Even so. A huge number of Teslas are being sold where they’ve already sold a lot of Teslas. The 405 isn’t the only driving scenario that’s abundantly covered by Tesla’s existing data (and therefore little marginal benefit to more data). Tesla already has mountains of data on the driving environment and behavior of Tesla owners in California, so it’s likely that getting more data about that has a pretty low marginal value. The second million miles just doesn’t give you as much new info as the first million miles did, and so on and so on. Most of that “parabolic” data is just going to be more instances of things they already have plenty of data on.
WRT geofencing, I’m not sure it matters. If my car can drive itself in every metro area and the roads between them, I don’t really care that self-driving might not be available on every country path or off-road route. If I’ve got a car that can self-drive 99.9% of the trips I take, I’d be perfectly fine with that - and if I need to take the wheel to get to one or two random destinations over the course of a year, I don’t care. And if AV usage is part of a TaaS operation, that’s not going to matter at all.
Correct. But they were the ones that figured out that the ~best way was to rank results by the links to it rather than just a list or a word search like others were doing previously
Okay. But I don’t disagree that if Tesla comes up with an AI driver that’s categorically better than any other offering, it will capture a vast amount of the market share in that category.
But as you can tell, I’m a little skeptical that’s going to happen. Or even if it’s possible. If we’re in a TaaS world, I don’t think it’s going to matter - when I hail a taxi or get on a bus, as long as my driver is competent and safe I don’t especially care whether they’re an average driver or an outstanding one. Even in a privately owned self-driving world, any AI that’s safe and effective enough to be allowed by regulators is likely to provide the same experience to passengers as any other.
So I can’t really see it being a situation like Google, where one program did a much better job of getting me where I wanted to go than all the other programs. Any AI that’s an actual self-driving car (and which is allowed by regulators) is going to correctly take me to my destination, safely and obeying traffic laws. Not a whole lot of room for differentiation.
When one is first starting out, getting the simple stuff, then the dataset doesn’t need to be very large to hold a significant number of examples of those simple cases, but as you progress to more and more special cases, the data set needs to be dramatically larger to contain any meaningful number of examples on which to train and test … and by “dramatically” read exponentially.
Well, anything out there that claims to be more that 2 is heavily qualified about the circumstances in which it can be used. To me, that is a false claim. It is not only not all driving, but is a very limited part of driving.
When one is first starting out, getting the simple stuff, then the dataset doesn’t need to be very large to hold a significant number of examples of those simple cases, but as you progress to more and more special cases, the data set needs to be dramatically larger to contain any meaningful number of examples on which to train and test … and by “dramatically” read exponentially.
Sure. But the converse isn’t true - just because you are making your dataset dramatically larger doesn’t necessarily mean you’re capturing more and more of the “special cases,” especially if there’s a lot of congruence between your later data and your earlier data. Once you’ve sold 50K Teslas to upper-income Los Angeles County residents (for example), you’re not necessarily broadening your exposure to different edge cases by selling the next 50K Teslas to upper-income Los Angeles County residents. They’re largely going to be driving under similar conditions to similar destinations.
That’s an exaggeration, of course - Tesla is broadening its geographic sales footprint, not just selling more cars in LA County. But it’s not not the case that an awful lot of Tesla’s sales involve large volumes in markets that they already have a lot of coverage in. Tesla’s going to sell a quarter million cars in California this year, close to 20% of their global total - that’s a highly concentrated customer base that’s duplicative of where they’ve already gotten a lot of their data from.
If Tesla owners aren’t seeing dramatic changes in the quality of FSD as updates roll out, that might be part of why the parabolic increase in data hasn’t translated into parabolic increases in functionality.
Really? AFAIK, the Waymo and Cruise operations are now available throughout the entirety of San Francisco - and expanding to various other cities. Those are the Level 4 operators: completely driverless and capable of operating through the entirety of a metro area.
I agree that the Level 3 systems (like Mercedes’) are really constrained. But on the TaaS side, you have systems that are technologically able to operate autonomously within the geofence.
Edge cases are not defined by geography. Different places and countries and seasons do create new cases, of course, but something that happens infrequently can easily happen in a place that already has a zillion miles recorded.
Geofenced may be fine for taxis, but is irrelevant to general car ownership.
Not really. If there were a car that could function autonomously in most metro areas and on major roads linking them, and you’d just have to drive yourself if you went outside those areas, I think that would be a pretty compelling product. Especially for an EV, which is more likely to be driven in urban areas. The analog would be the BlueCruise product - it’s a hands free level 2 system, like FSD, that can only be activated on limited access highways.
Living here in Miami, I don’t need a car that can navigate the backroads of the Tennessee mountains. I’d be fine with a car that could drive me where I need to go in Miami, the highways around the state, and in the other big cities (Orlando, Tampa, etc.). I’d be thrilled to purchase that as an option, even if I’d have to drive myself if I wanted to go off the Turnpike to a small town somewhere.
I don’t think Google is a particularly good example. For starters it was free, so there was no cost impediment to trying it or not. No it to mention it was platform agnostic, so you didn’t have to plunk down $75,000 for the platform and another $20,000 for the software.
It was also an arena which is widely seen as “winner take all” because the universe keeps expanding, and others couldn’t keep up. More searches begat more revenue, but eventual also rans like AltaVista or Yahoo chased the wrong squirrel down the wrong rabbit hole and fell away either because they weren’t as good, or they became not as good because they couldn’t keep up.
With self-driving there is a (large but) limited data set to master, even with edge cases. After that everybody gets the same playing field.
The fact that Musk is (allegedly) talking about licensing the FSD tells me that he recognizes that it’s a perishable advantage - unless he can put himself in a position to close out the other competitors (a replay of the Tesla plug scenario?) in which case he wins industry-wide, but at the cost of a hardware benefit.
If that was true every company in California would be level 5. I suspect every level they go up the harder it gets and the more data and better hardware they need. That is why you see Tesla continuously updating the software and the hardware.
I don’t see Cybertruck being exceptionally high volume. I’d be very surprised if they reached 500k.
As for the new model to be produced in Mexico, I expect volume to be multiple millions because it’ll be in the heart of the global market. And that they’ll be at that sort of level well before the end of the decade.
But, as I said, we shall see. It’s sort of pointless for a group of us to exchange our expectations otherwise. THere is no strong evidence to be presented beyond our perceptions of Tesla’s track record and competitive actions… and I doubt that either the optimistic or pessimistic camps will convince the other.
Rob
He is no fool who gives what he cannot keep to gain what he cannot lose.
It could also be that he really believes all the stuff he’s been saying for a decade or two. That he’s in it primarily to improve the world rather than to solely to improve his companies.
In the last couple weeks, Tesla reports a whole new way of processing FSD.
It replaces millions of lines of code with about 300,000. IIRC.
A couple YouTubers say that the process now does NOT train every road geography nor situation.
Instead, it trains the FSD the same way humans drive. If a human has Experienced a situation then that human “knows” how to handle that same situation, whether it’s in TX or TN, AK or AL.
I often drive roads, highways, neighborhoods etc, that I’ve never seen before. I’m successful cause I’ve driven others that are similar.
That explanation assumes I understand correctly, and the YouTubers also understand correctly.
Hopefully one of you can fine tune it?
The Level 2 conundrum. YouTubers, again, explained a year ago, that L2 allows every Tesla on the road to use the L2 level software, while running higher Level software behind the scenes, collecting those millions of miles of data.
If Tesla legally registered it as higher than L2, the number of cars collecting data would be similar to Cruise n Mobileye. And the range and be similarly limited. The data collected also would be dramatically limited.
FWIW.
ralph owns some TSLA core position, and a trading position.
BTW I’m currently navigating the wilds of Kauai. I’ve never seen these roads before.
Oh, I don’t think it’s pointless. Having a discussion about it helps us understand the assumptions that underlie Tesla’s sales goals. We don’t know the future - but we can unpack what would have to happen for Tesla to sell 20 million units in 2030.
It seems feasible that a carmaker might, one day, sell twice as many cars as Toyota. Heck, if Toyota merged with BMW today, you’d have a single automaker selling 20 million units. Achieving 20 million unit sales is something that an adroit and very competent carmaker could reasonably do.
But with only three models (in volume)? That’s another thing altogether, since each of the high-volume models would have to sell more than 6 million units annually. No single model of any car ever has sold more than 1.6 million units, globally. Heck, Tesla will have to sell more than 100% of the market segment those models are in to hit that goal. Decades and decades of consumer behavior exhibits a very fragmented and heterogenous preference set for a variety of different form factors, model types, design styles, and price points.
So while hitting 20 million units - in general - might be something that would depend on Tesla’s performance and competitive actions, hitting 20 million units with only three models depends on a sharp and dramatic change in consumer behavior.
??? Not at all. I don’t think that you can get to level 5 with a small data set - just that it’s possible, perhaps even likely, that you can’t get to level 5 with a parabolically large data set, either. Really super-big data might not help. Level 5 just might not be something we can build with our current knowledge and technology. We know we can build a geofenced Level 4 with back-up remote human for infrequent human intervention - and that might be what is available for the next decade or so.
My car? No. But there are cars in a half-dozen cities around the country that can. Cruise and Waymo have cars that can operate without a human driver, and have obtained regulatory approval for them to operate that way. And if that’s the direction that driver-assist ends up taking, at least in the near-term, then Tesla won’t have a differentiating feature that can allow them to move 6-7 million cars of a single model.
Right IF and everyone that is invested in Tesla is telling you that is not the direction we are going. IF Tesla does get to level 5 all of your IF’s go out the door. IF they don’t then mine do. It is simple as that. I like my bet on the future because Musk has already put so many IF’s behind him.
But that doesn’t mean a parabolically large dataset of human driving gets you to Level 5, either.
It may very well be the case that it’s not the size of the dataset of human drivers, but our knowledge and capabilities in designing the networks training on those datasets that’s the bottleneck. Or that we don’t (yet) have enough data on how AI drivers perform in the field in edge cases - which (ironically) Tesla’s approach may yield a smaller dataset than competitors. Building an AI driver might not involve feeding our current AI driving programs more data about how humans drive, but either a better neural net and getting more data about how AI drivers perform in the field.
FSD problems like the one in this article:
The author reports that his latest FSD update tried to swerve into a highway median.
But that kind of issue probably doesn’t stem from not having enough data about how humans drive on limited access highways. There’s no way that Tesla - of all companies - doesn’t have sufficient training data on that particular driving scenario.
It’s not just an IF. It’s also a WHEN. Tesla only has two high-volume models today, and will have no more than three (and possibly still only two) for the next several years. But it’s plans call for projected sales of more than 2.5 million next year, and 3.5 million in 2025. It’s not at all unusual for an automaker to sell 3.5 million units in a year. A half dozen did that last year. But no automaker has ever sold more than 1.6 million of a single model - and Tesla’s going to need to do that next year (probably) or the following (certainly).
You’re kidding right? Musk is well known for making wildly outlandish predictions that turn out wrong.
“These are just guesses, people shouldn’t hold me to these things.”
…
Musk is no stranger to making highly ambitious predictions about his business ventures. For instance, Musk has been promising that Teslas will drive themselves next year — every year since 2014.
That does happen too! But for all of the outlandish predictions, the ones he has made right have been amazing. Nobody thought he would be able to build Tesla into a 500 billion dollar company or land a rocket on a pad. Like I said, I like my chances. IF he gets to Level 5 Tesla will be at least a 2 trillion dollar company.
If Tesla “gets” to Level 5, and if Tesla gets to Level 5 first, and if no-one develops a “good enough” Level 4 alternative (either privately-owned or TaaS), and if Tesla gets regulatory approval for it to operate Level 5 in (most) of the markets they sell in, and if no one else gets to Level 5 for several years after Tesla does, and if the market price for autonomy remains high enough, then Tesla might be a 2 trillion dollar company.
But for all of the outlandish predictions, the ones he has made right have been amazing. Nobody thought he would be able to build Tesla into a 500 billion dollar company or land a rocket on a pad. Like I said, I like my chances. IF he gets to Level 5 Tesla will be at least a 2 trillion dollar company.