The a16z Show - Building the World's Most Trusted Driver

Episode Date: August 5, 2024

Waymo's autonomous vehicles have driven over 20 million miles on public roads and billions more in simulation.In this episode, a16z General Partner David George sits down with Dmitri Dolgov, CTO at Waymo, to discuss the development of self-driving technology. Dmitri provides technical insights into the evolution of hardware and software, the impact of generative AI, and the safety standards that guide Waymo's innovations.This footage is from AI Revolution, an event that a16z recently hosted in San Francisco. Watch the full event here:  a16z.com/dmitri-dolgov-waymo-ai Resources: Find Dmitri on Twitter: https://x.com/dmitri_dolgovFind David George on Twitter: https://x.com/DavidGeorge83Learn more about Waymo: https://waymo.com/ Stay Updated: Let us know what you think: https://ratethispodcast.com/a16zFind a16z on Twitter: https://twitter.com/a16zFind a16z on LinkedIn: https://www.linkedin.com/company/a16zSubscribe on your favorite podcast app: https://a16z.simplecast.com/Follow our host: https://twitter.com/stephsmithioPlease note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details please see a16z.com/disclosures. Stay Updated:Find a16z on YouTube: YouTubeFind a16z on XFind a16z on LinkedInListen to the a16z Show on SpotifyListen to the a16z Show on Apple PodcastsFollow our host: https://twitter.com/eriktorenberg Please note that the content here is for informational purposes only; should NOT be taken as legal, business, tax, or investment advice or be used to evaluate any investment or security; and is not directed at any investors or potential investors in any a16z fund. a16z and its affiliates may maintain investments in the companies discussed. For more details please see a16z.com/disclosures. Hosted by Simplecast, an AdsWizz company. See pcm.adswizz.com for information about our collection and use of personal data for advertising.

Hello everyone. Welcome back to the A16Z podcast. This is Steph. Now, one of my favorite podcasts we've recorded since I joined the team was just about this time last year. That episode was on autonomous vehicles, but it was actually also in an autonomous vehicle. That was my first ride in a self-driving car. And over the last year, I've seen so many others have their first as Waymo has expanded to the public in Phoenix and San Francisco, while also placing its roots in Austin and LA. In 2015, Waymo tested its first fully driverless ride on public roads. It then opened to the public in Phoenix in 2020, but it wasn't until 2022 that autonomous drives were offered in San Francisco. And by the end of 2023, it clocked in over 7 million driverless miles. Slowly, then all at once. So with this space moving so quickly, we wanted to give you an update on where this industry is today.

Passing the baton to properly introduce this episode, Here is our very own AI Revolution host and A16Z general partner, Sarah Wang. As a reminder, the content here is for informational purposes only, should not be taken as legal, business, tax, or investment advice, or be used to evaluate any investment or security, and is not directed at any investors or potential investors in any A16C fund. Please note that A16D and its affiliates may also maintain investments in the companies discussed in this podcast. For more details, including a link to our investments, please see A16C.com slash disclosures.

Hey guys, I'm Sarah Wang, general partner on the A16Z growth team. Welcome back to our AI Revolution series. In this series, we talk to the Gen AI builders who are transforming our world to understand, one, where we are, two, where we're going, and three, the big, open questions in the field. Our guest this episode is Dmitri Dolgov, the co-CEO of Waymo. Dimitri has led Waymo to solve some of the biggest challenges in bringing AI to the real world. And after tens of millions of miles of testing, Waymo's vehicles have shown themselves to be safer and more reliable than human drivers, myself included. Dimitri has a unique perspective, given that his work has spanned multiple AIML development cycles across decades.

He was an early pioneer in self-driving cars, working with Toyota and Stanford on DARPA's Grand Challenge before joining Google's self-driving car project. which then evolved into Waymo. In this conversation from a closed-door event with A16Z general partner David George, Dmitri talks about the potential of embodied AI, the value of simulations and building training data, and his approach to leading a company focused on solving some of the world's hardest problems.

Without further ado, here's Dmitri in conversation with David. Maybe to start, take us back to Stanford, if you will, and that was when you first started working on the DARPA project, and maybe give us a little bit of your history of how you ended up from there to here. My introduction to autonomous vehicles was when I was doing a postdoc as at Stanford that you just mentioned, David. This was during, I got pretty lucky with the timing of it.

This was when the DARPA Grand Challenges were happening. DARPA is the Defense Advanced Research Project Agency that started these competitions with the goal of boosting this field of autonomous vehicles. And the one that I got involved in was in 2007 that was called the DARPA Urban Challenge. So the setup there was, it's kind of like a toy version of what we've been working on since then.

It was kind of supposed to mimic the driving in urban environments. So they kind of created a fake city on an abandoned airbase and they populated it with a bunch of autonomous vehicles a bunch of human drivers, and they had them to various tasks. So that was kind of my introduction to this whole field, and it was a bit of a, I think, you know, DARP, these challenges are often by people in the industry considered kind of a foundational, pivotal moment

for this whole field, and it was definitely that for me. It was like a light bulb, light switch moment that really got me hooked. What was like the hardware and software that you guys had at that point? This is 2007? Yeah, yeah, I know. It's a very high level, not unlike what we talk about today.

A car that has some instrumentation so you can tell it what to do and you get some feedback back. Then you have what's called a pose system, a bunch of inertial measurement system, accelerometers, gyroscopes that kind of tell you, GPS, tells you how you're moving through space. And it has sensors, and radars, lighters, and cameras, the same stuff we still use today.

And then there's a computer that gets the sensor data in and then tells the car what to do, and a bunch of software. And software had perception components and decision-making planning components and some AI. But of course everything that we had, like each one of those things over that, how long has it been, 18 years more than that,

has changed drastically, right? So when we talk about AI today versus AI we had back in 2007, 2009, nothing in common. And similarly, everything else. has changed. The sensor are not the same. Computers are not the same. Yeah, of course. So then, okay, so take us, so at that point, that was the pivotal, that was like the light bulb moment. And then at that point, you said, okay, I'm at Stanford, I want to make this my career, right? Is that, and then it was Toyota, and then where did it go from

there? I don't know if I thought about it in those terms. I was like, this is pretty cool to work on. This is the future. I want to make it happen. I want to be building this thing. Career. Okay, you know, they can wait. But it was, that was the next step. That was the next big step. is a number of us from the DARPA challenge competitions started the Google self-driving project. It was about a dozen of us. Then in 2009, came together at Google with support and exciting from Larry and Sergey to see if we can take it to the next step.

And then, we worked on it for a few years, and that project then became Waymo in 2016, and we've been on this path since then. Okay, so we have this new big breakthrough in generative AI. Some would say it's new. Some would say it's 70 years in the making. How do you think about layering advances that have come from generative AI to what many would describe as more traditional AI or machine learning techniques that were kind of the building blocks for self-driving technology up to that point?

Yeah, great question. So maybe it can, you know, generative AI is kind of a broad term. Sure. So maybe you can maybe take a little bit of. of a step back and talk about the role that AI plays in autonomous vehicles and kind of how we saw the various breakthroughs in AI map to the space of our task. As you mentioned, AI has been part of self-driving autonomous vehicles from the earliest days.

Back when we started, it was a very different kind of AI, ML, kind of classical main decision trees, classical computer visions with kind of hand engineered features, kernels and and so forth. And then, you know, one of the first really important breakthroughs that happened in AI and computer vision, but really was important for our task was the advancement in convolutional neural networks right around 2012, right? Many of you are probably familiar with AlexNet and the ImageNet competition. This is where AlexNet, you know, blew away.

out of the water, all other approaches. So that obviously has had very strong implications for our domain, like how you do computer vision and not just on cameras, right? How you're on, you know, you can use ConvNets to interpret what's around you and do kind of object detection and classification from camera data, from LiDAR data, from your imaging radars. So that was kind of a big boost around that, you know, 2012-2013 timeframe. And then we played with those approaches and, you know,

try to extend the use of ConvNets to other domains, just be a little beyond perception, with some interesting but limited success. Then another big, very important breakthrough happened around 2017, when Transformers came around. It had a really huge impact on language, language understanding, language models, machine translation, so forth. And for us, it was a really important breakthrough that really allowed us to take ML and AI to new areas well beyond perception. And so if you think about transformers and the impact that they had on language, the intuition is that they're good at understanding and predicting and generating sequences of words, right? And in our case, we think about in our domain, think about the tasks of, you know,

understanding and predicting what, you know, people will do, like other actors in the scene, or the task of decision-making and planning your own trajectories or in simulation, kind of generating, generative AI, our version of genera AI, generating behaviors of how the world will evolve, that these behavior, like these sequences are not unlike sentences, right? You're kind of operating the state of objects, and there's kind of local continuity,

but then the global context of the scene really matters. So this is where we saw some really exciting breakthroughs in behavior prediction and decision making and simulation. And then, since then, we've been on this trend of models getting bigger. People started building foundational models for multitask. And most recently, all of the last couple of years, all the breakthroughs in large language models,

modern state, modern-day generative AI, visual language models where you kind of align image understanding and language understanding. And there's been, most recently, one thing I'm pretty excited about is kind of the intersection or combination of the two. So that's what we've been very focused on, that Waymo most recently, is taking kind of the AI backbone and all of the AI, the Waymo AI that over the years we've built up

that is really proficient in this task of autonomous driving and combining it with kind of the general world knowledge and understanding of these VLMs. One of the things that you just mentioned is the role of simulation and how that has been, you guys have had major breakthrough in the use of simulation. And this idea in, you know, the recent breakthroughs in generative AI

around synthetic data and its usefulness is somewhat in question. I would say in your field, this idea of synthetic data and simulation is extremely useful, and you've proven that. So maybe you could just talk about

the simulation technology you guys have built, how it's allowed you to scale, you know, build that real world understanding, and maybe how it's changed in the last few years. Yeah, yeah, definitely. It is super important in our field. Largely, if you think about this question of evaluating the driver, is it good enough?

How do you answer that? There's a lot of metrics and a lot of data sets you have to build up. And then how do you evaluate the latest version of your system? You can't just throw it on the physical world and then see what happens. You have to do in a simulation. But of course, the new system behaves differently from what might have happened in the world otherwise. So you have to have a realistic closed-loop simulation to give you confidence in that. So that is one of the most important needs for the simulation.

You've also mentioned synthetic data. That's another area where simulation allows you to have very high leverage. You just kind of explore the long tail of events, right? Maybe there's something interesting that you have seen in the physical world, but you know, you want to modify that scenario, and you want to kind of turn one event into thousands or tens of thousands of variations of that scenario. You know, how do you do that?

You know, this is where the simulation comes in. And then, you know, lastly, if you, you know, sometimes want to evaluate and train on things that you've never seen, you and are very vast experience, So this is where purely synthetic simulations come in that are not based on anything that you have seen in the physical world. So in terms of technologies that go into play,

I mean, it's a lot. And that is like a huge generative AI problem. But what's really important is that that simulator is realistic. It has to be realistic in terms of your sensor or perception realism. It has to be realistic in terms of the, behaviors that you see from other dynamic actors. If other actors are not behaving in a realistic way, like if pedestrians are not walking the way they do in the real world,

you need to be able to quantify the scenarios that you create in simulation to the realism and the rate of occurrence in the physical world. It's very crazy to sample something very easy to sample something totally crazy in a simulator, but then, you know, What do you do with that? So I think that that brings me to the third point of realism is that it has to be kind of realistic and quantifiable at the macro level at the statistical level.

So there's, and you can imagine, there's a lot of work that goes into building a simulator that is large scale and has that level of realism across those categories. And if kind of intuitively think about it, to build a good driver, you need to have a very good simulator. But to have a good simulator, you actually have to build models of like realistic pedestrians

and cyclists and drivers, right? So it's good, you know, you kind of do that iteratively. Yeah, of course. And then by having this simulation software that is very good at mimicking real world and very usable in the sense that you can create variables in the scenes, you can actually give the driver multiples of the amount of experience that they have on the road. That's exactly right. In real miles of right.

That's exactly right. I've driven tens of millions of miles in the physical world. At this point, we've driven more than 15 million miles in full autonomy, what we call, you know, writer-only mode, but we've driven, you know, tens of billions of miles of simulation, so you get, you know, orders of magnitude of an amplifier. Speaking of multiples of miles driven, one of the hotly debated topics in the AI world today

is this concept of scaling laws. So how do you think about scaling laws as it relates to autonomous driving? Is it miles driven? Is it certain experience had? Is it compute? Like, what are the ways that you think about that? So model size matters.

So we are saying scaling laws applied. A lot of typical old school models are severely under trained. So if you have a bigger model, you have data that actually does help you. You just have more capacity that generalize better. So we are seeing scaling laws apply there. Data of course usually matters. But it's not just counting the miles. right, or ours, it has to be the right kind of data that, you know, teaches the models or

trains the models to be, you know, good at the rare cases that you care about. And then, you know, there is a bit of a, then, you know, a wrinkle because then you have to, you can build those very large models, but in our space, it has to run on board the car, right? So you are somewhat computer restraint, so you have to distill it into your, you know, onboard system. But we do see a trend, which is, you know, common trend and we see that play out in our space, where you're much better off training a huge model and then distilling it into a small model than just training small models.

Yeah. I'm going to shift gears a little bit and I'm going to do a sort of simplifying statement, which is probably going to drive you crazy. But the DARPA School of Thought is, you know, there's sort of a rules-based approach, right? A more traditional kind of AI-based approach with a massive amount of volume

and you document edge cases and then the model then learns how to react, to those. The more recent approaches from some other large players and startups would say, hey, we just have AI from the start, make all the decisions end-to-end. You don't need to have sort of all that pattern recognition and learning, you know, like the end-to-end driving that is kind of a tagline out there. What is your interpretation of that approach and what elements of that approach have you taken and applied inside of Waymo? Yeah, I think it's kind of, you know, sometimes it's a, you know, the way people talk about it as kind of a, this weird dichotomy, is it this or that?

Yeah, of course. But it's not. It's that and then some, right? So it is, you know, big models. It is end-to-end models. Yeah, it is generative AI and combining, you know, these models with VLMs, right? But the problem is it's not enough. Right.

Right. So, I mean, like, we all know the limitations of. those models, right? And that's what we've seen, you know, through the years a lot of these breakthroughs in AI, right? Convness, transformers, you have big end-to-end foundation models. They're huge boosts to us. And what we've been doing at Weyman through the history of our project is kind of constantly

applying and pushing forward these data of our techniques ourselves in some cases, but then applying them to our domain. And what we've been learning is that they really give you a huge boost, but they're just not enough. Right. So the kind of the theme has. has always been that you can take your kind of latest and greatest technology of the day and it's fairly easy to get started.

Right. Like the curves always look like that. And they've been kind of the curves in the chirping, but the really hard problems in that remaining 0.001%. And there it's not enough. So then you have to do stuff on top of that, right? So yes, you can take, you know, nowadays you can take an end-to-end model, go from sensor to you know, trajectory as or actuation. You know, typically you don't build them in one stage, you build them in stages.

but you can do backprop through the whole thing. So the concept is very, very valid. You can combine it with a VLM, and then you add closed-loop simulation, some sort, and you're off to the races. You can have a great demo, like almost out of the box. You can have an ADESP or a driver system, but that's not enough to go all the way to full autonomy.

So that's where really a lot of the hard work happens. So I guess the question is not as it this or that, it's this, and then what else do you need to take it all the way to have the confidence in, so that you can actually remove the driver and go for full autonomy. And that's a ton of work. That's a ton of work through the entire kind of life cycle

of these models and the entire system, right? So it starts with training, like how do you train, how do you architect these models? How do you evaluate them? Then if you put in a bigger system, the models themselves are not enough, so you have to do things around them. You have to, you know, they have modern,

general AI is great, but there are some issues with, you know, hallucinations. explanations, explainability. Exactly, exactly. They have some weaknesses in kind of goal-oriented planning and policymaking and kind of understanding this, you know, operating in this 3D spatial world.

So you have to add something on top of that. We talked a little bit about the simulator. That's a really hard problem in of itself. And then, you know, once you have something, you know, once you deploy it and you learn, how do you feed that back? So I guess this is where all of the really, really hard work happens. So it's not like end-to-end versus something else. It is end-to-end.

And, you know, big foundation models. and then the hard work. And then all the hard work. Yeah, it totally makes sense. That is a great segue into all of the progress that you guys have made, right? Writing in the Waymo for those who have done it is an extraordinary experience.

It's not to say that you have solved all of these complex tasks, but you've solved a lot of them. What are some of the biggest AI or data problems that you still feel like you're facing today? The short answer is going to be, you know, taking it to the next order of magnitude of scale, multiple orders of magnitude of scale. And with that come additional improvements that we need to make it a great service.

But just to level set and in terms of where we are today, we are driving in all kinds of conditions. We're driving 24-7 in San Francisco, in Phoenix. a little bit, those are the most metro markets, but also in LA and in Austin, and all of the complexity that you see, you know, go and drive around the city, right? All kinds of weather conditions, whether it's, you know, fog or, you know, storms or, you know, dust storms or, you know, rainstorms down here, like all of that, all of those are conditions that we do operate in, right? So then I think about, you know, what makes it a great, you know, customer experience, right? Like, what does it take if you, you know, grow by, you know, next, you know, orders of

magnitude, there's a lot of improvements that we want to make so that it becomes a better service for you to get from point A to point B, right? Like, we ask for feedback from our writers. A lot of feedback we get has to do with the quality of your pickup and drop-off locations, right? So we're learning from users, like we want to make it a magical, seamless, you know, delightful experience from the time you kind of start the app on your phone to when you get in the decision. So that's a lot of the work that we're doing right now. Yeah. Pick-up and drop-off for what it's worth is an extraordinarily hard problem. right? Like, do you kind of block a little bit of a driveway if you're in an urban

location and then have a sensor that says, oh, actually I just saw somebody opening a garage door, I need to get out of the way, you know, how far down the street is acceptable to go pull? Or if you're in a parking lot, where in the parking lot do you go? Like, this is an extraordinarily hard problem, but to your point, it's huge for user experience. That's exactly right, right? And I think that's a good example of like to say, just one thing, one of the many things that we have to build in order for this to be an awesome product, right? Not just like a technology demonstrator. And I think you just like, you hit exactly on a few things that make, you know, something that kind of at the face of it might seem

fairly straightforward, right? Okay, you know, I know there's a place on the map and need to pull over, so like how hard can it be, right? But really, if it's a complicated, you know, a dense urban environment, there's a lot of these factors, right? Is there like, you know, another vehicle that you're going to be blocking? Is there a garage door that's opening, right? Like, you know, what is the most convenient place for the user to pick off? What is the, so it really gets into this, the depth and the subtlety of understanding the, kind of the semantics and the dynamic nature of this driving task and, you know, doing things that are, you know, safe, comfortable and predictable and it lead to a nice, seamless, pleasant,

delightful customer experience. Of course. Okay, so you've mentioned this stat, but 15 million miles. I know the number's probably a little bit bigger than that, but you just released it Tuesday. Yeah, it's growing by the day. 15 million autonomous miles driven. That's incredible. Even more impressive, and you didn't share this stat yet, it results in 3.5 times fewer accidents than human drivers.

Is that right? And I think 3.5 acts as the reduction in injury, and that's about 2x reduction in the police reportable lower severity incidents. This sort of comes to a question of both kind of regulatory and, you know, kind of business or ethical judgment. What is the right level that you want to get to? Obviously, you want to constantly get better, but is there a level at which you say, okay, we're good enough, and that's acceptable to regulators? Yeah, so there's no, you know, simple answer, super simple, short answer. Right.

I think it starts with that. It starts with those statistics that you just mentioned. Yeah. At the end of the day, what you care about is that rows are safer. So then you look at those numbers, where we operate today, and we have strong empirical evidence that our cars are in those areas safer than human drivers. So on balance, that means a reduction in collisions and harm. Then, actually, on top of the numbers, we've probably been publishing this,

you're quoting the latest numbers that we shared. Yeah. Consistently sharing numbers as our service scales up. and growth. If you can also bring in, you know, an additional lens of, you know, how much did you contribute to a collision? And we actually published, I think it was based on about 4 million miles, 3.8 million miles. We published a joint study with Swiss RE, which is, I think, the largest global reinsurer in the world. And the way they look at it is, you know, who contributed to an event. And there we saw like the same theme, but the numbers were very strong that Newfield was a 76% reduction

in property damage collisions, and it was in 100% reduction in claims around bodily injury. So if you kind of bring in that lens, I think the story becomes even more compelling. That is extremely compelling. Right, but there are some collisions where, you know, we'd be, and that's the bulk of the events that we see, it would be stopped at a red light,

and then somebody just plows into you, right? Sure. So, yeah. But then, like, we do know, it's a new technology, it's a new product, so, it is held to a higher standard. So when we think about our safety and our readiness, you know, framing of methodology, we don't stop at just the race, right?

We build over the years, you know, one of the huge areas of investment and experience over the years, like how, you know, what else do you need? So we have done, and we've done a number of the other different things. And we've published some of our methodologies. We've shared our readiness framework. You know, we do other things like we actually, not just statistically, but on, you know, specific events, we build models of, an attentive, very good human driver.

Like, not distracted, human, it's a good question whether such a driver exists, right? But that's kind of what we compare our driver to, right? And then in particular scenario, we evaluate ourselves versus that model of a human driver, and we hold ourselves to the bar of doing well compared to that very high standard. And then, you know, you pursue other validation methodology.

So that's my answer is that it's the aggregate of all of those methodologies that we, look at to decide that, yes, you know, the system is ready enough to be deployed in scale. I'd love for you to talk about what you think maybe today and in the future about market structure, competition, and what kind of role you envision Waymo playing. So the way we think about Waymo and our company is that we are building a generalizable driver. That's the core. And that's the core of the mission of making transportation safe.

and accessible. Right. And we're talking about right-hailing today. That's our main, most mature primary application. But, you know, we envision a future where the Waymo driver will be deployed in other commercial applications, right? There's deliveries, there's trucking, there's personally owned vehicles. So in all of those, you know, our guiding principle would be to think about the go-to-market

strategy in a way that... that accelerates access to this technology and gets deployed as broadly, while of course doing it gradually and deliberately and safely, as quickly and broadly as possible. So with that as our guiding principle, we're gonna explore different commercial structures, different partnership structures.

For example, in Phoenix today, we have a partnership with Uber and Wright Healing, both in River, Ray Healing, and Uber Eats. So in Phoenix, we have our own app. You can download the Waymo app and take a ride. And our vehicle will show up and take it where we want to go. That's one way to experience our product. Another one is through the Uber app.

We have a partnership where you can get through Uber app matched with our product, the Waymo driver, the Waymo vehicle, and it's the same experience, right? But this is another way for us to accelerate and give more people to experience full autonomy. And it gives us a chance to kind of think about the different and go-to-market strategies, right? One is us having more of our own app.

The other one is more of a driver as a service for somebody else's network. So we'll still early days, but we'll iterate and but all in service of that main principle. That's amazing. Yeah, that's going to be exciting. Maybe on back to the vehicle, what about the hardware stack that you use?

You and I have talked a bunch about, you know, you said like, hey, going all the way back to DARPA, you know, it's kind of the same stuff, right? It's, you know, it's sensor. They've advanced quite considerably, but, you know, you still use, you know, radars and LIDAR. Do you think that remains the future path for autonomous driving, LiDAR specifically? Yeah, no, I mean, the sensors are physically different, right? They have each one in cameras, Liders, radar.

They have their benefits. Each one brings their own benefits, right? You know, cameras obviously give you color and they give you high, you know, very high resolution. Lighters kind of give you, you know, a direct 3D measurement of your environment and they're an active sensor, right? So kind of bring their own energy, it's a pitch dark when there's no, you know, external light source. You know, you still get, you know, see just as well as they do during the day, you know, better in some cases. And then, you know, radar is, you know, very good at, like, punching through just, you know, physics, different wavelengths, right? to, if you build an imaging radar, which we do ourselves,

you know, it allows this to, you know, give you an additional redundancy layer, and it has benefits also an active sensor. It can directly measure, you know, through Doppler velocity of other objects, and it can, you know, it degrades differently and more gracefully in some weather conditions, like in a very dense fog, you know, are very dense, right?

So, you know, they will have their benefits. So if you, you know, our approach has been to, you know, use all of them, right? And that's how you have redundancy, and that's how you get an extra boost and capability of the system. And we are on, you know, today deployed and fifth and working to deploy

the sixth generation of our sensors. And, you know, over those generations, we've improved, you know, reliability, we've improved, you know, capability and performance, and we've brought down the cost very significantly, right? So, yeah, I think the trend, you know, for us, you know, using all three modalities, just makes a lot of sense.

Again, you might make different trade-offs if you are building a driver-assist system versus a fully autonomous vehicle where that last 0.001% really, really matters. Yeah, absolutely. One of the observations that we have from the very early days of this wave of LLMs is that there has been

sort of already a massive race of cost reduction, and many would argue that it's sort of a process of commoditization already, even though it's very early days. I would say the observation from autonomous driving over many, many years now is kind of the opposite thing. There's been a thinning of the field. It's proven to be much, much harder than expected.

Can you just talk about maybe why that's the case? You know, they always have this property that it's very easy to get started, but it's very insanely difficult to get it all the way, to full autonomy so that you can remove the driver. And there's maybe a few factors that contribute to that. One is compared to the LLMs and it's kind of AI in the digital world. You have to operate in the physical world.

The physical world is messy, it is noisy, and it can be quite humbling. There's all kinds of uncertainty and noise that can pull you up. out of distribution, if you will. Right, sure. So that's one thing, that makes us very difficult. And secondly, it's safety, right? Sure. These AI systems, in some domain, you know,

this is creativity and it's great. In our domain, the cost of mistakes, our lack of accuracy has very serious consequences, right? So that says the bar very, very high. And then the main, the last thing is, that it is, you have to operate in real time. And you're putting these systems on fast-moving vehicles, and you have to, you know, milliseconds matter, right?

You have to make the decisions very quickly. So I think it's, you know, the combination of those factors that really, you know, together lead to, you know, the trend that you've been seeing is that, like, you know, it's an and, right? You have to be excellent on this and this and this and then, right? It's all of the above. The bar is very, very high for, you know,

every component of the system and how you put them together. But there's big advances and they boost you and they propel the system forward, but there are no silver bullets. And there's no shortcuts if you're talking about full autonomy. And because of that lack of tolerance for errors, you have a very high bar for safety. You have a very high burden from regulators. You know, it's very costly to go through all those processes. And so it makes sense. And I'm very grateful that you guys have seen it through despite all the humbling experiences that you had a lot.

along the way. It's been a long journey, but it's, you know, for me and many people at Waymo, it is super exciting and very, very rewarding to finally see it become reality. Now we talk about safety and AI in many contexts, right? That's a big question, right? But here we are in this application of AI in the fiscal world. We have at this point a pretty robust and increasing body of evidence that, you know, we are seeing tangible safety benefits.

That's very exciting. Yeah, I always say to people, it was a long journey and very costly and expensive along the way. But this is probably the most powerful manifestation of AI that we have available to us in the world today. I mean, you can get in a car without a driver, and it's safer than having a human. And that's just remarkable. What were some of those humbling events along the way? And those early days in the first couple of years? Oh, I'm sorry.

I remember one, there's one route that we did that started, I think it started in Monterview, then went through Palo Alto, then went, you know, through the mountains to Highway 1, that took Highway 1 to San Francisco and I think, you know, went around the city a little bit and like actually finished for Lombar Street. So like in 2009, 10 people. That is really complicated. A hundred miles from beginning to end, right? You know, as human drivers would fail at that task, I think.

So, yeah, yeah. Yeah, yeah. So, you know, we're doing it one day. And we're driving, you kind of made it through the Monteree Palo Alto part. We're driving through the mountains, and it's foggy, it's early morning. And then we're like seeing objects. And, you know, objects seem like random stuff on the road in front of us. There's like a bucket and like a shoe.

And then there's like at some point we come across like a, you know, a rusty bicycle. Like, okay, what's going on there? And then we catch, you know, eventually. And then the car, you know, it doesn't, you know, handles it okay. You know, maybe not super smoothly. But, you know, we didn't get stuck. can we catch up to like this dump truck that has all kind of stuff on it and just like, you know, periodically losing things that present obstacles to the car.

It's like a, this is like a cartoon, you know, continuation of anomalies being thrown at you guys. That's pretty cool. Okay, last question. I'm going to tee you up to do some recruiting probably. But if you were in the shoes of the audience here and just kind of seeking your first job, I'm going to take something that you said, which is like, I can see your passion and excitement for doing the startup thing, right? And, like, you know, kind of longing back for those days is so cool.

What advice would you have for these folks in where to go, whether it's type of company, type of role, industry, or anything else? Way more? That's what I'm saying. It's the easiest. You just tee right up. Yeah, yeah. You know, I'd say find a problem, I mean, we're talking about AI today, but I'd say find a problem that matters.

You know, a problem that matters to the world, problem that matters to you. Chances are it's going to be a hard one. Yeah, many things, you know, we're doing have that property. So don't get discouraged by, you know, the unknown by what others might tell you. And, you know, start building. and then, you know, keep building and don't look back. A huge congratulations on all the progress you guys have made. And as a very happy customer, thank you for building it.

And we really appreciate you being here. All right, that is all for today. If you did make it this far, first of all, thank you. We put a lot of thought into each of these episodes, whether it's guests, the calendar Tetris, the cycles with our amazing editor, Tommy, until the music is just right. So if you like what we put together, consider dropping us a line at rate thispodcast.com slash A16Z.

And let us know what your favorite episode is. It'll make my day, and I'm sure Tommy's too. We'll catch you on the flip side.