Hacker News Reader

It's not complicated. Traffic laws are not perfectly aligned with safety. Plenty of legal things are unsafe and plenty of things that are illegal are not always dangerous.

This isn't surprising either. The process of making laws is political and politics are clearly not an ideal way to decide on a perfect and exhaustive and precisely detailed set of rules for anything. The laws do not constitute an algorithm for driving. They are necessarily simplified. And yes, a self driving car is going to need a set of driving rules that are much more detailed than anything humans have written down before.

That seems like an appropriate generalized take on "laws" but in the specific case of ignoring a stop sign, I don't think it applies. Stop signs are a traffic control device that other drivers rely on to predict the intent of a vehicle approaching the stop sign.

Stop signs are anything but "ignored". The actual behavior only applies when the car recognizes a stop sign and also assesses good visibility to ensure that no other cars are around, and the car still slows regardless. It's not just blowing through stop signs at top speed whenever it feels like it.

Pragmatic rulemaking for large numbers of people also needs to take into account the fact that the rule will naturally be viewed as a loose guideline by some people, and even as a thing to be defied for its own sake by others. Setting the limit stricter than the actual desired behavior can push average behavior closer to the target.

Yes, that's a great reason why laws made for humans may not be perfect when applied strictly to machines.

> Setting the limit stricter than the actual desired behavior can push average behavior closer to the target

I hate this so much. Selective enforcement only serves to enable overzealous enforcement against unpopular and disenfranchised groups.

One does one small thing wrong and they can then tack a large handful of other charges that would never otherwise be enforced. It's disgusting.

Make the rules things that should actually be rules. Enforce them as written. The end.

> Setting the limit stricter than the actual desired behavior can push average behavior closer to the target.

Absolutely not, at least in the context of speed limits. Where I live, most of the speed limits are at least 15mph lower than they should be (why is a 6 lane road 30mph???), so you get:

* people who follow the speed limit exactly: 30mph

* people who follow the speed limit the road was designed for: 45mph

* people who see the limit as a "limit:" 25mph

It's a mess. Roads need to be designed to the desired speed limit. Don't make a 6-lane straight road 30mph; people unfamiliar with the area will assume the speed limit's 50% higher than it is, and people from the area will feel like they're going incredibly slow.

In short, it should be uncomfortable to go >= 20mph over by design if you want people to follow your speed limits.

It actually is incredibly complicated. You can tell, because every individual person you speak to will have a different interpretation of which rules are there for safety and which ones can be ignored.

The fact that you start your post about how an AI can make such a determination with the words "it's not complicated" discredits you.

Obviously rules for self driving cars are very complicated. What's not complicated are the reasons why the road laws made for humans are not directly suitable as unbreakable principles for self driving cars to always follow.

No actually that's incredibly complication. It differs by culture, regional attitudes, quality of infrastructure and hundreds of other variables.

The way it's different is incredibly complicated.

The assertion that a significant difference exists is not complicated and is easy to prove.

Traffic laws are draconian/inflexible because:

1) they have to be simple enough to fit in an AVERAGE DRIVER's head (it is a miracle to me that highways in cities are as functional and low-error as they are yes I know the death statistics on driving)

2) the law makers know the laws will be bent, so they set a baseline of VERY VERY SAFE and hope that most people only bend the law to VERY SAFE or at worst MOSTLY SAFE.

It seems weird given how "bad" self-driving algs are right now, but with a detailed database of locations and context, I absolutely believe the "rolling stop" logic in the car would be safer and more effective than a human observing complete stops.

I definitely do not agree with Tesla's vision-only no-location-context algorithm approach. Such a thing is a baseline, but if you know the location to where you are going, then route-specific data should be downloaded that has been optimized with multiple specific AI computation passes.

That's how humans work: we are slow and dumb when we are driving in unfamiliar locations and routes, but for our commute routes we know where potholes are, cracks in the road, which lane to be in, dangerous sidestreets and intersections and congestion points, etc.

So I disagree that L5 is impossible. Well, it depends on "way above human intelligence". I think what you are trying to say is you need a supersapient intellect with IQ 195.

But AIs can be "dumb" but have access to a far larger memory space and database and respond to a far wider bandwidth of sensor information. The AI programs may not be solving Fermat's Last Theorem in the margin of a book, but they can be smarter in the sense that it's like having a million gnomes watching at once.

> Well, it depends on "way above human intelligence". I think what you are trying to say is you need a supersapient intellect with IQ 195.

I'd say even "normal human intelligence" would suffice for at least getting L5 to work (maybe not with perfect safety and efficiency, but working at all). But you cannot remove the need for actual "intelligence" from the equation, and that's the problem: our AIs are not "intelligent" at all. They are idiot savants, and driving a car isn't exactly a problem where you need savant-like capabilities. It is a very general, broad-range problem, which is much more applicable for the broad mental capabilities of even the not-so-smart part of humanity.

Case in point: there are human savants in areas like playing chess or Go or recognizing text in many different spoken languages, but there are no human savants in the area of driving cars on public roads.

There are public driving savants. They just don’t see much benefit from it.

> the law makers know the laws will be bent

Or possibly they just know that real world driving involves so many context-dependent decisions that trying to cover it all with laws is about as reasonable as making a self driving car.

I think it's the nerds that are delusional.

A human breaks the rolling stop rule on a case by case basis and is responsible for the consequences, while a rolling-stop-avoidance feature acts on your behalf and Tesla is responsible for it. An analogy might be that it's the difference between getting food poisoning from home cooking vs. from a restaurant. Even though you're consuming food in both instances, the liability is assigned differently.

> A human breaks the rolling stop rule on a case by case basis and is responsible for the consequences

Not really. You can fail to see a pedestrian and cripple them and you won't see a whole lot of consequences for your "accident" (negligence).

Breaking rules on a case-by-case basis is also how the normalization of deviance occurs:

https://www.youtube.com/watch?v=Ljzj9Msli5o

If you break rules when you think you don't need to follow them you'll find that you're not 99.9% perfect and you will inevitably cause a crash.

If you follow the rules all the time you won't make that error.

The reason why computers should ultimately replace human drivers is precisely because we're dangerously incompetent and overconfident and shouldn't be breaking any rules.

Oh, according to Tesla, not so. Their perspective is that you're absolutely and solely liable in either case.

Their legal department works studiously to handwave away every marketing statement.

"The driver is only in the seat for legal purposes. The car is driving itself." - "You are entirely responsible for driving the vehicle."

"Use Summon to bring your car to you while dealing with a fussy child." - "Pay attention to vehicle at all times."

Easy. There is no contradiction. If I can think faster, see better, and react faster than you, I can break exactly the same traffic laws as you and still be a much safer driver. Computers can do all of those things currently, and it is a matter of time before that is applied to driving.

> think faster

Computers cannot "think" at any speed.

Fine. "Plan" faster, if you want to be picky.

> because they never tire and they never break traffic rules and thus never cause accidents

No, because they never tire and almost never break traffic rules in dangerous ways, and thus cause accidents much less frequently than humans. Realistically, this means that instead of following the speed limit they'd follow speed limit + X, instead of keeping the prescribed distance they'd keep whatever distance human drivers typically maintain (while still keeping enough distance to react and stop), and it also means rolling through 4-way stops when they have sufficient confidence that it is safe to do so.