Just the other day my Tesla model 3 drove me home from a restaurant 10 miles away, at night, with plenty of other cars on the road, through side streets and on a highway, without intervention. Did it do weird things? Sure, but nothing that made me nervous enough to intervene. Compared to just a year ago when it was still phantom breaking on the highway and couldn’t do side streets at all. All the people saying it’s still decades away don’t realize how quickly things are progressing.
10 years is a long time and progress is not always linear. Think about what happened to the internet between 1995 and 2005.
We haven’t even begun to address emergent behaviors. You say your Tesla does “weird things.” What happens when all the cars on the road are Teslas and they do the same “weird things” at the same time (because they have the same programming)? This is an important aspect of systems engineering that we haven’t even scratched the surface of when it comes to self driving cars.
1. People "hacking" cars for fun/malice. This could be anywhere from "painting on the road/sign" to wireless comms penetration attempts or nation state attacks.
2. People who OWN the cars replacing firmware/hardware for perceived benefits. IE put back the "rolling stop" feature. Or "improve" behaviour/performance. (Car modders changing headlights in a Tesla might mess with the optical only approach) Or change the default decision in "trolley problem" events.
Okay, so you've got a private company explicitly writing down the value of human life in a big list, top to bottom. Let's pretend we fast forward the several years to decades for the public to accept this new reality. Now we realise that different locations operate under different value systems. In Saudi Arabia, one man is worth three women. You best believe they will require the car to run down the woman. So each country/location is going to require an ethics update, and it's all going to have to be public. Imagine the heat Tesla is going to take by complying with Saudi rules and explicitly programming their cars to mow down women instead of men.
Yet there has been no report of this exploit used maliciously in the wild.
There are definitely people in the industry who are thinking about this and working on it, but AFAIK none of them work at Tesla.
The cost, simplicity, and availability of self-driving cars will undoubtedly change the consumer.
Self-driving cars have been at 95% since the 90's. They've been at 98% since 2012. Humans are higher than 5 nines.
It turns out that human drivers, even poor ones, have a much better driving record than self-driving cars. With this latest Waymo experiment we may finally see some significant progress.
And then compare it to what happened to the internet between 2005 and 2022...
Progress isn't always linear but it tends towards diminishing returns more often than exponential progress.
And we're not talking simple things like "can billions of people and Moore's law result in something a bit more exciting than Facebook with a newsfeed and ads and YouTube without Flash", we're talking about progress in the form of monotonic improvement in complex intractable software to solve edge cases the dev team haven't even thought of where the failure mode involves corpses, and do it to the satisfaction of regulators across many jurisdictions.
The gap between "drive 10 miles without serious incident this time" and "consistently drives a billion miles between incidents serious enough to be fatal without any possibility of intervention" is enormous.
I hope I'm wrong though, I would prefer it if my kids and other road users didn't have to drive.
I am still hoping for fewer cars on the road overall. The car itself is inefficient and hopefully on its way out.
And then I think about all the potholes that I dodge when I drive around town. And how 4-ton electric vehicles are going to make road wear worse, not better.
Here’s an autonomous drifting Delorean https://youtu.be/3x3SqeSdrAE
The physics and control of cars are quite well understood, even in extreme conditions. Further, sensors allow a car to detect changing road conditions much more quickly and adeptly than a typical human.
The hard part of self driving seems to be understanding and predicting the humans in its environment.
If the car is "smart," maybe it will simply refuse to travel (because the calculated risk of traveling over an icy parking lot is above some maximum threshold risk).
What happens the first time a steering wheel free car strands a family in a blizzard due to a few feet of icy pavement, ultimately freezing them to death?
Traditional cars with drivers haven't solved the weather problem 100%. But they are here now. Sometimes we ban them in some conditions, but they are a game changer the rest if the time.
It’s not that traditional cars have solved for weather, it’s that self driving cars haven’t solved the problem of “being as good as human drivers in weather commonly encountered.”
And computers are better at reacting to unexpected things like icy road. And cameras are better at seeing through rain and fog. What's barely perceptible to you, computer sees clearly.
Other than random people saying this on the internet, I have seen zero indication that this will be the case. You cannot will something into existence just by strongly believing in it.
Driving in truly bad weather is still an open research problem, but it's also one that most of these companies haven't dedicated a lot of resources to, since they all have initial rollout plans in places that don't experience much extreme weather. But assuming these vehicles can drive well in good weather, I don't think it's a stretch to assume that progress will be made on driving in more extreme conditions once more resources are dedicated to it.
I dig the commitment and passion, but as a consumer, I want the thing to work perfectly first time, every time. If it doesn't, I don't want it.
Doesn't sound like much of a data trail. No one disputes that it at least sort works (while also doing "weird things", as you put it) in some conditions. As long as, you know, nothing weird happens out on the road. Like for example rain.
It's the fat tail of "weird things" (and how these systems currently handle them) that leave the question of whether FSD will ever arrive -- very much up on the air.
I know 5 people with the Tesla FSD Beta, and EVERY SINGLE ONE of them has admitted to letting Tesla drive them home when they drank too much and went past the legal limit. And not a single one of those 5 people has ever had a DUI. So FSD rapidly changed there drinking behavior.
Also I get that I have an atypical and shitty social circle. I don't think this represents all Tesla drivers.
Does anyone else remember that spyware?
Improved software, and maybe one hardware refresh should be sufficient to get us all the way there for 99.99% of drives.
Aside: some really old drive assist tech for snow plows on I-80 Donner pass - https://path.berkeley.edu/sites/default/files/advanced_snowp... /// https://highways.dot.gov/public-roads/janfeb-2001/safe-plowi...
I'm still waiting to see how an AI would handle Figure 5-1 from that. And no, the snow plow's solution of burying magnets in the road isn't practical (Cost of infrastructure installation for the test sites is approximately $11,000 per kilometer ($17,000 per mile), including surveying, installation, and magnets).
99.99% means 1 in 10,000 "drives" has an accident of some sort? So 100 accidents per million people/drives?
Not sure how to compare if that is better or worse than what we have now.
That opinion doesn't diminish the progress that has already been made or the accomplishments of the various engineering teams involved. It's just saying that when people talk about self driving cars they're talking about it as a drop in replacement to the general population of car owners' cars not as a curated location/digital track limited taxi service.
I'm glad that Cruise beat Waymo to driverless in SF because they are providing the swift kick in the butt that Waymo seems to require to actually make progress in a reasonable amount of time.
Waymo has 20 million miles IRL driverless, and 15 billion miles in simulated driving.
Fatalaties are not the only metric. Accidents are much more frequent and driverless cars have already proven they're safer in that regard.
Simulated miles are not convincing. Miles with a safety driver are better but still ultimately different. Also the type of miles matters too, and Waymo has been focusing on the easiest miles possible until Cruise forced them to up their game and try something a little more valuable.
Source? Because since 2020 Waymo has been primarily touting their "contact events" metric, which is what most people consider to be accidents.
Simulated miles are still better than no miles. Given the better accident rate it seems to contribute to a degree.
The point is that it's an argument from ignorance to narrowly focus on one metric and say we don't have enough information on whether autonomous vehicles are safe. We do have a range of metrics and information already that point to them being safer.
https://youtu.be/gW6Wt2WQotY?t=232
https://youtu.be/o8rCOKSDMcg?t=60
There are countless other examples. It's one of the first things they bring up in almost every presentation to the general public. It's how they present themselves to the world, as the answer to traffic fatalities.
You may think it "ignorant" but the fact is that fatalities are going to be the single most important metric that the public and journalists and governments are going to focus on, and that's why it matters. The first few driverless car fatalities are going to be a huge, huge deal and are likely to result in action from all three branches of government. There will be court cases, there will be regulatory action, and there will be a lot of noise from politicians about laws and maybe even some laws passed about it. The first one already forced Uber out of the game.
Provided your starting point is AV software with some ability to drive, this isn't necessarily true. Tuning on simulated miles can lead towards over-optimization for the conditions of the simulation.
Not with one version of the software. They are cheating: they should reset the mileage whenever they update the software.
Rider 1: I have 20 years experience riding a motorbike.
Rider 2: 20 years experience or 1 years "experience" repeated 20 times?
Their crash rate is 1 in 4.4 billion miles. That's crash, not a fatality. And it's getting better.
https://cleantechnica.com/2021/12/07/tesla-1-crash-per-4-41-...
As I understand it, they will only operate within a "grid". In other words, its use case will be limited to essentially a customized route city bus.
> widely rolled out within ten years seems very plausible.
Define "widely". We are decades away from a driverless car driving in snow.
I agree snow is a lot harder (I don't think it's decades, though). But the cases moritonal brought up mostly don't require that.
The mere fact that it won't panic on black ice is already a huge advantage. And it will have very detailed feedback to and from the traction control systems.
What's the basis for your imagining? Whenever I read some strong "computer > human" statement like that, I can't help but think it's substantially based on science fiction (which can make technology work fantastically well because it's not real, e.g. like Six Million Dollar Man vs actual prosthetic limbs).
Also, have you heard about the time I was driving on the highway in winter weather, and my adaptive cruise control would stop working every few dozen miles because I had to get out and scrape ice off the sensor?
That doesn't really matter, because I was responding to the the sci-fi sentiment that "of course the computer must be better because computer." It doesn't matter if the humans are doing a poor job if the computer is still worse.
Especially because the claim is only about snowy situations. Even in a reply about snow, someone that thinks computers are automatically better would normally just say that directly.
Parktronics routinely get covered by snow/slush/ice when driving.
When that happens I hit a button to turn on the defroster that then gets rid of the snow and ice.
Could they use a similar system?
Can I patent that?
Literally the two easiest places it could go into
So people should have been able to correctly predict how easy it would be for a computer to master go, right?
Said differently, I start from the assumption that some problems are obviously easy, some problems are obviously hard, and every other problem is somewhere in the middle. Self-driving is in the "obviously hard" bucket. Feel free to argue for where you think Go sits in the spectrum, but I would argue no board game is in the "obviously hard" bucket
Hard and easy defined as the ability to solve in a reasonable amount of time with the algorithms and computational models we have today or in the near future -- regardless of ~marginally increasing processing power.
Really? It seems to me that if people over-estimated the difficulty of an "easy" problem like mastering go, then they're even more likely to over-estimate the difficulty of a hard problem like self-driving. In fact, the over-estimation could scale up faster than linearly, if estimating two problems of size X is easier than estimating one problem of size 2X.
> with the algorithms and computational models we have today or in the near future
That's the thing. When predictions were being made about the difficulty of mastering go, people didn't have the algorithms and computational models that we have today. Similarly, predictions made today about the progress of self-driving cars may be lacking critical information about the algorithms and computational models that will be available in the near future.
Guess you missed this distinction.
If "customized route" counts any drive, then that's not really a limitation any more.
Yet we're in 2022 and the best of the best image recognition systems spew out absolute nonsense no matter how much money is poured into them.
It's just the same sort of issue. It can work amazingly well 90% of the time, but the remaining 10% is basically impossible without some kind of actual general AI that can understand the context of what it's looking at. I honestly believe that current approaches to this tech will not get us to general self driving tech that is safe enough for use on our roads.
The Earth is big and complex, I can't see the singularity on the horizon yet, but my vision is aging.
I don't know that creating super-accurate species trees is really the most important problem to solve in biology, though. Certainly DNA sequencing for health has been a really mixed bag.
Ah, the old why bother since it's a "mixed bag" argument. If it wasn't a mixed bag I'd be really worried, as someone has some pretty good snake oil.
I vaguely recall something about sequencing and its role in creating a vaciine that turned out to be pretty useful, something recent maybe?
What i'm troubled by is the projects that sequence a million people, using $1B to do so, and then just dump a bunch of genomes with some hand-wavy claims about how it's going to cure cancer. The costs here are rathre significant, but the health outcomes are not.
On the other hand, most of the commercial jets we do fly are basically 1950s designs with bigger engines and some computers in the cockpit, and HN thinks it was crazily irresponsible for Boeing to have decided to use sensor-triggered autonomous systems to do something as simple as temporarily adjust the trim system...
But: there are very large numbers of these edge cases and they are of the kind where new ones keep popping up all the time. This is a very easy problem for the first 99% or so and then the last 1% is super hard.
Indeed, and we aren't any closer compared to say 5 years ago. That is more or less my point: you can solve for some of this and then you are still left with a mountain more. And then at some point there is an even harder problem to deal with: how to ensure that fixing one problem won't regress one or more others.
That's just flat wrong.
And that may well require another leap of capabilities on the AI front, it might even require GAI. But you are entirely welcome to your own opinion, as I am to mine, but after watching the self driving industry since the 80's my timeframe is something like 25 years before fusion will arrive.
There may be some intermediary 'artificial artificial intelligence' solution where when stuck a remote driver can take over. But that has its own set of problems and I don't want to set this up as a strawman.
It's real progress. Anyone around Mountain View, or SF will have seen the progress with their own eyes. It's not a theoretical debate.
But that's the whole point: if this were a reality there wouldn't be a maybe there. So thanks for making my point.
"Materially closer" isn't the bar to cross the bar to cross is "It just works" which is for me what I'm getting at when I say that we are not 'closer', because it's a binary thing, arrival means 'it works' not that we have made some incremental change towards an ill defined goal.
The 'it works some of the time or even most of the time' bar was crossed years ago, but it needs to work all of the time.
In your view we are within striking distance of the moon, another couple of piles of stones and we are there. In my view, I'm not sure that it will just take more stones and not a better technology, and I'm beginning to lean more and more towards the latter because the problems that we were seeing 10 years ago we are still seeing today: more and more edge cases. And edge cases can be a symptom of a problem that is almost solved or it can be a symptom of a problem that is being solved using the wrong toolset, especially when there are lots of them and new ones keep popping up all the time.
Clear now?
And by the way: none of this is to imply that what has been done isn't impressive, the MB statement that they will accept liability if their software is in control of the vehicle is a step in the right direction and shows a lot of confidence. But it's incremental progress towards a goal that might not be reachable in that way.
Just come visit San Francisco and see the evidence. Look in the horse's mouth, so to speak. The cars are driving around a complex city, without a driver, and doing an amazing job. The progress is real.
I don't have a horse in this race, but I think what the other commenter was getting at is that: if there were 10 problems that needed solving 5 years ago, and in the course of solving those 10 problems you discovered 100 more than needed solving that you weren't initially aware of, then sure you got closer to the goal, but the progress was so tiny that it's almost not relevant.
I'm personally not arguing one way or another, but that's how I read the other person's comment.
I know the whole human flight is different to bird flight argument, so I guess we'll just have to wait and see how it all plays out. Exciting times :)
If the technology stagnates, the hardware stops improving, and the software stops getting better, then we would hit the limit pretty quickly.
> After all humans don't drive by analysing 1 frame at a time, we have a model of how the world works, we anticipate, we have intuition etc...
All self driving cars also maintain those models as well, and use frame by frame input to correct those models in real time (but they have already guessed where something is going to be 100+ frames from now).
Whether the grid is pre-mapped or not isn't particularly relevant. It's just how the technology works. The vehicles aren't driving on rails, they just have more semantic information about the world around them.
Fortunately, we also have a developing world where we can pay humans $2/hour to mechanical-turk tele-operate our 'mostly-working' AI.
A lot of people were expecting electric cars to kick off right after that, so OP's claim is not unreasonable.
Driverless cars in a European city center, during snow, ice and rain seems unlikely within 10 years. But we don't need that to see transformational change.
The fun thing about European cities is that many of them are extremely willing to make (to Americans) unthinkable changes to traffic patterns under their jurisdiction. I would not be the least bit surprised to see those downtown centers closed to all but driverless traffic (if they are not closed already).
BTW, humans are not ready for bad weather either, even though we drive in it all the time. However with driverless cars we will collect statistics proving it isn't ready and never noticing that it is better than humans.
I lived in Phoenix for nearly a decade. I moved there from the snow belt.
Teaching a car to drive itself in Phoenix is barely a step up from teaching it to drive in an empty parking lot.
It's borderline obvious we will have a taxi/Uber replacement for most of lower Manhattan, east San Francisco and southeast Phoenix before 2030. (Technologically, at least.) Less confident in long-distance interstate-only driving by 2030, but that's better than 50/50.
Beyond that, the edge cases are innumerable, so it's difficult to estimate. Nevertheless, we have clear line of sight to domains where, for all intents and purposes, a human-driven vehicle will be obsolete (if not necessarily yet unquestionably uneconomical) in well under ten years. That's far more than we have ever been able to say for fusion.
The FSD is really good.
I rode in a fullY autonomous car, with nobody in the drivers seat, about a year ago (we have had fully autonomous waymos in Phoenix for a few years now).
FSD/Autonomous driving isn't an all or nothing scenario for me. I use it when I deem it safe to be used, and drive manually every other time.
For instance: I don’t like how close it drives to the curb. I don’t drive close to the curb (and bias towards the center of a road) because I only have two eyes, and can’t really see the curb as I’m moving. I have an intuitive sense of how close I am, but that’s it.
The car knows exactly how close I am. So it driving in the center of a lame, while it makes me nervous because it’s now how I drive, is perfectly acceptable.
What I wish Tesla would do: set up some closed courses for people to come and experience what FSD is actually capable of. It would help the current FSD users gain confidence in the system, and it would show it off to people who haven’t ever experienced it.
I seriously don’t believe that most of the people who are critical of this thing have spent much time with it. I’ll admit after the first day or so I was skeptical too, in the same way I was skeptical of lane keeping and proximity aware cruise control.
But at this point: I hate driving without these things. They are an almost indescribable upgrade to the driving experience for me.
Cars that can drive everywhere, in every condition, without humans intervening at least remotely and occasionally in some limited way? Probably not.
Robo-taxis in most major cities and trucks that can reliably handle the full interstate network, to be picked up by a human driver for the last mile? Why not?
Sure, it‘s apples to oranges, but for self-driving, we‘re solving the last (and obviously hardest) percent, while for fusion, we‘re still stuck solving the first.
Look at advanced ports and inter-modal yards. That sort of human to vehicle ratio will probably move up the spectrum of complexity in which it can be deployed and down the spectrum of cost.
Honestly don't believe that at all. The tech is immature and works only in a very narrow set of circumstances with a lot of constraints - it's so far away from being functional that the whole legal discussion around it is almost academic at this point.
2014 - 90% capable of autopilot, for sure highway travel
2015 - autonomous driving for highways and relatively simple roads
2016 - S/X can drive autonomously with greater safety than a person - generally true, it still messes up for human-correctable situations but in general helps more than it hurts in terms of "crashes where autopilot was in control during impact or within 5 seconds of impact"[0]
2018 - incorrectly predicted - self driving will be 100-200% safer than a person by next year (this talking about their full self driving thing that will do turns and stop signs and whatnot)
That's sarcastic, right? Like "we'll have fusion power in 50 years"? Except that we might actually have fusion power in 50 years and I don't actually think we'll have truly driverless cars within 50 years.