Hacker News Reader

Comment by spookthesunset | Hacker News Reader

What about dirt forest service roads? What about the dirt parking lot at that wedding you just went to? How will it find a parking spot? How will it get into a ferry where people direct you to the correct spot?

How will it deal with an accident up ahead where some drunk bystander is trying to direct traffic? How will it know to ignore the drunk guy? What if it isn’t a drunk guy but a sober person directing traffic? Does the car obey in that case?

None of those are edge cases because every time it drives it will encounter some novel edge case that has never happened before and it will have to perform better than a human.

Don’t even get started with liability. Once you take away the steering wheel the manufacturer is on the hook for every single mistake and every single accident. You’d be insane to be a manufacturer and sign up for that.

Sorry, but self driving cars are a complete fantasy.

View on HN

Most people wouldn't instantly know how to handle those cases either. Many people would obey the drunk guy. Maybe that's the right thing to do.

None of those are edge cases because every time it drives it will encounter some novel edge case that has never happened before and it will have to perform better than a human.

That's why these things could never be rule based, there are too many small exceptions. When they're not overfitting/able to memorize all your training examples, neural nets learn heuristics, just as people do. Different people learn different heuristics. Granted, they don't have much of the same context about the world that people do, it will take a long time to build enough examples for them to infer all of that. But Tesla's fleet is getting more driving experience every day than you will get in your entire life, and every time they train on one of those exceptions, the entire fleet will benefit.

If drivers no longer carry their own insurance, this is probably going to be handled by insurance at the manufacturer level, and baked into the price. The insurance will demand certain processes to prevent large-scale bugs being rolled out.

I don't see any fantasy here, just a lot of work.

But Tesla's fleet is getting more driving experience every day than you will get in your entire life, and every time they train on one of those exceptions, the entire fleet will benefit.

Humans can reason about things that haven't to them happened before. Today's machine learning systems cannot. As you say, to react appropriately they must have been trained to do so using human annotated data.

The argument against FSD is that you would need an infinite number of annotated examples, and an infinite number of subroutines for behaving in any identified situations, because the space of driving is effectively infinite.

Until machines are able to do general reasoning about things they've not experienced before then FSD is not happening. By the way, Demis Hassabis thinks that this sort of transfer learning is the key to solving AI.

> But Tesla's fleet is getting more driving experience every day than you will get in your entire life, and every time they train on one of those exceptions, the entire fleet will benefit.

That’s not how machine learning works. You can “train on exceptions” as much as you want, and you have 0 guaranteed results. It can help, it can make no change or it can cause unexpected regressions.

It certainly can cause unexpected regressions, but of course that's being monitored during training. I'm not saying this is some real time learning system, nor that the system behaves perfectly on an exception case as soon as it's added to the training set. My point is that their engineers are getting an incredible amount of data on tricky driving situations to train their models on, far more than a human will see in their entire driving career, and when their model improves to deal with those, it improves for everyone. And at some point, these systems will be much better at driving than any human alive.

It’s one of great misunderstandings (or more likely - very effective way of getting money) to claim that machine learning is a data problem. We’re so far away from that point, that we have literally no idea what’s needed to make ML a data problem. Algorithms are extremely simple, and it’s all more or less curve fitting.

Media (and surprising amount of tech people as well) tend to claim that ML learning is like human learning - repeat something enough times and you’re done, you know how to do it. ML is no where close to that point.

You have any idea what the current state is of neural networks and their practical use? Most work only in very predictable controlled environments. Once you add one new ingredient things will fall apart. A 'complex' ai currently has less than 100 different things to deal with. And you need a hell of a fast machine to simulate millions of gens. In case of a self driving car you need approximately 1.000.000.000 different inputs. If you would the Moore law to predict the moment we deal with that amount you would have to wait about 500 years.

I do have some idea, I use them in my work. What are you calling inputs here? And gens?

> But Tesla's fleet is getting more driving experience every day than you will get in your entire life, and every time they train on one of those exceptions, the entire fleet will benefit.

Prove they are getting better. They run into the sides of trucks and off ramps on the freeway quite often (and don’t you dare blame the driver... it’s “full self driving” remember?)

Can you prove a machine learning algorithm does the right thing in novel cases it hasn’t encountered before? Nope.

And also, don’t reply with “well can humans”. That is a lame rebuttal. Computers will be held to an almost 100% non-failure standard before society accepts them. And that will never happen because of, well, reality.

The big reasons I want a self driving car is because I want to eliminate the boring drudgery and make the routine parts of driving safer. If I have to drive the last 1% myself, it's not a big deal. I suspect this is how most people feel.

I haven't done the math, but time spent on forest service roads, wedding parking lots, and boarding ferries is quite low overall. And the idea of driving in those instances doesn't bother me. Having the car take care of the other 99.999% of my driving live is what I do care about.

> None of those are edge cases because every time it drives it will encounter some novel edge case that has never happened before and it will have to perform better than a human.

If most of your driving involves drunks directing traffic, forest roads, and dirt-lot weddings then a self-driving car is likely not for you.

> If I have to drive the last 1% myself, it's not a big deal. I suspect this is how most people feel.

Which is all well and good until you realize a few things:

1) that “last 1%” happens every trip at any time. You will always encounter an edge case the machine cannot handle. Period.

2) as a result you have to always pay attention in order to immediately take over

3) you can’t because you (the royal you) are three sheets to the wind plastered drunk.

Sorry. If I have to pay attention for that 1%, it ain’t full self driving. And anything that encourages you not to pay attention 100% of the time is unsafe and shouldn’t be allowed in the road. And if I have to pay attention 100% of the time in order to take over, what the fuck is the point?

Not necessarily.

Take Japanese highway system. It is well maintained, has “rest” stations where you can park car for free.

If you could get an app that can drive drunk salaryman, or not even drunk just tired from Tokyo interchange to closest rest area to wherever.. you will win. Nobody will buy a car without that who uses car for highway driving, period. It is a killer app. Get off work at 9pm Friday, get the car to the IC punch the destination, wake up at a rest are 20mins from Ski resort, Onden, parents house etc

Snow, taiphoon coming whatever. Park at closest rest area.

I am very pessimistic about cars without steering wheels. I am quite optimistic about cars that have ability to drive well marked roads. Here is a crazy thing Charge fairs for highways like Japan does, then maintain them.

The safe disengage problem is solvable to match human level outcomes. Unexpected event? Slow down, avoid obstacles, flashing lights, slowly go to a spot that's ok-ish for parking.

Yes there can be a big hole in the road anytime, the AI has to watch the vehicle in front of it, if no such vehicle then it should chose a speed that let's it evaluate road conditions for the given weather/visibility.

Vehicle coming into our lane? The AI has to match human level maneuvering to evade the incoming car. It already has much better chance given it won't panic, will be always fully alert, and will be as accurate and precise as it can.

So the big categories are sudden road/environment changes (tree falls on road, hail, mudslide, earthquake damages road, animal crosses road), other vehicles, and pedestrians/cyclists/etc.

All are manageable with inferences from the environment (weather and roadside context determine visibility and how much space there is for maneuvering, how likely are unexpected crossings - eg. deer, kids) and surrounding traffic.

Are these hard? Sure, but none require human level cognitive reasoning.

> that “last 1%” happens every trip at any time.

Are you are going to randomly have to park in a wedding lot? The only example you gave which might happen in the middle of a trip, the Tesla can handle just fine for long enough to pass control over to the driver.

> you can’t because you (the royal you) are three sheets to the wind plastered drunk.

It's still the driver's responsibility to drive sober. Even so, I'd far rather someone who is drunk be behind the wheel of a self driving car than otherwise.

And if you have not been doing the driving, it will take precious seconds - too long - to assess the situation and react appropriately.

Either you're driving, or you're a passenger in a car with a driver who seems reliable, but really isn't totally so. Eventually, that will become a winning bet, but when?

Most realistic comment in the entire thread.

The "edge case" I worry about is driving on ice or snow.

Driving safely on frozen surfaces is not a solved problem for human drivers, but most of us insist it is a perfectly reasonable thing to do.

If the vehicle drives as slowly as it should in those conditions, it would probably frustrate a lot of people who really depend on their false sense of invincibility.

Ice or snow should be WAY easier for a self-driving car than a human (especially one with no or little experience).

A car has access to all 4 wheel sensors independently, it can apply brakes on each 4 wheels independently. It can always turn the steering wheel in the right direction, and it wouldn't panic.

Also, it would always drive the 'right' speed limit... sure, other human drivers might get annoyed, but assuming the true 'full self driving' future happens, there shouldn't be many of them on the road in time anyways.

> Driving safely on frozen surfaces is not a solved problem for human drivers, but most of us insist it is a perfectly reasonable thing to do.

What does that mean? That it actually is solved by humans? Or that it's not, and we just take the increased risk? If the former, then we can automated it. If the latter, then the self driving cars can also drive at increased risk, opt-in of course.

I meant the latter, but "equally safe as the status quo" might not be sufficient here.

If we built planes that were only as safe as highway driving, people would be outraged.

We aren't exactly rational about this stuff, and we expect a lot more from machines we don't directly control.

But I like your opt in solution. Maybe the UI loudly complains about the risk of current conditions and sticks to <5 MPH, unless the user enables "never tell me the odds" mode.

Most people never see a dirt road all year round. Don't let perfect be the enemy of good enough (ie SDCs in urban areas).

Reply on news.ycombinator.com