It wont. I am betting with Jeff Atwood. In any case with the current rate of inflation, in 2030, 10,000 dollars will the price of a Pumpkin Spice Latte...
I think by definition a level 5 system can do in winter exactly what it can do in ideal situations. See the chart on the linked page... "The feature can drive the vehicle under all conditions".
I am genuinely curious to what technology, based on my assumption it grew out of today's tech, a car would need to have onboard to safely operate at L5 under all conditions.
The bet says within cities, and humans struggle with those sorts of extreme conditions too. Natural disasters and emergencies are excepted. I suppose it just needs to do approximately as well as a human.
I'm with Jeff, self driving cars have been just a few years away for well over a decade now. I think it's one of those cases where the last 10% of the problem takes 90% of the effort. If we built cities and the road infrastructure around self driving capability that might work, but currently cities aren't built that way and that wouldn't meet the definition of Level 5 anyway.
That's what I'm thinking would be required, infrastructure changes and not solely 100% onboard technology. 8 years is not enough time for cities to plan for and build enough of that infrastructure (not to mention, what to build?) for an automaker to commit to a level 5 car when they are already struggling with L3.
> I suppose it just needs to do approximately as well as a human.
I think that's a fallacy. We allow humans to injure and kill each other in "accidents" with limited liability because we are human; I can't be expected to 100% reliably decide in a fraction of a second what to do when a car cuts me off, so if I hit it, or avoid it and hit another vehicle or pedestrian, I don't go to jail.
But a computer can be expected to make that decision fast enough; if it can't, it's because the designers or programmers made errors that the manufacturer missed because they rationally calculated they were not worth catching or fixing.
There's some level of safe driving it's unreasonable to expect from self-driving cars – but it's not the same level as humans.
> But a computer can be expected to make that decision fast enough
I don't think this expectation is reasonable. It makes sense only if you conceive of humans as flawed computers. But certainly it is an expectation people have, including a large segment (if not majority) of HN readers, and this expectation is absolutely slowing development and adoption of the technology.
> I suppose it just needs to do approximately as well as a human.
From a practical/logical sense this is true, but I'm not sure it would work out that way. People would never accept FSD cars with accident/injury rates comparable with humans. FSD would have to be at least 10x better than human drivers, maybe more. Psychologically we are much more risk averse to when not in control. It's why people are afraid of elevators but not stairs, despite a 50x difference in safety.
Historically computers go from "barely able to do X", to "better than any human could ever be" very quickly. I'm with Carmack on this bet.
The more inserting thing to me is will people accept FSD cars that tell them "no"? For example, we're about to get freezing sleet here, and a FSD car may very well decide "this is an impossible situation to drive safely" and refuse to move.
And it may very well be right; most people way overestimate how good they are at driving in very adverse conditions.
You can probably manage to have 10x less chance of a serious accident than the average human by being sober, calm, attentive and caring about braking distances anyway. "Approximately as well as a human", measured by a pool of human accidents dominated by serious driver failings is a pretty low bar.
> humans struggle with those sorts of extreme conditions too
People drive in a wide range of weathers. If you need to check the weather in a major city every day to see if your will drive itself, your car is not self driving (IMO). If that weather happens more than 5-10% of days every year, that’s not an “extreme condition”.
By "highway speeds" do you mean appropriate for the weather, or going at the speed limit?
A car should be able to see better than a human in those conditions, and at an appropriate speed it could be reasonably safe. It probably won't go at the speed limit, but a human can't safely drive at the speed limit either.
In a snow storm on icy roads it’s safe to assume the speeds would be as slow as safety dictates, just as the speeds should also be for humans in the same conditions.
Also, computerized traction control is amazing, especially with electric drive and no transmission lag. When driven by a human it’s not infallible, but if the car doesn’t have the flawed yet behaviors that humans sometimes have, it could be pretty competent. They’ll have to get it to the point of being able to do without lane lines too of course.
As someone who lives in a region where snowfall and icy roads are common I know for a fact that traffic regularly flows at highway speeds in those conditions.
Level 5 implies "all conditions" so it follows an autonomous system should be able to safely operate along side human drivers. If a level 5 system cant keep up it would become a major hazard to non autonomous vehicles.
Driving the speed limit or higher on an icy road is wildly reckless, it would be insane to program a computer to do that, regardless of whether you could.
I am not so confident that even in 8 years the outcome will be clear. Just like today Tesla is selling hot garbage and calling it "self driving", I am sure manufacturers will be selling cars that claim to satisfy the requirements of the bet, and even do so... sometimes. I am not confident they will work as reliably as the bettors here are intending, even if it is marketed as L5.
Edit: What I'm getting at is even though it is a friendly bet, they should define the terms of success / failure a little more clearly. Unless there is some governing body that grants L5 status?
Given that these gentlemen are known to have at least ordinary prudence, even in the face of a $10k loss, an operational definition might take the form
L5 = the person asserting that L5 has been achieved is willing to be driven by the vehicle, without access to the controls, through mixed conditions for X hours.
I feel that is pretty bad. It doesn't require system to actually work too well. It doesn't name anything about effectiveness of system. Just safety. Still, effectiveness might come to abstract things like:
L5 = vehicle can achieve similar travel times and destinations to average human driver in mixed conditions.
What if the car "works" under all conditions, but moves extremely slowly, takes huge detours to avoid tricky roads etc. In fact what is stopping this "technically correct" approach: car calls a human driven tow truck, gets hitched and towed to destination?
Of course. I meant this in response to dharamon above talking about manufacturer's misleading claims. But I see their edit and I think I also misunderstood your point - in the context of this bet the selected criteria are probably good enough.
If the car moves extremely slowly in a major city, it would get banned after a handful of traffic jams, and so wouldn't be commercially available. And anyways no company would risk the bad PR of launching a car that tops out at below the speed limit.
If the car calls itself a tow truck, it has obviously failed to drive itself.
If the car is programmed to avoid tricky roads, it can't "drive everywhere in all conditions" per SAE.
Exactly. "All conditions" means all conditions. Once a self-driving 4WD truck can be relied upon to get me over Echo Summit pass on US-50 during chain control / whiteout blizzard conditions, then maybe, just MAYBE I'll start to take it seriously. "Full" self driving, at least as of 2022, relies too heavily on all cows being spherical.
I don't really care for the L5 definition. Humans can't drive on all roads in all conditions even if their vehicles are theoretically capable of it. There are certainly weather conditions I don't want to be driving in. And there are certainly some unpaved roads I don't want to be driving on even with high clearance 4WD.
The realistic definition would be that if a human could not reasonably be expected to drive in it (fog/smog with less than 6 inches visibility, hurricane, blizzard, military invasion) then the cars are not required to drive it in. And perhaps the cars can accept a circumstance or two where humans can't drive, in exchange for one or two where humans could, but the cars can't. But it would need to be circumstances that humans can predict/understand. Like if the car won't drive if the countrywide car-to-car comms network is down or something.
If a company falsely claims level 5 they can be disproved pretty easily by some posts on twitter. If the stoppages are too rare to even show up there, then it might as well be level 5.
There's no way. I'm basically trailer trash, so take this with salt and tequila, but Level 5 is gonna be a 95% problem. You can get to 95% by spending gobs of money, but to get the other 5%, you're going to need exponentially more money PER PERCENT. And nobody is going to want to invest that much.
I am not so sure. Think of the potential value to be unlocked: It will radically change the way real estate is planned. If cars drive themselves, you can switch over to a "Car on demand" model for most current vehicle owners. Why would you need to waste space on parking lots, single garages etc. You also unlock the ability for more people to live further away from work.
It has the potential to extract a massive amount of value out of other parts of the chain as well. Aside from software(which in my opinion is just stealing existing value from everyone else) pretty much everything else is minuscule growth. What is the next leap in maintaining growth of the economy? Extract the value out of these blue collar jobs(trucking, taxi driving, other logistics).
Combine this with the rise of simpler designs in EVs + less car ownership you also can bring in mechanics and repair into your value chain and extract that value as well.
You just outlined the advantages to be had should such come about. But am I missing the part where you explain why you think this has any chance of coming to fruition? Because that's what parent is on about. I mean, if a frog had wings it wouldn't bump its ass when it hops. But no one is explaining how frog wing technology has taken (if you'll pardon the pun) some big leaps the past few years.
>> "Car on demand" model for most current vehicle owners
I don't see that. Cars are not just transport. They are fashion statements. They are virtue signaling. They are portable storage lockers. I just don't see everyone who owns a car now giving all that up just because the car can drive itself. The traffic/congestion implications of "car on demand" are also staggering, basically doubling the number of trips made as each car delivers itself empty to the next customer. And further reducing the cost of car ownership will only pull more people away from mass
transportation/biking/walking.
> They are fashion statements. They are virtue signaling. They are portable storage lockers.
Flip it around the other way - if you had a L5 car at your fingertips, would you give it up so that you could make a fashion statement, virtue signal, or have a portable storage locker?
Yes. Instantly. I keep all sorts of stuff in my car, from shopping bags to a trauma kit. When i am on alert (military) i keep a duffle bag full of my stuff so i dont have to go home before a snap exercise. And in a town of pickup trucks and suvs i enjoy being the smallest car on the road, especially when i see trucks in the ditch needing help. Ever seen a 2door honda pull an f150 out of the ditch? I have. I got a picture of that one in my office.
There is so much wrong with your statement. First, Gen Z doesn’t care nearly as much as previous generations about driving and what vehicles mean for status. Expect that trend to continue especially when ownership goes away.
And as far as “doubling trips”, the car won’t have to drive far to find its next pickup and it may not have to do empty pickups at all with the equivalent of Lyft Line.
If we don’t need parking lots and spaces we can use that real estate for more mass transit.
Also there are numerous other complications that get glossed over. How many L5 cars are going to be available in my area on a popular holiday travel weekend that have a roof rack for kayaks and a 7 pin trailer connector and electric brake controller for my camper? Not to mention giving up the other customizations I enjoy with my current vehicle - removed rear seats for more space/less weight, dash cam, winter tires, etc.
I suspect if self driving comes to fruition it will be a mixture for many people. I can definitely see our 2 car household just owning one car that fits our niche use cases, and then one on-demand cars for daily work trips and filling in the gaps.
You are in a privileged position. Regulations will follow the introduction of self driving cars to make owning a car a much more expensive proposition. With wages stagnant for decades, the mass market will be forced into the rental model. That will be able to cover their use case as well as your use case because there won't be as many of you.
> How many L5 cars are going to be available in my area on a popular holiday travel weekend that have a roof rack for kayaks and a 7 pin trailer connector and electric brake controller for my camper?
Weekend reservations. Just like people do with kayaks today.
> Not to mention giving up the other customizations I enjoy with my current vehicle - removed rear seats for more space/less weight, dash cam, winter tires, etc.
None of this is necessary with robotaxis. You have way more flexibility in terms of space. You won’t need a dashcam. Weight and tires won’t make a difference to the rider.
Anyone who kayaks frequently is much more likely to own than to rent.
And you are far too dismissive of the customization/equipment/storage aspect. For just one example, there are vast numbers of parents who have child seats, strollers, diapers, wipes, etc stored in their cars. You can be sure they have no interest in hauling all that stuff around during their shopping or whatever because the robotaxi is off getting another customer.
I think part of it has to do with the possibility that they are just poorer on average compared to their parents. Circumstances help define the culture so rallying against excessive consumption is more in vogue when you cannot consume in the first place.
Think about the possibility of changing cars on demand. In the morning you can rent a mobile bed like car to get to work(take a nap while it takes you to work or get some tasks done), when coming home you can rent an suv to pick up groceries/pick up your kids and on the weekends if you need a pickup truck/van, you can request one of those for the occasion.
Yes there will be massive traffic implications. This is probably something that will have to be addressed when it happens. If there are dedicated self driving car lanes, can the cars communicate with each other to help reduce traffic? At the very least pollution will be significantly reduced from today's smog filled cities like LA thanks to the transition to EVs(I don't expect this revolution to occur before the transition to EVs).
Finally, in regards to public transportation: it is already a failure in the US outside a few cities. It is a lost cause in a society that focuses on individualism vs collectivism. In terms of walking/biking, I can see improved access to these services. Can you imagine how a car that never gets tired/drunk will reduce accident and subsequently insurance rates? If the cards are played right, there is an opportunity for a massive transformation of how bicyclists and public walkers are treated in relation to cars.
It would be way cheaper and more practical to just to have everyone work from home that can or live close by to where they work or have cities invest in better public transport and biking infrastructure.
Sure none of those are as cool as the JohnnyCab but they already exist so require no R&D.
Sure. All it would take is razing LA to the ground and rebuilding it from scratch.
I don't think I'm exaggerating, either. In a city plan like that one, any public transit route is going to, even at optimal configuration, have simultaneously too many stops and too little:
* Too many stops, because the train or bus spends more time stopped than it spends moving, so rides across the city take forever.
* Too few stops, because you still have to walk a mile to actually get where you're going.
Fixing this requires a walkable city, which can be summarized as "don't spend half the land area on parking and stroads."
Lots of people -- including me -- don't believe that it's possible without significant technological advances that look to be extraordinarily expensive. ie open R&D problems.
I will say that if anyone can do this, it will be Waymo/Google. Unlike Tesla, it's not a fraud.
That said, it's notable they're doing this in Arizona. Limited rain, no snow, clear weather all the time.
> I am not so sure. Think of the potential value to be unlocked: It will radically change the way real estate is planned. If cars drive themselves, you can switch over to a "Car on demand" model for most current vehicle owners. Why would you need to waste space on parking lots, single garages etc. You also unlock the ability for more people to live further away from work.
none of that value accrues to the people who make self-driving cars though (or their investors). It's not like if waymo gets rid of your car for you, they get your parking spot or garage.
It's a huge social benefit but investors are looking for return-on-investment, not making a multibillion loss to improve society.
Maybe I didn't word it properly. I discussed the value gained as well as how the advent of self driving will transform existing stores of value like parking spaces.
In this case: the value is captured by the car rental company (which will likely be the car manufacturers themselves). Part of the value comes from the rent that the parking space would have extracted.
>It's a huge social benefit but investors are looking for return-on-investment, not making a multibillion loss to improve society.
Its hard to predict how much value they can gain unless we have more information but I'd imagine if the majority of cars become on demand, then there has to be at least a few billions of dollars in potential value.
In regards to if people will actually take up rental: Since everyone else is moving to a rental model today(streaming services vs owning, cell phone "refresh" contracts, potential apple offering a monthly subscription that always gets you a new phone, car companies offering all included monthly car subscription) by the time this thing rolls out it is plausible that there will be high uptake on a rental model.
That was the whole point of my response. I am not saying every piece of software steals value from everyone, I am saying software in general steals value from everyone else. So in your example, self driving software steals value from real estate, taxi drivers, truckers, other logistics services. By making them redundant their margin is now owned by the software makers that made them redundant in the first place.
This is what is driving the valuation of all these FAANG companies. They are just vacuuming up all the existing value from the rest of the country/world.
Sure, but that is more of "reassignment of ownership of capital" and that makes it less likely and not more likely that such an endeavor will be funded (for political reasons). Unlike say nuclear fusion (safe and practically unlimited energy at an order of magnitude previously unseen is worth the spend and most countries would bite I think), is self-driving worth hundreds of billions in expenditure?
>Sure, but that is more of "reassignment of ownership of capital"
ha ha your phrasing is funny. It is more like "sudden unexpected reassignment of ownership of capital".
>and that makes it less likely and not more likely that such an endeavor will be funded (for political reasons).
Yes the powers that be are finally starting to slowly turn against the software people. You see Europe getting tired of them losing every last bit of the little value chain they have left. Meanwhile in the US, elites are getting tired of software empowering people other than themselves and are starting to make moves to try and put the genie back in the bottle. We saw all these tech CEOs being brought up in congress but all that it showed is that they still don't fully get how they have been robbed yet, just that they know that they have been robbed. If they manage to stop the software industry (a big doubt) yes I can see them putting a stop to self driving tech(although China seems determined to steam ahead no matter what). If they do manage to stop self driving that means they have also managed to kill a lot of the software industry as well and with that one of the last few paths to a solid comfortable middle/upper middle class life in this world.
>Unlike say nuclear fusion (safe and practically unlimited energy at an order of magnitude previously unseen is worth the spend and most countries would bite I think), is self-driving worth hundreds of billions in expenditure?
Now you are pulling an exact number out of thin air. Will it really be hundreds of billions in cost? When you talk about "hundreds of billions" you are talking about a noticeable chunk of the US's GDP and the complete GDP of most countries.
Since when has software or even advanced AI research ever cost that much to develop? Don't get me wrong: I believe it will be in the billions or even tens of billions over many years, but this is a cost that that can be absorbed by the companies/government entities that are pursuing this and even tens of billions can be justified by the potential gains.
Maybe hundreds is too much but 100 billion spent over years is not necessarily an exaggeration. Google and Cruise together have spent like 20B to get to where we are right now (easily googled) which IMHO is not the 95% mark where the returns get much harder. The cars are heavily geofenced and need help from service personnel regularly.
Correct me If I am wrong but getting to the point where humans are completely unnecessary is still a research problem. We need more research and more breakthroughs which costs a lot of money and is non-linear (you are not guaranteed results). It could very easily take 100B to get to a point where a car can ferry people anywhere they want with zero human assistance (no service personnel to guide out of traps etc).
>Since when has software or even advanced AI research ever cost that much to develop?
Since now? If software is eating the world, you should expect software research to also eat up a lot of money. As the low hanging fruit is picked, further gains are going to cost a lot more.
I see lots of self-driving car advocates talk about this "on-demand car" model, but I don't see any way it works in practice.
No actual sharing: At rush hour, I and most of the people around me will need a car to get to/from work. Therefore the rental fleet will have to be large enough to meet this peak demand, and effectively there will have to be a car in the fleet allocated to me (to get me to and from work). So (amortized over time) I will have to compensate the company for the cost of the car, plus the cost of bringing it to me every day, plus overhead, plus profit.
Reduced convenience: A car may not be available immediately when I need/want it. I have to either risk not having a car for some important event or plan in advance and book what I need. The car I'm allocated on any given day may be too small to meet my needs. I have to wait for the car to arrive, can't store my stuff in it, have to take time to install/uninstall car seats, get charged a fee if I leave my empty soda can in it.
IMO, for a regular car user (i.e. someone who currently uses a car for multiple trips per day) moving to an on-demand model will be at least as expensive as and less convenient than owning a car. There might be exceptions in very dense urban environments where parking is very expensive, but those places usually have a good public transit alternative.
It's hard to coordinate and doesn't save money since you still need your own car to do errands. Pre covid many people absolutely were using uber pool to carpool to work.
Trailer trash or not, sometimes it takes the kid to point out the nudity of the emperor. This hick from Indiana would take Atwood's side of the bet against any that care to pony up the money for the other side. The only thing about this that surprises me is that this post isn't from, like, five years ago. But Carmack took this bet within the past week? I just don't think we have the evidence to suggest this is realistic, and plenty to suggest that it's a long way off. We do seem to have the evidence that if it happens, it probably won't be Tesla doing it.
EDIT: ah, perhaps Carmack knows better and is simply willing to write off $10K for the lulz (from the TFA): I'd like to thank John for suggesting this friendly wager as a fun way to generate STEM publicity.
We don’t at all have evidence that Tesla can’t do Level 5 within the timeframe of this bet. Tesla is the farthest along of anyone who’s trying self-driving. Waymo and Cruise are geofenced, and still don’t work much better than Tesla. Aside from those, Comma.ai is doing seriously impressive stuff (sneaking up on Tesla tbh) on a way, way smaller budget, with less hardware, and notably, they’re going for a vision-only stack this year. Judging by what Comma’s been able to accomplish with basically smartphone hardware (maybe one step up from that), I’d be shocked if Carmack loses this.
why does everyone think this is a pipe dream when waymo has been already offering driverless rides? they were sae level 4 already IIRC.
the key part of the bet is the "in major cities". you may have complications on some road without much data, but major cities will be tested to the inch multiple times by then.
of course waymo has huge sensors but 8 years to shrink those doesn't seem impossible.
> Waymo has a staff of fleet operators that oversees the cars and provides guidance—such as confirming that the car has chosen a safe route through a construction zone (Waymo says that its remote operators never drive the cars directly).
Companies like Waymo and Cruise are both working on making L4 scalable to multiple geos and it's going to take a ton of time and money. But L5 is another beast altogether.
yeah I reckon they started on the easiest place to implement. but it's still impressive seeing the thing driving around a shopping mall parking lot with pedestrian traffic, they put a lot of trust in this thing: https://www.youtube.com/watch?v=tBJ0GvsQeak&t=3820s
Yeah I've wondered about testing conditions for a while, at least for a human snow certainly adds a big twist to things. Off the top of my head the most challenging driving conditions I semi-regularly encounter in the Northeast US are:
- Heavy snow falling at night on an unplowed, narrow road with ill-defined shoulders. The falling snow ruins visibility, and the smooth, white, unplowed surface of the road makes it super hard to identify where the pavement actually ends.
- In traffic on a busy highway in snow. Berms of slush make sudden movements very dangerous, and there is lots of unpredictable behavior from other cars nearby. Often there are cars/trucks nearby getting stuck, sliding off the road or spinning out. Going slowly enough for conditions might mean others start passing you dangerously. Other times taking a more aggressive approach is required if cars around are going too slowly and don't have the momentum to avoid getting stuck on hills.
In either of those cases I'd likely decide just to not make a trip at all if I knew in advance what it would be like, but sometimes you get caught or otherwise simply have to push through. I guess that's an interesting question: at what point does the car tell you it's too dangerous to go or keep going? Hopefully not while on a back road in the middle of nowhere at the beginning of a two day blizzard.
> to get the other 5%, you're going to need exponentially more money
Yeah, but you're going to be making exponentially more money in the markets you have already entered. Investors will be able to do the math rather than envision dreams. If the math adds up, they will be willing to deploy a truly staggering amount of capital to secure the growth.
I don't think the growth is unlimited; there's got to be an upper bound of "value" even if it's literally "cost of a chauffeur for every single car on the road today".
Also, we are presuming that this will always be a one-sided affair, and that self-driving cars will always have to adapt to roads and not vice versa. I'm almost certain that once self driving cars become a thing somewhere, they will be so desired by everyone else that people will become willing to clean up messy infrastructure elsewhere that would otherwise preclude their deployment.
How long have we had commercially available voice recognition for? And why does it still not work ~25% of the time? I'm with codinghorror on this, there's so much interpretation that needs to be done of other drivers that I don't see how this will work.
If you can sell the car at 95% codinghorrer wins the bet. It's only gated by the ability to buy a car. You could even implement a remote driver where a human drives your car using the internet and that would fulfill the conditions of the bet.
Ditto, but I think that failure will be regional, and that this will push the fervor. Neighboring regions have L5 but we don't? Suddenly we have a monumental impetus for change.
I ask because I think that there are useful points short of "never makes a mistake."
The alternative to self-driving cars isn't perfection. It is humans, humans who not only vary in their peak performance, but in their context-specific performance. A car that drove better than me when I'm tired is a good thing for everyone.
There's lot of talk about liability, but it seems both wrong and incomplete.
One incompleteness comes up with an "accident" happens that a less-than-perfect self-driving system would have avoided. The legal system might not be willing to penalize in that situation now, but the upcoming Breathalyzer mandate is moving us that direction.
One wrong focuses on "will the author of the self-driving system be held liable?" I think that the liability will go to the driver/owner, as today. If it's better than me and my insurance company is willing to insure me, then why wouldn't my insurance company be willing to insure my use of a self-driving system?
My heart says bet yes because I want one of these cars so badly, but my mind say bet no because that level of technological change inside 8 years seems unlikely to occur.
I think the intent is to clarify that the vehicles should be available to the general public, rather than just in some DARPA lab somewhere or something.
> L5 in major cities is a bit strange for this, since by definition L5 isn't bound by any restrictions, including geography.
I suspect it is just easier to define a "win" that way. L5 on a test track or L5 on a quiet freeway are very different challenges than what most people think of when they think of L5: Driving in downtown [major city] without intervention.
L5 driving downtown for an entire day with zero interventions, in a consumer vehicle, is really the gold standard for this tech. You nail that, you're done it.
It's also just a weird definition in general. I'm a human, and seemingly a competent driver (zero accidents or moving violations), yet there are still regions, specific roads and definitely road conditions where I will say "nope, not gonna try that." Does that mean I am not L5?
Yes actually, I don't think humans are L5. L5 automation is supposed to be the holy grail, it's supposed to be far superior to human driving. Widespread L5 automation is expected to save countless lives every single day.
This is an easy thing to bet against, so much so that I would raise the bet too. I don’t think the current technology is capable of L5 at all but obviously lot of people feel differently
I think the hardware is capable of L5, but the software is still going to take another 20 years.
Humans only have two eyes that can only look in one direction at a time, and we achieve L5. Yes, we get into crashes, but because humans do stupid things like text while driving, drive hyper-aggressively, drive drunk, or do simply careless things like change lanes without even looking first, none of which are mistakes a computer would make.
On a side note, I am happy to see that all this hubub over self driving has resulted in some really sexy hardware designs. Just thinking about how the Tesla computer is so optimized that they are measuring instruction per pJ just really satisfied the inner side of me that wants to hyper optimize everything.
> I think the hardware is capable of L5, but the software is still going to take another 20 years.
Doesn't this inevitably mean that the hardware isn't capable too with extra steps? Like in 20 years with tens of millions more lines of code, we're going to need more hardware resources to run it (in particular at near real-time) than the hardware would provide today.
I agree with you, to be clear. I am just pointing out that 20 years worth of software is going to be a monster to run at the timings that will be required. So we could hit a hardware limitation between then and now on top of the hard software problems that must be solved.
> Humans only have two eyes that can only look in one direction at a time, and we achieve L5.
Humans use parallax to perceive depth via the overlap of both eyes, which as far as I know isn't used by automated vehicles for reasons unknown (patents?). Automated vehicles using visible light cameras are mapping the world in 2D, humans perceive the world in limited 3D.
Stereo vision is oversold in humans. You can easily drive with one eye. We use a lot of information to estimate depth including relative size of things, as well as “structure from motion” which is the relative movement of things over time.
Tesla does estimate depth using structure from motion (and probably other cues). They call it virtual LIDAR in presentations.
Who does? Parallax requires the cameras to be very carefully angled (toed-in), so they overlap in a very specific way, it isn't simply two cameras recording the same image straight.
Most automated vehicles companies using visible light cameras have them pointed straight forward, that isn't therefore using parallax like a human does.
Parallax only requires you to know the separation and angles (and fields of view) of the cameras. Two parallel cameras at 1 meter separation and 90 degree fields of view are fine. You could change the separation and angles every frame, if you wanted, as long as you always did your calculations from the correct position data.
For that I'd purpose a "taxi Turing test" - a vehicular automation engineer rides in the back wearing augmented goggles that block out where a driver may or may not be sitting & distort all voice audio to sound robotic. The engineer requests navigation thru various routes & environments or whatever testing they'd do for normal L5 evaluation. Not sure I see the point, but doable
It is a tough space, but Magic Leap is a poor reference because that’s a very different space.
I have no idea how close we are to level 5, but I suspect the challenge is closer to “invent a new algorithm that works fine with less data” than it is to “make new hardware”.
I think the requirement that the "vehicle performs all driving tasks under all conditions – except in the case of natural disasters or emergencies" will enable both sides to declare victory. I'm sure that there will be some other unforeseen edge case where it doesn't work even if for all intents and purposes L5 is reached.
Edit: Or to put it more accurately, enable Jeff to declare victory even if he loses.
Yeah i mean they are both honorable guys AFAIK and wouldn't try to welch in the event of a technicality but I can easily see the industry get to a place where there's a long period with debates as to whether L5 was actually attained.
8 years seems like a pretty short time, especially considering all the technical and legal hurdles that are still basically just being kicked down the road for the time being. I'd like to see fully autonomous cars (I'd like to see cheap, accessible, a ubiquitous public transit first) but I'm not holding my breath (for either of those things)
Cost estimates are all over the place, but I could easily see LIDAR that adds less than 30% cost to a typical car model, and that's within the range of many many middle-class households.
This was my first thought, especially since they clarified that it has to happen in a "major city" -- the biggest traffic hazards are going to be other drivers and pedestrians behaving erratically/unpredictably and creating unsafe situations the car has to react to.
You don't even need to have a crash; if you're trying to drive from Manhattan to New Jersey on any given Friday afternoon, you're going to find roads that are "in principle" navigable blocked off by cones and traffic cops waving you through reds etc.
I don't think that there would be much need for a 100% L5 car.
Remote driving where a single driver controls a few cars, possibly by people from emerging countries, will solve this problem, will be quite cheap, and will always add an extra safety element.
Given that, i don't think the last 5% will be an obstacle for the technology making it.
If it's not legal to drive drunk, why would it be legal to drive up to 5 cars at once through a computer, where latency & limited views would likely create similar impairment?
Sure, 95% automation w/ remote drivers can work most of the time. If there's an obstacle blocking the road an automated car can stop for a remote driver to carefully pass by. But it would be a huge liability in any situation where "stop & wait for a remote driver" isn't a safe option.
I do think there will be a significant market for this kind of thing, but it will be more of a premium subscription for L4 vehicles, mostly sold to older folks & less confident drivers. Using it to approximate L5 directly seems dubious
Isn't the restriction to availability in major cities contradicting the requirement of fully autonomous driving in "all" conditions? Driving in a major city with (mostly) intact infrastructure is much easier than e.g. driving in sand or mud.
What do you think the regulatory timeline would be for this? If my startup had a working L5 car today, how long would it take regulators to approve so the general public could buy it? I'm guessing optimistically 3 years but probably 5+. But my opinion has no experience behind it.
If they are not too expensive, I can see families own them (to transfer kids to/from school), disabled or frail people, and people who like getting drunk often.
Why not? Apartments solve the housing problem, why would home ownership even be necessary?
If I live on the outskirts of a city I don’t want to wait an hour for a robo taxi that might smell like vomit to arrive and overcharge me because it’s “high demand time”.
You use your primary home 100% of the time either to live in or store your things in. It makes sense to own it outright. A vehicle is only necessary for short periods for the vast majority of people.
Of course if it is successful it won’t take an hour because there will be far more and I would expect different price points for different levels of service. I don’t check into a Marriott expecting it to smell like vomit but at Motel 6 it wouldn’t surprise me.
To be clear, I think this is one scenario of many possibilities.
Solving self-driving in a select group of cities with permanent 5g connectivity, hi resolution gps, and 30k worth of sensors is a massively easier problem than solving “Tesla fsd”.
For this bet cars can deal with unusual circumstances in all sorts of suboptimal but harmless ways. Pulling over, stopping, going very slowly and cautiously is all allowed. You’ll still get to your destination safely.
> Pulling over, stopping, going very slowly and cautiously is all allowed
Not according to the law though. If these "L5 cars" are obstructing traffic on the regular there's going to be a big problem. And generally, this wouldn't be L5 driving.
> Solving self-driving in a select group of cities with permanent 5g connectivity, hi resolution gps, and 30k worth of sensors is a massively easier problem than solving “Tesla fsd”.
And yet I don’t trust that we’ll be there by 2030 either.
> For this bet cars can deal with unusual circumstances in all sorts of suboptimal but harmless ways.
Stopping or slowing down isn’t always a “safe state”. Stopping in some situations (like on a highway) is the opposite of harmless.
That is a fun bet, I consider myself WAY more pessimistic about self driving than most people interested in the topic, but 2030 doesn't seem totally implausible to me. I would not be surprised if either person won.
Only if the self-driving is augmented by a remote driving human. Think mechanical turk, but for self-driving cars. 95% of the time, the self-driving car will self-drive, but for that remaining 5%, it needs to detect that it is out of its depth and engage a remote driving human to navigate it back into the 95% space.
I have seen the ,,last 95% effort’’ mentioned in many comments here, but the experience with deep learning in the last 10 years is that algorithms are getting exponentially better (training the same task at a specified precision), but as researchers need to run lots of experiments, we just need time and the same effort for the algorithms to improve at the same pace as they have improved before, similar to Moore’s law (just faster).
Regarding the bet I wouldn’t bet against John Carmack (although I’d love to see human on Mars :) ).
Seems like we have a bunch of ultra realists on this thread, so I will take the opposite position. Where's the optimism fellas? We have 8 years - we can do it.
> (My take on VR is far more pessimistic. VR just… isn't going to happen, in any "changing the world" form, in our lifetimes. This is a subject for a different blog post, but I think AR and projection will do much more for us, far sooner.)
Can Carmack toss in a free Quest 2? It looks like the last time he tried VR was back in 2015. The state of the art is a lot better now.
Interesting bet. I don't really know much about self driving car technology. That being said, there are construction zones I've driven through where they didn't really cover the old paint lines, such that its a tossup which lane people choose. There are busy roads I've driven through with the painted lines faded out of existence, on an S-bend curve that simultaneously adds more lanes that still captures people by surprise even after it was repainted. Or the NYC free for all where the all the cars drive together like a school of fish, avoiding all the obstacles jutting into the road, yet not quite block the lane. And snow == anarchy.
Is there any indication self driving cars can handle these scenarios?
Last year, I sat by a construction zone along a popular self-driving circuit and watched. At least a few times the car made it through without the safety driver taking over, but most of the time the safety driver took over.
This was in the summer, but it was a poorly marked lane shift situation.
The terms of the bet require the car to navigate through New York in the winter. Definitely a lot of challenges there that aren't present in Austin. I think 2030 is a nice contentious number to bet over. Far enough in the future that there could be key breakthroughs to make it work.
Well, Waymo is already working in San Francisco. Winter in New York adds traction control and additional obstacles. Traction control is already mostly the vehicle's job anyway, what with anti-skid systems. It can go way beyond that. Stanford has an autonomous, electric, drifting, deLorean.[1][2]
2030 seems impossible to me. In NY, you might not be able to see the road surface at all due to snow. Curbs and other landmarks that are necessary for Waymo to work (they map _everything_) will be covered. Seems unlikely to account for all conditions necessary for level 5 in 7 years.
Yeah. The against side of this bet seems like a no brainer. Is it possible (maybe even more likely than not) that there will be a fully autonomous system that will be available in that timeframe that can navigate many limited access highways in certain weather conditions by then? Absolutely.
And that actually looks very interesting to me for long drives. I actually care less about handling local driving including urban driving near me. I want the long highway drives to be handled.
Yes, 100% autonomous is another kind of interesting and opens up use cases like dropping me off at the airport. But that seems really hard.
Yes, snow is worse in every possible way than rain (unless the rain then freezes to the road). We are so far away from that though.
Don't forget, for L5 it also needs:
- Do the right thing in construction zones where signs and/or barriers may indicate that following the lines is wrong (including flag-men)
- Detect when an out-of-site emergency vehicle with sirens running may be approaching the next intersection and slow down
- Go 25 "when children are present" and 55 when children are not present (though TBH, humans are mediocre at this, and L5 only requires human-level ability)[1]
To my knowledge, there's nothing in the prototype stage that can handle the above in sunny weather. Going from "No protytpe can do this in sunny weather" to "Commercially available in 8 years" seems unlikely even if we restricted it to "only in areas with no snow"
My reading of the terms of the bet is major cities, which means more than one city with a population in excess of 250K. In my reading just Phoenix and Tucson would be sufficient.
The terms appear to actually suggest any of the top ten most populous cities in the US, plenty of which are in places where it doesn't snow.
Atwood's bet is extremely pessimistic, and I don't blame him. I don't even think that weather is the biggest issue to self-driving cars: dealing with roads full of human-driven cars will be harder to handle. It's not that AIs can't learn to aggressively accelerate onto a crowded highway or aggressively turn out of a busy stop sign, it's that the companies running the AIs will be unwilling to accept the legal liability that comes with being responsible for a fleet of aggressive AI drivers.
What do you mean by "operate a self driving car"? That's the point of SAE5: You can't really operate a SAE5 vehicle other than telling it where to go. No steering wheel, no gas pedal, no brake pedal, no nothing. There is no way you could argue that the owner of a SAE5 car is responsible for it. You might as well be legally responsible for the bus, train, airplane you ride in.
Operate as in, put fuel in it, turn it on, tell it where to go. You can operate a dishwasher without washing the dishes yourself.
People are required to buy insurance to operate a vehicle. I don't see how why this needs to be any different.
Companies go out of business. How can a defunct company be held legally liable? Companies and employees can easily die over the 20 year lifetime of a car. Liability can only be placed on the operator.
Companies release new models every year. They get safer every year. Older cars will be more expensive to insure. Will a company want pay the ongoing insurance cost of your 20-30 year old self driving vehicle? What if they can't afford to? Will you still be able to use the car you bought despite the company not being able to insure it? or refusing to insure it (and wanting to force you into a new vehicle purchase)?
All these problems magically go away when you make the operator responsible. They are able to make the choice every month whether they want to continue to pay to insure it, and are legally required to have insurance even if they die in the accident.
Some brands will have better safety, and buying insurance for them will be cheaper. Everyone with the same car model will have pretty much the same insurance rate, unless you are negligent when it comes to maintenance.
Insurance companies are more than capable of calculating these costs. Especially over 10-20 year time frames.
If this sounds crazy to you, it's not that much different than buying liability insurance for pets. Owners are legally responsible.
If you don't want to be legally responsible, don't buy a self driving car. Or a dog for that matter.
No sane person would take that risk. And even if regulations made it a purely civil covered by regulation thing even if the car runs over some kid, I'd fully expect the insurance company to go after the manufacturer.
Drivers are not qualified to assess the safety of a self-driving system. These cars should simply not be allowed on the road unless manufacturers take liability.
Interesting. I interpreted "work in any of the top ten cities" as meaning I could pick any city and it would work. I can see the other interpretation. Hopefully they agree on the parsing!
> The terms of the bet require the car to navigate through New York in the winter.
By NY, do they mean manhattan, new york city or the actual state? Because much of manhattan is gridlike and almost made for self-driving cars. Though the sheer amount of pedestrians, cars and road work could make it especially challenging.
Navigation in much of Manhattan is straightforward. Driving at high traffic hours is certainly not in large areas of the city. I hate driving in Manhattan and will do just about anything to avoid it.
As someone who has worked on self driving for a few years, and holds some patents in the space - I do not feel at all qualified to make a guess. It is always a little odd to me, then, to see how many folks are. Can we all tone down the confidence? Skepticism is healthy, but armchair expert-ing is not.
Carmack is a genius, and 8 years is a long time. If we look at what cruise is doing in SF, it certainly does not seem impossible.
I like that people are seeing this as 8 years from now when the original bet was from 2003. That there is still a reasonable chance that this could happen by 2030 and by that point was predicted 27 years in advanced. It will just add to the Carmack myth if it goes well.
Personally I'm not sure we will see it but there is nothing that jumps out and says that it is impossible at the same time. 8 years ago I would have said it was a no brainer that we would be there, nowadays not so much. It feels like with a lot of things, the last 10% takes 90% of the time.
I have no candle in the game and am an arm chair speculator at best, but it is fun to wonder.
Let just agree to reconvene on this in 8 years time. It isn't that far away.
Wait where did that “original bet was from 2003” fact come from? Atwood’s post doesn’t mention 2003 at all, and looking at Carmack’s twitter thread didn’t bring anything up for me either.
303 comments
[ 3.4 ms ] story [ 355 ms ] threadTo Downvoters: I'm sure you all know what happens when you don't archive the tweets or one (or both of them) get their accounts and tweets deleted?
[0] https://longbets.org/
Can someone school me on how a level 5 system might work in a snow storm on icy roads at highway speeds in 2030?
I am genuinely curious to what technology, based on my assumption it grew out of today's tech, a car would need to have onboard to safely operate at L5 under all conditions.
I'm with Jeff, self driving cars have been just a few years away for well over a decade now. I think it's one of those cases where the last 10% of the problem takes 90% of the effort. If we built cities and the road infrastructure around self driving capability that might work, but currently cities aren't built that way and that wouldn't meet the definition of Level 5 anyway.
I think that's a fallacy. We allow humans to injure and kill each other in "accidents" with limited liability because we are human; I can't be expected to 100% reliably decide in a fraction of a second what to do when a car cuts me off, so if I hit it, or avoid it and hit another vehicle or pedestrian, I don't go to jail.
But a computer can be expected to make that decision fast enough; if it can't, it's because the designers or programmers made errors that the manufacturer missed because they rationally calculated they were not worth catching or fixing.
There's some level of safe driving it's unreasonable to expect from self-driving cars – but it's not the same level as humans.
I don't think this expectation is reasonable. It makes sense only if you conceive of humans as flawed computers. But certainly it is an expectation people have, including a large segment (if not majority) of HN readers, and this expectation is absolutely slowing development and adoption of the technology.
From a practical/logical sense this is true, but I'm not sure it would work out that way. People would never accept FSD cars with accident/injury rates comparable with humans. FSD would have to be at least 10x better than human drivers, maybe more. Psychologically we are much more risk averse to when not in control. It's why people are afraid of elevators but not stairs, despite a 50x difference in safety.
Historically computers go from "barely able to do X", to "better than any human could ever be" very quickly. I'm with Carmack on this bet.
And it may very well be right; most people way overestimate how good they are at driving in very adverse conditions.
We already have things like this “what if I had a beer and need to drive mom to the hospital” say those with a enforced breathalyzer.
They are hoping to be 2x or 3x as good, and the amount of lives saved by that should be enough to influence lawmakers and the public.
Cities have fast roads too.
> humans struggle with those sorts of extreme conditions too
People drive in a wide range of weathers. If you need to check the weather in a major city every day to see if your will drive itself, your car is not self driving (IMO). If that weather happens more than 5-10% of days every year, that’s not an “extreme condition”.
A car should be able to see better than a human in those conditions, and at an appropriate speed it could be reasonably safe. It probably won't go at the speed limit, but a human can't safely drive at the speed limit either.
In a snow storm on icy roads it’s safe to assume the speeds would be as slow as safety dictates, just as the speeds should also be for humans in the same conditions.
Also, computerized traction control is amazing, especially with electric drive and no transmission lag. When driven by a human it’s not infallible, but if the car doesn’t have the flawed yet behaviors that humans sometimes have, it could be pretty competent. They’ll have to get it to the point of being able to do without lane lines too of course.
Level 5 implies "all conditions" so it follows an autonomous system should be able to safely operate along side human drivers. If a level 5 system cant keep up it would become a major hazard to non autonomous vehicles.
But, what you describe does not meet the basic criteria for doing it safely.
Regularly flowing at highway speeds in those conditions does not imply doing it safely, as much as you and those drivers might like it to.
We read about pileups from time to time and the consequences of these are not trivial.
Edit: What I'm getting at is even though it is a friendly bet, they should define the terms of success / failure a little more clearly. Unless there is some governing body that grants L5 status?
L5 = vehicle can achieve similar travel times and destinations to average human driver in mixed conditions.
What if the car "works" under all conditions, but moves extremely slowly, takes huge detours to avoid tricky roads etc. In fact what is stopping this "technically correct" approach: car calls a human driven tow truck, gets hitched and towed to destination?
Decency and reasonableness. Neither John nor Jeff are going to try to lawyer they way out of losing.
If the car moves extremely slowly in a major city, it would get banned after a handful of traffic jams, and so wouldn't be commercially available. And anyways no company would risk the bad PR of launching a car that tops out at below the speed limit.
If the car calls itself a tow truck, it has obviously failed to drive itself.
If the car is programmed to avoid tricky roads, it can't "drive everywhere in all conditions" per SAE.
It has the potential to extract a massive amount of value out of other parts of the chain as well. Aside from software(which in my opinion is just stealing existing value from everyone else) pretty much everything else is minuscule growth. What is the next leap in maintaining growth of the economy? Extract the value out of these blue collar jobs(trucking, taxi driving, other logistics).
Combine this with the rise of simpler designs in EVs + less car ownership you also can bring in mechanics and repair into your value chain and extract that value as well.
I don't see that. Cars are not just transport. They are fashion statements. They are virtue signaling. They are portable storage lockers. I just don't see everyone who owns a car now giving all that up just because the car can drive itself. The traffic/congestion implications of "car on demand" are also staggering, basically doubling the number of trips made as each car delivers itself empty to the next customer. And further reducing the cost of car ownership will only pull more people away from mass transportation/biking/walking.
Flip it around the other way - if you had a L5 car at your fingertips, would you give it up so that you could make a fashion statement, virtue signal, or have a portable storage locker?
Not all car drivers' total cost of car ownership works out less than the Uber fares for the rides they take either....
And as far as “doubling trips”, the car won’t have to drive far to find its next pickup and it may not have to do empty pickups at all with the equivalent of Lyft Line.
If we don’t need parking lots and spaces we can use that real estate for more mass transit.
Also there are numerous other complications that get glossed over. How many L5 cars are going to be available in my area on a popular holiday travel weekend that have a roof rack for kayaks and a 7 pin trailer connector and electric brake controller for my camper? Not to mention giving up the other customizations I enjoy with my current vehicle - removed rear seats for more space/less weight, dash cam, winter tires, etc.
I suspect if self driving comes to fruition it will be a mixture for many people. I can definitely see our 2 car household just owning one car that fits our niche use cases, and then one on-demand cars for daily work trips and filling in the gaps.
Weekend reservations. Just like people do with kayaks today.
> Not to mention giving up the other customizations I enjoy with my current vehicle - removed rear seats for more space/less weight, dash cam, winter tires, etc.
None of this is necessary with robotaxis. You have way more flexibility in terms of space. You won’t need a dashcam. Weight and tires won’t make a difference to the rider.
And you are far too dismissive of the customization/equipment/storage aspect. For just one example, there are vast numbers of parents who have child seats, strollers, diapers, wipes, etc stored in their cars. You can be sure they have no interest in hauling all that stuff around during their shopping or whatever because the robotaxi is off getting another customer.
Environmentalism seems to be "cool af",to the point where not owning a car is virtue signalling.
Yes there will be massive traffic implications. This is probably something that will have to be addressed when it happens. If there are dedicated self driving car lanes, can the cars communicate with each other to help reduce traffic? At the very least pollution will be significantly reduced from today's smog filled cities like LA thanks to the transition to EVs(I don't expect this revolution to occur before the transition to EVs).
Finally, in regards to public transportation: it is already a failure in the US outside a few cities. It is a lost cause in a society that focuses on individualism vs collectivism. In terms of walking/biking, I can see improved access to these services. Can you imagine how a car that never gets tired/drunk will reduce accident and subsequently insurance rates? If the cards are played right, there is an opportunity for a massive transformation of how bicyclists and public walkers are treated in relation to cars.
Sure none of those are as cool as the JohnnyCab but they already exist so require no R&D.
I don't think I'm exaggerating, either. In a city plan like that one, any public transit route is going to, even at optimal configuration, have simultaneously too many stops and too little:
* Too many stops, because the train or bus spends more time stopped than it spends moving, so rides across the city take forever.
* Too few stops, because you still have to walk a mile to actually get where you're going.
Fixing this requires a walkable city, which can be summarized as "don't spend half the land area on parking and stroads."
Lots of people -- including me -- don't believe that it's possible without significant technological advances that look to be extraordinarily expensive. ie open R&D problems.
I will say that if anyone can do this, it will be Waymo/Google. Unlike Tesla, it's not a fraud.
That said, it's notable they're doing this in Arizona. Limited rain, no snow, clear weather all the time.
none of that value accrues to the people who make self-driving cars though (or their investors). It's not like if waymo gets rid of your car for you, they get your parking spot or garage.
It's a huge social benefit but investors are looking for return-on-investment, not making a multibillion loss to improve society.
In this case: the value is captured by the car rental company (which will likely be the car manufacturers themselves). Part of the value comes from the rent that the parking space would have extracted.
>It's a huge social benefit but investors are looking for return-on-investment, not making a multibillion loss to improve society.
Its hard to predict how much value they can gain unless we have more information but I'd imagine if the majority of cars become on demand, then there has to be at least a few billions of dollars in potential value.
In regards to if people will actually take up rental: Since everyone else is moving to a rental model today(streaming services vs owning, cell phone "refresh" contracts, potential apple offering a monthly subscription that always gets you a new phone, car companies offering all included monthly car subscription) by the time this thing rolls out it is plausible that there will be high uptake on a rental model.
This is what is driving the valuation of all these FAANG companies. They are just vacuuming up all the existing value from the rest of the country/world.
I don't know.
ha ha your phrasing is funny. It is more like "sudden unexpected reassignment of ownership of capital".
>and that makes it less likely and not more likely that such an endeavor will be funded (for political reasons).
Yes the powers that be are finally starting to slowly turn against the software people. You see Europe getting tired of them losing every last bit of the little value chain they have left. Meanwhile in the US, elites are getting tired of software empowering people other than themselves and are starting to make moves to try and put the genie back in the bottle. We saw all these tech CEOs being brought up in congress but all that it showed is that they still don't fully get how they have been robbed yet, just that they know that they have been robbed. If they manage to stop the software industry (a big doubt) yes I can see them putting a stop to self driving tech(although China seems determined to steam ahead no matter what). If they do manage to stop self driving that means they have also managed to kill a lot of the software industry as well and with that one of the last few paths to a solid comfortable middle/upper middle class life in this world.
>Unlike say nuclear fusion (safe and practically unlimited energy at an order of magnitude previously unseen is worth the spend and most countries would bite I think), is self-driving worth hundreds of billions in expenditure?
Now you are pulling an exact number out of thin air. Will it really be hundreds of billions in cost? When you talk about "hundreds of billions" you are talking about a noticeable chunk of the US's GDP and the complete GDP of most countries.
Since when has software or even advanced AI research ever cost that much to develop? Don't get me wrong: I believe it will be in the billions or even tens of billions over many years, but this is a cost that that can be absorbed by the companies/government entities that are pursuing this and even tens of billions can be justified by the potential gains.
Correct me If I am wrong but getting to the point where humans are completely unnecessary is still a research problem. We need more research and more breakthroughs which costs a lot of money and is non-linear (you are not guaranteed results). It could very easily take 100B to get to a point where a car can ferry people anywhere they want with zero human assistance (no service personnel to guide out of traps etc).
>Since when has software or even advanced AI research ever cost that much to develop?
Since now? If software is eating the world, you should expect software research to also eat up a lot of money. As the low hanging fruit is picked, further gains are going to cost a lot more.
No actual sharing: At rush hour, I and most of the people around me will need a car to get to/from work. Therefore the rental fleet will have to be large enough to meet this peak demand, and effectively there will have to be a car in the fleet allocated to me (to get me to and from work). So (amortized over time) I will have to compensate the company for the cost of the car, plus the cost of bringing it to me every day, plus overhead, plus profit.
Reduced convenience: A car may not be available immediately when I need/want it. I have to either risk not having a car for some important event or plan in advance and book what I need. The car I'm allocated on any given day may be too small to meet my needs. I have to wait for the car to arrive, can't store my stuff in it, have to take time to install/uninstall car seats, get charged a fee if I leave my empty soda can in it.
IMO, for a regular car user (i.e. someone who currently uses a car for multiple trips per day) moving to an on-demand model will be at least as expensive as and less convenient than owning a car. There might be exceptions in very dense urban environments where parking is very expensive, but those places usually have a good public transit alternative.
I think most people will carpool to work in self driving vans. It'll be so much cheaper than having your own car.
EDIT: ah, perhaps Carmack knows better and is simply willing to write off $10K for the lulz (from the TFA): I'd like to thank John for suggesting this friendly wager as a fun way to generate STEM publicity.
Looking at a busy street makes me almost certain, that true self driving cars will be too expensive to develop.
However, if we accept random deaths caused by autonomous vehicles, there is a chance.
But everyone wants guarantees that cannot be given.
the key part of the bet is the "in major cities". you may have complications on some road without much data, but major cities will be tested to the inch multiple times by then.
of course waymo has huge sensors but 8 years to shrink those doesn't seem impossible.
> Waymo has a staff of fleet operators that oversees the cars and provides guidance—such as confirming that the car has chosen a safe route through a construction zone (Waymo says that its remote operators never drive the cars directly).
As someone who doesn't live in a city with Waymo, do they have backup drivers?
If they do, hopefully they are paying attention and not watching Hulu.
- Heavy snow falling at night on an unplowed, narrow road with ill-defined shoulders. The falling snow ruins visibility, and the smooth, white, unplowed surface of the road makes it super hard to identify where the pavement actually ends.
- In traffic on a busy highway in snow. Berms of slush make sudden movements very dangerous, and there is lots of unpredictable behavior from other cars nearby. Often there are cars/trucks nearby getting stuck, sliding off the road or spinning out. Going slowly enough for conditions might mean others start passing you dangerously. Other times taking a more aggressive approach is required if cars around are going too slowly and don't have the momentum to avoid getting stuck on hills.
In either of those cases I'd likely decide just to not make a trip at all if I knew in advance what it would be like, but sometimes you get caught or otherwise simply have to push through. I guess that's an interesting question: at what point does the car tell you it's too dangerous to go or keep going? Hopefully not while on a back road in the middle of nowhere at the beginning of a two day blizzard.
Yeah, but you're going to be making exponentially more money in the markets you have already entered. Investors will be able to do the math rather than envision dreams. If the math adds up, they will be willing to deploy a truly staggering amount of capital to secure the growth.
Also, we are presuming that this will always be a one-sided affair, and that self-driving cars will always have to adapt to roads and not vice versa. I'm almost certain that once self driving cars become a thing somewhere, they will be so desired by everyone else that people will become willing to clean up messy infrastructure elsewhere that would otherwise preclude their deployment.
And, as experienced in similar issues with uptime guarantees, the costs are quite different.
But so is the technology used! You don’t just scale-up a single MySQL server for all of Twitter.
I feel this is what should be happening, but our economy and our culture prefer Lone Ranger style innovation.
I ask because I think that there are useful points short of "never makes a mistake."
The alternative to self-driving cars isn't perfection. It is humans, humans who not only vary in their peak performance, but in their context-specific performance. A car that drove better than me when I'm tired is a good thing for everyone.
There's lot of talk about liability, but it seems both wrong and incomplete.
One incompleteness comes up with an "accident" happens that a less-than-perfect self-driving system would have avoided. The legal system might not be willing to penalize in that situation now, but the upcoming Breathalyzer mandate is moving us that direction.
One wrong focuses on "will the author of the self-driving system be held liable?" I think that the liability will go to the driver/owner, as today. If it's better than me and my insurance company is willing to insure me, then why wouldn't my insurance company be willing to insure my use of a self-driving system?
For me, that means there are only three years left to have something like that on a prototype level.
Did they mean the ability to take a ride between two arbitrary points in major cities with no limitations on time of day and weather instead?
I suspect it is just easier to define a "win" that way. L5 on a test track or L5 on a quiet freeway are very different challenges than what most people think of when they think of L5: Driving in downtown [major city] without intervention.
L5 driving downtown for an entire day with zero interventions, in a consumer vehicle, is really the gold standard for this tech. You nail that, you're done it.
Given that this is what normal people think of when they think “self driving car”, I’d say that’s exactly what they were going for.
Humans only have two eyes that can only look in one direction at a time, and we achieve L5. Yes, we get into crashes, but because humans do stupid things like text while driving, drive hyper-aggressively, drive drunk, or do simply careless things like change lanes without even looking first, none of which are mistakes a computer would make.
Doesn't this inevitably mean that the hardware isn't capable too with extra steps? Like in 20 years with tens of millions more lines of code, we're going to need more hardware resources to run it (in particular at near real-time) than the hardware would provide today.
I agree with you, to be clear. I am just pointing out that 20 years worth of software is going to be a monster to run at the timings that will be required. So we could hit a hardware limitation between then and now on top of the hard software problems that must be solved.
> Humans only have two eyes that can only look in one direction at a time, and we achieve L5.
Humans use parallax to perceive depth via the overlap of both eyes, which as far as I know isn't used by automated vehicles for reasons unknown (patents?). Automated vehicles using visible light cameras are mapping the world in 2D, humans perceive the world in limited 3D.
Tesla does estimate depth using structure from motion (and probably other cues). They call it virtual LIDAR in presentations.
https://youtu.be/NSDTZQdo6H8
Most automated vehicles companies using visible light cameras have them pointed straight forward, that isn't therefore using parallax like a human does.
I have no idea how close we are to level 5, but I suspect the challenge is closer to “invent a new algorithm that works fine with less data” than it is to “make new hardware”.
Edit: Or to put it more accurately, enable Jeff to declare victory even if he loses.
Available for a middle-class person in 2030? No.
The goal is a highly flexible, reactive, all-electric public transit network where vehicles of varying sizes get routed exactly where they need to be.
Officer with a whistle in an intersection, manually directing traffic through the intersecting and around a crash.
Remote driving where a single driver controls a few cars, possibly by people from emerging countries, will solve this problem, will be quite cheap, and will always add an extra safety element.
Given that, i don't think the last 5% will be an obstacle for the technology making it.
Sure, 95% automation w/ remote drivers can work most of the time. If there's an obstacle blocking the road an automated car can stop for a remote driver to carefully pass by. But it would be a huge liability in any situation where "stop & wait for a remote driver" isn't a safe option.
I do think there will be a significant market for this kind of thing, but it will be more of a premium subscription for L4 vehicles, mostly sold to older folks & less confident drivers. Using it to approximate L5 directly seems dubious
No marks and no hard surface: Much more challenging.
Seems like there are still a number of issues but it's promising and as open to the public as Uber/Lyft (if your trip is in that area).
https://www.cnbc.com/2022/01/08/heres-what-it-was-like-to-ri...
I’m not sure private ownership of L5 would ever even be necessary.
If I live on the outskirts of a city I don’t want to wait an hour for a robo taxi that might smell like vomit to arrive and overcharge me because it’s “high demand time”.
Of course if it is successful it won’t take an hour because there will be far more and I would expect different price points for different levels of service. I don’t check into a Marriott expecting it to smell like vomit but at Motel 6 it wouldn’t surprise me.
To be clear, I think this is one scenario of many possibilities.
For this bet cars can deal with unusual circumstances in all sorts of suboptimal but harmless ways. Pulling over, stopping, going very slowly and cautiously is all allowed. You’ll still get to your destination safely.
I’m with Carmack on this one.
Not according to the law though. If these "L5 cars" are obstructing traffic on the regular there's going to be a big problem. And generally, this wouldn't be L5 driving.
And yet I don’t trust that we’ll be there by 2030 either.
> For this bet cars can deal with unusual circumstances in all sorts of suboptimal but harmless ways.
Stopping or slowing down isn’t always a “safe state”. Stopping in some situations (like on a highway) is the opposite of harmless.
Regarding the bet I wouldn’t bet against John Carmack (although I’d love to see human on Mars :) ).
Can Carmack toss in a free Quest 2? It looks like the last time he tried VR was back in 2015. The state of the art is a lot better now.
Is there any indication self driving cars can handle these scenarios?
This was in the summer, but it was a poorly marked lane shift situation.
[1] https://www.youtube.com/watch?v=WNIDcT0Zdj4 [2] https://www.youtube.com/watch?v=3x3SqeSdrAE
And that actually looks very interesting to me for long drives. I actually care less about handling local driving including urban driving near me. I want the long highway drives to be handled.
Yes, 100% autonomous is another kind of interesting and opens up use cases like dropping me off at the airport. But that seems really hard.
That is to say, you could have L5 in Miami without handling snow, but rain...
Don't forget, for L5 it also needs:
- Do the right thing in construction zones where signs and/or barriers may indicate that following the lines is wrong (including flag-men)
- Detect when an out-of-site emergency vehicle with sirens running may be approaching the next intersection and slow down
- Go 25 "when children are present" and 55 when children are not present (though TBH, humans are mediocre at this, and L5 only requires human-level ability)[1]
To my knowledge, there's nothing in the prototype stage that can handle the above in sunny weather. Going from "No protytpe can do this in sunny weather" to "Commercially available in 8 years" seems unlikely even if we restricted it to "only in areas with no snow"
1: For the non-Californians, I drive by 3 of these on my way to work: https://www.pressdemocrat.com/article/news/what-does-when-ch...
- everything is some shade of white/grey/brown
- snow is incredibly reflective
- winter is cold; sensors are almost certainly affected, dirty, or covered
- pedestrians are bundled up, looking less like people
- road markings are obscured or buried
- signs are dirty
- stopping distances are increased
- speeds must be slower to compensate for lack of traction
- "traction control" only helps when accelerating; stability control and ABS only help so much in snow/ice
"By "major cities" we mean any of the top 10 most populous cities in the United States of America."
Phoenix is in that list.
Atwood's bet is extremely pessimistic, and I don't blame him. I don't even think that weather is the biggest issue to self-driving cars: dealing with roads full of human-driven cars will be harder to handle. It's not that AIs can't learn to aggressively accelerate onto a crowded highway or aggressively turn out of a busy stop sign, it's that the companies running the AIs will be unwilling to accept the legal liability that comes with being responsible for a fleet of aggressive AI drivers.
If you don't want to accept that risk, don't operate a self driving car.
People are required to buy insurance to operate a vehicle. I don't see how why this needs to be any different.
Companies go out of business. How can a defunct company be held legally liable? Companies and employees can easily die over the 20 year lifetime of a car. Liability can only be placed on the operator.
Companies release new models every year. They get safer every year. Older cars will be more expensive to insure. Will a company want pay the ongoing insurance cost of your 20-30 year old self driving vehicle? What if they can't afford to? Will you still be able to use the car you bought despite the company not being able to insure it? or refusing to insure it (and wanting to force you into a new vehicle purchase)?
All these problems magically go away when you make the operator responsible. They are able to make the choice every month whether they want to continue to pay to insure it, and are legally required to have insurance even if they die in the accident.
Some brands will have better safety, and buying insurance for them will be cheaper. Everyone with the same car model will have pretty much the same insurance rate, unless you are negligent when it comes to maintenance.
Insurance companies are more than capable of calculating these costs. Especially over 10-20 year time frames.
If this sounds crazy to you, it's not that much different than buying liability insurance for pets. Owners are legally responsible.
If you don't want to be legally responsible, don't buy a self driving car. Or a dog for that matter.
By NY, do they mean manhattan, new york city or the actual state? Because much of manhattan is gridlike and almost made for self-driving cars. Though the sheer amount of pedestrians, cars and road work could make it especially challenging.
Carmack is a genius, and 8 years is a long time. If we look at what cruise is doing in SF, it certainly does not seem impossible.
Personally I'm not sure we will see it but there is nothing that jumps out and says that it is impossible at the same time. 8 years ago I would have said it was a no brainer that we would be there, nowadays not so much. It feels like with a lot of things, the last 10% takes 90% of the time.
I have no candle in the game and am an arm chair speculator at best, but it is fun to wonder.
Let just agree to reconvene on this in 8 years time. It isn't that far away.