Cheeky Pint - The 20-year journey to fully autonomous cars with Dmitri Dolgov of Waymo

Episode Date: March 24, 2026

Waymo is now doing nearly 500,000 rides a week across 10 cities. Co-CEO Dmitri Dolgov came to the pub to discuss how they moved from scientific research to massive global scaling. He gives a masterclass on the sensor stack (and why you still need Lidar), how they use "Simulation" and "Critic" models to train the AI, and why he believes cars that require human supervision will never naturally evolve into robotaxis. They also cover the new custom-built vehicle that feels like a living room, the economics of ride-hailing in rural Alaska, and the "Russian math nerd" diaspora that seems to run the UK tech scene.Timestamps(00:00:22) Russia(00:02:51) Waymo architecture(00:09:59) Why now?(00:19:46) Driving nuance(00:29:37) Stripe Agentic Commerce Suite(00:30:17) Hardware(00:40:20) Emergent behavior(00:46:36) Scaling(00:57:56) GoogleArticle:EMMA: End-to-End Multimodal Model for Autonomous Driving – Waymo Research: https://waymo.com/research/emma/

Dmitri Dahlgaov is co-CEO of Wemo. He joined Google's self-driving car project in 2009 as one of its first engineers and was repeatedly promoted until he took it over in 2021. Waymo is Google's most successful moonshot, and now provides over 500,000 fully autonomous rides each week. Cheers, by the way. Yeah, cheers. You grew up in Russia, right?

I grew up in Russia. Then I was actually a Soviet Union. Right, exactly. My dad is a physicist, so the Soviet Union started falling apart, And then he had a visiting position in Kyoto University for a year. We moved there as a family. And then he went to Berkeley and I kind of tagged along. And then I ran out of, you know, I graduated from high school.

I was thinking about the next thing I wanted to do. And I really liked that technical school in Russia. The Russia is a serious better for this. They are. They are. So I went back to Russia and I got my bachelor's and master's there. What year was since that you went back to Russia? 1994.

Okay. So that was kind of almost peak Russian optimism in a sense where he was opening up. It was. Yeah, yeah. No, actually I remember talking to my mom about it. And, of course, my parents grew up in the Soviet Union. They've seen it. You know, they were worn right before the war.

And then they saw, you know, they lived through some really tough times. And I remember talking to my mom. And she, you know, in fact, I got my green card here in the U.S. before I went back. I went back and she insisted that I do it. And I was actually at the time, I wasn't thinking of coming back. But then I was pretty excited about where Rasha is and trajectories on. And, you know, being young and naive, I was like, there's no turning back. And so why did you decide to come back?

There's more of a playback made. Yeah, yeah, no. School, it was pretty clear to me. Like, I wanted to continue studying, you know, math and computer science. And while the undergrad and master's that I got in physics and applied math, that I think was still an incredibly strong kind of foundational school of Russian math and science. Graduate school, it was very clear to me that the best way to do it was in the U.S. So I came back.

I'm struck by the founders of the two most valuable U.K. companies are Russian math nerds, who both went to the same school, Nikolai at Revolut and Alex Gurkow at XTX. But yeah, it's a strong diaspora. There is a company not far from here where one of the founders also has, you know, a similar pedigree. Right, a company that we're close to. Exactly. You're the classic engineering interview question of, you know,

what happens when I type Google.com and hit enter? as, you know, talk me through, you know, whatever you like, you know, HG and D&S and BG, you can go down to whatever level of the stack you want. Do you want to maybe just describe, when I take a ride in a Waymo today, what's happening at a technical level? Like, what is the architecture? Let me answer your question. What's happening in real time?

But this is going to be only a part of the story because we're going to be talking about kind of the inference, the real-time inference part of it. And if we want to have a deeper reference. Richard, technical conversation, I think it would be interesting also to zoom out and talk about the entire ecosystem of what goes into building, evaluating, and deploying the way and we're driving around. But when you're driving around or being driven around, say, we think about what we're building as a driver. Obviously, it's not a car. So it has a number of sensors that are positioned around the vehicle. We use three different sensing modalities. There's cameras, there's lighters or lasers. and they're Raiders. You know, those are the primary ones. They're, you know, also microphones, directional, you know, microphone arrays. But those are the primary three for sensing the world.

They all have very nicely complementary physical properties. They all have 360 degree coverage around the vehicle. So the way more drivers sees kind of 360, you know, all the time. So all of the data goes into a computer, you would expect. and they're the software that process, now it's all AI, I'm kind of specialized AI in the physical world. So it processes the sensor data. Nowadays, we talk about it in the, you know,

using AI terminology as, you know, encoders that, you know, take this data in. And then there's the kind of the decoder, the action, you know, the generative part, if you will, in the car. And the generative task there is to, you know, figure out how to drive. Right. And that is, of course, connected. through a specialized interface to the car where we can actuate the vehicle.

And that's why you see the steering wheel turn and it drives you around. Okay, so I get into my car. There's three main families of sensors, LiDAR, radar, and cameras. And then it is using that to first build a model of what's going on in the world, you know, where are all the other cars and things like that. And then you say make decisions and then actuate that with the car. That is the system that you're living in. And is all that inference done locally or presumably yes, nothing's in the cloud?

Nothing real time. Nothing real time, the cloud. And there are some things that can happen in the cloud, but they're not required. Got it. What's an example of a nice to have that happens in the cloud? You can imagine a situation where we do, you know, some of it is not directly related to the test of driving, but I can say after you leave the car, we want to check that, you know, the car is not dirty. You didn't leave anything.

there, if you did leave an item, well, if you left in a mess, then, you know, we want to send the car to one of our depots, get it cleaned up. If you left an item there, you know, your phone, all right, we want to detect that, and then, you know, send it to our lost and phone and let you know. So that, you know, we do with kind of a, by asking a model that actually lives off board as opposed to having to put it on the car, right? Because it's not a real-time task related to, you know, the driving. So that's one example of something that... There are all these debates that go on on Twitter around self-driving, so I can think of, you know, end-to-end versus the more kind of modular approach.

There's cameras only versus array of sensors. And I can't tell, are these debates actually interesting to an expert in the field? Or do you think these are just settled matters and they're just grist for the algorithm? I understand one of the questions that come from. I do find that often the way they're posed and the way the debate happens is losing a lot of the nuance and a lot of detail that really matters where to me the most interesting technical

questions are in that level. Because the way we think about the building the Waymo driver, it starts with a large off-board foundation model. I can imagine building a big model that understands how the physical world works and understands the important properties of what it means to drive, the social aspects of driving, and what it means to be a good driver

as opposed to a bad one. So that's the foundation. Then we specialize it into, let me call it, three main off-board teachers. There are still large, high-capacity off-board models. There's the Waymo Driver, there is the simulator, and then there's the critic. And those then get distilled into smaller models

that you can run inference on faster. So the Waymoid driver becomes the backbone, the male backbone of what's in the car. The simulator, of course, is what powers our synthetic generative environment that can run on the cloud, for training and for evaluation in close over the system. And the critic, they value. Is this simulator ever run locally?

No. No, it doesn't. However, what I think is interesting, in a way, the way the decoder works, the way the model works, if you think about the generative task in the simulator, of kind of creating those realistic worlds and how other people behave, how cars, pedestrians, cyclists, in order, and the task that you have to solve,

on the car in real time, there is this fundamental shared capability of understanding how these objects relate to each other and predicting what they might do in the future if you are running on the car and then generating some sampling those probabilistic behaviors in the simulator. So it's different models, but there is, you know, this is why the shared foundation model is able to power both. And similarly if you think about the critic, like the job of the critic is to fine interesting events and then, you know, be opinionated about what's good behavior. Yes. And what's bad behavior.

Similar fundamental understanding, right? If you're running, you know, inference on the car, you still have to, like, figure out which of the multiple hypotheses of these future worlds you want to, you know, take action to steer it towards. Yes. Okay. And these are all downstream of the same foundation model. So start with the foundation model. Yep.

Then you specialize in fine-tune, still off-board model. Those are the teachers. And then you distill, each one of the teachers kind of distil, trains its own student. Yes. The driver, the simulator, the critic. Yes. You started working on self-driving 20 years ago. Yeah.

As you think about the tech evolution, is this just a scaling law story where we had to be able to throw enough compute at it? Were there architectural approaches we needed to wait to, have been invented. Was it just a story of we needed 20 years of going down the wrong cul-de-sacs before we eventually arrived at the right approach? You know, could you, knowing what you know now, could you have a successful Waymo in market in 2015, or was there some enabling technology? No. Technology breakthroughs that happened over the years were critically important, primarily in AI, but also in other areas. like compute, you know, have a compute than you don't run.

Now, I wouldn't characterize it as like going, you know, a thousand different dead ends and then having to retract and then finding like the one right bat, I would characterize it as, you know, iterative learning and evolution. And then, you know, transformers came around. But, you know, transformers, for example, are very general architecture, right? Powers of LEMs, powers, you know, our models.

But how you apply them to that space, I think this is where... It didn't just fall out of transformers. Exactly, right? And of course, people like to talk about architectures, but architecture is important, but really a lot of it comes down primarily to your metrics, to your evaluation mechanisms, to all of the training recipes,

and, of course, data. Yes. LMs are good as text or tokens specifically, and obviously perform best at domains that have some kind of single corpusive text they can work on, like coding, where it's very helpful that everything was just kind of textual already. And part of the success has been creating textual representations for domains, such that we can then put al-LMs against them.

Can you describe how you encodes the world that you're seeing? I mean, are you just building a 3D map, like a 3D bitmap, essentially, or? So this is where I think we can get a bit into the, question of what is the interface between the encoder and the decoder parts? And I think that touches also on the thing you flagged earlier where people like to debate end to end or not into end. So the way, let's make you know talk a little bit about end to end and then get back to like what is the interface between those two, right? So when we say end to end, what do we mean? I mean that it is some large ML model.

Typically, you don't build them monolithically. You have different parts and different subgras. But what's important is that you can propagate backprop the gradient and the loss function all through the different layers. So every layer you can learn the weights and the representations that matter for the final task. You don't force it through some narrow funnel between, let's say, the encoder and the decoder.

Yeah, I think of a simple view of. of end-end being, you know, pixels go in and car actions come out, which maybe a bit of an oversimification. Yeah, that's exactly right. And if, like, this is kind of the basic vanilla version of it, right? There, if you think about the, you know, what will it take to build the driver that's capable of fully autonomous operations? If you think about this entire ecosystem of the driver, the simulator, the critic,

If that's all you do, pixels in, trajectories out, it becomes very difficult to do all of those three and achieve the high level of safety and performance that we require, and it becomes very difficult to kind of do it at scale. However, it's kind of a very easy way to get started. You collect some data, kind of like an answer to the LLM world, right? The easiest thing you can do is have, you know, pick a model. The easiest way to get started nowadays would be just, you know, take a VLM. It already has a kind of a language-aligned camera encoder. And then it has a decoder that, you know, can predict, you know, generate text.

And you can fine tune it and say, hey, instead of text, generate trajectories. You know, very, very doable. In fact, a while ago we published a paper called Emma, that did. Exactly that. Yes. And it will actually, in the nominal case, drive pretty darn well, which is mind-blowingly impressive. That is very funny, yeah. And, I mean, there's some intuition.

You're saying you can take it off-the-shelf model, which has nothing to do with driving to start with. And you'll get these good results. That's right. You get it in the normal case. I just want to be clear. It's orders of magnitude away from what you need. Yeah, you should not try it on the streets, but it works. But for example, it's like a hawking horse.

It's impressive that it's talking. Exactly, exactly. And you can actually, the product that you wanted to build was maybe a driver assist system, not a fully autonomous system, then maybe that's all you need to do. And then for that, you don't need all this other machinery of the simulator and the critic because the number of nines is drastically. But this is interesting because there is, you know, there is some intuition behind, you know, why that works. If you think about the hard parts of driving, it's, you know, not on. unlike, you know, having a conversation. Except if in the LLM world, right,

having, you know, you're modeling language or maybe modeling a dialogue in the space of sentences and words, what makes driving hard is also this kind of multi-agent social interactive part of it. And if I do something that's going to affect you, it's going to affect somebody else. And the history matters. It's not local and just geometric.

context matters, semantics matters. But it's in a different, you know, it's not in the language of words, it's in language of body language, if you know, right? So, and we see that empirically validated if you, you know, do this approach. Okay, so then let's say we build this thing.

Just cameras, camera encoder, pixels go in, trajectory, go out. The quality is sufficient to, you know, drive in the normal case, it's not sufficient to deal with the long tail of, you know, the edge cases, and hit the high bar of superhuman safety that we require. So then you start asking a question,

what else do you need? Yes. And if all you did was kind of observing how other people drive when you trained this system, maybe observing just passively how people drive and how they interact, maybe also driving the car yourself and then using imitative learning to train it, Mind that that's not enough.

You have to do something in closed loop. You have to, you know, you have to do things like RLFT, which is also, you know, parallel to what we see on. RLFT. RELFT. Reinforcement learning based fine tuning. Okay, yeah, yes. So you know, similar to the reinforcement learning with human feedback in the LLM world, right?

You want to do maybe close loop, proper closed loop driving, where you explore all kinds of different situations and then you give it a reward signal to keep it in distribution. For that, then, you need a realistic simulator. Right? You also, if you want to have a good RRL system, you need to have an opinion for the reward function.

This is where the credit comes in, right? If you have a purely end-to-end system, let's look at the simulator. Now, what do you do? You're then constrained to just go from pixels to trajectory, right? That's all you can. you can run the system on, right? And it's a very high dimensional space.

So it's a hard problem to generate everything. But even if you solve that, it just becomes incredibly inefficient to run it in the full way of pixels to trajectories and simulation for training or for evaluation. So this is when intermediate representations come in. There are some intermediate representations in the world in this task of, you know, in the physical world.

we know are correct. They're not sufficient, but they're not generality limiting. There's an object here, there's a concept of a road, there's signs, there's speed limits. So this is where augmenting that learned representation, those learned embeddings from the encoder decoder with that more structured representation is what we do. And we find that this kind of gives us additional knobs to simulate

in that space, just pixels to trajectories, it allows us to have additional safety validation layers in real-time. And it also allows us, you know, it gives us additional mechanism to specify the reward function for evaluation of critic or, you know, for training. So this is again, like we've gone kind of full circle of it. Is it antenna? Yes, it is. Yes. But if you want to do it at scale for full autonomy, it's augmented with all of this other stuff. That's very interesting on the simulating point. It's just very hard to simulate for an entity. model because it's easier to deal in intermediate representations rather than coming up

to the pixel-perfect view of the world. You need both. Yeah. So, you know, having an antenna architecture that's augmented with that structure allows you to kind of play in both of those worlds. Yeah, yeah. What are you looking to do as a self-driving car? I mean, it sounds funny, but I think people maybe don't realize that there are many different

things that you're looking to solve for where you're looking to get the person to their destination, you're looking to get them there reasonably promptly, but also drive quite smoothly and also have many lines of safety, and also not annoy other drivers and get honked at, and, you know, and then, and. So what are some of the reward functions or kind of things you're optimizing for that maybe are not obvious to people? So safety is the primary focus, right? But of course, we also want to be a smooth driver, so that for both. both people in the car and other actors. And I also want to be a predictable,

well, behave, one so that it can nicely fit into the whole social ecosystem of our roadways. It seems like one of the issues that has quickly emerged with self-driving is the fact that people can't have nice things, or not everyone is nice to the robots. And so, you know, whether you're driving through dodgy area or getting blocked or, you know, maybe I'm not going to drop you off here. Maybe I'm going to go around the block and, you know, drop you somewhere better.

But all of these, as you say, kind of other human issues, how do you go about solving this? A lot of the ones that you mentioned are just things that, you know, we need to work on. And understanding, honestly, you know, said that if we're not dropping you off, where exactly where you wanted to be dropped off, or, you know, we don't give you kind of a good interface to tell us, That's on us. Right? I just, you know, how to make it better.

Yeah. It feels like the drop-off is actually a pretty nuanced part of the self-driving journey. Like the highway stuff and the, you know, the 35-mile-an-hour roads. Like, that is all nailed, but there's just like a lot of nuance in the drop-off experience. I'd say they're all hard. You picked freeways and you picked drop-loss for different reasons, right? For drop-ups, there's, you're absolutely right. There are, you know, a few things that are maybe not obvious.

you know, you just think about this problem. But it's understanding where you want to go, and making it as convenient as possible for you, and pickups from drop, it's not exactly symmetrical, right? But then I was also understanding the context of the situation where you, you know, where do you stop? You don't want to block a driveway, you don't want to, you know, double part, although in some cases where if it's a quick one, maybe it's okay.

So there's a lot of nuance that goes into doing that well so that it's a kind of smoothless, frictionless experience for the rider, as well as other folks. Freeways, for most of the time, not much happens. They're very well structured because we design them that way. But there is still that long tail

of really complicated stuff that happens where the consequences of better event are much more severe, right? Speed is much higher. Everything is quadratic in speed. But we see a lot of stuff.

Imagine grills falling off of freeways. Imagine people getting into accidents and kind of spinning out of control. You see one of those flat-ed trucks but just like a bunch of stuff piled in it and you're driving behind it? I don't know. I always find it very nerve-wracking. Looks bad.

I know. Yeah. And we've seen them leave a trail. Yes, yes. Yeah. Okay. So it's a different set of problems. But I feel like the general sentiment with Waymo is that the driving has mostly now been solved by you guys. And it's kind of a question of scaling up and maybe some super long-tailed stuff, really snowy condition.

Like, is that your sense internally or is there actually much more nuanced with than that? I would say the, yeah, it's not like we're done with engineering. I would say that we've clearly moved past the stage of scientific research and kind of, deep core technology development to this new phase of accelerated global scaling and deployment. We still have work to do, but I don't see today any limitations or any gaps in the core technology. The driving is good enough now. Well, the core technology, I think, is good enough that I can't think of any aspect of driving that is not supported by the fundamental technology.

Now, that said, there is a lot of work to do in specialization and in validation before we can deploy responsibly. We're not driving everywhere in the world. We are planning to start operating in London and in Tokyo this year. We just do have a driver that you're using today in San Francisco that we can just plop down in London and go? No. But what we're seeing is incredibly encouraging from the perspective of, like, is the core technology there?

So now it's a matter of collecting the data, doing some specialization and validation. And you can use science are different. You know, in both of those places, people drive on the other side of the road. But, you know, that's actually not that hard for computers. Right. The core technology generalizes really well, but you still work that you have to do. What generalizes least well? Increasingly we're finding, especially, you know, now that we're able to, you know,

to kind of hook the Waymo AI to the AI in the digital world, the VLMs, and kind of inherit the general world knowledge from VLMs. We're seeing really strong results from like zero shot or few shot learning because of that general knowledge that we bring in. But there are a few things like, let's say, cold weather, cold winter weather, where it affects the entire stack. So it's not just the AI, we actually have to. The hardware, you know.

You need the hardware, you need to have the proper cleaning solution, you know, heating elements in it, and then you think about things that are completely solvable, but computers like motion control and slippery surfaces, right? So that takes a bunch of work. You don't get that for free from just, you know, pulling it some, you know, VLM decoder. Was it the case, I mean, my impression, not knowing anything,

is that in the early days, there was maybe a lot of San Francisco-specific work or Phoenix-specific work in the early markets, whether it be mapping or something else, and that you guys seem to either have solved that in generalizing it or just scaled up your ability to do the city-specific work. What enabled the rapid city expansion?

We usually think about it, the capability of the windward driver as well as deployment, not primarily and directly in that space. of cities or zip codes, I think about the operating domain. And then that's just the freeway, includes freeways, cold weather, snow, rain, fog, density, etc., and then that's what we are building,

that's where we're valuating, and then that maps to a city, like particular city, be within the operating domain or outside of it, right? So where, you know, if we provide history a little bit, our initial deployment in, where we start, offering a fully autonomous commercial service for the first time was in 2020 in Chandler, Arizona. So and that was on what we called the fourth generation of the Waymo driver. This was the, if you remember, the Pacifica minivans with different hardware, different software.

There, you know, we were super focused on kind of doing the whole thing end to end. learn how to build the driver, evaluate it, deploy regularly, operated end-to-end 24-7 with customers, learn from the customers, and then we're very focused on that operating domain of mostly Chandler, which is a medium, low complexity one. Then when we made the jump to the fifth generation of our system, this is what's on the eye basis today,

we really wanted to take a huge bite out of that operating domain. And we collected data all over the United States. all different states, different cities. When we chose to deploy in the hardest parts of San Francisco, hardest parts of Phoenix, we made a big jump on the hardware side and most importantly on the software, the AI side. And I would say that was the big discontinuous jump. And that's what you're saying.

Now after we've scaled up and iterated in all of the aspects of building and deploying the driver, this is now why you're seeing us kind of go, in parallel and scaling in the US and... So driver version 5 was just a much more generalizable stack than version 4? And what was it about just that it had been trained on a much wider data set? It was when we made this big bad on AI. There was a lot more, you know, kind of little AI models and ML models in the fourth generation.

Got it made a much bigger bet and jump to kind of AI as the backbone. for the fifth generation. AI is the backbone as the core engine, as in you're saying that Gen 4 had lots of small little AI subsystems. Okay. Yeah. And that's been, so we made that jump, and we've been iterating and improving the model since then.

As we're seeing with Waymo rolling out widespread autonomy, it has second order changes on the entire system, in this case, traffic patterns or other drivers' behavior, or eventually how cities are laid out. And autonomous systems are coming in many domains. In commerce, soon agents are going to be transacting without human intervention. We're basically getting driverless commerce. And Stripe is building the economic infrastructure for AI.

And as part of that, we're letting payments be initiated by humans or by agents. So if you want to sell to agents or if you want to let your agent spend money all around the web, check out Stripe's Agent and Commerce Suite. Can we talk about hardware a second? So lots of hardware questions. But one is maybe everyone in this space has a very charismatic demo of a vehicle that is custom made for self-driving. And so, you know, it's often the van with the, you know, no steering wheel, seats facing in both directions. You know, you guys have one.

Tesla has the steering wheelless cybercab. You know, crews have the cruise origin. And yet, we're still driving in Jaguars that have steering wheel in the front and are pretty similar to consumer cars. And it's interesting to me, because, you know, if we were talking about this 10 years ago, we might say, well, yeah, developing a custom car, like, that's relatively straightforward. We know how to put a bunch of sensors on a new car. But the software will take a long time.

And what's interesting is we've made huge progress in the software, but interestingly, the cars are still derivatives of cars that people are driving. And so I'm curious why you just think the custom hardware has not happened as of 2026. It's obviously it's a small improvement compared to Waymo is the big improvement, but it's just interesting that it still hasn't happened. Well, let's say our sixth generation of the vehicle and the driver is our version of that. Oh, no, I know it is. Oh, high platform, right?

So that is, you know, it still has the, you know, we can talk about. But whether you want to have the seats pointed backwards or not, I actually think it looks nice in a demo, but practically speaking, it's not the way to go. But that is, it is a custom design vehicle, and it is, we put a lot of thought into moving away from a car that's designed around the driver, to a car that's designed around passenger.

And it's much more spacious, and it's happening. It's, you know, we're not, it's not open to the, the public yet. But, you know, I took a ride in it the other day fully autonomously. And that's coming this year. Yes. How much better is it as a passenger experience? You'll tell me, once you give it a try.

I love it. Okay. So it's, yeah, it's all about the space. Yeah. And the convenience of, you know, ingress and egress and the screens and the interface of the pasture. So we put a lot of thought into every aspect of it. Yes. So it has sliding doors.

It's very easy to get in. It has a flat floor. It is, yeah, if you sit in the back, you can like fully stretch out, and there's so much space there. And it looks, you know, from the outside, you know, it looks fairly big. Yes. Right. But the actual footprint of that is barely, barely, barely larger than the eyepace. So it's kind of amazing that, you know, you walk in and it feels like you're in the living room.

Yes. I guess my question is just, you know, WAMO does, you know, 25 million rides a year, run-ride-ish. with the Jaguar IPase, and it's interesting that so much scaling has happened with self-driving so far on the old retrofit. Maybe that's to be expected. Well, it matches the high, I don't think of, you know, as a given. You're right.

But if you think about the value proposition, right, of course there is the safety of it. You have to worry about. There's also the privacy. being in the car by yourself, maybe with other folks, but not having to share that space with another human, right? No, Wayne is a great product, yeah.

But I guess this is why we're seeing such, you know, consistency in car, you know, drives well, you know, very predictable. And, you know, you can go beyond that, right? And you specialize even more to make the experience even more magical around the writer. But I guess it would have been disappointing if, you know, without the specialized car. And I think I would have been surprised if we leveled off, at some other much lower level of customer adoption.

Because a car seems like more of an optimization and improvement, but the core of the value proposition comes from those other factors. Yes, yes. I guess it just take risk on one thing at a time. We'll start by doing the software layer and then we'll build a specialized car or something like that. That's right. Yeah, it's also, as I said, it's a big investment.

So you have to do you risk the fundamentals? Yes. And throughout our history, we were very focused on setting the most, you know, the biggest goal for the company to de-risk the most important questions, right? We talked about, you know, the third generation where, you know, we wanted to deploy something and go into it. We talked about what was the goal with the fourth generation, and then, sorry, the fifth generation. And then there's the sixth generation, right? So it was a sixth generation where it's made sense to go out, you know, spend all this, you know, effort into

the custom. And sixth generation is both the custom vehicle? Is it also a new generation of the driving stack? Yeah, it is the new hardware. Yep. The sensors, you know, the hardware, the software and hardware they're putting on the Ohio vehicle is the sixth generation. It is very different from the fifth generation. It is simpler. It is more capable. It is much lower cost. It's thinking a fraction of the cost is, you know, comparable to what you would get, like a fancy A-DAS system nowadays, the driver assist system. The software is pretty much the same. So when we talk about generalizability of the Waymo driver, we talk about weather conditions, we talk about cities,

but it also generalizes well to different vehicle platforms and different sensor configurations. Okay, so Gen 6 is a new vehicle and a new sensor stack, but it's almost a TikTok cycle happening here. It's a similar software. That's right, that's right. And then we're going to put the sixth generation Waymo driver on other vehicle platforms like the Hyundai Ionic.

that's coming, you know, later in the year. What is different about the sixth generation hardware stack, and how did you make it cheaper? So it still has the same three sensing modalities, but we've made significant optimizations in all three. So unification, simplification, and there's just, you know, the kind of just riding the commodity curve. Is it a classic case of, you know, manufacturing scale where we're not going to

well, scale hasn't fully come in a place, But all of those, if you think about the supply chains, the industries, cameras is pretty mature. Radars, way many years ago, used to be bulky, complex, very expensive. You know, when we're putting them on planes. But then we start putting them on cars. Now you can get a decent automotive radar for, you know, tons of dollars. There is a variant of the automotive radar. and it's called the imaging radar,

it gives you a richer, something. So that is also, you know, has come down and caused drastically, but it's a little bit behind your standard automotive raters. Liars are following the same, very predictable, very well-known trend. So we're writing that, and we're also, you know,

learning from the previous generation to just make improvements and simplifications and optimizations. So I have a very silly question. What are lighters versus radar is better in a self-driving context? Lider. Are they complementary?

They're very complimentary. Yeah. You know, it's all blasting, you know, effectively like you know, blasting photons out there, and then they bounce off of something, they come back, you know, you measure what comes back. The frequencies are very different. So laser gives you its very, very high resolution.

So you can think of it as like a laser beam that goes out, spins around, it shoots out millions of these laser pulses per second, and then each one comes back and you can, you know, you're kind of sampling the 3D structure of the world with very high resolution. So, LiDAR for very fine-grade mapping. That's right. Radar has much lower resolution, but because of the physics of it, it can, it degrades much better in adverse weather conditions.

So fog, snow, heavy rain, you know, So it's not included by particles between it and the target. So imagine driving in super dense fog. We're close to San Francisco, so probably don't have to think that hard. It can be really hard to see. So cameras degrade. Laser, depending on the size of the particulates, can degrade better or worse than camera.

Radar is not well-affected. So you can imagine driving on a freeway, then radar will give you really good returns for cars that are absolutely, invisible in the camera space. Oh, that's interesting. So does that mean there are some environments where you'll be relying significantly more on radar? But the performance is good enough? Well, it's a combination of the sensors, right?

So we rely on, you know, each one is noisy, right? How the noise characteristic show up in different environments is different, but it is, I mean, it's not like we switch from one to another. Not like we estimate what's happening with the world through cameras and for raiders and for lighter and then we compare. No, they're like there's an encoder for camera. There's an encoder for lighter. And they all go into the system that gives you jointly the best view of what's happening in the world. So if you're, you know, if it's a nice, bright sunny day, cameras are very valuable.

If it's pitch dark where you have like sun in your face or you're blinded by the headlights from, you know, an incoming car, then camera will degrade. There's still some noisy signal, but it will degrade. Yes, yes. And lighter lighter is completely unaffected. Are there technical problems that are your white whale or you're still chasing or you are particularly interested in solving, even if they're kind of niche for the, you know, we really want to have, you know, driving when it's actually snowing, nailed, or steep hills in San Francisco or, you know, are there problems you've been very interested in historically or still are. I'm super excited right now about the accelerating global expansion.

More cities in the United States and going internationally. So being, I understand I'm not answering your question about the knowledge. I'll come back to that. But really, that's the thing that I'm today most excited about. Just getting to a place where any major metropolitan area you can fly into the airport and then take away. or go anywhere you want to go. That is insanely exciting to me right now. So then technically, what I'm most excited about is all of the rapid progress in AI and the world models,

the foundational model work. And it is just such a massive boost to how much we can see. simplify the system, how much we can bring down the cost, and how we can scale globally. And there's just some magic that happens that I don't think I would have anticipated a few years ago. So that I find from the technical perspective, just insanely thrilling. Yes. When you talk about kind of the progress in AI, what are the most fun parts of it for you these days? I think it's seeing the capability and the scaling laws from,

from this approach of starting with that cornerstone of the foundational model and then specializing to T-shirts and then distilling, you get such big wins in performance across the board. I understand you invest something into the architecture or get better at data or a training recipe. And then you invest in that early stage and it just has massive amplification and ripple effects.

So that is in some ways is kind of magical. And then you, I guess, then you see it in the car. And I've had some moments where, you know, a car does something and you look at a log. And I've been surprised. Like it does things that I didn't think it was capable of doing. All right. So it's that, you know, it's, you know, it's, you know.

When you see emergent behavior, that's kind of a proud moment. One example, yeah. When you build a system and then you think you understand how it works and you understand fully the limits of its capability and performance and then it does something kind of almost magical, it's exhilarating. Yes. One example I can give you, I think I've shared some videos of those that publicly in some talks was this example where the situation happened in San Francisco are, you know, fairly benign situation where at an intersection, our light is red. There's cross traffic, a bus goes by, and it stops partially blocking. Our light turns green, so we start to go.

We're nudging around the bus, and then you see a pedestrian being detected on the other side of the bus. And then your car responds appropriately, it slows down, goes a little bit wider, and then a pedestrian. actually emerges from the bus and we go on our own way. So the first time I looked at that log, like, what's going on here? I know we have pretty darn good sensors, and the software is very capable. Like, we don't see through stuff, right? That's not how cameras or lighters and radars work, right? It saw the pedestrian through the bus.

It saw the pedestrian on the other side of the bus. And it's not like, you know, you look at the windows, you're like, okay, you know, radar shouldn't, it's a massive metal box. Yeah. Yeah, you know, look at the sensor data. Yes. And, like, it shouldn't, radar shouldn't be able to go through it, right? You know, camera, like, you can't see in the camera because, you know, there's reflections

and there's people on the bus, so it's not like you can see through the windows. Right. So, like, what is going on? Maybe it's, you know, noise or some coincidence. And I, you know, first time I saw it, like, I couldn't actually believe it. It's like, no, no, there's something. It doesn't spell right. So what actually turned out was happening is that our peripheral lighters and,

bounce under the bus, and there was just a little bit of very, very noisy reflection of the movement of the person's feet that was enough for the AI models, hey, I likely there's a pedestrian there, and I'm going to, you know, I detected a site. Yeah, yeah, yeah. And moreover, there's enough data there to predict what they're going to do. Yes. And it's kind of like, blew my mind. Is this the perfect example to explain what we were talking about earlier, the value of one fusion across a sensor suite, but then secondly, building, I mean, relatedly, building an intermediate representation of what's going on, where if you're just dealing with pixels, I mean, the person behind the bus does not exist in pixel space,

and so you need to have some representation of the world that exists to be able to reason about the person behind the bus. I think it's an example where giving it kind of using that intermediate representation to boost the level of performance of all parts of the model is what's happening here. Just imagine solving this problem with a black box, purely open loop, imitative system. Is it impossible? Is it impossible? No. In practice, what would it take to achieve that level of performance? Yes. Very, very difficult.

What metrics can you share in just where the business is at today in terms of rides, revenues, cars on the roads? We have about 3,000 cars on the roads. We're doing about half a million rides per week. That translates is about over 4 million fully autonomous miles. per week. We are operating in a fully autonomous mode in 11 cities in the U.S. And 10 of those, we have riders, public, you know, riders. What's the Go City? The Go City is Nashville. We just started there.

So we just opened it up to riders in four new cities in one day. That was one of those little but super exciting moments where I thought back to the history. How long did it take us from the first time we started fully autonomous writer-only operation to the first time we had external riders in four cities? It's about eight years. And then the other week we just launched four in one day. Yes, yes. It seems now clear that in 15 years, most miles that are driven,

will be autonomous. There'll be some burning period, and there's lots of old cars in the road, I think it'll actually take a little while. And some of that will be by level four, level five systems expanding in new cities, and that expansion continuing. Some of it will be, you know,

you referenced existing driver assist systems and kind of getting up to level two and level three and existing systems in across current car brands, getting more and more capable. What do you think that's working your way up from the lower levels versus working your way expanding from existing products like Waymo? What will that convergence look like? Because we're going to eat it from both sides.

I don't believe we will. And I actually think this. That's a great answer. Cars will get smarter. There's going to be advances in driver assistance. And if there is at the same time, from level four autonomy, you know, there is simplification and, you know, the sensors of today are not going to be the sensors of tomorrow, so they'll be much more integrated, they'll be simpler,

there will be much lower cost. So from that perspective, they're going to, you know, there is a path of convergence. Yep. And there's also, you know, a path of convergence from, you know, the product lines. You know, right healing and what, you know, you can take, you know, a ride through the Waymo app today, you know, eventually they'll be on your personal car. So that I see. You can talk about the technology and I see it just as fundamentally two different problems. There's driver assist systems and then there is full autonomy. And I think it's deceptive to think of them as kind of incremental on one spectrum of complexity. Okay, but you think one cannot work one's way up from driver assist systems to full self-driving. You think you have to start building a full self-driving

system. You have to tackle, if I think about the hardest parts of building a fully autonomous, you know, rider only system, they are very different from, you know, what you do for a driver assist system. Yep. And, of course, you know, some work in the space helps you, right? So, you know, I don't want to say you can't make the jump, but it is a qualitative jump.

Yes. When can I buy a Waymo so that I don't need to wait for us when I want to. to go, I can just like, when I'm ready, I can walk out the door and it's there. I'm not going to give you a date to date, but you're not the first person to bring this up as a product request. Do we note it. Okay. I'll add it to the list. You know, that's waiting for the car.

It should be nice just in the garage there and keep your stuff in it and everything. It's not the first time you've heard that request. So how, it seems to me operationally very intensive and very hard. Like a self-driving car is actually not self-driving. It takes a village. You have all of the human operator ready to step in. And there was that thundering herd incident that you guys talked about in San Francisco. That kind of highlighted that for people.

And then there's just like keeping the cars clean and keeping everything running in that regard. And so, can you describe just what the operational infrastructure that sits behind Waymo looks like? Sure. And I will say that we are overall, in all of those areas, on a path of increasing efficiency and automation. Yep. Right. So, you know, the number of manual steps that, you know, one had to do, you know, five years ago to, you know, launch AWAMO versus where we are today is drastically different, right? say, yeah, but nowadays, if you look at one of our depots as like a fully automatically orchestrated, you know, dance of autonomous vehicles.

So the way it looks, you know, what it looks like today is cars will automatically, you know, go on their, you know, to pick up their riders, you know, serve their trips. if for some reason, you know, they need to come back. You know, maybe they're low on energy. Maybe if somebody left the mess on the car, they will, you know, automatically come to the depot. If it is, so cleaning today is a manual process, right? So it'll get flagged in the car, you know, we have fleet management systems. Say, hey, you know, car, you know, 378 needs cleaning. And we'll actually on the sensor dome, we're able to, you know, display icons.

So we'll show you like a little emoji. Don't put their handouts. Yeah. And, you know, there's people whose job it is to clean cars. They'll come and, you know, clean up. If that's, you know, cleaning is not required and it's just charging, you know, we'll also, you know, pull automatically pull into a charging stall. And we'll say, hey, you know, I need charging.

We don't yet have automated charging. In the future, you can imagine that being fully automated, right? But, you know, a person will come in and, you know, plug in a cable and the car will charge and then say, hey, you know, now I'm ready to go. And, you know, it will get unplugged and the car will, you know, pull out of the, you know, it's parking stall and then go on its merry way. One of the new portions, I think it is, has inductive charging, just like your iPhone where you just drive over the charging mat.

I was amazed that that works at car scale, but presumably in the future they'll just be able to drive on with the charging mat. Or do you think just robotic plug-in will be easier? We'll see. I don't know. I think there's some questions about efficiency and how that plays into the overall cost and which one will be most cost-beneficial remains to be seen, I think. How well-behaved are the Waymo riding population in terms of not living a mess in the car?

We have wonderful riders. We have the most amazing customers in the world. Generally, I would say they are very good. I think, you know, there is something about, you know, I talked about not having, you know, a person in the car, it's not somebody else's car. In some ways, you kind of like want to preserve the, I think generally people want to kind of preserve the nice aspect. It's a broken window thing where it's so clean to begin with. I know, yeah, it's kind of like, you know, I think that that's the general trend that we see, right? And it's like because there's not somebody else's space, you know, you're in it, it feels like it's your own.

Yeah. So you don't like want to mess up, you know, your own space. So I think, I mean, I don't want to speculate too much on the, you know, psychology of things. However, I will say that it varies. And you can imagine, you know, a college town on a, you know, Saturday night. And, you know, that's a different distribution. Yes, yes. Will I be able to get away mode any address that has USPS service in the US,

or will there be some head-tail dynamic where Ketchkin, Alaska is just never worth it? Eventually you will, absolutely. There's no doubt in my mind. I think it's just a matter of when and what modality would make the most commercial sense. This is for your right-chair versus privately owned. For right, you know, it's not a technical problem. I mean, you know, technology is solved. Yes.

But then, you know, if you're in the middle of nowhere and there's just not enough density of the trips, does it make sense for the right healing service that, you know, Waymo is running to, you know, have cars on standby? Yes, yes. Probably not, right? They can be deployed, you know, somewhere else. And you probably don't want, you know, a horribly bad ETA.

And this is where, you know, personally owned vehicle that is equipped with the Waymo driver is maybe, you know, how you will see it materialized. Relatedly, what will the second order affect of say majority autonomous traffic be like it feels like a lot of things will work better where as you say you know when someone merges into a lane very poorly and everyone all the way back has to you know slam on the brakes that's kind of anti-social behavior and so it feels like higher quality and more pro-social driving will just I mean basically reduce traffic a little

bit even for the same number of cars on the road but presumably there'll be other second-order effects like we'll want higher throughput traffic lights and yeah how else will things change So the first thing I think, you know, that you mentioned is, you know, that's a huge deal. I just think about traffic jams. Yeah. And what was that saying? The Navy SEALS, slow is smooth and smooth as fast. Right?

That's what, like, you know, traffic jams are like, you know, you accelerate abruptly. Then you come to a stop. And, you know, sometimes you have a traffic deal, like, what happened? Yes. Well, like, you know, an old lady crossed the road three hours ago, and we still have the standing way. there, right? So if everybody, you know, was a kind of a smooth, predictable driver and a consistent driver, and you would still have those, you know, traffic jams at the time off, but then the time constant to clean it out, I think would be very different. But longer term, and, you know, things like parking lots. Right now, if you look at, you know, what is our most interesting, you know, pieces of land allocated to. It's parking lots, it's garages, and why is that? Well, because, again, your car is just sitting there 90% of the time, right?

If more cars become fully autonomous, then there's no need of that, right? And then imagine, just imagine what you can do with your favorite city in the world if you don't have to spend that much, that huge fraction of it on, you know, just keeping these chunks of metal sitting around. Yeah, I don't think people often realize how big a deal of parking minimums are, for the layout of the urban landscape, the coffee shop here where I am, would like to have outdoor seating but can't

because it would reclaim parking spots. Yeah, wouldn't have a few more questions, but I'm curious to talk about Google's relationship with self-driving where, again, it feels like right now, Waymo is, aside from everything else, AI-related, kind of the most exciting thing happening at Google. But it was a very long journey to get here. I mean, I feel like you could say that Google almost started working on it too early

because you were saying there's been a bunch of recent enabling technologies. And so did it require Google starting when it did so early? Or could one have spun up this project in 2015, 2020? And then how did Google keep the faith when it always felt like it was, perennially two years away. Yeah, no, I have another latter part. I just have to give credit, huge kudos and gratitude

to Larry and Sergey and, you know, Alphabet leadership in the center company. It is part of the culture and the DNA of the company is to have that vision and have the stamina

and conviction to go the distance. So to the other part of the question, was it too early? I don't know. I think what we've been seeing clearly all of the breakthroughs that we've seen over the years have changed how we're building the system. But the complexity of the problem is such that you need to go through these ether of cycles, right, it's not, you know, still, and we've seen many waves of technology, right?

There's, you know, breakthroughs in, you know, 2013, ImageNet came around. And there's, okay, like, that is the right time to start a BSLDRI company. And then, you know, Transformers came around, and your VLMs and that, and all of those are super powerful and you have applications in other spaces. In the digital world, they certainly have an impact on, you know, our AI and in the physical world. but there are no silver bullets. They kind of, they drastically reshape that early part of the curve. It's always been the nature of this problem.

It's very easy to get started. It's deceptively easy to get started. But it is super hard to go, you know, the full distance. It's the number of nines, right? There's the standard engineering rule of thumb that, you know, every next nine takes, you know, 10x more. So I, yeah, Maybe there is a more optimal path, but I don't see that there's some magical moment where the true complexity of the problem goes away,

and then you can just take some off-the-shelf components in your business. If that were the case, then I think the industry would look very different today. Yeah, yeah. Last question I have, you've been promoted a lot at Google. It feels like Google really recognized your talents. Just what do you think Google does, like Google is famously one of the very best in the world at technical talent and say, you know, the current AI wave more broadly happening, you know, is either stuff happening at Google or generally Google alumni.

But just what have you observed firsthand from how Google does this so well? Yeah, I'll say Google, you know, that culture of Google of not accepting the status quo, having a big vision and investing in technical talent, the people who can go the distance and realize the vision, that is part of the culture. I think this is what you're seeing. And with the breakthroughs in AI in the digital world and all of the early investments in transformers

and other fundamental technologies, quantum computing, And I guess we are not unlike those efforts as well. To be sure. Thank you. Thank you.