Hacker News Reader

Comment by JCM9 | Hacker News Reader

Fully self driving cars are far further off in the future than most in the sector would publicly admit. In controlled environments there’s a lot of potential today and things like self driving trams at airports are nothing new. But self driving in the “real world” mixed with humans and all the crazy one-off things that happen in such scenarios is not something that will happen at scale anytime soon.

There are still far too many edge cases to work out and for all the amazing things that “AI and ML” can do, handling edge cases is something this technology is actually quite terrible at relative to humans.

View on HN

Small, route-constrained mass transit vehicles would solve 80% of this with 20% of the cost. We could call them trollies or buses.

And 10% of the convenience, comfort and utility of having your own car.

Cities already have buses. Rural areas don’t want them.

I wonder how good “public transport” could be if it matched the money inflows of private cars. Possibly amazing.

Think customisable bus routes, cheap subsidised taxis, a fast bus every small-n minutes and so on.

Clearly we won’t find out as it’s a massive sunk cost, but I think this alternative reality could be great.

Having lived in places with both amazing and terrible public transport, I understand the skeptics, but also know just how amazing public transport can be.

In London the cost of public transport is comparable to running a car. The end result is meh. It works ok but with the usual drawbacks.

Yes, I wouldn't put London on the "great" pile. Adequate. Tube is OK for longer distance, but the friction of getting on and off is a pain. Buses have to crawl through heavy traffic and pedestrians everywhere.

Warsaw in Poland IMHO has very good public transport. Reliable, timely, used to be cheap when I lived there (not sure how much it went up). Never struggled to get from A to B.

Idk what prices you are paying.

It varies but a zone 1/2 annual travel pass is £1,444 presently.

Why don't we have an Uber for buses yet? Has this been tried anywhere? Not necessarily a for-profit company, but an app that people could use that would offer potentially better routes / trip times. So the bus wouldn't have to hit locations where no-one was getting picked up / dropped off, etc..

It has been tried in a few places. Here's one: https://www.cbc.ca/news/business/uber-lyft-ride-hailing-on-d...

I'm using these a couple of times a month: https://www.moia.io/en

> Rural areas don’t want them.

Da'fuq? Rural areas love and use the heck out of buses - school buses.

As for me, I'd love to have a bus which came within 3 miles of my "rural" neighborhood. Save on gas, insurance, and be able to read when I have to commute to the office?

I can't remember the last time I used a school bus to get to the grocery store or park

I know. It's strange how they have such great coverage for kids, but decide that it's not worth it for adults with adult money.

I think that's the key. If self-driving cars ever reach a level of ubiquity, it's going to be on large, open roads and highways. Urban areas will still require a human driving.

Excellent vehicles for facilitating a pandemic - packing many passengers from different social circles together in an enclosed space.

Not if the air is filtered.

https://jamanetwork.com/journals/jama/article-abstract/27714...

Full self-driving cars limited to highways is probably a much more reasonable goal. A lot of the issues seem to be with intersections and pedestrian crossings - things that are absent on highways. Plus, infrastructure to assist self-driving cars like beacons on off ramps could feasibly be added.

"infrastructure to assist self-driving cars"

That would seem to be a good investment for the industry as a whole, yet the trade group doesn't seem very visible. Possibly the sort of thing that should have been in BBB.

I'll go a step further and say they aren't just further off but impossible to achieve until someone can reasonably prove otherwise. We simply aren't going to see cars without a steering wheel and manual controls (i.e. Level 5 automation) for general purpose use on public streets in our lifetime.

Yep. I would argue that driving is AI complete - in the sense that only true Artificial General Intelligence can do it, and so do everything else to.

Perhaps this is because it requires cooperating with other General Intelligences (i.e. us) in an open world.

I'm still mystified as to why no country has switched over to ultra-light rail. The costs of asphalt projects seem to dwarf the cost of doing anything reasonable and leave AI as an exercise for the over confident.

https://www.youtube.com/watch?v=tBJ0GvsQeak

can't you get a ride in phoenix like today with waymo? or is not open to the public?

I mean I understand that their AI is not perfect but it doesn't seem like a pipe dream

> Fully self driving cars are far further off in the future than most in the sector would publicly admit.

Full self driving have NOT been solved for highways in a desert.

This is literally the easiest real world scenario imaginable. Highly and strictly regulated environment, and no obscurities like snow happening. I don't think there is a hope we will get any close to self drive with current technology.

Is the sudden object in front of me a block of concrete that fell of or stray page from newspaper floating by on a wind gust? Hard bake or ignore it?

We cant possibly capture all edge cases of real life. You need systems that have cognitive capabilities to deal with real life

Or optionally different type of transportation that is not so complex to navigate and understand. There are already fully automated metro lines. Maybe solution is to change the problem itself.

Average human is actually quite terrible in driving cars, as evidenced by the global death toll numbering in dosens of thousands every year. Even halving this number would be a win, losing a lot of people to such edge cases. However, the nature of the problem would not allow for it, because every death from a self-driving car will get excessive coverage and public outrage, and people will prefer losing far more people to human errors than fewer to computer errors.

Put into scale, there were 1.11 deaths per 100,000,000 miles driven in the United States in 2019. The average human is actually surprisingly good at driving cars, we just drive 3.9 trillion miles per year so the absolute numbers of fatalities are enormous.

I say this as a person who genuinely feels like people are generally terrible at driving.

Should we improve these numbers? Of course. But AI-driven cars will have to beat a pretty high floor of capability here. It's entirely possible they're already there but it's difficult to know for sure given the biased way the numbers are calculated. The manufacturers claim that their systems are much safer drivers than humans, but those systems also hand control back to the human driver during difficult or unpredictable situations. It's easy for the statistics to swing in AI's favor when they only have to count the "easy" majority of miles while shifting blame back to humans for the hard parts.

I don’t think that’s the case- I don’t think there’s any evidence that self driving cars are better drivers than human drivers in real world conditions.

If this were only a human psychology problem, someone would just dump a fleet of self driving cars on the road and prove they have the tech. The rewards are too good to ignore and surpass any potential fine/regulation, if the tech is good enough.

But it’s not there yet, it might never get there. We’re so far off general AI it’s unreal, e.g. a human can drive a car with a single eye open, and one hand on the wheel, and another on a throttle- all from the drivers seat. Even with sensors and data coming in from all over the car, AI can’t manage it.

I didn't take the parent to be saying that self-driving is better, just that humans are really surprisingly bad, and that's the bar we should be measuring against, not an ideal world where no one dies in traffic accidents.

The main ethical issue of 'deaths due to self-driving cars' is because of who being at fault?

If the car is self-driving, and truly self driving - no steering wheel, when it causes an accident, who is at fault?

Humans make mistakes, but humans also can change their response easily to a given input. Is Tesla at fault? Do we just issue them a fine for a person's life? Should the car do its best to preserve the life inside the vehicle, or the lives outside the vehicle? Automobiles do significant damage to pedestrians in regards to their dangerous design.

Surely corporations should be held to a higher standard than people, as they're not bound by the same restrictions a single person is.

I don't see how that's any different of an issue for when any machine malfunctions and kills someone. Was the design or manufacture of the machine negligent in some way? Was it a failure mode that we've decided is rare enough that no one is at fault?

Self-driving cars may be new, but deaths caused by machines are as old as machines.

Machines in a business setting when they cause death, often it is due to improper safety protocol for said machines.

Consumer devices don't have this level of control over a person's life anywhere near as often, in the sense that the device has your life at stake and makes decisions for you.

"no one is at fault" is entirely counter intuitive to engineering principles. When a rare accident happens, the engineer is often responsible.

We don't 'build' AI in anywhere remotely the same way as buildings or legacy software products, there isn't a 'designer' for said AI system other than the training model - which can certainly not account for every possible edge case.

Self-driving cars can't 'only be \better\ than humans'. Because only humans can currently drive. Self-driving cars need to be nearly 100% safe when you entirely remove the human driver.

When machines malfunction and kill people, the manufacturers pay. If it was gross negligence that killed someone, someone goes to jail, unless they're the dead one.

Every day, our homes don't catch fire despite all of the electricity running through the walls, our combustion engines don't explode, our batteries don't explode, our buildings don't collapse. The modern world is full of things that could kill us and don't.

I didn't say no one is at fault, the manufacture is at fault if they built the vehicle in a way that we deem was wrong enough to deserve fault.

And we get to decide who's liable, just because you can act like the AI is sentient doesn't mean you can't make the company who wrote the code responsible for the actions of the AI.

And if we wait until self-driving cars are 100% safe, hundreds of thousands of preventable deaths will happen in the meanwhile.

Isn't that a bit of a false dichotomy? GP is specifically calling out that self driving technology has limitations (and specifically as they pertain to car accidents being edge cases in the first place). Self driving tech hasn't proven itself to be better than human drivers outside of controlled environments, nor are these two options the only options available (e.g. more investment in public transit and alternate forms of transport come to mind as being very closely related to the problem space)

By all means, improve the tech and show us that it can in fact put out better-than-status-quo safety numbers, but it's important to not confuse potential with wishful thinking.

Humans are quite good at driving cars, as evidenced by the death toll per million km. In fair comparisons self-driving cars aren't even close.

Yes people would prefer to have agency in their lives and die based on their own decisions instead of a nameless programmer (a large fraction of the victims have themselves to blame for their death, this would change with self-driving cars).

> In fair comparisons self-driving cars aren't even close.

Please reference a fair comparison.

AFAICT, we only have the unfair comparisons from Tesla, which say that Autopilot has an accident rate 1/5th compared to when it's disabled, and that Full Self Driving Beta has 0 accidents total. Neither of those comparisons are fair, but that's the best we have, AFAICT.

All the more reason for Tesla to avoid sleazy sales tactics like calling adaptive cruise control "self driving", and claiming that cars will be able to drive themselves coast to coast before 2020.

People _might_ overreact to deaths from true self driving cars, but we don't know that yet because such vehicles don't exist.

Come back and make this comment when we actually have something better than human driving.

Human walking, human biking.

Humans are actually really good at driving - we had just assumed it was easy because anyone could do it.

Driving plays to our two key strengths - tool use, and cooperation. Indeed, the car becomes an extension of our body, and the road is our forum.

Don't forget all the sensory processing stuff that comes with being a predator that eats other mobile animals.

The main barrier is the state of the tech, not media outrage.

There is a gap between the current state of the tech, and computers being better at driving than humans, that may be impossible to bridge. In order to bridge it, as a society we would need to accept a period where these things start to become more common, but they make our roads less safe.

Only through trial-and-error in blood can this technology improve. I don't think we will collectively see it as a worthwhile experiment.

Must there be blood? How many people have died because of Waymo driving? How many because of Tesla or Uber?

The annual US death toll alone is dozens of thousands; the global number is well over a million.

This is a very underappreciated point. SO many people die from driving each year, but a lot of people treat a single death from self-driving cars as a reason to scrap the whole endeavor.

Reply on news.ycombinator.com