HONEST Thoughts on Tesla's Robotaxi with James Douma
02:12:12 Share

Alright, it's day 6 of Robotaxi launched by Tesla. Today I'm going to be meeting up with James Dalma. He's been a frequent guest on my channel. We talk about Tesla AI and FSD and a bunch of other topics. So I'm super excited because it's been like a long time since we've been talking about this.

And I remember actually I test rode I think version nine FSD city streets with James Tama a few years back. And Tesla has made such big progress since then. And it's been an amazing journey. And I'm especially excited to hear James's thoughts on Robotaxi. There's lots of questions like, you know, what are the weaknesses? What are the faults?

What are the challenges to scaling? What's the timeline that we could expect for Tesla to scale RoboTaxi? How does it compare with Waymo? Tons of other questions. So let's go ahead, meet up with James and take a bunch of RoboTaxi rides and hear what he has to say.

Alright, so James just knocked on my door and we've been talking. We're about to take our first Robotext ride. So how long have you been in Austin for? This is my third day of driving here. And how many rides have you taken so far? That was just my 24th. 24th, nice. Alright, so the plan is we'll take a bunch of rides. I'll record our conversation during the taxi rides and probably in between. And yeah, we'll have a fun time. This is

The Montopolis Recreation Community Center. It's kind of a longish ride and it takes us to a park. All right, let's do it. Okay. Call it. It says it guessed four minutes. Now it's found one six minutes away and then that's kind of the... So this is a route. So you can see that it shows you immediately once it picks a taxi which like where the taxi is, what its route is going to be to you. This is us over here at the pin. Of course, we're sitting in the car. The pin drop...

is a little bit away from where we are. Actually, we are right over here. So the GPS dot is a little bit off. But it's frequently the case that the pin drop-- - Wait, it's making us walk across the street, huh? - We're not across the street, we're on this side. - Yeah, we're next to Starbucks. - The dot is right here. This dot is wrong. The dot is over there. We're actually here. So it's gonna come to the parking lot where it just dropped me off in Starbucks. So anyway, we-- - Wait, isn't Starbucks right here?

Oh wait, you're right. It is. It's making us cross the street. Oh man. I think the initial GPS location it sent was to Proxman or something. I don't know. I've had this happen a number of times. Where, like I went to HEB and it won't drop you off in the entrance.

inside parking lot and there's no outside parking lot so it drops you off at a bank across the street yeah but the thing is i've actually had been dropped off and picked up here multiple times at this starbucks yeah yeah well it clearly can come it just dropped me right there yeah yeah so it's picking us up yeah okay um so anyway we'll do that um it'll be here in five minutes i guess we could wait two minutes before we go stand in the heat yeah so james i have a question my seven-year-old was asking this i was like telling her that you know there's this

geofenced area and it's pretty big and originally I thought that Tesla would have to go through each intersection multiple times to make sure you know it's safe but then and then my 7-year-old was like no they probably just like

ran it through simulation a lot. Like all of this, the intersections to make sure that, you know, it's safe. It would be super easy for them to build a simulator. I mean, they've shown us simulator technology that could, you know, basically span this whole area so that they could have a simulator of the area if they wanted to do that kind of stuff. Do you think mapping, doing LiDAR mapping would help?

Some, yeah. I mean, it gives them some more data for registering stuff. So yeah, they could be doing that. And then, yeah, because I would imagine they could go through

all the streets initially, maybe run it through simulation, you know, whatever times they need to do. It just boosts up the confidence in this area, perhaps. Maybe. I mean, the system that they have where they run a bunch of cars through a bunch of intersections and then they build a map based on lots of overlaying, you know, 3D perspective stuff, that gives them a pretty good...

map but maybe maybe they want to map all the potholes maybe they want to know how high all the speed bumps are i mean there's some things that are difficult to resolve you know i noticed that it's so good with speed bumps here yeah like it's crazy good with speed bumps so maybe they measured the height of the speed bumps you know and they know now yeah i'm not sure really yeah like part of me thinks like what kind of magic had they done you know with with robotaxi um it just seems like it's

There's so much stuff you can do. It's uncannily amazing. It's crazily smooth. I was skeptical that it would be... There are certain aspects of the dynamics that have always been significantly better than Waymo because the end-to-end planner is just much more smooth and human-like.

But there are lots of little things, too, where it seems like it's outperforming. My experience with Waymo in San Francisco, right? Which is different. Have you done a lot of Waymos? Like 100 rides. I use them in my regular day-to-day life in San Francisco over the last 18 months.

Yeah, so I haven't tried using them in Austin. I've heard, I suspect, they started up in Los Angeles. And if you look at the incident statistics, like they were having way more incidents in Los Angeles after they started up there. So I think that's just growing pains, right? They're getting out of the city. Because it definitely got a lot better in the time I've been using it in San Francisco. So there...

I don't know if they're settling in, you know, whatever it is. The longer they operate in the city, the better the service seems to get. The driver used to be, it started out being pretty timid. In fact, Karen and I got stuck in the Waymos a couple of times because like you'd get into a situation it couldn't get out of and you're in the middle of the street and it doesn't want to let you out in the street or whatever. And then so they'd have to intervene. But the, you know, the Waymo cars, they're notably more aggressive now.

and more willing to like squeeze into little holes and stuff like that. So I think they're just not getting stuck in everyday situations as much as they used to. And that, you know, which is just a sign of them getting better as they run their service. Yeah. Anyway, we've got two minutes. We should go. Absolutely. Yeah. Thank you so much. There you go. Thank you. Oh yeah. You should pay attention. Okay. On top of us, 16 minutes. So this is a little on the longish side.

for a time of day when there's not a lot of traffic. So you've done 29 rides. 24. This will be my 25th. Have you had any interventions or anything? No interventions. Okay. No, they've been really... And no shortage of complicated situations where you might anticipate an intervention. Actually, on the ride over here, there was a pickup truck that just came right out of a...

like it was in a gas station or something, and it pretty much cut the taxi off. And you could see that the monitor was like...

you know getting ready to do something but nothing was necessary i mean the car slowed it left plenty of space it was actually kind of funny we we had three pickup trucks in a row of like human dry human pilot kind of things and the car handled them all really well so i've been quite impressed with i mean i'm generally impressed with 13. i've over the last few years i've had a bunch of opportunities this is my third trip to austin since january

And third trip with, you know, the truck dragging me here. So I get plenty of practice in the Cybertruck around town on FSD. So do you think this version is like...

on the similar-- let's say, is it a 13.3 version, or do you think it's an early version 14? The dynamics aren't enough of a-- it's hard to be confident in making a statement because it's not

you know, we're in that zone right now where it's good enough that it's hard to find things to criticize. And so if you're if if the metric is just like, you know, what's the rate of interventions once they're down? Like for me, serious interventions, probably. Man, I don't know.

I got to be in the, you know, one per 10,000 mile kind of zone right now. Because I don't think I've had a critical intervention in six months on 13, and I'm sure I've got 15,000 miles on it in that window of time. So the thing is, when they're that rare, like if it goes from one in 15 to one in 20, like how can you tell, right? You have to drive it forever. Yeah, exactly.

So, but so what I use as my metric is like how smooth is it? You know how a thing that's been a pretty reliable metric for me is how much hesitance do you see? So when when when the planner isn't, you know, when it's this when when it's making a choice like turn or not turn or, you know, turn the wheel left or right. It's the output from the network is probabilistic. Right. And so it's.

When, if it's at like 50% and jittering back and forth, you can see that in the car and it kind of manifests as a kind of hesitation. So as, you know, as the system has matured, there are fewer and fewer situations where it hesitates. When it does hesitate, it hesitates for less time.

So, the biggest thing I noticed on the version that I've experienced in the Robotaxis so far versus 13 is that the regular 13 is a lot more kind of hesitant and, you know, it seems to lack confidence in parking lots.

Like once you get off the road and you're getting into these kind of less constrained driving environments. Whereas like that was the first thing I noticed on the robotaxi thing. But the flip side is I can't tell...

If they're like running on a grid where, you know, they've essentially they've mapped out where they wanted to go in the parking lots in most of the parking. Like, you know, you go to a grocery store parking lot, you look in the map and there's, you know, there's like lane lines and roads and stuff like that. So if they're doing that and they're using GPS, then it can it can be behaving kind of as if there's lanes in that space as opposed to.

You know, you can certainly find parking lots where there doesn't seem to be any mapping at all. And so the planning and perception systems, they're just on their own when you get in there. But yeah, it feels smoother in parking lots significantly. And that would make sense, right? For a robo-taxi, you know, pickups and drop-offs are a pretty important part of the whole experience.

And that would be something where you could criticize the version of 13 that's getting pushed out to the retail customers right now. So do you think the version they're talking about with three or four more times parameters, is that version 14, you think? I mean, they can name stuff however they want. I don't think that... If the question is, do you think that's a major rewrite, kind of? Like...

I mean, maybe not. I mean, if they keep the architecture the same and all they're doing is increasing the number of parameters in some of the layers in order to increase accuracy, that's not really an architectural change. You're just going to get more accuracy by throwing more compute in the systems operating at the edge.

They could do that, but you wouldn't think of that as like a major upgrade. It's more like kind of a feature change that might lead to qualitatively better behavior or it might not. It's kind of interesting that they haven't increased the parameter count. I think Elon said something about increasing the parameter count.

about a year ago he was talking you know that was when we first started getting i think early versions of 13 were saying and coming soon it had the you know 3x more and uh we haven't been told that the parameter count's gone up now there's you know talk about this 4.5 x maybe parameter up increase you know i mean who knows exactly what's going on what it feels like is

they were able to continue getting better performance without having to increase the parameter count significantly. So they just didn't. They just kept refining the architecture that they had. So now when they think about going up, they're thinking about something different than they were thinking about before. Or, alternately, maybe they did do the thing and it just never was reflected in the notes.

It doesn't feel like there's you know in at any point since 13 came out that there's been like this discontinuous improvement and smoothness Which I think I think if if they kept the architecture the same and all they did was train it harder on A model with more parameters so they basically get higher actually the main thing that you would notice is it would just be smoother like you know decisions there would be a lot less

a lot fewer decisions where it was uncertain about what it was doing. And I have, it's definitely gotten better, but there hasn't been this discontinuous step where, oh, suddenly the, you know, it felt like the network would be 3/4x bigger. So that would be my guess. And another thing that's interesting to me is that hardware 3 has significantly less resources in terms of, I mean, our understanding when 13 came out, initially they only pushed it on hardware 4.

And, you know, the thinking was that it took some additional work to get that down to so that you could make it work on hardware 3. Hardware 3 never got a version that was named V13, but at the same time, 12 got a lot better during the window of time that 13 was out. And, you know, I spend a bunch of time in hardware 3 cars, and my sense of them is that

you know aside from the smoothness thing and there's some features that you don't get in 12 you know uh that you get in 13. but in terms of like you know if you're using it for road trips if you're using it for driving around town that there's not actually a huge difference um in the in the true functionality i think

13 is significantly more confidence-inspiring. Yeah, that's what I see. It's like we have a hardware 3 Model Y, and it's good, especially the latest versions of 12. But on the 2026 Model Y with the hardware 4 with 13.2.9, it just has that extra smoothness, extra confidence, extra responsiveness. It's just, yeah, it just feels like a more refined version. Yeah.

We have a 3 that has hardware 4 in the latest 13 on it. And we... I have an S that has... We have a 3 with hardware 4 in it in the latest 13. And I have an S that has hardware 3 in the latest 13 on it. And I do notice...

you know, there's a noticeable difference, especially at the difference is small enough that like, if you'd only ridden in the cars, like one or two times, you might not notice. But, um, but if you drive them regularly, you, you definitely notice. On the other hand, I don't feel like there's any routes that I am comfortable in on one car versus the other, you know, or any there, like I said, there's some features like, you know, um, probably some in that kind of, you know,

corner case stuff might be different qualitatively. Do you think just in hardware 3 they're just running a quantized version of the same neural nets with maybe a couple less features like I don't know summon and other stuff? I mean they're both going to be quantized to death. I mean they're probably 8 maybe 4 bit. The

The neural network processor in the hardware 3, if they use the same architecture they had in the patent, it's optimized for doing 8-bit quantized networks, which is what they were doing on the GPU, the GP106, before they got hardware 3. So it was optimized for that. I wouldn't be surprised if they're quantizing at 4 bits now.

you know, the field as a whole has kind of found that quantizing at 4 and 6 bits you can get, I mean, depending on exactly what you're doing, you can get performances very comparable to 8. You definitely wouldn't go higher than 8, but the networks, they're probably not training at 8-bit fixed point. They'll be training at either BFP16 or FP8, maybe, depending on the hardware they use in the training cluster. So...

The dynamic range of the D types that they're using is definitely going to be more quantized in the cars. So in that sense, hardware 4 and hardware 3, they're both going to be quantized. Hardware 4's got the capability to run networks with more parameters in them. And I'm guessing that they probably do.

So the difference would be not so much quantization as distillation would probably be the way you would go. Although it's pretty likely that they train a reference network and then they distill it down to something for hardware 4. They distill it separately for hardware 3. And both of those targets are quantized. So the rumors are AI5, hardware 5? Mm-hmm.

Probably coming out sometime. Who knows exactly when? Well, the rumors are it's going into production and it starts with the CyberCab whenever they start production on that, right? Yeah. I mean, do you think they'll start CyberCab production this year? 25? Yeah. I mean, they'll...

they've got a whole new production line that they have to bring up for it. So I imagine that the ramp, you know, the super, the early parts of the ramp are probably really long. And I think Lars has said something along the lines of, you know, the production line itself is in pretty advanced stage of development. The

Also, if they're doing cyber cab on an unbox process, then that's going to be a radically new production process and you can imagine the The early production ramp is going to be it's going to take quite a bit of time So we could there could be a really long tail at the front end of production So at what point do you say it's in production like when they built a hundred? When they're building

you know, a thousand a week. It would be cool if, you know, they built at least 20 or 30 by the end of the year and actually deployed them on, you know, the streets. You know, at prototype phase, you could probably build

you know you could build most of the vehicle and you know for the few odds and ends that aren't out of production right now you like hand built those or you you build them on a prototype line or something and so you've got you know maybe a 90 production car if if they even think it's worth it i mean they've got limited prototypes that they can use for doing stuff like you know getting reference data for cam and we've seen model threes that are outfitted with

cyber cab camera mounts, right, for gathering data. So they've got other ways of doing that. So I don't know, you know, how much weight there would be on trying to-- - Do you think Tesla will retrofit hardware three cars with AI five eventually to skip four? - My guess, well, okay. I think I'm on the record at length on this topic, but if, I don't think the jury is out on being able to make a software package for hardware three.

that will work for FSD as it is. And, you know, my napkin math tells me that if, for instance, they wanted to upgrade the whole fleet of hardware 3 cars, that's roughly a $10 billion expenditure. Oh, wow. And something like a year of work to do it. Like, even once you get...

once you develop an upgrade kit that they can install in 30 minutes or something, it's still a lot of money and a lot of technician time to outfit that fleet of cars. So I say this not as like a suggestion that maybe it wouldn't make sense to upgrade the cars, rather that if you can do it with software, you're going to save that $10 billion expense. I think they probably have...

a decision point coming up once once they start seriously ramping the robo taxi i think there's there's going to come a point where they're very much limited in terms of the number of vehicles that they can field and uh

you know, the number of hardware for vehicles that they can either produce or that they can, we don't know yet how many, you know, retail customers are going to lend vehicles to the fleet. That, you know, I'm sure they have their best guess on that kind of stuff, but it's the kind of thing that you won't know until you start doing. So there's going to come a point where they're, where their capacity constrained in terms of number of vehicles that they have. And at that, if you get to that point,

It's a huge... There's a lot of incentive to start retrofitting hardware 3 cars if you don't feel like the software at that point is ready to do it. So... And of course, the upgrade package, it's not something that they're going to be able to do in three months. You're going to have to plan it out well ahead of time. If AI5 is...

built to run in bursts as opposed to flat out continuously. So right now, as far as we can tell, both hardware 3 and hardware 4 run continuously at a constant frame rate on their current networks. If AI5 has significantly more sprinting speed than say thermal envelope that it can run at flat out, you might have a situation where

you've got an architecture that naturally you can just dial the frame rate up and down. And so you could retrofit that into hardware three cars and run it at whatever frame rate the power and thermal envelope allow. The thermal envelope that you can get in a hardware three vehicle is really high because it's liquid cooled. The power, they're limited to something like 100 watts, so they might have to run auxiliary power for it. Thanks so much, guys. Thank you. Appreciate it. Have a great day. Thank you.

Why do you think the robot taxis pause for like 20, 30 seconds, sometimes a minute? Yeah, that's an interesting question. I don't know if they always do. I've seen pauses up to two minutes. Okay, yeah. So I think I saw one really long pause, and I was actually timing it when it dropped me off at Home Depot the other day. I'm dying. Yeah, yeah, yeah.

Yeah, so the hardware 3 AI5, you know, the hardware 3 upgrade thing, that's a pure economic decision as far as I can tell. I tried to look at it a couple different ways. They can definitely build an upgrade package to upgrade those cars. Like, that's, it's definitely not, there's no technical obstacle to doing that. The obstacles are, you know, it's going to cost money.

and then it's going to take time, you know, because all these cars have to come in. You can get the install down. You can make it like the hardware 2.5 to 3 upgrade. If they have to add auxiliary power, like if they feel like they need more than the 100 watt power envelope, and there's a decent chance that even if they did a hardware upgrade, they wouldn't need the power because...

You know, the original 14 nanometer chip, so now, I mean, we don't know, but, you know, the rumors are building at 3 or 5 nanometer. If they're building at 3, 5, 7 nanometer, they get a lot more compute inside the same power envelope. So they could up the compute maybe five or six times in the same power envelope, and maybe that's enough. These are all...

lot of these decisions, you know, there's this trade-off between the amount of compute you use and the accuracy and that trade-off between compute and accuracy as the technology gets better, you just you get better and better accuracy at a fixed amount of compute. And you don't know exactly what that number is going to be. So

There's an economic incentive to put off the decision until you understand what the trade-off is going to be as well as possible. But you want to be able to make the decision soon enough that you can do the thing and not, you know...

My math says that when Tesla starts scaling, they're going to net something like $50,000 per vehicle that they put on it per year that they have. And so a million vehicles is $50 billion a year of profit, not revenue. I mean, the margins are so high that revenue and profit aren't so different. But it would be super painful to not have...

to be able to have demand that you could be satisfying for 2 million vehicles that you can't field, right? That's a hundred billion dollars in profit. So the economic forcing function is really powerful, but it's not super clear externally from what we know about what's going on.

Like, when is the point where, you know, they could really use those 2 million extra vehicles and they don't have them and they wish they had built the upgrade package so that they could have upgraded the hardware-free cars ahead of time? Like, right now, I would guess that that...

the point when they would need to start doing the upgrades. Like they will certainly once Robotaxi starts ramping they'll have good visibility into like at what point that needs to start happening but at that point the point when it needs to start happening might be a year away right but it's not far enough away that they can wait to start designing the upgrade package like I would guess they've already designed an upgrade package and maybe they've started fabricating or whatever and it also kind of seems like

Oh wow that's not it. Is that it? That's it yeah yeah. We're doing good. So you need to see this 18 minute ride it says. Okay we're off. We're gonna go right now to Merritt Coffee. I've been here many times maybe four or five times. This has a really challenging parking lot situation. Emmett, it was the highlight of Emmett's ride yesterday. Oh yeah? Yeah Merritt Coffee. Why was that? It's just a very

Most of the time it'll go into the parking lot, but it's very narrow at the end and you don't see the exit. And so, Robotaxi has to decide whether or not to peer into the corner to see if there's an exit or to back out. But in order to back out, it has to do like a multi-point turn. So yesterday when I was doing rides with Hans,

I navigated to Uchi, which is a really awesome Japanese restaurant here in Austin. So Uchi is like a house that's been converted into a restaurant. Oh, really interesting. Wait, it just backed up? Yeah. Okay, it's having a little bit hard time, but no, it's getting there. That was kind of an interesting indecision. Yeah, that was interesting.

I've seen it succeed that situation a bunch of times. Yeah, yeah. Yeah, and I see, you know, the 3 and the Cybertruck do that kind of stuff all the time, and it's generally not an issue. So that's actually, so this is what, ride 26 for me, and that was the most indecision I've seen on any. Oh, interesting, yeah. Yeah, actually, that was the most indecision I've probably seen. I've done, like, 30-plus rides before.

I don't know what was that park called that we just went to. I'm trying to keep my notes updated. Let me move this microphone a minute. It had seemed like the routing was avoiding particular... There are certain sections of the...

city that I was going through repeatedly that were close to the hotel as I would navigate to different things and it kept doing similar kind of roundabout routes so it seemed like it was trying to avoid a particular neighborhood and I was thinking well maybe this is there you know there's some

I had this hypothesis that the routing was constrained in that area. And then yesterday, I ended up taking a drive that routed right through there, and I realized there was a huge construction zone in the middle of it, right? So it was probably just that the traffic was terrible there. So I saw your post on X saying that you were kind of

surprise that this isn't I mean it's not a beta or a pilot or a test it's actually like a real service like what did you mean by that? it's a vibes kind of comment right? going the other way um

you know, you, in you, in the back of your head, you have this sense of like how ready a particular, if you do demos, you know, if you've worked in the tech industry and develop products and whatnot, you go through these phases of confidence in what you have built, like how much of it you want to, how much, you know, you're willing to expose potential weak points in your product. Uh,

And, you know, when you have a really complicated product that's going through a long development cycle and there's lots of uncertain stages, you know, if you've watched products go through that development cycle, you kind of get the sense that, you know, when you're looking at a new product that the, you know, the organization that is producing it has a certain degree of confidence based on,

you know, on what they're doing. So my general vibe sense of the thing from driving 13 a lot on, you know, a bunch of other vehicles was that because there were, there's plenty of like corner cases and if you're going to turn a service loose on the public, you have to have super high reliability and everything. And, you know, Elon said they were going to be super paranoid and whatnot. I was expecting it to kind of have the vibe of,

of something of a product test as opposed to a service start. But it feels more like a service start than a product test to me. And it definitely doesn't feel... Demos, when you see demos, they will generally have a lot of constraints, right? Because the entity that's demoing the technology, they only want to show it to you in situations where they know exactly what it's going to do, where they've mapped it out pretty well, right?

But the Robotaxi service, it doesn't seem to have as many constraints as I would have expected if this was like demo phase. And I wasn't expecting that, just to be clear. But I was kind of sort of in the mind that this would be more like testing, data gathering, validation kind of phase of thing. But it feels more mature than that. And I'm not, that's not...

a comment on the driver because my sense is the driver's been good enough to do this for a while at the level that you could tell as a retail customer. But rather, kind of the way that they have organized the Robotacti service, it feels like the start of a service as opposed to just a data gathering exercise.

What about it in particular? What's the standout thing that really gives you that feel? When you look at the app, they've put a lot of effort and thought into what's the app you would want to have for a service. There's a lot of things that you don't develop until you need to. And so there are lots of aspects of the service that you wouldn't expect them to be

polished to the point where you could tell they had worked on them a lot until they were getting really close to pushing the service. But I don't really see anything in the service setup right now. Like, I can't think of a single thing that they're missing right now to be a service. Maybe having monitors kind of breaks the economic model for robo-taxi, but not that that matters in the short run. Like, until you're getting to, you know, thousands of vehicles at a minimum, maybe tens of thousands of vehicles,

you're still in the ramping phase, so maybe that isn't super relevant. But they started with a constrained area, but it's a big enough area that it's got the full gamut of things that you would, I mean, it's half of Waymo's area. And Waymo's pretty clearly planning to run a service, like they're ramping a service here. They're not checking to see if it will work in Austin, right?

So that's how it feels. But like I said, that's a vibe check kind of thing more than based on any real data. What's your kind of idea of from now until the end of the year? Like do you think safety monitors go away or do you think they keep them like until the end of the year? I think the things that are going to determine like the right point to take the monitors out are things that –

Outside of Tesla, we don't have any visibility into right just to be you know Brutally honest about it like I wouldn't be surprised if they took the monitors out You know in three weeks and I wouldn't be surprised if monitors are still in the car in three months Right and like I can imagine rationales for having the monitors in the car longer than that depending on how fast they ramp the thing is, you know, the guys sitting in the operation center are seeing

you know, certain things happen with a certain frequency that they have a certain comfort level to. And they have to, like those things have to get down to a threshold where they're willing to like take the next stage. And we don't know what they're seeing. We don't know what the frequency is and we don't know what their comfort level is, right? Because I mean, the comfort level is actually a big variable. Like that's something having watched the Waymo effort for a really long time and having, you know, tried to watch the Tesla effort for as well as I can.

you know, I definitely get the sense that Waymo has been, they've moved really slowly and they've been very timid about stuff and that Tesla has been less timid. And so the thing is, if you're less timid, you're willing to move faster. You're less fearful of some kind of outlier event creating problems for your rollout plan.

You know, so Tesla is starting with no safety driver, right? Like, Waymo had safety drivers for, like, a long time, for years. They had safety drivers, right? And they started out in Chandler, Arizona, which is kind of, you know, it's kind of ridiculous. Chandler, Arizona is just a ridiculously easy place to drive around in. And even so, it was years before they took the safety drivers out of the car and that kind of stuff. So, like, Tesla's clearly being quite a bit more...

I think aggressive is the wrong word. They're being less fearful is probably the right way to put it. Because I don't see Tesla as aggressive. I see Waymo as fearful. I mean, my sense is they're moving prudently.

As I said earlier, like I've got tens of thousands of miles on recent versions of the software and it has felt good enough to do this to me for a while. Like I rode Waymo a lot, you know, 18 months ago and Waymo was, when I was riding him, I was seeing all kinds of, you know,

things that I did not consider to be safety violations, but they were the kind of inconvenient things. You know, you see people like it turns the wrong way onto a road or it gets stuck in a parking lot or something like that. Like nobody's going to die over that.

But, so to me, the frequency of those, the occurrence of those kinds of events is not an argument to not get going and start ramping. Because the sooner you start, the sooner you get to scale. And getting to scale is really important. Like this technology is going to have a profoundly beneficial impact on the world when it gets to scale. So...

uh so you know my preference is to you know be less fearful and uh you know trust that the public will be understanding and that they understand the benefits yeah but i mean you on the flip side you have tesla is such a public like pariah i think they're not actually really no i i think that's actually that the

you know the public perception that we get exposed to there's a very small number of people who talk a lot who have very strong opinions and so your impression of what's going on gets skewed by you know the same i mean who takes gordon johnson seriously just to toss one example out yeah right like he's just a clown he's not even he's not a serious analyst right similarly am

MSM articles about Tesla, like if you actually know anything about it, they're just ridiculous. I suspect the public is rapidly getting to a point where they treat the New York Times like they treat the National Enquirer, right? You just, everything you see, you take with a grain of salt and you don't, not that they don't have influence, they do have influence, but I don't think it's as much as

I don't think they can define the narrative in a sustainable way. I think they can define the narrative in a transient way, but that people, when they get direct experience of something, they completely discount what media says and they go with their direct experience. I mean, it seems like with Waymo, it just feels like Waymo is so...

I don't know, boring, but people just aren't as interested. Like a Tesla accident on Robotaxi. It's going to be like, everyone's going to find out. Not just the first accident, but it just feels like every single accident. Waymo got beat up a lot in San Francisco. Waymo, in San Francisco, Waymo got a lot of pushback from local papers. I don't know, the San Francisco Chronicle has run so many kind of... Is that a Zoox there?

Yeah, Waymo right behind you. So, Waymo was not welcomed with open arms into San Francisco, if you read the newspapers. The flip side is Waymo has taken share, I mean you've probably seen the analyst report on Waymo has taken in over the span of 12 months, they took a third of the ride share market. Essentially, they're roughly the size of Lyft right now.

And they have, you know, the ride share market didn't get bigger while they were doing this. They like took share away from Uber and Lyft. If you, you know, if you just consider rides that are within Waymo's operating area. So on the one hand, you've got media, you know, being super negative about Waymo and its impact on the city and all that kind of stuff. And on the other hand, you look at the actual data and the adoption is like,

they're totally limited by the number of vehicles they have and they're trying to get more vehicles in the city as fast as they can. So, you know, I mean newspapers talk, but at the end of the day, what people actually spend their money on, I think is what they, what they actually feel. - That's true. I mean, in the app store, Waymo gets like a Waymo one app gets like five stars. Like everyone loves it. - Yeah, it's great. It's, I never take any other ride share service if I can get a Waymo in San Francisco.

It's just a qualitatively better experience and I would be willing to pay significantly more. Are the wait times pretty short?

It's very time dependent. Like, you know, ride share is a super peaky service in the middle of the day, like around noon. The demand for ride share is 3x, maybe 4x lower than it is at rush hour. You also get this dynamic that actually where I live in San Francisco, the bigger problem with the wait times, it's not the vehicle availability. It's how the traffic gets so bad. There can be a car like four blocks away and you're still waiting 20 minutes for it, right?

So wait times during rush hour and like Friday evenings and Saturday evenings, they can get pretty long. But I think that's true for Waymo and it's true for Lyft also. I think Waymo, for a long time, they limited the number of people they invited to the service so that the wait times wouldn't get too long. They stopped doing that several months ago now.

And now I feel like they're just fleet size limited. The wait times would be lower if they had more vehicles. But I think there's a significant demographic of people who, what they do is they look to see if they can get a Waymo, and if they can't, then they get a Lyft, right? Because they want to go where they're going. So Waymos, I think they're servicing everything they can, probably. And I don't know how much that is. The vehicles are expensive and complicated. Mm-hmm.

So they may have a lot of downtime on those vehicles. I also, one time I got in a Waymo and they had left, usually they set the displays in the vehicle so it's not giving you any operational data on the vehicle. But one time I got in and the battery range remaining indicator was there so I could see how many miles it had driven.

since it started at shift and what the battery remaining was and i went and did the math because that allowed me to calculate because i could look up the jaguar eye pace like range to figure out what the burn rate was and it let me figure out how much power the compute and sensors was using which is i imagine that's why they don't leave the data on but you know

you know from that calculation the Waymos they use roughly as much power for their compute and sensors as they use for driving the vehicle so that I mean I didn't sit down to figure out what the what the consumption rate was based on average I could I could figure that out but

I think they're comfortably burning an order of magnitude more power. They're burning enough power that it meaningfully impacts the range that they can get. They may have to recharge those vehicles significantly more frequently than this vehicle would need to recharge. Have you noticed in your 25 or so rides, what's the quality of comfort differences between, let's say, the Robotaxi and Waymo?

Like ride quality. The ride quality. And this is better now, I think. But, you know, which is...

The Waymo ride quality is pretty good. It's kind of funny. A year ago, my biggest complaint... None of it is bad. It's always been a good ride. And there are degrees of good. But the biggest complaint was probably hesitancy. The ride could be a little bit jerky in the early days. And now, I think...

you know, they had this learning cycle where they found that the cars were getting stuck a lot because they were too timid. So they dialed up the aggression. And so now, like, weirdly enough, my complaint about the Waymos would be that they're kind of like needlessly aggressive. Like, to the extent, none of this is a serious complaint. Like, I think they're great. I feel like the Robotaxi software right now, it has a much better balance between being aggressive when it needs to and not being aggressive just gratuitously.

Thank you. Thanks. Here's the merit coffee. We just got off. So check this out this Park, do you see the exit? That's an exit there, but you can't really tell it's an exit from here. Like right? Yeah, I bet it's on the map Yeah, we can look inside and actually it's gonna try sometimes actually it just backs up but actually well if it's too tight Yeah, right. Well, it did it. Yeah. Yeah I did I've had it

back up do a multi-point turn two times because it didn't want to go through that but then three times or four times I mean depending on like if there's a pickup truck parked there then that's way less accessible right yeah yeah so Merritt Coffee on the map you know they've mapped a through route for that so yeah yeah okay okay so uh should know there's something yeah I was saying yesterday with Hans we

It was interesting that we got dropped off in a place that was completely covered with trees. And the way maps generally get made is they take satellite images and then they post-process them, right? To basically pull the streets off and the connectivity right now, which is why sometimes you get a thing where there's... We actually...

Yeah, when I was with Matt last night, we had a thing where we were in a parking lot that was separated by another parking lot and it had a road that connected through, but there were bollards that had been installed. So that's the kind of thing where a satellite image tells you these two parking lots connect. So the robo taxi came in.

expecting to be able to cross through there, but then the ball ends there. So like it went around, it's this huge parking lot with this big roundabout thing. And it went around and then it did, made the same mistake again. But the second time, I don't know,

I don't know if it decided we were close enough that we could walk to it because the pin was on the opposite side of this barrier, which, you know, we could walk right through the bollards. It was no problem. But the second time, it stopped on the other side of the bollards, and we just, like, walked through them to it, even though it was supposed to come. And then you had to go, like, literally, like, a block back to get to where the two connected. It was just, like, a really pathological parking lot. I can show you the...

What do you think? Thoroughbred is a breakfast place. It has sweet chops and that kind of stuff. It's really close. Let's do it. Thoroughbred. So it says we get in five minutes and it looks like it's a four minute ride. Okay. We'll ask for it.

Let's see. Okay. I love how fast it just gives you a text. Is Waymo, the Waymo One app similar, like in terms of how quick it gives you a, right? Well, I don't, I've got the Waymo One app here. I mean, it doesn't work in Uber, in Austin. I know, but we can, I don't know if the app will work and show us stuff. It might just say, I got to open the Uber app. Yeah, so the Uber app, it'll be the Uber app. Okay, so in San Francisco, the...

It's a really similar experience is the short answer. You have to pick your destination, they assign a vehicle to you. Although an interesting thing is I noticed with the Tesla system you get your vehicle like that. The Waymo system often takes a really long time before you get a vehicle, which seems weird. It seems like their system should know where all the cars are and be able to allocate one. But sometimes it takes two or three minutes for the vehicle to get routed to you.

And we do, I mean I've used Waymo a lot more so it's probably not as fair a comparison but we'll get the thing where the car, where you know, the Waymo gets allocated and then you wait a while for it to come and then they allocate you a different vehicle and then the timer starts and I have no idea what's going on here but it's happened to us a few times.

We're still in the zone with Waymos that we might want to use them. You need to budget some extra time. One of the hopes of Robotaxi when it gets closer to saturation is that I think the wait time should get down to like a minute. And they should be super reliable. So you're never thinking, oh, I need to budget time. I need to see what the wait time is right now. Because the wait times will always be short. You'll always be able to get a vehicle off.

Except in weird situations. Like it'll be like rush hour in LA right where most of the time you don't worry about it But you know at five o'clock you need to check the traffic before you figure out when you're gonna leave. I remember this guy from yesterday. There's nobody driving. It's a robotaxi. Yeah, it is. No, I don't. I'm just a fanboy. Have you heard about the robotaxi? You have not? Yeah, it's a...

Yeah, there's no one in the driving seat. Hey, how's it going? Oh, let's see. You need this, right? Perfect. Okay. A guy who was at the pumps, like, walked over and said, I don't know how you can trust those things. I was surprised that he recognized, you know, that it was a robot taxi. It gets through these tight spaces very, very impressively. Oh, man. The Cybertruck is kind of nutty.

I'm still not used to the limited visibility and the size of the thing. So it frequently squeezes through these holes that have gotten me grating my teeth. And then I'll look in the rearview mirrors, and of course, there's plenty of space. But it just reminds me that the vehicles are so much more aware of their dimensions than we are. Are you doing OK? Yeah.

Was it too cold? No, because it turns down the fan noise. Oh, I see. So I could record. It has a fan. It has two settings. Okay, so is that what you adjusted? Yeah, low, and then I keep the temperature a little bit high to keep the fan a little bit lower. We're going seven minutes, pretty close. I kind of like these short rides because...

It kind of feels like more the robotaxis more useful like I just do like if I do like three or four or five minute rides in a row and do it kind of more in these populous areas like Le Mar and Congress whatever it just feels like you're just on the go in and out in and out and it feels like In the future people are gonna use your robotaxis they kind of like this just gonna eat lunch go somewhere go coffee go somewhere you know in and out in and out quick

Bunch of Waymos here. Yeah, I feel like the meantime between Waymo here is in the San Francisco zone right now. 60 seconds maybe. I haven't seen many passenger carrying Waymos. Oh, really? Yeah, there's none in here. The ones that went by us, they didn't have passengers in them.

There was a really long period when the I-Pace Waymos were operating in San Francisco. And you, the passenger, like you expect Deadhead, just all the things being equal, Deadhead should be about 50%. So you would expect like every other taxi to have passengers and every other one to be empty. But like in San Francisco for a really long time, it was like 90-10 or something. They were...

They were pretty clearly, the fleet was running them to vet their performance. Because I can't think of another reason why they don't need that vehicle for mapping. They've mapped the hell out of it. What do you think RoboTaxi does between rides? Does it just roam aimlessly or does it rest a little bit? At the current density, what you would do long term and what they are doing right now are probably two different things. You...

In order to respond to your riders, like if you know what the geographic distribution of potential riders is,

I think, you know, the RoboTaxi app right now, it's asking for background GPS or whatnot, I think, because they might be doing that. If you know what the geographic distribution of riders is, then what you want is the distribution of the taxis to match that. So you want the taxis, when they're not carrying a rider, to be repositioning themselves so that you get the right distribution to match the expected pickups. Right.

So they would, you know, tend to reposition themselves, but, you know, they might reposition themselves and then park. Or they might not. You know, it's... I did this really interesting model of a robotaxi pickup in Chicago, which was... So there's an interesting thing that you get where there's a big... There are big chunks of the day where the movement is kind of random, and there's no net flow in or out of the city or anything. But then during the morning and evening rush, there's a net flow. Yeah.

So because you've got passenger carrying stuff going in one direction, in order to maintain the distribution, you have to have non-passenger stuff going the opposite to maintain the thing. Big bulldozer up here. Construction worker's not looking. Another construction worker. Yeah, that was really excellent. So this is a thing where I feel Waymos are super clunky. They're much more...

I live, my San Francisco place is right at an intersection which is a popular pickup spot for Waymos and people also stop at the same corner to load and unload trucks and stuff for houses. So you'll see this thing where the Waymo will come up to the corner and it'll want to turn right and there will be some guys standing near the corner because they're working on something and

If they're too close to the vehicle, it won't move. It just becomes frozen in position. And then you'll see the guys, they'll walk up to the car and be doing hand gestures and stuff to try to get it to move. But the thing is, the fact that they're close to the car is what's locking it in place. So it'll be this weird standoff, and sometimes it'll last for like five minutes before they get fed up and they just walk away. And then that way, I'm like, it's over. But...

For the most part, the Tesla's robo taxis, they don't seem to be exhibiting that sort of absurd sensitivity to pedestrian proximity.

What do you think Waymos are doing in SF for charging? Do they have-- Yeah, they have a couple of garages around the city where they overnight the vehicles. And they have charging infrastructure there. Do they have a person there plugging in and all that? Yeah. How about cleaning? Do they do that there too? Yeah, I don't know. I mean, I've walked by a couple of those garages. So you can walk by one of these garages, and you look in, and it's just completely full of Waymos. And they have a couple of chargers.

But the times I've walked by, there haven't been any staff doing anything. But I imagine that they're manually cleaned. You know, they only have 600 vehicles in San Francisco right now. So the fleet is not so big that there'd be a lot of merit in automating that kind of stuff. All right, guys, we're coming up to our destination here on the right side. Okay.

Where did we just go? Coming to Austin for 20 years and I actually don't know because I'm always just hanging out with my family. And so like this has been an opportunity to explore Austin. Yeah, it has for me too. Especially the south, you know, Austin area. I'm amazed at how many like really cool and funky and interesting places there are. Yeah, a lot of character here. All right, so we just called our next ride. It's coming in about three minutes. All right, James, so...

How is Waymo gonna compete with Tesla going forward? Let's say in a year or two, let's say in Austin, in this area, let's say Tesla has a thousand or whatever or more robo taxis running around. Waymo has a lot but...

Can't Tesla just undercut Waymo's price because they're cheaper to run, cheaper hardware, cheaper everything? If Tesla has more density, wouldn't it just give shorter wait times, better quality service, more demand? What can Waymo do, you think? Okay.

I think that your question misstates the magnitude of the problem that Waymo is facing. The problem isn't like how do a thousand Waymos compete with a thousand Tesla Robotaxis when the Tesla Robotaxis can undercut them. I think the problem that Waymo faces is how do a thousand Waymos compete with ten thousand Robotaxis

Actually, so Austin probably needs to satisfy rideshare in like the urban parts of Austin. You probably need like 2,500 taxis to saturate rideshare or something if it was a wholly owned fleet, right? So imagine that Tesla fields 2,500 vehicles enough to satisfy that or 5,000. Like they could easily do 5,000 taxis from a supply standpoint. There's no challenge for them. And then Waymo's got 1,000.

Right, they're not just getting undercut. That's not the biggest problem The problem is that the service quality is going to be radically different because you're going to be looking at you know eight or eight minute weights for Waymo's and 30-second weights for Tesla's right so Tesla is Like they have it's just it's a network effect on steroids So like it's a from a competitive standpoint Waymo has a really big problem if they don't figure out how to scale really seriously and

I just don't know how they do that. I mean, I don't... I'm super averse to knocking Waymo. Like, I'm thankful for Waymo. Those guys have worked hard. They developed really great technology. I think they, in a sense, they were betting that...

that what Tesla was doing was just not technically possible, right? And that that was what they were saying. And so Waymo, they're kind of in this situation right now where if that bet that Tesla's approach doesn't work, doesn't pan out, then plan B for them is they have to find a way to radically reduce their platform cost and

and radically ramp up their ability to get vehicles, which both of those are just fundamentally very big challenges. Like coming up with vehicle supply is a huge problem. You know, it's just, it's a really big problem. The other thing that they could do is they could try to position themselves as like some kind of premium service, right? They could try to just not compete head to head. Now, that said, there is going to be enough demand that

It's going to be hard to saturate the market for a while by just throwing too many vehicles at it. Because...

If you start driving the price down to a dollar a mile, 70 cents a mile, 50 cents a mile, then the TAM is going to expand from, you know, the TAM right now in the United States is a million vehicles worth of miles. The TAM will go to 25 million as you crank this down. You know, Tesla cannot produce enough vehicles to saturate that before 2030 or whatever the deal is. So that creates this operating window for other providers to also, for there to be business for them.

But if the question is like compete in the sense of like become a dominant player or take a big chunk of the market, I think that's, you know, it's challenging. It's like tough. Like I want them to figure out a way to stay in the race.

But when I think about it as a commodity service, which is how I fundamentally see it, you know, I see it going to a commodity services, high quality, short wait times, very inexpensive, like less than the cost of owning your own vehicle, kind of inexpensive and 30 to 60 second wait times being the typical experience. I don't like I just I have a really hard time seeing how Waymo can get there and less than like, you know,

Nationwide in less than I don't know 20 years, you know, like that's the scale whereas I can see here's our ride I can see Tesla getting there in a few years. Yeah Just fulfill the requirement All right, we talked about Waymo a bit I mean, yeah, it's like

That whole issue of like wait time, to me that's like so important. Like you need the short wait time. It changes the use case. Well, there are two things. One, getting the wait, getting the, I think the higher...

Sort of economic value thing is to get certainty like if you know exactly how I mean you don't want to take too long But you could argue functionally there's there's not a huge functional difference between like a 60 second and a five minute wait time You could argue that right? But the thing that's the big problem is a five minute wait time and a 10% chance that it doesn't show up because then

You know, if you've got a if you have a restaurant reservation, if you're going to the airport, you know, there are things where now you're having to build a lot of buffer time in and the economic utility starts to go down significantly. So I think the first thing is getting certainty. And then the second thing is driving the wait times down. And both of those, both the certainty and the wait times are satisfied by getting the fleet relative to the ridership up.

Now, you can regulate demand by just adjusting the price. So that's less of a problem at any given fleet size. But obviously, lowering the price, increasing the fleet size gives you more revenue. And long term, that's where you want to be is drive the revenue up so that you can afford to invest in the business. I was thinking one advantage Tesla has is

outside the city, like in the suburban areas, it's going to be hard for Waymo to get coverage. But Tesla can just access owner's cars, you know, in those areas and wake them up for people who are willing, you know, and get rides quickly in the suburban areas as well. Yeah, it's kind of a hidden, I think, advantage that only Tesla has. The flip side, I think...

that is actually not important. And my argument for that goes as at the end of the day, those are both just like total fleet size constraints because

Like, the margin on operating the service is so large that even in rural areas, it makes sense to, like, just park vehicles out there to satisfy demand. You know, once you have enough of it, like, once there are 25 million vehicles satisfying U.S. demand or something like that, the, you know, you're paying...

you know it's going to cost twenty thousand dollars to to like feel the vehicle like in terms of one cost one one time expense uh and in a in an urban area you have this challenge of uh you can't use your whole fleet most of the time because the fleet has to be very big to satisfy peak demand but in rural areas you're going to have a flatter kind of use curve relative and uh

And you still are able to get more than enough margin to justify putting vehicles there. The best you can do in a city is have a vehicle which is busy

Less than 50% of the time like you can't hit 50% utilization because of how peaky stuff is like you're gonna satisfy peak demand You can't get 50% utilization. So and if at 50% utilization, you've got 80% gross margins so in in a rural area if you you know if you if you feel that enough vehicles that you're getting good saturation in cities and

Putting a vehicle in a rural area and, you know, only getting 50% gross margins or only getting 30% gross margins, that's still an incredible business. I mean, the taxi business today is 10% gross margins, right? I'm talking more about, let's say, like, you're 15, 20 minutes outside of... Let's say you're outside of Austin. Let's say you're, like...

you're like 25 minutes from downtown, but it takes like... The wrap on that truck. It takes like, let's say, 15 minutes for an Uber to get to you usually because it's just you're out about. It takes you like 15 minutes to just get on the streets to where you are, right? And so it's like if Tesla can access your neighbor's cars in that neighborhood...

wake up a car, get to you in like two minutes versus 15 minutes or 20 minutes. But the accessing an idle car versus...

having a car of your own that's on standby, the difference of those is like the investment costs. In the short run, the difference is fleet size, right? Until enough cars have been built out to satisfy like whatever demand, and that certainly is going to take a while. So in the short run, accessing customer cars allows you to increase your fleet size, and that's just flat out useful all the time. As much...

as much usage as you can get, you can deploy. So there's merit in that. But the idea that it's especially useful in rural areas is, I think, basically...

it's conflating this, oh, we need to make good use of these assets in order to get them to pay for themselves kind of thing. And the reality of the business model in the foreseeable future is that the margins are so crazy high. You want to access every single resource you can as much as you can because you can't satisfy demand. That's true. I mean, I understand that you're saying in rural areas just you won't make as much, but still it's worth it to stick some urban taxes out there. Yeah, that's a better way of saying it. Yeah.

Like the margins lower, but the margins are so insane that even a lower margin is still a really good business.

Yeah. So Uber, what do you think is going to happen to Uber? I mean, okay, so first question is why is Waymo partnering with Uber? Why is this such an interesting topic? It's an interesting question. I understand people like the horse race thing of like, oh, what company is going to fail first or like who's going to be on top? But I think the much more interesting question is like what value is this going to create for the world? Like when do we get it? What form does it appear in?

like you know uber put a lot of effort into building a fine business and i don't want to dance on their grade but like if you ask me like how are they going to respond i'm at a loss to come up with you know a strategy that's going to work i mean there may be a premium business where you know where you know robos it doesn't make sense to field robo taxis because you know you're

you're going to substitute for like super fancy limo drivers that are like, I don't know, it's hard to come up with. Or maybe there's a market of people who is willing to pay extra for handcrafted drives by humans or something. Kind of like people will pay more for handmade furniture. But in terms of, you know, just bulk, satisfying bulk demand for transport at high service quality and low cost, like I said,

You know, if you have a competitor that produces a better product at 10% of your cost, you're just kind of doomed, right? It's very difficult to respond to that. Antique, right? Yeah.

So, okay. But why do you think... Maybe they'll switch over to horse-driven carriages, right? Yeah. So why do you think Waymo, for example, they have their own app in San Francisco, but in Austin and now Atlanta, like, you have to use the Uber app. Like, what's the rationale behind that? I think...

I mean, my first thought when I saw that, you know, you'd have to ask somebody at Waymo to really know, but my first thought was that it's a way to compensate for the fact that they can't field enough vehicles to satisfy demand. So, you know, people pulling open the Uber app, you're basically contributing to this large pool of vehicles that Uber already draws on. And they can, you know, they can prioritize you to keep you busy and that kind of stuff. But

you know if they if they want to compete against uber as a separate like if if you don't have enough taxis to be able to provide decent wait times on a separate service like people are going to get out of the habit of opening your app they're going to open the uber app first and then you're just maybe you're just at a significant disadvantage by partnering with uber and going into that pool they get away from the uh

the challenge of not having a big enough fleet to be able to provide timely service. So, I mean, you're saying it's basically easier entry, you have enough density or enough supply, right? Extra supply with Uber to cover. One of the things that occurs to me though is that in LA and San Francisco, I think one of the reasons you don't see them expand their area, one of the constraints on expanding their area faster is they have to have enough vehicles to service the bigger area.

at whatever like demand price point that they're operating at. And, you know, and vehicle numbers are a constraint for Waymo and they're going to continue to be a constraint. I mean, we'll see how the new, the sixth generation vehicle

works out for them if they're able to get that in numbers. So, I mean, the question though is if Waymo just expands, let's say in Austin, they just get more cars, more cars, what's keeping them in Uber? Why not just jump to their own app? Well, it may well be that Uber, that one of the...

terms that uber asked was you know some constraint that doesn't let them jump jump off from a practical standpoint if they got enough density they could come off but imagine like they're not they're current they're currently just serving you know kind of the downtown area like there's nothing to stop you from from servicing you know the suburban areas i mean they're not operating on freeways yet and i still don't i don't really understand that i i tend to interpret waymo's reluctance to get on freeways as uh

another sort of sign that they're fearful. When you get on freeways, the big freeways are easier than driving on surface streets, but they're high energy. Right? So the potential for having a really ugly fatality. Yeah. Things, if you do make a mistake, it's higher. And that's, I mean, the odds of having an accident are lower, but the consequences of accidents can be higher. So they seem they're like, they've been pretty dimmited about moving that way. That said, yeah,

even without using freeways, you could expand out to the east and the west. You could go a lot farther south. And they would very quickly have the problem that they didn't have enough vehicles to provide decent wait times in the larger area. Got it. Do you think Tesla's experience with autopilot FSD on freeways is going to help them a lot to get to freeways quicker than Waymo?

I think Tesla's lack of fearfulness, the lack of less timidity is probably the major thing there. Yeah, I mean, if you're going to talk about experience,

you know, a thing that's worth pointing out is that it's not that Tesla has more experience on freeways than Waymo does. Tesla just has like, you know, a thousand times more experience total than Waymo does. Right. Also, Tesla has a data source that Waymo doesn't have, which is Tesla has,

millions of human-driven vehicles. Remember the reference data for doing end-to-end path planning isn't robotaxis driving around, it's human-driven vehicles driving around, right? So Tesla has this huge resource of human-driven miles and Waymo's got, you know, Waymo only basically gets that if they have a human drive one of their Waymos around, right?

So once again, they're fleet constrained on that thing. And it's just egregiously expensive to actually pay people to drive around for you. Tesla gets all those miles for free. So it's just an incomparably larger database, not just on highways, but like everywhere. So there are some people who would argue that

Tesla is not very far ahead that there will be four or five major players who reach autonomy since it's a similar time period and therefore you know there's nothing special about let's say Tesla, FSD, RoboTaxi like how do you what I mean what's your what's your view on that?

I want to ask somebody, what technology do you think these people are using? Let's say they reference Waymo and Zoox and Pony.ai and something else. I'm not as familiar with what Pony.ai does. The approach of Waymo and Zoox are pretty similar. They use very heavily sensor-laden vehicles.

in an attempt to lower the burden of how sophisticated the software has to be. Which I actually feel like

I feel like that actually doesn't work. It looks good on paper. We have a lot more sensors, and we do sensor fusion, and we have all these modalities where we feel like we can rely on stuff, but then you get the problem of complexity, and the system gets complicated, and now you have bugs. You have to do insane amounts of validation to make sure that you don't have... It's not just the edge cases in the world. It's the edge cases in your code.

you now start having problems with because the system itself gets you going. But that said, they made a bet that having a platform with a lot of sensors on it was going to be able to get them to market sooner because it was going to lower the software development burden. I don't think that worked out for them, but that was what they bet. So, I mean, can you explain what you mean by it won't work? I mean, do you think it just won't work as well as Tesla's?

Solution enough to compete with them or do you think it's there's something fundamentally flawed about its performance of driving for the for the High sensor platform. Yeah. Yeah, I mean they obviously you can make them work It's the operational overhead and the and the capital investment are you know

you know, let's say uncompetitive compared to the approach Tesla is using. So like, that's a problem. You know, they can get the vehicles on the road, driving, providing service. It just, you know, the way that those two businesses evolve over time is just starkly different. So, you know, if you, if you, if somebody wants to argue that, you know,

other players are going to come out with a competitive technology. What I hear is, you know, say Mobileye or...

or BMW or Mercedes internal development programs, or Ford's going to get Blue Cruise polished up and it's going to get working. If those people are going to do that, they have to use a completely different technology than Tesla is using because Tesla's neural net based technology, it leans very heavily on their access to miles and none of these other players have access to the data, so they can't use that approach.

And to, you know, I follow this area very closely from a technical standpoint, and I am unaware of the existence of a technology other than the one Tesla's using that can do it. So if they're not investing in, you know, giant clusters and huge training arrays and a giant fleet of vehicles gathering data, like, I just don't see how they can do it. To me, it's magical thinking that there's a technology out there that we don't know about that is capable of doing this.

But would you say, though, that it would be... Obviously, it would be easier, but you probably wouldn't have to collect as much, maybe? As the tech gets better, you can get by with less data. But remember, the trade-off isn't... There's not a threshold of possible versus impossible. It's quality versus the amount of data that you have. So if you wanted to say, well, yeah, they can use 10...

They can use well, I mean you can't say 10x less data because these guys they've all got like a thousand X less data But you could say well they use a thousand X less data and they have a system which is you know 10 times worse in terms of its performance Okay. Yeah, sure. It's possible to field that system. Will people accept it? Are you competitive? like it

Are timid corporate organizations willing to take that risk? And I think at least the evidence so far is no. Thank you. We're taking a break, just walking right by the Gosseton River here next to downtown. Downtown is on the other side. You know, walking this path,

It's kind of cool because with robotaxi you can just get dropped off at one side, walk a mile, get picked up on the other side after dinner or something. I've been, I'll have the taxi come down here drop me at one thing and then I'll walk until I don't feel like walking anymore and then I just call a taxi. Yeah, that's the thing when I hike with my kids it always has to be a loop or else you're stuck. But just robotaxi on one side, get picked up on the other side. The whole getting picked up, like I totally misjudged

the weather the first time I came down here a couple of days ago. Like I had it, I had a, it dropped me off down at the end of the boardwalk, which like that was a point that I knew I could get to. And I started walking after I'd been going for like 30, 40 minutes. I'm like, I'm like, oh my God, this is killing me. It's just because it was like, I was just dripping in sweat and no time at all. I'm like, I got to leave. And I just called the taxi and like got it. So that was pretty cool. I didn't have to walk back to the starting point, you know?

Essentially. Yeah, what a nice walking path here. Yeah, it's great. It's awesome. It's like shaded too. Yeah, it is. That's the cool thing about RoboTaxi launching here. South Austin is like...

all the people filming or getting to know the area and then publicizing different parts of Austin. There's so much fun stuff that I didn't like. You know, I come here a lot over the years, but I just haven't spent that much time in South Austin. And there's...

There's so many funky cool places to eat and it's got a ton of parks for doing outdoors like I found we went to this public pool the other night and I can't think of public pools as these like little places where you know little kids like go in the summertime whatever the deal is but apparently it was super popular because there were all of these like 20 and 30 year old people who were there like well into the evening. Did you go to Barton Springs?

Yeah, that big one. Okay. Yeah, that's amazing. Yeah, I mean we didn't go to the pool I just drove there but I saw people coming and going like it it wasn't You know, it wasn't this thing that they closed out at night. It was like popular. It's a happen in place. Yeah Have you done a kayaking out here? No, that's awesome activity. Yeah kids and I we do a paddle boarding sometimes. Uh-huh out here you rent a paddle board and kayak and

so yeah it's a fun thing yeah this is really nice like i'm surprised as hot and humid it is right now how pleasant it is to even like walk here yeah so what do you say to some skeptics who are like well tesla's only level two still because the safety monitors like the supervisor so still supervised fsd they're still way behind ramo because they don't even have level four out

Like, what's your response? I mean, the levels, they're not, people think of them as a technology metric. This is something George Haas pointed out, which I thought was really clever and insightful, is

is that the level thing, it doesn't tell you what the system can and can't do. It tells you who's taking responsibility. Level 4, the OEM, the vehicle maker, the system operator, they take responsibility. Level 2, you as an individual take responsibility. And level 3 is a non-existent... Level 3 is a unicorn thing, which is kind of a mix of the two, where it has a mode...

where the operator control is responsible and it has a mode, but because that boundary isn't well defined in the standard, nobody ever did it. So like level three sort of doesn't even exist. Tesla is...

level two because they're requiring owners to take responsibility for it. So, you know, pretty much by definition, the Tesla robotaxis are level four because Tesla's, you know, we're the passengers. We're not taking responsibility. Right. The monitor in the vehicle, he's not taking responsibility. It's Tesla's responsibility. So is it a level four to technology report? Well, it's a dumb question. It's level two when it's in the vehicles, when it's in private vehicles and it's level four, not because the technology is different, but because

You know Tesla's taking responsibility, you know, that's what makes it level four So it's not people talk about it like a technology distinction and it's not a technology. It's an operating distinction Yeah, it makes sense. I mean you could have a level four so-called system that performs way worse than a level two, you know I mean, yeah or vice versa you get a lot of a level two. That's perfectly which is I'd argue. That's what FSD is right now FSD is a

quote unquote level two system that performs at level four level. What do you think the safety monitor, actually no, I want to talk about that a little bit later. How much do you think Tesla can remotely operate the vehicle, if at all? Like in what functionality do you think they have remotely? Okay, so as I understand Waymo's operator, and this is from direct experience, like my interpretation from direct experience, in addition to like what I've heard from third parties, is

is that they don't have like a steering wheel and pedals kind of arrangement for taking control of the vehicle. When the vehicle gets in a jam, an operator gets kind of a bird's eye view of like the vehicle and its surroundings, and they can basically tell the vehicle, okay, go here and then go this way or whatever, you know, so they touch a couple points, they instruct the vehicle to do that. So they're not

remote operating it in a way like you might do with a video game. They're remote operating in the sense that they're providing support to the path planner, but the vehicle drives itself at all times, right? And in fact,

My understanding is that's why Waymo doesn't have to list those as interventions because no human is taking control away from the path planner. The path planner is doing it. All they're doing is supplying supplemental instructions to the path planner to do it.

I don't see why Tesla wouldn't do the same thing. It's a good idea. The Teslas, they're perfectly capable of driving themselves. They don't need any help. They can maneuver parking lots. The biggest problem they run into when they do run into problems is deciding what the appropriate path to take is. It's not like, is that hole big enough for me to drive through? The cars know that perfectly well. It's just, should I go forward? How do I resolve this problem?

this path planning. Have you noticed like the Tesla monitors they have their right hand on the door handle and it seems like there's something under their index finger. I don't think it's the top actually there's maybe like a trigger button. Here's an observation. Yeah. If you're in a robo taxi and you want it to stop right away open the door. It'll stop right away. Yeah. So it may be like

I wouldn't be surprised that if I open the door from the back, the car would stop right away. So maybe... Actually, Emmett did that yesterday. It calls support, but it doesn't stop. Oh, really? Yeah, it calls support right away. Oh, wow.

He's a brave man. Yeah, he said he did it by accident. So my theory is I wonder if it's actually not an emergency stop because we like today we saw one guy monitor. He was like his finger was on the stop and lane button. Like that's the emergency stop, I think.

I think that they might have an extra button on the index finger. I don't think they modify it. Like, why would you need to modify the card? You can reprogram the card to just use the thumb button for something else. Like you double click it for a-- I think it's maybe like a kind of a snapshot slash telling maybe the AI team that this is a challenging situation or is a little bit like hesitation or something that they're just

basically marking that you know like that time that whatever for later it could also be that like if you're sitting in your car for a really long time your arm gets tired yeah i don't know it's just every single monitor has it it's not just that they have their hand there it's the way they had it i had a monitor who didn't have his hand on them i did people pointed that out yeah and i was kind of looking at the things like yeah they did like when i sit in the back seat i don't put my arm when i sit in the front seat i put my arm in so it does seem a bit unnatural but

I did see one monitor who didn't seem to be keeping his hand up. Like, not that you can pay close attention from the back seat. - Yeah. - But yeah, the overwhelming majority of them do it and it does seem kind of unnatural. - Yeah, yeah. - But like I'm not sitting in the seat, so I don't know. Maybe if I rode around in the passenger seat of a Juniper for a while, I'd have a sense of whether that's the natural place to put. There isn't an armrest on that side. - Yeah, that's true. - Right, so like if you put your hand on the armrest, which is on the door, that's where your hand is gonna end up. - Yeah.

to not let it get super low. Have you seen the back trunk area where they have the extra module? Yeah. Okay. The thing that's stuck in the machine. Well, I hadn't noticed it. I don't know the Juniper that well myself or the Refresh Model Y. Yeah. So I wouldn't have noticed it, but then I saw somebody mention it and then, you know, I saw... Actually, open the trunk. We'll take a look at it. Okay. Some people were saying they think it's like a mini Starlink, but I don't think so. I mean...

I kind of wonder if it would just be like extra cell receptions or GPS or I don't know. All right, guys, we're headed out. Oh, this is a long ride. Our next ride is 18 minutes. Wow, 18 minutes. Okay. So I've got a couple more questions here for James here. The traffic is picking up too. That might be a factor. Make sure that I recorded this here. Okay. All right, so what do you think...

Tell me what you think about the option of expanding here in Austin first versus another city or a few other cities. Would you expect Tesla just to focus on Austin, you think, for the immediate future in the next few months? Or do you think there's value in expanding to other cities in the early stages as well?

I would think that they would expand, that it's not an either or thing, that they'll, you know, they'll expand to other cities if other cities are receptive. And, you know, I mean, they talked about starting in San Francisco, I think, or he said California, right? Yeah, they said. So starting, you know, in like on the peninsula in the Bay Area, that would be,

that's a place where they have a lot of resources and it would make, you know, that would be a convenient place to do it. They could do it in LA, right? They have Hawthorne and they've got a big presence there so they could do that. I, it, I mean, starting in,

Launching simultaneously in two places, that probably seems like a bit of a stretch. Because you do the, you know, 10 vehicles to 20 to 40 to 80 kind of thing. Do that in one place so that you can get some density. But, you know, when you get to the 200, 400, 800 kind of scale, there's probably not a lot of distinction to starting in a second city. They do...

I mean, cities are pretty different and San Francisco has all kinds of... It's not that Austin doesn't have challenges. Austin has all kinds of interesting challenges in the road system that are kind of unique to Austin. San Francisco certainly is like a certain kind of extreme as far as the United States is concerned. So that's another kind of interesting stress test.

I mean, it's nice that they didn't start in Chandler, Arizona or some, you know, sleepy bedroom community that is totally not a challenge at all. But, you know, I think their ambition is to rapidly scale both within cities and across cities. So the stuff that they, you wouldn't want to do a thing where you like,

start up in a city where you don't have the infrastructure to be able to provide a decent service early on. So until they get to the, you know, thousands of vehicles kind of stage, they're probably not going to want to do anything small. But otherwise, I don't see a good reason for them not to go any place that they can. Is there like a certain number of, let's say, safety drivers that they have before it just gets like...

too much headache to manage all these people and you just the monitors? you know like for example 100 monitors or 1000 monitors like is it at a point where you're just like okay either we stop expanding because it's a headache to you know enroll more you don't want to build a big employee base that you can't maintain yeah exactly now it you know

What I expect is for them to get the monitors out of the cars quickly and before they get to the 100 vehicle stage.

And, I mean, they might have reasons that I don't. I haven't anticipated for, like, keeping monitors in the vehicles beyond that stage. But their ambition is to get the monitors out of the car. So I would expect them to be, you know, moving toward that as quickly as they can. Once they can get monitors out of the cars, they can scale much more aggressively. So I would imagine that they probably have...

limit on how large they want to go with monitors in the cars. So, you know, what we saw other companies do, what we saw Waymo and Cruise do, I mean, now Defunct Cruise, was, you know, they were using safety drivers, not monitors, but they gradually, you know, introduced vehicles without people in them and then gradually expanded that. So there wasn't like this black and white thing that they hit. So, you know, you could see

The thing is, I don't know if monitors are just monitoring or if they also have a data collection function aside from just ensuring safety. To the extent that they really are ensuring safety, then you can't really expand without them. My expectation is that the current presence of monitors is a belt and suspenders thing, right? Yeah.

Until they get enough experience to know that at least in this area that they're good So maybe once they're comfortable in South Austin, you know They expand the surface of the service to cover South and North Austin and maybe they only have monitors in North Austin right for a while until they're comfortable there and then you know they add east and west and you know and you have monitors in newer areas and maybe not so much in others and

That could be hard to arrange, of course. Because you can't, like, you know, if you have north and south Austin, you'll have rides going back and forth between them. So that's a complication. But then, you know, you go to a new city and maybe, you know, you bring monitors in when you're in the, you know, the warm-up 10, 20, 50, 80 vehicle range. And then you get them out as quickly as you can so that you can scale. Yeah.

But I think get them out quickly so you can scale is, that's what I expect. That's what makes sense to me as a strategy. So we'll see. So with Robotaxi, I'm seeing a lot of similarities with...

the advantages of Starlink in a way. For example, Starlink's big advantage was if they can get more satellites out, they get more coverage, better service, it increases the value of the service, you get more demand, they can put out more satellites, get better service. And basically, I mean, I wouldn't say you have a near monopoly until

for the next competitor to come out, they have a high hurdle now to cross or to jump over because they need to provide an internet service that is somewhat competitive with Starlink or else no one's going to sign up for that service. So in a way, what Starlink has done is it's created an incredibly difficult hurdle for a new competitor to come up with or to come out.

And it just gets harder and harder because Startlink is just growing, has so many satellites, their service is just getting better and better. Eventually there will be some player but it's just so hard. When you compare that kind of thought to Robotaxi, it's if Tesla can put out basically these super cheap cars

and saturate an area, give super low cost, high availability, like super low wait times. And it just takes over an area with autonomous driving.

it provides this high hurdle now like for the next competitor to come in and to be a legitimate let's say competitor against tesla you have to provide rides that are somewhat competitive to have even a chance so you need the either the wait times and the cost right i mean

Or you have to differentiate your service. Yeah. So you're not... I mean... Yeah. Now, there is this thing which is worth considering, which is that at any given price point, there's a certain amount of demand that you can generate. So imagine, you know, two competitors. Say, I don't know, Tesla and Moima, just to put some names on it, right? If Tesla's got...

you know, if they've got 10,000 vehicles in an area, there's a certain TAM at, you know, there's a certain price point that keeps 10,000 vehicles busy, right? So a competitor, like, if they could, if they can offer service under any kind of justification, even if it's unprofitable, at less than that price point, then they can get rides, right? Sure, sure. It's not a fixed TAM. The TAM expands as you drive the price down, and there's a lot, at least in, you know,

With, you know, CyberCap, for example, the operating cost is very, very far below what the current ride share rate is. We don't know exactly how it grows, but, like, if you were to imagine that, like, if you get the mileage rate down below the cost of, like, a privately owned vehicle, that you could take maybe half of private vehicle miles or something like that, right? Even if people are using their own vehicles for use cases, like...

you're going someplace and you're taking a bunch of kids or you're going someplace and you want to leave your stuff in the car, you know, or you're a plumber. I mean, there are going to be people who have use cases where,

Moving over to like CyberCab or Rideshare isn't trivial. But even so, even if you have your own vehicle, there are going to be nights when you want to go out on the town or you need an extra vehicle because some family members go in some place where the main vehicle isn't going. Some of your miles start to move over to those services. It's not an either-or proposition, right? Almost everybody is going to have some use cases where it'll make sense to use that stuff.

And so I think of, you know, a more reasonable way or maybe a better framework for thinking about it is what fraction of total miles of light duty vehicle miles can you get? And it seems to me that the TAM is certainly half of light duty vehicle miles, which, you know, if you use 50K vehicles, if you use 50K miles per year as, you know, expected revenue miles per

then that means for the US you need like 25 million vehicles. So to the extent that there are people who can operate with some justification, until you get to 25 million vehicles, it's not a zero-sum game. Anybody who's willing to come into the market and operate and meet or exceed the quality and price of the existing services can take some market share.

and they can operate at least that to me the bigger question is like people don't start businesses if they don't see a justification for it down the road which you know usually the justification is i want to make money and so i need to be able to operate you know i need to be able to charge revenue which is above my cost at scale when i get to some kind of scale but you could imagine other things right there are

there are countries that, you know, favor their own unprofitable industries because they consider them strategic. And so you could imagine municipalities that do that, right? You could imagine car makers that decide that,

you know, they're going to be in the robotaxi business even if it's a loss leader for them because it's an important component of a bigger strategy of like not having your car business go away. I'm just making stuff up now, but I'm imagining that the space of motivations for doing it might be bigger just than the obvious commercial one, right? Yeah, I just think there's going to be tremendous value to get to market like to...

to get to the predominant place in the autonomous ride hailing market to establish kind of saturation and satisfy the demand. I just think you could, the first mover advantage in terms of, I mean, I think if you're profitable, you give like a ton of, you have a ton of cars,

um people know you you're gaining demand there's a network effect for sure yeah the network effect it just seems like we got a u-turn here yeah okay this is the first time i've had a u-turn on a ride so far yeah but the network effect um it's like a flywheel i i i think that is underrated by by a lot of people i think

that's going to be one of the big reasons why it says it's a it's a great business long term for tesla it's just if they can get it established and and up and running the way that i think they can um yeah it's like and again it's

With Starlink, right now, there isn't a second place, in my opinion, with RoboTaxi who has the same potential in terms of cost structure and all that, right? And supply to do what Tesla can do over the next few years. Like, I don't know who number two is. Who was it that Yvonne had that line about we can't see number two with a telescope? I think you couldn't see number two with a JWST right now. Yeah.

Don't forget the bag. There you go. Cafe? Yeah, let's go for it. I am really reluctant to just bag on other players in this space, but it's simultaneously true that like I don't... You know, the simple commercial, what I was calling a simple commercial motivation, and you know, to the extent that there is competition, like

Competition is weird until you start to satisfy demand, right? We're so far from satisfying demand that there's a sense in which competition can exist, right? There can be other people also offering a service that looks similar. But at the end of the day, when you look at the simple commercial thing like, can you offer a compelling service at high margin, you know,

Like we don't know who number two is. If there is a non Waymo approach, a non high def, you know, lighter plus high def camera player out there who's capable of fielding it, like that could happen. I mean, that will eventually happen for sure because the tech is getting better. But I don't find credible claims that there are other players that are, you know,

that are at Tesla's quality of service. Right now, if you're willing to settle for significantly lower quality of service, I think there are other players that can feel that

And there may be places where you can offer a lower quality service. It's also possible that once the market gets proven out, that you may find that people lower their guard, they don't care as much about quality of service, especially if another entrant is willing to come in and undercut you. So hypothetically, say Mobileye has a service, say they field one, and say it's just not as good as Tesla's service. There might be a market in a lower quality service. Given that

that Mobileye is willing to absorb the risk of accidents, right? That's the big scary thing that nobody really understands right now. We don't know what the public's tolerance for it is. We don't know what people will pay for safety. For the most part, airlines don't say, "Oh, this is a plane you're flying on people." I mean, until recently, people weren't like, "Oh, that's a Boeing plane? I'm not flying that." Right? So...

In mature markets, we see that retail customers, they don't seem to put a premium on safety. So will they not do that in cars? Are people willing to ride in a lower quality service and accept a higher risk of an accident? Accidents are going to be super low, right? I think generally you're not going to see...

like Mobileye hypothetically, come out with a service until they can get significantly below human accident rates. But say they could. Say Tesla is at 100x less or 10x less than human and Mobileye is at like 1.5x less. If Mobileye was willing to charge less for the service,

you know if they could i mean this is kind of the funny thing like the cyber cab is going to be like one of the cheapest vehicles made yeah so the idea that you're going to take you know a buick escalade or something you know some some vehicle like that made by some other uh maker and you're going to slap mobilize sensor suite and compute package in and

be able to undercut, like be able to profitly undercut, that's another kind of like weird space that we're going to be in. Like that there literally aren't, maybe a Chinese vehicle could be made, you know, at a lower price point than cyber cab. It's not outside the realm of possibility. I just, I don't think the idea that, you know, a week after robo taxi rolls into a street that there's going to be a competitor, that's just not credible.

A week, a year, two years, five years, even five years. I think if there were players out there building robo-taxis that were planning to get to scale, like scale being two million vehicles a year kind of scale, I think we would have heard of it. In the US you'd hear about it for sure because people would be raising money to do it. I guess there might be some justification for trying to do it in secret, but...

I think it'd be hard to keep secret. I don't follow the China market enough and they can scale faster, so maybe that would happen, but it would probably end up being a China centric service. I've sort of been assuming that Tesla is going to be overwhelmingly competitive when I say assuming, like when I look at the landscape,

of players and what I understand of the requirements of the markets. The Tesla is a really strong potential in the US and Europe, especially the US, but maybe not in China, depending on the regulatory environment. I mean, China might build the rules around their service such that they favor local

countries do that, right? Wouldn't be surprising if China did that. And of course, you know Chinese OEMs, they're super competitive. So maybe that happens there. Also, Chinese consumers might turn out to be super more cost sensitive and more willing to do the safety versus quality trade-off, right? So I haven't assumed anything as far as success for Tesla in China because it's like I have a hard time.

it's not that I don't think they can succeed. I totally think they can succeed. I just don't understand the landscape well enough to have any confidence in a prediction. But that said, I wouldn't be surprised if worldwide, 50, 80% of all the margin dollars in the world come out of the United States too, because of the combination of what Americans are willing to pay and how big the market is. I think you just don't see that anywhere else.

How long do you think it's gonna be until Tesla can do autonomous driving or robotaxi, let's say, in India? I don't think India is that hard actually.

Yeah, it's kind of funny. There's human perception of what's hard and the robots' perception of what's hard are very different. I noticed in early versions of-- there were versions of 11 where whole Mars was demonstrating how it could navigate crowds of people in busy downtowns and whatnot. And to a human, that seems fairly impressive. I think driving five miles an hour through a crowd, that's the easiest thing ever because it's super low energy. Cars got plenty of frames to evaluate the thing.

The thing that's intimidating to humans in a situation like that is that you can't see everything. It's a field of view, but the cars, they see everything all the time. So that's not a challenge for them. Similarly, they have much higher dynamic range in their visualization. So scenes that seem dark and indistinct to a human, they're super clear and sharp to a machine. It's got much higher dynamic range on its cameras. It's much better in low light.

So we tend to think of night as difficult. I think night driving is probably easier for the machines than day driving. To us, illumination really matters. To the machine, flat contrast is actually a big win. It makes processing it easier.

India, similarly, I think it's like the driving through a crowd of people. Like it seems intimidating if you don't drive there because there's so much going on and it's hard to keep track of things. But, you know, a human being pays attention to two or three things at a time and then we shift, shift, shift, shift, shift. The car, it's watching all hundred things all the time and that's just normal everyday things. Like you're in India, there's a hundred, you know, there's eight cows and four rickshaws and 35 pedestrians and the machines, it's not getting overwhelmed.

So I think India isn't that hard. A bigger challenge might be a place like, where you have people drive very fast with very low margin. So like, I haven't driven in a long time, but driving in Italy was, you see there's lots of little cars and drivers seem perfectly willing to do stuff like, drive fast the wrong way down a one-way street, kind of unexpected. Or mopeds will like,

go past you at speed in lines of cars and things like that. And I think the combination of like short reaction times and high energy, high consequence stuff, those might be more challenging, right? Because I feel like the machines in general want more safety margin than...

There are situations where the machines have to compromise on safety margin because humans do and that's the only way you can blend in with traffic. But in general, I feel like machines are going to want larger safety margins. So the more you put them in machines where there is no safety margin and they're high consequence, like hitting somebody on a moped is a super high consequence event, that that's actually the spot where the machines end up struggling. Interesting.

Guys this might be one of the last rides of the day See if that works for you. I don't know. I don't think it is actually yeah Yeah, if you worried about wind noise just direct that so it's not blowing on that Yeah, it's pretty hot here in Austin. So we're trying to turn down the AC but it's still pretty hot Yeah, James was commenting. I'm starting to make videos again. I think um, I

It's just such a historic moment. Robotaxi, you know? Yeah. So that's what pulled you out of retirement? Yeah, it's like, how can you just sit home, you know? You've got to get in one of these things and try it out and kind of ruminate on the possibilities of the future. And it's in the town you live in. Yeah, exactly. I'm not too far from it as well. So I was thinking...

it's not too long since Tesla released their end-to-end V12. I was looking back and I think it was like February of last year they released it or something. Yeah, it's been a little over a year, four months or something. And I was like thinking the progress since that time

Till now, it's just been absolutely phenomenal. I was chatting with Hans, and he kept saying things about like, oh, yeah, in the last year. And I'm like, no, no, that happened in the last four months. The thing stuff you're talking about, it's like we're in the... Do you remember the expression internet years? Yeah. Or whatever it was like. It was the, you know...

People originally, when the internet boom started, people kept saying that time is passing. Is it in dog years now or whatnot? And it became internet years. Now I feel like we're in the AI years kind of zone, which is like, it's like internet years on steroids. The whole, everything changes every 12 months. Yeah, I remember when we were talking back in the day with V10, V11, it was like, there were just these like,

these the jerkiness you know the real like hesitation like all that stuff in different situations there are just so many like cases where it just felt like Tesla would take them it just for me it felt like it just

It would take a while, but then once V12 came out, I remember talking with you and we were like, this could be a game changer in terms of improving the rate of improvement. And it really has. The rate of improvement just drastically changed with V12. But it's simultaneously kind of a cautionary tale. For the longest time,

Pretty much everybody in the space that was developing technology thought the big deal was perception. And once you have perception down, that the planning is relatively straightforward. If we can do path planning in video games, how hard can the real world be? That's part of the reason why

You've got the Christmas tree thing with Waymo and cruise cars and whatnot. Those things don't help with planning at all. They're 100% perception. So everybody thought perception's a problem. If your perception's good enough,

problem solved, right? But then Waymo built out their thing and they discovered, "Oh no, I guess planning is kind of hard." Well, so end to end with Tesla, it wasn't a huge change in the perception performance, but it was a radical change in planning. So Tesla also basically went through this phase where they got the perception down, but then it turned out planning was a really big problem. And then end to end seemed to be the

the breakthrough that made planning work really well. And it ends up being a game changer because the thing that they were plateauing on prior was not that their perception wasn't good enough. It was that the planning was not there. So coming up with an alternate approach

that you know well for lack of a better term solved the planning problem all of a sudden it was you know you can start making a lot of progress and that's what we've seen yeah i noticed this watching you during these rides today you're not really like previously when i've ridden with you you're

analyzing like you know the car's moves yeah and looking at the mistakes and all that stuff but well you can't analyze the non-mistakes exactly i see a different completely different james dalma here you're just basically just hanging out and just resting and just like

There's not much to really fault. My truck drove me here from LA and I spent 26 hours watching it drive. And there's no interventions. I'm thoroughly aware that the mistakes are just like...

they're not really there anymore. Obviously, it's not zero. It's not that there's never something that you might have been a little more comfortable with, but the... Man, did I say this earlier? But, you know, I think I've got, like,

Man, I must have north of 15,000 miles on V13. No, probably north of 10,000 miles on V13. I don't think I've had intervention on V13. Right? It's just like in a totally... I mean, V12 was an amazing improvement on V11. Yeah. Just like shockingly huge step up. But, you know, V13 just continued that trend in a way that...

sort of eliminated, you know, a large... almost all the rest of the planning issues. And now there's just... It's kind of a domain thing, right? There are domains that Tesla is focused on, you know, like... You know, and like parking lots were not a domain that were getting a lot of attention. I feel like on the...

The current consumer version of 13, that's why we still see that it leaves something to be desired in parking lots. But then what we're seeing in the Robotaxi version is that, well, that's a solvable problem. Yeah, for me, driving around 13.2.9, it feels or felt like Tesla could have solved...

non-edge case driving, basically 99.9% of driving or so. It feels like they could have, but I wasn't sure because you're always being nagged to pay attention and all this stuff. But in the Robotaxi, it really feels that. They've solved the non-edge case driving, basically 99.9% of driving. It's just these rare freak situations that perhaps who knows what would happen. We just don't know if they've solved it or not. I don't know.

And again, it's the whole thing of how can we evaluate it unless... Because it's like you're not going to sit in the car for tens of thousands of miles. But I think it's going to be harder and harder to evaluate because the basic driving is just going to be perfect. You're just going to sit in the car and there's basically almost nothing to improve. But yeah, it's remarkable. Here's my...

Last month of safety score, right? Yeah. And, like, I have 3,155 miles on FSD and .02 miles on autopilot and 3,186. So I have less than 30 miles. I have 31 miles of manual driving in 3,100 miles. So, yeah, 99% there. Yeah, yeah, yeah.

Overall, I'm just surprised at how well Tesla Robotax is able to do in this area. Yeah, this isn't even the vehicle I put most of my FSD miles in. We have a Model 3 that gets way more. So, I'm wondering, at this point, what do you look for right now? I mean, you've done 30...

or so right, you're here for another few days. Is there anything else you want to see or experience or just kind of waiting to see what it does in some strange situations? I'm not, like I'm over the edge cases. They still exist. I don't think they're common enough to like, I don't think they're common enough to be something that you can like as an individual expect to, you know, get any insight into.

You just don't spend enough miles in the car to develop statistics on edge cases anymore because they're too far apart. The thing that I said before about how I noticed a while back that you can get a sense of

the confidence level, like how well the nets are performing by basically looking for hesitance. Yeah. Which isn't, it gives you an idea about like, you know, basically when you have a neural net classifier, you know, you have some probability for the thing that it thinks is, you know, number one, and then you have probability for number two and three and four and that kind of stuff. Right. And ideally, you know, for in, if all the situations are like really well categorized and everything, you know,

you know the number one probability is always much higher than the number two and the situations where one and two are like neck and neck are rare and that's what i think you see in the hesitation like how many situations is the second highest uh choice high enough that there's ambiguity because that in the in the ambiguity there's risk um

So that's a thing you can kind of feel, see to the pants, like the sense of confidence. Do you think, you know the jittery steering wheel that used to be like there in V11, maybe a little bit in V12 sometimes.

Is that how they got rid of it? Is just having more confidence, like less second place, close second place options? Or is there something else you think they did? That's certainly a factor. I think Ashok had a comment on that one time about how the class of, you know, if you're using a classifier for path planning, like you're either going to go straight or go right. What was the deal? He had a comment about like how they were going about trying to make that

more discrete and less continuous and I don't remember what now what the comment was on the thing but you know that the thing that I just said like it's a simple interpretation that you know when you have these decisions which are you know

digital go straight or turn right like at an intersection of the turn or right. I mean in reality there are more choices than that. Do I go wide right? Do I go narrow right? Do I go you know do I go through the on the left side of the lane right side of the lane and whatnot but there is kind of a binariness to the decision there and on those binary points you can get a sense of whether the system is

well characterized for this particular test case. Have you noticed this going over speed bumps is just like... It's really good. Really good? It feels like it's better than 13.2.9. Yeah. Although, I gotta say the latest version of 13, I really have not noticed problems. So I took my truck out after my first couple of road taxi rides and I had it drive around a little bit. And it was similarly doing really well on the speed bumps here in Austin, so...

Maybe the speed bumps are solved and I'm just remembering that they were a problem in an earlier version.

It also gets dips. Like, 12 started solving speed bumps pretty well, but it didn't, you know, when you have those, like, the water channel dips where you turn, and those can also be pretty harsh. They're not, I don't think that they're there functionally to try to slow you down, but you have to slow down for them. And 12 wasn't getting the dips at all. And 13 does them just fine.

I guess one benefit of the monitor is if there is an accident, say even with a parked car, the monitor knows what to do. You know, we'll get out, report it, do whatever. Versus if there's no monitor and, you know, the car hits something, then it's like up to the passenger. I mean, you know, support. I don't know what you do. I guess Tesla might have to send someone out or...

You know Waymo just traps the passengers in the car and waits for the police to arrive. That seems to be the fix. Alright, we're almost here. Oh interesting. Is this it? Alright guys. Thank you, appreciate it. We might be sitting outside. How about... Man, it's so really hard inside.

We're having some ramen here. Ramen! For those who don't know, James Dama, one of his passions is Japanese food. Japanese food is great. I just can't eat very much of it. We have a good time talking about language acquisition and other things like that. Awesome. Can't wait for my food as well. We're stuck between a truck and a...

Oh, and there's another one over there. Here we are waiting for our robotaxi. Waymo's having a hard time here. But actually, everyone's having a hard time here. Humans are having a hard time. Yeah, yeah. Hey, James, so I wanted to ask about Optimus. So I was thinking, like, one of the things we've talked about the past, I don't know how many years, four or five years, is autonomous driving. And the...

the whole thing of Tesla making the right moves, the right decisions, being on that path, right path. And like as we drive RoboTaxi, it's like, oh my gosh, it seems like it's here, you know, it's like it's rolling out. What do you think about Optimus? I mean, like it seems like Tesla's making all the right moves with

the right focuses, the right strategic decisions, the emphasis on the hands, the right AI software decisions. It just seems like it's following, I wouldn't say it's the same, but similar systematic pushing it the right way. How are you viewing Optimus development? Are you as excited as you were with autonomous driving with Tesla?