저자: Kevin · 외국인
요약 보기
테슬라 강세 시나리오의 핵심 쟁점은 HW3 카메라의 체인(장애물) 인식 한계와 이를 소프트웨어·HW4로 해결하는 과정이다. 사이버캡 스케일링 확인(지역 확장 속도, 엣지 케이스 매핑)과 금리 인하가 동시에 충족되면 $775 목표에 도달할 수 있다고 주장. 로드스터 10/1 출시는 주가 촉매가 약하고 '뉴스에 사고 팩트에 파는' 이벤트로 봄. 사이버캡 생산 효율 90% 개선은 목표치이며 실제 달성은 50% 수준이라고 정정. 장기 관점의 인내를 권고.
다듬은 글
Is the bullish scenario for Tesla ruined if the chain problem is not solved ? What about the Roadster launch? How about what Arc Invest just said? Or how does Uber fit into all of this? Friends, there is a lot to sort out here, including a lot of ambiguities. We've talked about some of this before, but there's a lot we haven't touched on yet, and it's finally time to clarify the situation on some issues.
So the first thing people worry about is they say, “Kevin, if the bullish scenario for Tesla is to get to $775 per share and above as soon as possible, then we need to scale Cybercab. But how do we scale Cybercab if we can't even solve the chain problem? » This is an example of how Electrified describes a chain problem.
And I'm going to simplify this problem for you as much as possible. Then you will see that it is not a problem where it is important, but for some it is still a problem. This is a problem for those with 3rd generation hardware cameras (HW3). You see, HW3 cameras typically perceive horizontal objects and can actually process them through data compression with a convolutional neural network.
That's a lot of information, but essentially, when you have this artifact on the camera for the Model 3 with HW3, it frankly disappears. Here is the Astra output, which roughly shows what it looks like. This is a chain captured on a camera with 3rd generation hardware. So it's a combination of both older cameras and an older computing unit.
And what the computer actually sees after processing is that those 1-2 pixels of the data chain disappear. You get this image, which essentially assumes that the car continues to drive on the chain, and therefore there is a problem with the chain. People say, “Oh my god, that means it can’t be solved.
This means we won't be able to achieve full autopilot." The solution, well, I mean, there are two ways to solve it. There is a correct solution, but there is just an example of how it can be solved quickly. One way you can instantly solve this horizontal chain problem for Tesla's full autopilot is to literally turn the camera sideways, and now suddenly these horizontal problems disappear, and the chains become very clear vertical artifacts that are surprisingly easy for artificial intelligence to recognize.
But then you no longer have a single data stream between the vertical camera and the horizontal one, and your computer is like, "What is this?" This won't work. So you solve this using employment networks and time context. Okay, in simple terms, what you do is this : when you see tiny objects, you turn them into large blocks, Minecraft-style, and they send a signal that that area is occupied.
But it doesn't just happen once and then disappear in the next frame. The temporal context that Tesla uses in image processing is that if they see something in one frame, they remember the context of where it is and what was in subsequent frames. So if you see a chain in one frame , you expect it to stay there in the future.
This is a very realistic solution to the chain problem. Plus what are called skip connections . Through connections . Again, an oversimplification. This is actually a way in which you literally take high-quality image data from cameras. You skip the compression entirely, and so the algorithms at the end get both high- resolution and low-resolution context to understand what's what.
You know, 99% of the work can be done in low resolution, but then you do, so to speak, double- check with the high- resolution version, and then the probability of spotting the chain is higher. The problem with these three solutions— time context, busy networks, which are Minecraft-style blocks , and of course, bandwidth connections— is that they require a lot of hardware resources that are not available in Hardware 3.
You see, Hardware 3, and people forget about that. I think we should just write this down, because there are so many examples of people criticizing Tesla, saying, "Oh, Tesla's FSD sucks." But they use cars from Hardware 3 as an example. That's like referring to a seven-year-old iPhone. Do you know what this is?
It's like an iPhone 9 or 10, or something. And I'm not saying they don't work. They work. For example, my black Model X runs on Hardware 3 and it drives great. Just not as cool as the newest and most powerful Cyberbeast from Hardware 4. And of course, the Cybercab. This is quite understandable. When you look at Hardware 3, you see 2019 technology : 12 cores, 1.2 megapixel data, and 2x8 GB of RAM.
Okay, this is 2019 technology that people complain about, like, "Oh , why can't he see the chains?" Because you need better hardware to be able to see them. That's how it is. Besides, Tesla already comes with a backup, right? This is multiplication. These are actually two systems with 8 gigabytes of RAM side by side, it's like an airplane: two engines, two pilots, two fuel pumps, and two brake systems.
I had to use this example with my own plane, we've already talked about it . We're like, " Wow, one of our brake systems is failing." " We hope the emergency brake will work." I'm still here, of course, so you can guess that it worked. But anyway, you know, there are two chassis release systems. Oh, you know, this hydraulic system is n't working.
No problem. We have a backup. OK. So, redundancy is built into the design. It is laid down at different levels. But think about version 20 or hardware 4. Hardware 4—that's 2023, and it really scaled to 2024 technology. So in the 4 hardware, you get almost twice as many cores—a whopping 20. You get 5-megapixel cameras and 16GB*2 of RAM, right?
This is 2023 technology . And know that in Cybercab, the amount of RAM will probably double, at least. Again , we don't have exact RAM details for the Cybercab yet, but we believe there will be two 32GB slots. This allows for a much more powerful peripheral computing device, i.e. the car's neural network, compared to some older generations.
You can run larger models, and you know what? More context, more power reserve. Hey, why does AI hallucinate so much less today than it used to? Because we have broken the memory barrier with high-speed memory and increased on-chip space. If you are currently using a GB 6000 with 96 GB of video memory and are running a model that previously ran at 32, you can run it at 96; but now you have a bunch of free space—another 64GB for context, which would be like the chat context we were talking about, or the size of a contract or codebase that I can upload there.
For Tesla, it works the same way. So I think it's a little unfair when people say, "Oh, you know, 2019 equipment doesn't see chains. This is the end. This is an existential threat to Tesla.” No, it's just technology. Although, to be fair, one of the problems, and this is probably a more real issue for Tesla...
The real problem is not that they, you know, are significantly improving the Cybercab hardware, although that is the case, compared to what's on sale now, but that's okay. We expect that with each generation, each year, the hardware will get better and better. The bigger problem is not the circuit itself, which I can show you.
There are two ways we can solve a chain problem graphically. One is what I was talking about: the hardware 4 gets significantly more detail and a headroom of computing power. And what do you see? I see the pixels of the chain. Wow, I see a whole chain here. Oh, guess what ? So, we don't drive through it, but we stop in front of this chain, right?
But Cybercab does more than that, Cybercab also does something else, and that raises one of the issues of risk. Cybercab knows how to avoid known problem areas. Maybe they have a fancier word for it, but that's basically what it looks like. So, Cybercab uses Hardware 4+, that is, those 32*2 GB of memory , which is again significantly more, well, it's four times more than we had in the previous generation.
But for on-device AI, this is a huge step; you can run a much larger model and, again, with more context. Context matters, especially when driving. The chain just doesn't disappear in front of you. But anyway, Cybercab probably also uses the phrase " known entry restriction," where you know historically that it's an area that you shouldn't drive on, and that may have been confirmed by humans, so the car sees that area and doesn't care if there's a chain there or not.
He just won't take this little road. So, with or without a chain, Cybercab will simply stay away from this area. This is, by the way, Astra, which I think is a pretty cool little model. I mean, I put this together in about four minutes. This probably cost $50 worth of tokens. I don't know if that's true, but you know, when we put this together, we start thinking, " Okay, so what does this mean for Cybercab?" It will have better power reserve and better data.
But the known entry restrictions go back to the “edging” theory, and that’s my biggest concern about the speed with which Tesla will be able to bring these “toys” to market en masse . You may already know—we talked about this before—that Electrified, together with Perplexity, created a great dashboard based on Vercel.
I'll leave the link in the description. You can also find it on their channel. But, whatever it was, I found it quite interesting. This is just a small map of the areas where these services are expanding . But the most important thing, which we have already talked about, is right here. This is a timeline, real-time counters, right?
It's a simple counter that says, "Look, we've been actively operating Cybercab in Austin for 14 days, and the entire robotaxi network has been operating there for about 448 days," which actually seems like 100 days longer than I remember, but I guess time sometimes passes faster. Whatever the case , the San Francisco Bay Area is starting to see some support for these Cybercabs, but many other cities—Tampa, Orlando, Miami, Dallas—don't yet support these vehicles.
So I think we really need a dashboard that shows the number of days— maybe I'll make one myself. It's not that hard to demonstrate how many days passed from the appearance of robotaxi in Austin to the appearance of Cybercab, but let's imagine it. So it took 0 to 436 days in Austin to get a Cybercab, right?
We need the next place to look like this, and then the next one after that. And once we get to the point where we can master the territory in 90 or 30 days, that's when you should be excited about Tesla stock, because that's when the real scaling will begin. Right now we're still in this unconfirmed hype of, "Dude, we just put Austin on the map." And I'm not talking about Waymo-style mapping.
I want to emphasize this. I don't think they do mapping because they themselves said they don't do it. But I think they test and validate the area so thoroughly that they've actually covered all the edge cases. That's why I call it "edge mapping." And again, the speed of entering a new market will be , so to speak, the best proof for you of when the company will finally bring cyber taxis to the mass level.
Now I want to touch a little on what Arc Invest said. And then, of course, we have another catalyst ahead of us—version V15. That's great, but again, all of this will always require more and more equipment. This is understandable. Let's focus on this interview. It was actually a conversation between Sam and Brett from Arc Invest.
And I'll just give you a few of the main conclusions I've drawn from this. So, some of the conclusions I've drawn about cybercabs are that I didn't initially factor this in when I modeled my numbers for Tesla's optimistic $775 per share forecast. It makes sense that in addition to selling the car and getting a margin from it, you would also charge a platform fee for people adding their taxis to the robotaxi service through the Tesla app.
Well, through the robotaxi app developed by Tesla. The interesting thing is that you will have to take it really seriously , and we have already mentioned that if you get a low rating due to poor fleet maintenance, you will simply be removed from the system. What did I mention? But I didn't mention that Tesla will probably take a percentage, like the franchise fee at McDonald's.
You know, give me 6% of the revenue. This is another potential product or profit vertical where Tesla gets a large margin at the start from selling cars, and then residual income from full autopilot and platform collection. It's almost like owning all of McDonald's real estate and collecting 6% from all of the franchisees who run the actual business.
So, you know, they deal with all the mess that happens when people mess up their cars. And Tesla—no. They also noted, in particular, that Brett believes that the cost of a cybertaxi ride will be about 25 cents per mile. I estimated the cost higher because I factored financing into my calculations. So I think there is still an element of interest rate sensitivity in this business model, even when it reaches mass adoption.
So, the second driver after scaling begins will undoubtedly be interest rates on financing. Who knows, maybe Tesla will offer business financing at a lower rate, but then they will cover the costs and discount points themselves. So this is worth knowing, because when determining Tesla's place in your investment portfolio, you need to understand: we need to confirm scaling to map edge cases and we need interest rates to fall.
If both of these things happen at the same time, the company could take off, because that's when you can really get the " cybercab" business going and, as Brett himself says, you can push Waymo. The only thing you won't be able to click on is the price of the products on meetke.com, because we don't have a discount coupon right now.
And I apologize for that. I just want to be frank about this. So, Brett argues that if Tesla can get its costs below $1 per mile , which they expect and I do, then Waymo will be too expensive to compete at that price. Waymo is a great experiment and they're doing a good job, but it's not a scalable model.
This is a classic Tesla argument: Tesla's strategy is scalable, while Waymo's strategy may not be, so we'll see. Because right now it seems to me that we are doing mapping in the same way as Waymo, not necessarily the map itself, but the edge cases. Oh , there's the chain again. " Cybercab," to be honest, would probably still recognize the chain, but there are still quite a few edge scenarios that need to be mapped.
And Tesla's advantage is that it can involve crowdsourcing in this. That's why now, every time you turn off FSD, the system asks whether you turned it off for navigation, parking, a critical error, or something else, and Tesla's validators can prioritize software improvements based on the feedback they collect on a massive scale .
This is truly impressive. Well, Brett, good thoughts. I liked it. He also considers the V15 release to be very important, as it somewhat constrains their expansion capabilities. Although I assume that " cybercabs" already take all this into account . And this, in fact, is the difference: there are a lot of Model 3s among the public , there are a few Model 4s, that is, I meant Hardware 4.
So, Hardware 3, there are a lot of them, some of Hardware 4, mostly these are sales from about 2024, partly from 2023. And besides, Cybercabs have an advantage because they are all equipped with the latest Hardware 4+ equipment. You see, that's why we had to quantize Hardware for V14 to create a lighter version for Hardware 3, because we have to squeeze all this context into a smaller amount of memory.
Otherwise, Hardware 3 will slow down. And this is even worse. Because you can get an FSD that brakes and literally drives into a park bench or something else silly, or into another car, just because of the delay. You do n't want that. Just to address the edge case, you don't want your main system to fail.
You don't want this. So, honestly, Hardware 3 will eventually just be left behind. And that's going to disappoint some car owners, but I think a lot of people...well, there's going to be some people who are going to be really upset, and, you know, there's going to be lawsuits or something about it .
But I think the vast majority of people would say, "Well, of course, the car is 7 years old." It's probably time to buy a new one. Just let me transfer my " lifetime" FSD license, so to speak, from my first machine. And I think that's probably the case , most reasonable people would agree. And I'm not saying you were lied to or anything like that .
Okay, this discussion can be held separately. I understand. But , you know, Elon has always been very optimistic. We recognize this. So I'm not here to take anyone's side. I'm here to say, rather, is this an existential problem for Tesla? No. And that's where we come in with the launch of the Roadster and the other data we have .
Roadster launch on October 1st, all this " launch readiness" talk. This is really cool. We have already covered this in reports by The Information. According to reports, it will be a stunt, a complex stunt, there will be no one inside the vehicle, and it will be able to fly thanks to cold gas engines developed in conjunction with SpaceX, but it is so loud that it can damage eardrums.
So, it will be controlled remotely, probably from the SpaceX facility in Texas. Some materials were distributed here as an example of what this could look like in the end. Let's take a look. Here is the outline of the car. And you'll actually see it go up 20 feet. I think they'll launch this thing over the ramp, and it'll fly over it .
Let's see. Watch him rise. And then there he is, here. And a vortex forms beneath it . Almost like a tornado underneath him, and there will be some sound in the next passage. Oh, I ca n't really hear anything. It doesn't matter. It's just so-so , but most importantly, it's cool. I don't think this will actually be a big catalyst for Tesla .
This may be a buy-on-rumor, sell-on-fact event, but if you believe in CyberCab scaling and that it is sensitive to rates that will decline in the long term , then there is some value in this. Especially because I like the recent Wall Street Journal reporting on this. Tesla is reclaiming the electric vehicle market in the US, while traditional automakers are retreating.
I find this quite interesting. The company did not escape the general downturn in the electric vehicle market. Tesla sales fell 16 % year-over-year . However, the rest of the electric vehicle market fell by 30%. So yes, they are slowing down, but slower than everyone else. And that means there are fewer alternatives as traditional automakers are somewhat abandoning electric vehicles, so if you want an electric car, you're left with Tesla.
And Tesla still has the best fully autonomous driving system. That's why I think we've been seeing headlines like this for years. Hyundai says it will overtake Tesla in autonomous driving within 5 years. I've been hearing people say this for what seems like forever . I mean, I bought my first Tesla in 2017.
So, my god, it's been almost a decade, man. Well, man. It was magical when I first tried autopilot in December 2017 and then bought stocks and it made millions of dollars. And I'm like: This is just amazing, you know? But some of that magic has already disappeared. Of course, a lot of time has passed.
You get used to everything. It's still impressive. I feel like I'm not driving the car anymore. But since then I keep hearing the same thing . They say that this or that company will soon release FSD or some competing software, but I don't see it yet . I'm sure it will happen, but it's obvious to me who's in the lead right now.
And I hope Tesla can maintain this position. Oh, I already deleted everything. I meant to say that it would all come down to edge compression. This is exactly what is needed. Pinch the edges. The sooner we move from edge processing to completion, the better . Okay, here's someone else claiming that the JP Morgan memo says that almost all the billions from robotaxi will go to Tesla, not you.
Essentially, this is an argument that you were lied to, saying you could put your own car in a fleet, rent it out, and make money overnight. You know, I think most of us saw this as a bit of an impossible dream from the start. I don't know, maybe 10 percent of people will say, "No, no, I really expected that." I think there is a difference between expectation and hope.
We can hope for that , but let's be realistic: they will need the latest and greatest equipment, and first it will be cybercabs. You will be able to buy them in the future when we fix the edge issue and add them to the network. But will we just give our regular cars to the network? Very doubtful. And it will also create serious maintenance difficulties, I mean controlling charging and cleaning.
You don't have the means to clean these cars regularly. Are you going to run out to your car every time someone complains about the smell in there and drop your business to clean it up? This is something that will work well except for fleet operators. And I suppose Tesla could have some cleaning stations or something like that, but I don't think that's where the profit is.
So there's this utopian scenario of, "Oh, I can make money while I sleep," and then there's the reality of, no, Tesla will offload the worst part of this business to fleet operators who can do it at scale, and they'll be happy to get a little profit and a big tax break, you know, a 100% tax write-off on capital expenditures thanks to Trump.
And then Tesla will take a huge profit from Cybercab, taking a license fee for the app and, uh, getting paid for FSD subscriptions, which have, you know, 90+ percent margins or something like that, uh, additional profits because, you know, you're still selling a car; even for it to be completely autonomous, we can say that the additional margin is about 99%.
This is quite impressive. So I think that's great. Now I will say that I was a little disappointed with Tasha's script. Tasha made this little video for Arc Invest and they put a bunch of footage on top, which was cool. But she had this little two-minute video. And she claimed, I think somewhere in here, that this unpacking process would be 90% shorter or something like that.
Let's listen. And I thought that was a little careless. I wanted to add a little context. Everything else is fine. This is a kind of summary about Cybercab. Let's listen to this part. a process that is much more efficient compared to traditional car manufacturing. They were able to reduce the production line, they said, by 90%.
Hmm . Okay, let's clarify this. No, they couldn't reduce the production line by 90%. At the Tesla Cybercab presentation, they said they were able to reduce the duration of the “unboxed” production process by 50% compared to the traditional process . Which is still fantastic, but it's not 90%. The only place where you could claim 90% is in the prediction that maybe the car could depart their general lines every 5 seconds compared to every 47 seconds for the Model Y.
That's where you can talk about 90%. But she just told you that they were able to achieve 90%, you know, 90% more efficiency. No, they couldn't . That's the point. It's just an aspiration. We're not there yet. And that little word or phrase really matters, because you don't want to buy a stock thinking it's already reached that scale.
Those of us who are really paying attention know that all of this is still to come. But not everyone understands this. Listen again. A process that is much more efficient compared to traditional car manufacturing. They managed to reduce the production line by 90%. Ahem. These are her words, not mine.
they managed to reduce it by 90%. Again, not 50, they said it was aiming to reach 90%. But that's not how it's happening now. So you know, sometimes I hear this and think: is this why people ask why stocks aren't going up tomorrow? OK. Of course, we must remember that there will always be negative press around Tesla.
For example: a 14- year-old teenager died in a Pacific Beach accident on an electric bike involving a Tesla. I don't even understand why there was a need to mention Tesla's participation. I think they just understand that it's clickbait. But everyone knows that 14-year- olds on electric bikes are just crazy these days.
They don't have driver's licenses. They don't know the rules of the road and feel invulnerable. And when you read the details, it turns out that what do we have here ? According to San Diego police, the teenager was riding an electric bike on Zoo northbound at 9:20 p.m. and ran a red light at an intersection.
So, he ran a red light at 9:20 p.m. Seriously. Unfortunately, the guy died, but you have to admit, the situation with electric bicycles has gotten out of control . This is enormous power in the hands of children who don't know the rules of the streets. This is very sad and terrible. But unfortunately, this will continue until there is more regulation around the e-bike craze.
In any case, it's interesting to see what happens with this Roadster. Do I really think we'll be able to order a Roadster anytime soon ? Honestly, no. I think this is a showpiece event. and the Roadster presentation. If you want to buy more Tesla stock for the long term, I think you'll probably have a chance to do that, you know, this is not personalized advice, after the Roadster launch, because people will see the launch and say, "So , when can I buy it?" And then people say, "Oh, you can't ." It was just to show how cool our engines are." And we're, you know , we're still in the development stage, and people are going to be pissed off about it right now.
People are going, "Oh yeah, it's coming soon." That's great. The bottom line is, with Tesla, you're going to have to be very patient. I think that's the main takeaway. Let them finish their work, and then we'll be much further along. That's my take. Bye . Bye. Why don't we advertise what you've sai thoughts.
Comments 0