no'.

> They have products that did, before or at the same time, similar functions. I don't think that's accurate. Autosteer is the fundamental innovation of Autopilot, and no other manufacturer released a system that "did, before or at the same time, similar functions" as Autosteer. Other manufacturers have since released systems that technically check the box of turning the wheel. But the value of a system isn't how it's…

Taking lane keeping assists and calling them autosteer is innovation?

Re: Self-driving cars are headed toward an AI roadblock

#218

Post by rohit2412 » Fri, Jul 06, 2018, 1:17 AM UTC

zwieback wrote: post #38

How much better than human drivers do self-driving cars have to be to succeed? If you take a metric like fatalities or damage per mile I'm guessing we ask our robots to be much better than humans, even though it doesn't really make sense. Over time the differential will come down but a better intermediate strategy would be to keep humans away from robot cars until collision avoidance techniques improve substantially.

The driverless technology has to be better than an average human assisted by the same technology. In over words, the human has to add no value in safety, and then can be allowed to not drive.

Notice it is not the average human on the road right now, but the average human assisted by the same magical driverless technology that is developed.

Re: Self-driving cars are headed toward an AI roadblock

#219

Post by notahacker » Fri, Jul 06, 2018, 1:25 AM UTC

vntok wrote: post #206

notahacker wrote: post #168

Earlier quoted context omitted.

They don't look remotely similar to me, especially considering that black, blue, navy, grey and khaki (and reflective yellow for people on roads) are not unpopular colours for clothes worn by people who are not police officers or otherwise involved in traffic direction, many of whom may have cause to move their arms in ways comparable to signals. Humans are just a lot better at gauging intention and even simple stuff…

> They don't look remotely similar to me All of these agents follow a similar pattern of clothing (as I said, combination of similar garnments and colours) and behaviour (placement on the road, gesturing with authority, directly facing the car, etc.). Machine learning algorithms are especially good at recognizing patterns and storing their abstracted form, so it should come as no surprise that understanding what a po…

> All of these agents follow a similar pattern of clothing (as I said, combination of similar garnments and colours) and behaviour (placement on the road, gesturing with authority, directly facing the car, etc.)

Repeating an assertion does not make it cease to be false. A very small handful of pictures you linked to shows a wide variety of coats, vests and shirts of many different colours, all of which heavily overlap with general garment types and colours used in everyday clothing which tend to indicate police only with small and greatly varying trim detailing (and sometimes hats). And is the clothing and trim designed to convey authority? isn't the sort of abstract pattern recognition computers do better than humans, or even at all well. Sure, you could certainly create specific police uniform training sets for every jurisdiction and possibly even cut down false positives in other jurisdictions by geofencing them (so you don't get people stopping in California for commuters wearing the distinctive er... blue shirts and black trousers of the Hong Kong police) but it's a non-trivial undertaking even if there are bigger problems for SDVs to tackle

More importantly, unlike humans evolved to have an intimate understanding of human mannerisms, machine learning has no concept of "gesturing with authority", beyond whether moving human shapes fit very specific patterns within its calibration parameters, and police officers often don't have scope to place themselves in a particular position in order to get the car to understand them.

> Google seems to understand these intentions well enough, and at a much higher level than mere word parsing.

The video shows examples of predicting possible directions of travel of moving road users based on maps and movements (i.e. its fundamental driving model) and a shot of it recognising two arm gestures in an idealised front on positions. Neither fall under the scope of being able to understand how the traffic policemen intends to clear the blocked intersection from his shouts and gesticulations at you and various other vehicles. Humans also don't need to be signalled to go again if the black-jacketed man they've stopped for was actually trying to hail the taxi behind them.

Re: Self-driving cars are headed toward an AI roadblock

#220

Post by dang » Fri, Jul 06, 2018, 1:28 AM UTC

phyzome wrote: post #139

dang wrote: post #123

Earlier quoted context omitted.

Please don't post like this to HN. It degrades discussion and encourages worse. Downward-spiraling is the internet default, so effort is needed to avoid it. https://news.ycombinator.com/newsguidelines.html

Can you tell me more about downward-spiraling, and how this relates to the guidelines? The guy in the article seems to be implying that pedestrians don't have the right to be in the road. I think it's pretty OK to be dismissive of that.

"Seriously fuck that guy" is low-grade internet sludge. We don't need that here. If you post it, others will post more, and we will get all sludgey. That's point #1.

Point #2 is more subtle and more important. You picked an uncharitable interpretation in order to get snippy at it. That's a self-referential activity, not the thoughtful engagement we're looking for. This is why the site guidelines say:

"Please respond to the strongest plausible interpretation of what someone says, not a weaker one that's easier to criticize."

That applies to articles and quotes as much as to fellow commenters. If people just rail against weak versions of what others say, the only ones interested in reading it will be others in the same mood. Everyone else will get bored and leave, which in the long (or maybe not so long) run amounts to the death of this community. That's a downward spiral we really don't want, so please make an effort not to do that.

In short, please follow the rules: https://news.ycombinator.com/newsguidelines.html