Hacker News Reader

The way I think about is that when we drive as humans and an obstacle appears, we have over 15 years of knowledge that we have to determine our next step. And we coalesce a lifetime of knowledge in a matter of seconds.

These modern algorithms do not try to implement general AI. They do not work with 15 years of knowledge. They just recognize some patterns with no understanding of what they’re are looking at. It’s a different and easier approach but I think it has some severe limits that will never be overcome.

To me it’s like a software programmer that copies Stack Overflow snippets but doesn’t really understand programming. They get only so far

I would feel much safer having a more advanced AI (not the current ones) driving cars.

It's not like humans get better at driving with each year of driving. Some become more reckless, people become old, some might think they can have some drinks before sitting behind the wheel and it takes 15 years to get 15 years of experience. What about that over-confident 19 year old who wants to show off their "skills"?

And 15 years of one driver's perception of safe driving is nothing compared to swarm intelligence of AI for driving. If you have 10,000 autonomous vehicles driving for a single day and combine that data at the end of that day to improve driving, that's collectively over 25 years (assuming the human equivalent drives each day of the year) of "knowledge" in one single day. Every day.

Being a father with 2 small kids who drives their kids by bike to kindergarten / school - with a large section of that road is shared by SUVs, vans and bikes (while each side is blocked by parked cars or narrow garage exits built into the building), I can't wait for safe, autonomous vehicles to become the norm.

> I would feel much safer having a more advanced AI (not the current ones) driving cars.

You shouldn't, there is no AI capable of doing better than a human, and there does not appear to be any currently known method of creating one.

> I can't wait for safe, autonomous vehicles to become the norm.

I wish I could tell you different, but the reality is you could wait the rest of your life and it would not happen.

I remember people saying no computer could ever beat the best humans at chess.

Would you be willing to play a computer at chess if there was a possibility on any given move that it would squash you like a pancake?

There was a kid with a broken finger recently...

I mean, you aren't wrong. But we can almost certainly do better than where we currently are.

There's a possibility that all the particles in my body might leap simultaneously one foot to the left, or that I might transform into a bowl of petunias. The question is how likely playing chess or riding in a car might squash me like a pancake.

We accept a not insignificant risk by driving ourselves and riding in cars driven by humans. Happily, it's not hard to estimate the risk that a machine-controlled car poses by letting them roam around. We can compare that with human-controlled cars and start shifting between controllers according to their risk profiles. I'd much rather have a machine-controlled car than a human asleep at the wheel.

When do you remember this from?

In the mid 1970's, computers playing chess had in 20 years gone from being utterly trivial to beat if you know the rules, to easily beating amateurs and playing mildly interesting games against top humans like Greenblatt vs Bobby Fischer in '77.

And it was clear that the power of the chess computers would keep scaling with Moore's law as they could look further into the future and store more of a complete opening book.

In the late 80's the first versions of Deep Blue were playing serious games against Kasparov under standard time control, even though it took until 1997 to beat him - one can argue that Kasparov was also still developing during this time, since his rating peaked in 1999.

Can you find a citation from after 1980 from someone with a bit of understanding of chess and computers, saying that computers would never beat humans?

In the 70s, there was widespread misunderstanding of what artificial intelligence would be good at. Take a look at Hubert Dreyfus' _What Computers Can't Do_ (1972). To be fair, he was correct in some respects, but I think with flawed reasoning.

> You shouldn't, there is no AI capable of doing better than a human, and there does not appear to be any currently known method of creating one.

I don’t care that much if it’s better than the best driver, I care if it’s better than the driver checking their phone, blowing through stop signs, etc. As a cyclist, I reckon that it’s the most selfish 10% of drivers that represent 90% of the risk; if they’re replaced by self-driving I’d feel safer even if it also means the good drivers aren’t quite as good.

The ones which are not checking their phones can't detect semis or cyclists for quite some time. So You're replacing the 10% with their electronic versions.

But if AI can be better than humans at driving a car, then it can also be better than you at driving a bicycle. Will you accept AI-augmented tools on your bike? I see automatic braking and speed limiters as very easy first steps. Then geofencing to prevent bicycles on pedestrian paths. All of that is dead simple with curret tech.

Bicyles aren't very dangerous. Cars, on the other hand, are a leading cause of death in basically every country on earth.

Yeah, you only have a ~seven times higher rate of death per mile traveled as compared to a car. It’s just there are significantly more car miles traveled so more car related deaths.

statistics…

... 'death rate', as in, the cyclist is killed by a car.

Check your stats. In most of the largest countries (India, China, Russia etc) buses are actually more dangerous than cars. One bus crash can be more dangerous than dozens of car crashes.

And my phone just had a weather alert. -35c tomorrow, possible -50c with wind chill. I'm sure I'll see not a single bicycle ... but definitely a few snow mobiles. They are ridiculously dangerous and polluting, but at least they aren't cars.

I would imagine that a very large portion of all deadly crashes (buses included) involve a car.

Cars are pretty safe for their drivers; they are just very dangerous for everybody else. So, if you split 'passenger deaths' by 'vehicle category' you're going to have a lot of situations where a car hits a bus/cyclist/pedestrian, the driver survives, and whoever they hit does not. This creates misleading statistics.

> I care if it’s better than the driver checking their phone, blowing through stop signs, etc. As a cyclist, I reckon that it’s the most selfish 10% of drivers that represent 90% of the risk

We have a(n imperfect) mechanism to catch and penalise those drivers.

An ex-colleague of mine was incapable of sticking to the speed limit (she drove me round Paris once, that was an exciting ride...). Eventually she collected enough speeding tickets [in the UK] that she was within one ticket of losing her driving licence completely. At that point, she [finally!] slowed down.

Q: What's our proposed mechanism for detecting errors made by AI "drivers" and enforcing improvement?

> What's our proposed mechanism for detecting errors made by AI "drivers" and enforcing improvement?

Lawsuits by the families of the fallen?

Have be a big lawsuits though — an 80 year old taken out by a robocar is just “the cost of doing business” so there would have to be multiple school busses on fire to make a dent.

> Lawsuits by the families of the fallen?

Umm, maybe.

Except we all know that that's a truly terrible way to persuade corporations (with deep pockets, lots of lawyers, and even more lobbyists) to do the right thing, isn't it?

I didn't say "best driver" I said better than a human - even an average human.

I'll copy/paste something I wrote before:

The computer would have to be 99.99999% reliable to simply match humans.

The accident rate is around 74 per 100 million miles (and fatalities is 1.13).

It's unclear exactly how to turn that into a percentage, but no matter how you do it it's quite high.

Say an accident takes 5 minutes, and people drive 30 miles/hour. Then that works out to 99.999% for humans. If you use the numbers for fatalities then it's 99.99999%.

I.e. 99.99999% of the time, as whole across all [US] humans, people drive in a way that does not cause a fatality.

Do you know any computers that can simply run with 5 9's of reliability? Never mind actually do something complicated?

Your line of reasoning (which is very common on this topic) is just fascinating to me.

You start with the assumption that it's possible to build something almost perfect, far better than we have today. Then you proceed to say that if we had this thing it would make life much better than today. Therefore we should work hard to build this thing.

I mean, if you replace "self-driving cars" with "anti-gravity technology", your line of reasoning can be applied just the same. But nobody is pouring billions into anti gravity startups?

No, it doesn't have to be 'almost perfect'. Significantly better than what we have now (people with emotions and imperfect decision-making, ranging from 19 year-old newbie drivers, 45 year old over-confident CEOs with their SUVs driving in the city to 82 year old drivers who can barely walk so would rather drive) is already a HUGE improvement.

"self-driving cars is" not the same as "anti-gravity technology". People - including many innocent people - get killed every day because of reckless driving or human mistakes. People-driven cars also needs more infrastructure - more signs, more roadblocks / infrastructure, cluttering up the city, making it less human-friendly. Cars clog up the city, because people don't want to drive too far to work.

Concerning safety and making cities more livable, I only see upsides for autonomous cars. People can work in their cars coming from outside the city. The cars are safer and more reliable. There's less cluttering in the city (e.g. an intersection with 10 signs for each corner), because cars can also communicate in non-visual ways.

Plus we've made significant headway and it's not a fantasy anymore.

"anti-gravity technology"... is just a random technology. A solution looking for problems to solve, I suppose. With positive and negative consequences.

> "anti-gravity technology"... is just a random technology. A solution looking for problems to solve, I suppose.

If cars drive above people instead of next to them that would also solve that problem.

I'm saying, an AI significantly better than what humans can do today, is just as much a fantasy as anti-gravity technology.

People get injured and killed every day because they don't have anti-gravity tech and are trying to lift or move heavy objects.

You can make all the same arguments.

And most of the "people are stupid" problems with cars are better solved by other tech like driver monitoring - Volvo is already deploying this.

Self driving cars/buses is just a fairytale used by Tesla/Uber/etc. tech bros to try and defend their valuations.

> I'm saying, an AI significantly better than what humans can do today, is just as much a fantasy as anti-gravity technology.

I think this is a key thing where we have differing opinions. I believe an AI taught to drive safely will be significantly better than humans at it. It's not a general purpose AI, just one that is taught to be better at driving. An AI doesn't get distracted as easily. They don't look to the sides for 2 seconds while driving 125 mph on the freeway. They don't check their emails while on a busy city road. They don't need 1-2 seconds to react to a baby stroller emerging from a line of parked cars.

My mind has slowly accepting the idea that "specific purpose AI" is already here. Working with an "AI-infused" coding assistant (Github Copilot), it took me a few weeks to get used to. But now I feel it can complete some of my coding assignments at a junior coder level (good effort, but I still need to check the work, sometimes it's non-sense, but many times it's pretty good). I can already off-load some of my routine work to this assistant.

> an AI taught to drive safely will be significantly better than humans at it.

That would depend on which human you have in mind. Some are 100 x more dangerous than others. (See the link below.) Or 100 x safer depending on what you consider normal

You don't happen to know about any statistics about fatalities per 1 billion vehicle-km for (partly) self driving cars?

To compare with https://en.m.wikipedia.org/wiki/List_of_countries_by_traffic...

Edit: i found:

> Overall, autonomous vehicles (AVs) were involved in more crashes: 9.1 crashes per million miles traveled, compared to 4.1 for conventional cars. However, compared to injuries experienced in traditional vehicle collisions, the ones involving injury were minor.

https://1800injured.care/self-driving-car-accident-statistic...

If that's reliable info, then they'll eventually get better than an average human in the US (or are already).

And in some countries, they could be an amazing improvement -- see the insane amount of fatalities, in some countries (Wikipedia link above).

"Significantly better than what we have now" is the hard problem to fix.

We can fix autonomous driving today if we give them dedicated roads not shared with humans (no cars, no bikes, no pedestrians). Works perfectly fine with dedicated tunnels under cities.

An emotional and imperfect decision-making 19 years old can pass a driving exam where they drive around for 45 minutes in a mixed traffic environment without the inspector being forced to take control of the car. This is kind of absolute minimum standard is one that autonomous cars struggle with, especially when people try to make it cheap by removing radar and lidar.

What we have made slight headway on is that under very strict environments, with more restrictions on the driving than human drivers has to follow, with expensive sensors, and with a extremely good access to maps and information, then some autonomous transportation is possible. They still causes accidents and people will still get killed. It is debatable if they will kill less people than if human drove under similar restricted environment with access to equally expensive hardware and information, but it is some progress.

Couldn’t agree more. The quest for self-driving cars seems to be a capitalistic endeavor to offer the convenience of chilling in your car, and platform for all the services one could consume in the time they’d otherwise be driving.

If safety was the first principle, then traffic science and urban planning would lead the way until systems are simple enough for self driving cars to take over. It doesn’t seem we are in the place (in the US) where this type of private-public collaboration can take place at meaningful scale.

I’m not knocking all of the brilliant work that has been done but peddling self-driving cars for safety (not saying OP is doing that) just seems like a version of “making the world better (to be rich).”

It’s closer to 3 decades today starting with European project: https://en.m.wikipedia.org/wiki/Eureka_Prometheus_Project

Looks like, that some problems can’t be solved just throwing money at them. Modern hardware is light years away from the one in Prometheus Project. Radars, lidars and cameras are superior now. Heck one can design and build dedicated ASIC for cheap nowadays. Maybe taking Nvidia Drive and tweaking it is just not enough to get there.

Autonomous driving is in "principally works but lawsuits haven't started yet" phase.

I love how you asked people to share a smart summary and analysis, but all you got was hackernews comment pontifications. Folks: pretty sure they meant a researched article that takes longer than a poop to write.

I'm interested in this too.

The Starsky postmortem 2-parter published on Medium reveals that: 1) true autonomy is worth less than a cup of coffee a day 2) safety isn't sexy 3) nobody knows which features sell.

Mathematically, zero human supervision driving is close to impossible.

It's simply almost impossible to guarantee, and convince society's institutions, a computer car will perform better than a human.

And if a human still has to be sitting in the driver's seat paying attention?

The financial justification flies out of the window.

More importantly, if the robo car IS ACTUALLY safer than average human drivers (which is a stupidly low bar because that includes drunk and distracted driving), but has extremely weird, unpredictable failure modes like suddenly driving off a bridge with no warning 1 out of 1,000,000 times, how safe do you feel riding in this car? People don't like working with systems that have utterly inscrutable error conditions. The devil you know is much better.

I've heard it said the single best thing that could be done for airline safety is getting rid of pilots since most crashes are the result of pilot error. Then ask yourself if you would get on a plane without anyone flying it.

Driving is one of the most dangerous activities people undertake on a regular basis, yet we are all numb to the risk since it's just normal. Changing a major variable in the risk profile of driving by introducing autonomous cars is a big deal since the failure modes are so foreign.

I think the autonomous vehicle resources should be more focused in the short term in making "uncrashable" cars as we're a long way off from true autonomy. Systems like automatic emergency braking that live in the background and act when the driver is making a serious error to prevent or dramatically reduce the impact of an event the car is certain will occur.

> I've heard it said the single best thing that could be done for airline safety is getting rid of pilots since most crashes are the result of pilot error. Then ask yourself if you would get on a plane without anyone flying it.

The first one? Probably not. But going but a lot of comments here and else where, there are plenty of other people willing to be early adopters. People already do some pretty stupid shit with Tesla Autopilot.

5 years in with statistics to show? Yes, I would.

Of course, that logic of removing pilots could be completely flawed based on the chance that the only reason the pilots failed at all was because they had to step in when the automated systems could not handle the situation.

The problem with autonomous driving is if you were to swap all cars on the road to autonomous ones and banned pedestrians and cyclists except for traffic light controlled crossings, it would go swimmingly.

Barring that, it's absolutely impossible. It's not a kind of problem where more powerful hardware in five or ten years will just magically solve it.

Cyclists, pedestrians but even cars in a parking -- it's not just a moving object, it's a human giving clues which we can read based on experience. You are veering into full artificial vision if you want to read this.

> The problem with autonomous driving is if you were to swap all cars on the road to autonomous ones and banned pedestrians and cyclists except for traffic light controlled crossings, it would go swimmingly

Would it? Teslas, the cars with most widely deployed autonomous driving system (unlike say Waymo which is still geographically fenced), get confused by regular things such as lights (stop lights getting interpreted as a red light) and vehicles in front.

Presumably with a full replacement we would also replace all the stop signs and lights.

Manifestly statistics and massive data does not replace reasoning