You are looking at this situation with a human perspective. You think the car is going to be freaked because a human would be freaked, but the car has no reason to be freaked unless you substitute traffic cones for the penguin. There is either no discontinuity or an equivalent discontinuity. Not to mention that the whole through it to the human thing is a way to handle discontinuities, not a difference in the perception of the discontinuity, As the L2 system gets better and better, at some point one will decide stop is better than human.
I don’t think you can get a few bolts for $0.49 each. But a box of 50 or 100, yes, next day on Prime. And how much for drone delivery?
Mike
1 Like
You keep missing the point. Both the L2 and L4 system can’t proceed as they were going to without a human intervention. But the L2 system stops immediately - it can’t do anything more and disengages. The L4 system doesn’t stop immediately. It keeps driving until it can come to a safe stop. It needs human help before it can continue on the journey to completion - but it doesn’t need instantaneous help to drive the car to a safe stop, because it’s got the capability to do it on its own.
It doesn’t matter if it behaves like an L4 system most of the time. This isn’t making a social media app. Cars can - and do - kill people if they’re not operated safely within their specifications. If you build an L2 system that is safe with a person behind the wheel, but will crash an additional one time every seven years if there’s no driver, then it can’t operate as an L4 system - because something like that will be 7x more dangerous than a human driver. You can’t take the human out of the loop if the L2 system only acts like an L4 system “most” of the time.
It’s not just nomenclature. Disengaging in the moment back to a human is a different functionality than never disengaging back to a human ever.
I think that’s wrong. Imagine my alarm clock example above, but that instead of alarm clocks we’re talking about life support systems at a hospital. If the machine loses power, the patient dies. In my L2 life support system, the patient’s dead if the power goes out unless there happens to be someone physically in the room with him to start CPR until an emergency generator can be wheeled in. So the hospital has to maintain an orderly in the room 24/7 in case that happens. But my L4 life support system has a battery back-up - if the power goes out, the life support system stays running for a half hour. So there doesn’t need to be an orderly physically there.
It’s foolhardy to think those two systems are the same. Or that to claim that since they both do the same things when the power is on, which is 99% of the time, there’s no real difference. Or that the issue involved here is trying to minimize how often the power goes out, so that the patient in the L2 system still dies every time the power goes out but it doesn’t go out as often. It’s true that if the power never ever ever went out, the battery back-up wouldn’t matter - but in a world where we haven’t perfected keeping the power on 100% of the time (L5), it’s what happens when the power goes out that is the fundamental difference between the two systems (and their operating costs).
1 Like
I’m not assuming the car will be “freaked.” I’m not anthropomorphizing it that I believe that it has feelings.
What happens is that the car is encountering something in its environment that is so outside of the data it was trained on that it can no longer engage in next token prediction. The inputs that it has been given (car on road with sign and tree adjacent and clouds and oncoming traffic and Crazy Freaky Pelican Mating) don’t generate an output. The car isn’t freaking out, it’s just had inputs that don’t generate a set of driving instructions in an L2 system. The car can’t generate what needs to happen in the next second - there’s no next word in the autocomplete - so it disengages.
No-one’s saying that. But thanks for inventing a strawman to attack.
Perhaps I misunderstood what you meant when you suggested that “L2” and “L4” were merely labels.
My position is that these designations reflect important functional differences in the capabilities of the two systems, especially in the critical capacity to respond to unusual scenarios that present safety problems without having to have immediate assistance. Do you disagree with that?
1 Like
That wasn’t me.
Sure, L2 and L4 systems are not equally capable.
Many of us are talking about things that help us track the capabilities of the various systems in development and use, because that helps us invest more wisely. You now just seem to be shouting that L4 is better than L2, which is pretty obvious. Not sure how that helps any discussion or investing thesis.
Cruise was shut down even though it was labeled L4.
Tesla’s label has zero bearing on the progress it may or may not make. You are too hung up on labels.
Winning an Oscar does not make a movie great, the same way winning a Nobel price does not make the invention great. The movie and inventions are great regardless. Shawshank redemption did not win any Oscar. Theory of relativity by Einstein was never awarded a Nobel prize. Gandhi was not awarded the Nobel peace prize.
Reality is more important.
1 Like
Not at all. Part of the reason I talk about these things so much is that I think this is the most, perhaps the only critical determinant in true AV tech that remains. Which makes it essential to the investing thesis.
Look - AV technology has advanced to the point that how the vehicle performs when things are going well is almost a solved problem. Which is amazing. But you have several systems (FSD, Waymo, God’s Eye, and perhaps Cruise’s restarted program) that are capable of having the car perform all of the driving functions for a complete trip most of the time.
The most important problem these systems need to solve is keeping their cars from hurting or killing people in problem cases, if they’re driving unattended. That’s the gate that determines who gets a commercial home run in these systems. It’s not smoothness of ride or who does PUDO with the fewest glitches or which car can figure out where to wait at the airport if a flight is delayed. It’s which car gets to the point where it can safely do all of the driving with no real-time human. Which car can regularly drive 700,000 miles without a real-time human and without a crash. Improvements on all the other parts of the ride aren’t germane to the business question any more, IMHO - the tech is already so good among so many companies that that’s not going to be the distinguishing factor going forward.
Waymo has a car that has passed through that gate - but it requires a lot of expensive equipment and is very tightly geofenced. Tesla doesn’t have a car that has passed through that gate - it’s more capable on the 95-99% of rides where everything goes well, but they haven’t solved the “safely drive with no human” part yet. God’s Eye seems to be content being a very advanced ADAS.
I focus on what the car does in the very rare circumstances where things have gone badly because that’s what matters in this space right now. Not the 20,000 trips where everything goes smoothly - the one in 20K trip where things have gone pear-shaped, and whether that results in people going to the hospital or just a passenger annoyed at the delay. Because that’s the issue.
Yes. But L2 and L4 aren’t merely labels. They are categories of what the capabilities are of systems in reality.
The Shawshank Redemption was a movie (and also a short story). It wasn’t a painting. Those categories - “movie” vs. “painting” are not ‘labels,.’ They’re based in reality.
An L4 system has certain attributes in reality that an L2 system does not have. Critical to this conversation, an L4 system can operate safely with no human driver in the car, and an L2 system can’t - which has enormous implications for the economics of a system using those cars.^^ So the reality is that it is critical that an AV get to the point where it doesn’t need an employee/operator sitting the vehicle - and that’s a description of the difference between L4 and L2.
^^ Key thing - an L4 system can operate safely with no human operator in the vehicle. It doesn’t mean all L4 systems do that: it’s a necessary but not sufficient condition. Someone can make a bad L4 system, just like someone can be a bad licensed driver. So Cruise’s failure as an L4 system doesn’t mean that it’s irrelevant whether something is L4 or not, any more than the fact that there are some bad licensed drivers means being able to get your drivers’ license is irrelevant. In order to have an actual robotaxi system that can be economically relevant, you can’t have a system that requires you to have an employee physically in every car. Which means L4, not L2.
1 Like
Is there anything behind this, other than the obvious? The way you keep posting this, it’s like you think people here don’t realize that L4 systems are defined to do more than L2 systems. No-one here that I’ve read says that.
The mistake I believe you’re making is to extrapolate what people say about Tesla’s L2 system and/or about Waymo’s L4 system to some broader L4 vs L2 context. While Tesla’s FSD is officially an L2 system, it not only does more than the minimum requirements for L2 under SAE’s guidelines, it is the basis for Tesla’s Robotaxi software, which is an L4 system IN DEVELOPMENT. And while Waymo’s system can be considered an L4 system, it still gets in accidents, still needs remote drivers.
Perhaps more importantly for investors, Waymo seems to have had a difficult time scaling their system up. We’ve recently seen some progress in that area, but for years now the rate of expansion has been slow. We’ve speculated on potential issues ranging from remote operation infrastructure, to the hiring or remote operators, to the economics of running the service, to availability of suitable sensor-equipped vehicles.
Waymo is also now starting to try to adapt its system to be suitable for personally-owned vehicles. What’s driving this? Does it indicate that the taxi business isn’t viable economically, or just the company doing an expansion thing since it has Google’s resources behind it? Tesla is going at this from the opposite end - starting with software in presonally-owned vehicles and adapting it for taxi service. They’re both trying to end up in both spaces.
No one said anything of the sort. Yet another straw man from you. You keep oscillating between what L4 systems should do, must do, and then bring in that some “bad L4” systems don’t do. Since every L4 system operating today has had a recent at-fault crash, maybe you’re trying to say that no-one has a true L4 system. But we can’t tell because it appears to us that you want to both say that all L4s are better than any L2s while at the same time saying that some L4s are “bad,” which opens up the possibility that a good L2 system might actually be better - in some ways - than a bad L4 system.
But, how that helps us invest better remains a mystery. OTOH, talking about how Waymo may or may not be struggling, how Tesla may or may not be improving its L2 and L4 systems, etc. does help us invest better.
But the reason I keep bringing it up is exactly that - Tesla hasn’t yet developed an L4 system. That’s hugely significant. If you speculate that they’re on the cusp of developing an L4 system, then you might think that’s not very relevant. I happen to think it’s a massive gap between where they are now and where they need to be, making a materially significant Tesla robotaxi program in the near term very unlikely.
I think it’s very reasonable to look at where they are now when trying to assess where they’ll be. And I think that the relevant aspect of “where they are now” has nothing to do with how the system functions when things are going well, but the circumstances surrounding disengagements and what Tesla currently has to do to keep those disengagements from resulting in crashes. Despite their claims of how advanced FSD is, and despite the fact that they’re putting in some bespoke software for the Robotaxi rollout, we’re still several months in (they ran on-street testing for at least a month before bringing in customers), and they have a very limited number of vehicles and still have their safety drivers. This tells us something important about whether Tesla’s FSD program is anywhere close to the “flick the switch” moment of turning it on for private car owners generally, and something indicative about how ready even an upgraded FSD is for driving by itself.
Or perhaps that they think their software stack would be really marketable as a high-end ADAS system, either directly to consumers or to automakers?
Ubn says that sort of thing all the time. That’s who I was responding to. He continually brings up Cruise’s failure (though they’re restarting again) as a reason why it’s meaningless whether a system has L4 capabilities or not.
Why is that a mystery? Assessing whether Tesla is very close or very far away from having an L4 system is critical in investment decisions involving that company. Because whatever other problems or issues might arise in driverless cars, if your car can’t function without an employee in it then it won’t be viable for use in a robotaxi service.
That’s what you say at odd hours of the day.
At even hours of the day you talk about “bad L4 systems.”
Fine. Talk about that then. Where are Tesla’s “massive gaps?”
Are you forgetting that Waymo did that for years, with safety drivers behind the wheel even? That Tesla was able to start with passenger seat remote drivers (SAE term), says a lot above being able to leverage FSD for robotaxi use.
You do understand the massive differences between an ADAS system and an autonomous system, even a partially-autonomous system, right?
Well, you invite that by talking about “bad L4” systems.
But most of the time you’re talking about L4 being better than L2 without talking about anyone’s actual implementation.
They don’t have a system that can operate without an employee in the car, at least not on public roads for more than a demonstration.
That’s a “massive gap” from where they are to where they need to be if they want to use their system as part of a robotaxi or similar TaaS process.
Nope, I haven’t forgotten - but Waymo was doing that in 2015. Ten years ago. If you said that Tesla’s ten years away from having a workable L4 system in a few cities, I’d say that was actually too pessimistic. Because, of course, they’ve been working on it for the last ten years - and the general state of the technology is vastly higher in the entire industry (both AI generally and autonomous vehicles specifically). But being where Waymo was when Waymo started out is not a positive sign. If Tesla were to safety drivers for as long as Waymo had safety drivers, then that would be very bad news indeed for Tesla.
Absolutely. Which is why I suggested that Waymo’s software stack wouldn’t be sold to third parties as an autonomous system, but as an ADAS. Right now, Waymo is able to run as an L4 autonomous system. But if you stuck a passenger behind the wheel, you could disable the “phone a friend” feature and just let them operate the car as a super-good L2 ADAS system.
Because L4 is a necessary but not sufficient condition for a robotaxi. It’s necessary - if you’re not L4, it does not matter at all how good your L2 system is - because you can’t operate it without a human driver in the vehicle. But it’s not sufficient - if your system is L4 but is deficient in a lot of material ways otherwise, you will not go to space tod….you will not have a workable robotaxi.
So one of the most fascinating things to me about Tesla’s robotaxi open beta in Austin is that despite being years and years ahead of where Waymo was when they first started going on public streets, Tesla’s system still needs drivers. The most fascinating thing about Tesla’s CA “robotaxi” venture is that they didn’t convert their permit to start having CA assess driverless miles with a safety operator. Both of those things scream that all of Tesla’s years and years of developing FSD has still left them a fair bit away from being ready to either start running without drivers or even to “show their work” on how often the system needs intervention. If you’re looking for information to help your investing decisions in this area, the “driver vs. no-driver” question is vastly more important than “can it handle PUDO at the airport”?
1 Like
What we’re seeing, and obviously some here and many in the world at large, don’t recognize or outright deny, is that because Tesla’s and Waymo’s approaches are different, it’s difficult to compare progress towards a shared end goal or goals.
This isn’t the best analogy, but maybe it’ll help. Imagine needing to paint the inside of large home. One approach would be to paint each room, one at a time. Another approach would be to paint common items in all the rooms at the same time, like to all the ceilings, then do all the doors, then all the baseboards, then all the window trim, etc.
Without saying that one is more efficient, how do you compare mid-way progress? For the first method, do you count rooms? Well, some rooms are larger than others. Some rooms have more doors/windows. Some have taller ceilings requiring ladders/scaffolding. For the second method, maybe you started with the ceilings first, getting maybe the hardest parts out of the way.
And then throw in the technology you’re using to paint. Spraying covers more area faster, but there’s lots more prep time in masking, mixing, and cleaning tools afterwards. So, early on with sprayers, you’ve probably spend days if not weeks just masking everything - and not a single drop of paint has been applied. So, how do you compare progress versus roller/brush that has two rooms out of 20 painted, or versus just ceilings painted with rollers on poles?
It’s hard to compare Tesla’s and Waymo’s progress. One has had data gathering and testing over all roads. The other has had detailed data gathering in the areas they’re offering service. One requires specially-equipped vehicles, adding to cost and reducing availability for expansion. Is leveraging FSD in all those Tesla vehicles on the road an actual advantage or not? Is having all those sensors in Waymo vehicles an advantage or a disadvantage?
Just saying Tesla still has drivers or Waymo covers more cities/road area doesn’t really tell us which is closer to the shared end goals or taxis and autonomous personally-owned vehicles.
It does matter if you’re using that L2 system as the base for your L4 system. And so why it matters if your L2 system does more than the minimum to be called “L2.”
You declare L4 as something that can’t be built on top of L2, yet you insist that ADAS for personally owned cars is easy to build from an autonomous taxi. It’s not as easy as “disable the phone a friend feature.”
I don’t entirely disagree. But I think it’s easy to overlook that this isn’t just a discussion about approaches and which one works best….but also about what goals are being pursued.
To use your analogy, the “paint by room” vs. “paint by component” approach might make it more difficult to assess “who is closer to finishing the house.” Hard to measure mid-way progress. But if you’re assessing a mid-way goal, then the measurement gets a little easier.
So if there’s a midway goal of, “who’s got a livable room yet,” it might be much easier to get an answer. The company going room by room has a pretty obvious headstart there, and probably will be the first to have actual livable rooms. And if the other company tries to take one of their partially-painted rooms and say, “no, we’re very close to having livable rooms - we just need to lay down tarps and put up butcher’s paper and have people be willing to move out so we can come in and paint during the day,” well that’s not a really great argument.
What if both companies are at least ten years away from finishing the entire house, but the room-by-room company is likely to have the 100 most important rooms finished within the next five years, while the other company will still have no fully finished rooms? Well, we might find that to be more important to us than the final completion date.
I suspect that Waymo and Tesla aren’t even pursuing the same goals. I’m sure Waymo would love to be able to provide AV service to every square inch of the country, down to the tiniest hamlet or village - but I don’t think they care if they do. If they can set up an operational system in the 75 largest metros - basically from NY down to about Dayton, OH - they’ll cover about two-thirds of the population. That’s a huge win, and it really won’t matter whether Tesla comes on board with a system that can also provide service to the people in Twin Falls, Idaho. Because that won’t really be a relevant competitive advantage.
And to be honest, until you get to actual L5 full autonomy (which I think we’re very far away from), you probably can’t run a robotaxi service in places like Twin Falls anyway. There’s just not enough taxi demand to pay for the fixed cost of having the facilities necessary to do all the non-driving stuff to the cars (clean, return lost items, shut the doors if the passenger leaves them open) when they’re not being done by a driver.
I’m not assessing a mid-way goal. Are you?
Both want robotaxis and both want their autonomous system in personally owned vehicles.
Tesla doesn’t agree with this. The Tesla Network is purportedly designed so that owners can put their cars into service, so anything physical with the car is taken care of by the owners.
As for shutting the doors, Tesla already has cars whose doors can be opened and shut electronically. And CyberCab is designed to also have that feature. It’s not a big deal.
Oh, yes. I thought we were talking about robotaxis in this thread? Having a robust Robotaxi system is definitely a narrower goal than Musk’s “flip a switch and every car is fully autonomous everywhere” scenario.
It is entirely possible (and IMHO very likely) that both of these things are true: that Tesla is far closer to getting to either full Level 5 or near-universal Level 4 autonomy than Waymo, and that Waymo is closer to getting a broadly adopted robotaxi system established than Tesla.
Those are two different goals. They’re not unrelated, but they’re different. And I think Waymo’s business model is far more focused on the robotaxi version than the personally-owned autonomy version. The latter would be nice, but it’s not really their focus.
I know that’s their vision, but I don’t think it’s very plausible. I’m not going to take off work in the middle of the day to tend to the Tesla that I’ve loaned into the network - and even if I was willing to do that, I’m not going to want to have to shlep all the way into the boonies if someone left the car door open and it needs to be shut, or if the car has a flat tire, or if it gets into some situation it has to be driven out of. Especially since the car that I would normally take to get out there isn’t available to me. So any robotaxi service is going to have some practical geographic limits due to the fact that there’s no one in the car to handle these things. Even a Level 5 car isn’t likely to be venturing out too far away from its supporting human.
Really? Do the Model Y or Model 3 close their own doors? I’ve been in them both and I don’t think they did - but maybe after the upgrade? If not, then this is still a big deal. Because there aren’t any CyberCabs yet, and probably won’t be until after 2026.
1 Like
X does, and there are aftermarket kits to do that on 3/Y.
CyberCab in 2026/2027.
Note also that Waymo apparently still doesn’t do trips on highways:
But, that may be out of date as we know they’re testing on highways in Calif.
The next 12-18 months should help make comparisons clearer.
Yeah, they’ll partner with whomever to try to develop their product. Lots of these strategic partnerships in the space (and in lots of other cutting edge space). Their main business line, though, is the development of the robotaxi.
But none (or virtually none) of the 3’s and Y’s that might make up a “Tesla Network” today have self-closing doors…so that’s a real substantial barrier to relying on that feature to solve the “close door” issue. And of course, the CyberCabs won’t be available for quite a while - certainly not in any volume until 2027 (if then - Tesla doesn’t have the most consistent track record of meeting it’s projected time frames for bringing new models to volume production).
An excellent illustration of half-way goals vs. full goals. That’s certainly a limitation of the Waymo service right now. A “complete” robotaxi would travel on highways. But I’m struggling to remember the last time that I took a cab/rideshare - either here in Miami or as a tourist somewhere else - that ended up going on a limited access highway. Might be a product of the places I’ve done that lately (NYC, Boston, DC, Charleston), which don’t really have highways integrated so much in their **intra-**city trips.
But you could probably have a fairly robust robotaxi system in many (if not most) major metros while avoiding highways….