Nvidia 和 Uber 支持的86亿美元自动驾驶 AI | Alex Kendall, Wayve

Autonomous driving is all about looking at the AV problem with an AI approach. There is an opportunity there to drive accidents down to near zero.

What we did was raise $1.5 million, get some friends together, rent a house, put our car in the garage, and start hacking away. We took it from expensive retrofit vehicles, which relied on compute, HD maps, and infrastructure, to mass-market vehicles that you can buy or manufacture for $30,000, $40,000, or $50,000 each, with built-in hardware that’s in global supply chains, doesn’t need HD maps, and can drive anywhere.

We’re now the first company to have driven zero-shot in over 500 cities throughout Europe, Asia, and North America. It’s also driven in over 10 different cars, from electric vehicles to vans to SUVs. We went north of the Arctic Circle to test the driving in 22-hour darkness and snow. We were even in Tokyo during a typhoon, when local trains and buses shut down. There were no disengagements—perfect autonomy the whole day. It demonstrated that we truly can generalize to any vehicle anywhere.

I think now the question is: How do you bring it to global scale? That’s the shift that we want to take the industry through, where every vehicle is capable of driverless operation, which is clearly the steady state of where we’re going to go.

All right. Here we are at the Core Weave House at NVIDIA GTC. I just got done interviewing Alex Kendall, who's the CEO of Wave, a company that I've admired for a long time and has done phenomenally well in self-driving. And I think you can actually see him behind me messing with the Core Weave F1 car. So, it was a fun interview. I hope you enjoyed it.

1. The story of Wayve

Well, I’m honored to be here with you, Alex, at GTC. This is the first podcast I’ve recorded anywhere except my basement or the office, so this is pretty exciting. I have to start by asking you to tell me the story of Wayve, because I think it’s an especially interesting one.

Yeah, actually, it’s fun. I just got done giving a speech here at the event, and I put my face up on the big screen in front of a 50,000-person stadium about 5 minutes before Jensen came on. It reminded me that I’ve only spoken in front of 50,000 people twice in my life.

The last time was at a conference called Web Summit back in 2018, where we had this seed-company pitch competition. I got to the final and had to give, I think, a 3-minute pitch in front of the crowd. Long story short, there was a crowd vote that I ended up coming last in. Yet the judges gave the competition to me because they liked what I said.

2. End-to-end deep learning for autonomous vehicles

It just gives you an idea of what we were doing 10 years ago when I started the company: It was completely contrarian. Wayve started with the thesis that autonomous driving is all about looking at the AV problem with an AI approach. It’s a decision-making and complex-reasoning challenge, and we took the approach that the best way to tackle it was with end-to-end learning.

A decade ago, we started working on building a single model that could reason, drive, and generalize at the largest scale possible, without the AV 1.0, or first-generation, approach, which relied on retrofit sensors, compute, HD maps, infrastructure, and really limited scaling.

We started that a decade ago, when this was really laughed at and dismissed by the industry. We’ve been quietly building away, and we’re at the point now where not only is the industry excited about AVs again, but we have an AI model that has driven zero-shot in over 500 cities. This means the model has driven throughout Europe, Asia, and North America. It’s also driven in over 10 different cars, from electric vehicles to vans to SUVs. It’s now emerging as an AI model capable of driving any vehicle anywhere.

We’ve got a sprint ahead of us over the next year to get this deployed in both robotaxis and consumer vehicles you can buy.

3. Driving zero-shot in 500+ cities

I remember having a really amazing experience driving with you in London, where you let me decide where I wanted the model to go. I hadn’t spent much time in London, and I was astonished by how challenging it seemed compared to San Francisco, where people, I think, obey the traffic laws much more thoroughly than in London, as I understand it. There are just many more surprising things coming at you, and I couldn’t believe how well your model worked at the time.

4. Growing up on a farm in New Zealand

But before we get to the quality of your model, I want to go back in time a little bit and ask you about your history. You grew up on a farm, I think, in New Zealand. Is that right? And how do you think that’s affected your leadership or your thinking about autonomous vehicles?

Yeah, in Christchurch, in the South Island. I think there are a few things that were really important to me growing up, on reflection.

I spent half my life building things, whether it was video games, different Lego projects, robotics, or other contraptions. I spent a lot of time messing around with that kind of stuff. The other half was in the mountains: climbing, surfing, mountain biking, and all this kind of stuff.

5. Why the company is called Wayve

I called Wayve “Wayve” because I wanted riding in a car to feel as good as surfing a wave. But I really think that, when it comes to pushing forward frontier technology, it’s all about an adventure. It’s all about doing something where you have to plan, risk-manage, be ambitious, and push boundaries. I think I got a bit of that from those experiences and those values that we can take forward today.

What we’ve been doing is one of those situations where, when we started, this wasn’t possible. But I thought and hoped that, in the future, building this technology would become possible. I think we’ve been lucky that so many things have come together to make that a reality.

I was fortunate to get a scholarship that took me from New Zealand to Cambridge University. Turning up to that place was like arriving at a castle from the Middle Ages, full of the most amazing, brilliant minds, where I could learn and explore. Again, optimizing for adventures has somewhat led me to where we are today.

6. Building the first end-to-end prototype

You’ve got to tell me about building the first end-to-end prototype, because I would imagine that would be particularly hard with the way you’re approaching the problem. You can’t really cheat, right, if you’re trying to do end-to-end machine learning? Tell me about that experience and how you made it work.

In 2017, I had just finished my PhD thesis, and I had been able to build the first deep-learning model for different perception tasks: localization, stereo vision, semantic segmentation, uncertainty estimation, and all of these kinds of systems. They could take high-dimensional, multi-million-dimensional images and turn them into compressed representations of the world, learning end to end.

At the same time, friends down the road at DeepMind had just produced AlphaGo. That’s a system where there are more states of the game of Go than there are atoms in the universe—extraordinarily complex. They showed that, through self-play, you could learn in simulation and learn an agent that could beat the world champion.

I saw these things and thought, “Okay, we can now understand the complexity of the real world through computer vision. If you have unlimited data through a simulator in a very low-dimensional state, you can solve the hardest game there is, Go. Now it’s time to bring these together into the physical world,” which I’ve always been excited about: physical mobility, adventure, and robotics.

7. First lane following with 10 interventions

So I thought we could go try it. We raised $1.5 million, got some friends together, rented a house, put our car in the garage, and started hacking away.

The first thing we tried was on-policy reinforcement learning, where we put the car on a road and said, “Okay, go drive with the reward of driving as far as you can without any human intervention.” Your reward is distance traveled. The car starts driving randomly, and essentially, over a period of months, we developed a Q-learning algorithm that lets you intervene and grab the wheel every time it tries to drive off the road.

8. Waymo's impact on San Francisco

The big breakthrough was the moment when, with just 10 bits of intervention, we got it to learn how to lane-follow. I remember getting that breakthrough on a Saturday or Sunday. I was pushing really hard, and nothing was working. These kinds of algorithms tended to work 1 out of 10 times.

Wait, let me make sure I understand. You only intervened? It had no prior information about the world. It was just looking at a camera. After 10 interventions, it understood that you wanted it to follow lanes, and it could follow lanes?

That’s right. The input was a front camera. There was a GPU onboard doing optimization, and it was essentially trial and error. After a few different trials, it was able to learn a policy.

There’s a video on YouTube that shows this. We muted the audio because the audio is me whooping and cheering as it starts to drive, but it was essentially just able to lane-follow quite robustly after 10 bits of feedback with no prior knowledge of the world.

Were you afraid when you got in the car the first time and it had no prior information? Were you actually on a road?

It was going slowly, so you had time to intervene before it drove off the road.

9. Scaling beyond on-policy learning

That brings up a good question. That method obviously doesn’t scale. You can’t do on-policy tests at scale around the world. That’s pretty unsafe. So we quickly moved to an imitation and offline reinforcement learning paradigm and started scaling up world models so we could learn through dreaming and understand through their experience.

Later in 2018, that’s where we started to take things. The growth there has brought us to this point today, where we can drive any vehicle anywhere, all around the world.

What happened after you got the lane following? What was the next step?

The next step was starting to scale it up. We tried so many different things. We tried injecting priors around computer vision. We tried building world models and learning in dreaming and in imagination states. We tried imitation learning, and we tried sim-to-real learning—transferring from simulation to the real world. We tried everything.

We wrote all these blog posts that ended up being industry firsts in all those different areas. Then we settled on a recipe that seemed to work and scale, and started really pushing. We had a couple of years of building infrastructure, fleet buildout, and actually bringing up this technology to a point where it could start to drive in London.

10. Why London was a great training ground

Retrospectively, learning in London was great because it forced us to build something that could deal with the chaos and complexity of your average London street. You have about 100 pedestrians and cyclists around you at any one time. So it forced us to build something that didn’t rely on mapping infrastructure or anything like that, but could actually drive in a city that’s 2,000 years old.

11. Handling regional driving norms

I remember 10 years ago, when you were getting started, the self-driving vehicle space was super crowded, with lots of different approaches and lots of VC money flowing in. It seems like it’s consolidated at this point, but it also seems like it’s really working. I use Autopilot in my Tesla all the time. I ride Waymos every day. I feel like it’s kind of quietly snuck up and become a technology that, at least here in San Francisco, we use all the time.

I’m curious how you think about the competition at this point. Is it becoming a commodity? Is there something different about what Wayve can do or how you approach this?

12. The autonomous vehicle market then and now

That’s a good question, because I think, broadly speaking, when we started in 2017, that was probably the peak of the hype cycle. Raising $1.5 million to go tackle what was about 10 companies that had just announced billion-dollar ambitions was a bit lopsided. But it’s fair to say we saw a trough of disillusionment, and I think largely that was from the sheer expense—the $100 billion-plus that’s required to build that AV 1.0 paradigm of mapping, rules, and infrastructure, and to get that to maturity.

There was a lot of consolidation there, where really only 1 or 2 companies had the appetite and capital to push through and make that work. At the same time, Tesla FSD and ourselves matured an AV 2.0 approach—a next-generation stack that’s more affordable and scalable—and have now gone from a demo, which we had 5 years ago, to a point where it’s actually ready for a global automotive product.

I think the most exciting thing today is that we’re also seeing huge validation of the product. You and many others are now buying FSD and paying for it, and it’s actually loved by a lot of customers. When you go to Shanghai and San Francisco, places like this, people will pay—even pay more than for a human taxi experience—for a robotaxi, even though it has longer wait times and is limited by a geofence. It’s a product that people really love.

13. Mass market vehicles versus expensive retrofits

So now the question is, how do you bring this to any vehicle anywhere? I think that’s the step change that we’ve built and can offer: taking it from expensive retrofit vehicles to mass-market vehicles that you can buy or manufacture for $30,000, $40,000, or $50,000 each, that have inbuilt hardware in global supply chains, don’t need an HD map, and can drive anywhere. They have the flexibility to be in a fleet that you ride-hail or in a car that you own.

By the way, this ambition and this business model are only available because we’ve built this generalizable AI driver. But I think now the question is, how do you bring it to global scale? That’s the shift that we want to take the industry through.

To me, I’m thrilled that we’re now seeing the conditions in the market come together. People are now paying, and there’s real commercial validation of this as a product. Regulation is in place, not just in the US, but we work really closely with the UK, Europe, and Japan. We co-chair the UN committee on DCAS regulation for autonomous systems, which later this year is going to come through and actually legalize these products.

14. Algorithm improvements versus logistics

Then, of course, manufacturers are now building vehicles at a scale of millions that have centralized compute, surround sensors, and the ability to update over the air and get data off them, which makes a fleet-learning product possible. All of these factors together give me a lot of optimism for going through a massive commercial inflection point in the coming year.

It’s kind of interesting that you didn’t mention improving the algorithm in that list. It’s very hard to judge the quality of an algorithm or the safety of a self-driving car when you’re in it. But I have to say, when I was inside one of your cars—and this was, I think, 2 years ago—it really felt ready for prime time.

I think it felt safer than the Cruise cars that were on the street in San Francisco, which I don’t know if I should trust my own judgment or not. But are there still things to do with the algorithm to get it to the point where you feel good about using it globally, or is that no longer the challenge and you’ve moved on to logistics and regulation and things like that?

Absolutely, there are still things to do. We don’t yet have a driverless service today, but this is no longer the critical path. We’re seeing that the quality of innovation coming from our frontier embodied AI science team, plus the data and compute growth that we have, is just driving compounding scale. Then you layer on deeper integration into vehicles with our partners, and you get sensing and compute on the vehicles. These things compound. Performance is no longer the critical path, but we still have work to do there.

We’ll keep pushing that and pushing it into a generalized state. I think a large challenge is evaluation and validation. How do you prove the level of performance is sufficient to yourselves, to regulators, and to consumers?

Then, of course, there’s commercially deploying and setting up these systems. The last couple of years for us have been a big period of growth for the company: building team strength, partnering with regulators, automotive companies, and fleets in major markets like Germany, Japan, and the US, getting the integration going, and navigating the multi-year sales cycle of automotive. Now it’s going to be the deployment and bring-up of these systems.

I think that’s where the critical path is, but we’re not taking our foot off the accelerator in terms of performance, which clearly has orders of magnitude still to go in terms of the opportunity ahead.

15. Plugging perception into an LLM

One of my favorite demos I’ve ever seen was someone on your team who showed me plugging your perception model into an LLM, so that you could actually see what the car was thinking as it was driving, which seemed surprising and delightful. Was that a toy thing that you realized you didn’t need, or is that something that you’ve continued to invest in?

In 2021 or 2022, way before ChatGPT, one of my team, Vijay, came to me and said, “Hey, Alex, I want to start working on language for driving.” My first reaction was, “No way, man. We’re going to stay focused. We’re never going to product-ship this. Come on, stay focused on driving.”

16. Adding language to the driving model

He laid out the argument that the knowledge you can get from text data is enormous, and it might produce new, interesting products and interaction modalities. By bringing in language, we can improve the representation, the performance, and the product experience. So we said, “Okay, let’s start exploring this.”

Very quickly, we got up a prototype. It wasn’t plugging together an LLM, because LLMs didn’t really exist at that large scale yet. We trained a vision-language-action model—a single model that could see the world, drive a car, and understand language.

It started off as a gimmick. You could drive along and have it explain its driving: “I’m going around a double-parked bus. I’m stopping for a red light.” It would just explain it. You could cherry-pick some really delightful things, but then again, there were also some things where it was just hallucinating and wrong.

17. Learning social cues like flashing lights

From that point, we started to train it up. We actually got it to a point where it became very robust. We saw one interesting thing: if you’re driving and you have an oncoming car, and you want to make a turn across that oncoming car to go into a side street, if that oncoming car stops and flashes its lights at you, it’s a signal for you to go in front, right? Even though you don’t have the right of way.

Our AI wouldn’t do that. It would stop and yield and just wait for that car because it had the right of way.

When we started bringing language pretraining into the VLA model, it actually demonstrated that behavior. When it started flashing its lights, it would now turn. I saw that for the first time, and that was an example of how we could improve the representation and reasoning capabilities by bringing language into the system.

18. Personalization and driving styles

Today, it’s a core part of our training. It gives us a boost in performance, opens up new interpretability capabilities, and, of course, enables product experiences involving personalization and interaction with the vehicle. You can now get in our car and prompt it to drive in different driving styles, and I think that’ll be a neat experience.

When you get into an ADAS or a robotaxi, if that car is driving too conservatively, you’re going to get really frustrated. Or if it’s driving too fast for you, you’re going to get freaked out. Different people have quite a wide variety of expectations there, as do different brands. You can go to some brands that want to be sporty, or some brands that want to be safe and reliable. I’m not saying the sporty ones won’t be safe and reliable, but you know what I mean. Providing that personalization is quite a nice outcome of this work.

It’s a funny moment for self-driving, at least in San Francisco, where we suddenly have Waymos everywhere. I love it. I think it’s an amazing experience, but we’re also witnessing all kinds of crazy traffic issues being caused. People, maybe for fun or maybe maliciously, are messing with them in ways that you wouldn’t have expected. It does seem like that capability might actually be completely critical to making a self-driving car a system people trust.

Yes, although a couple of things. Firstly, our cars, over time, you won’t notice anything different because they’re just normal cars with built-in sensors. I hope that people don’t behave differently around them compared to other traffic.

Secondly, our system is designed to drive in a very human-like way. The way you see it picking up and dropping off in a robotaxi experience in busy cities, or the way you see it nudging through crowds of pedestrians and doing things that are very human-like instead of just being stuck—I think these kinds of things are really going to improve public acceptance.

19. Human-like driving behavior

We design the system to operate in the messy cities we live in and not require infrastructure or behavior change. I think that’s really important for adoption of embodied AI. You need to use mass-market hardware, you need to be able to leverage the availability of off-policy data, and you need to be able to operate as a drop-in to existing infrastructure, not requiring massive capex changes. If you get these things right, then embodied AI is a delightful deployment experience.

20. Why Wayve doesn't build its own cars

How did you make the decision not to build your own car and to license your technology to other car companies?

That was very clear to us from the beginning. I think the mindset I’ve always had, and we’ve always had as a company, is that self-driving is an AI problem. We want to work on the hardest problem first.

We started working on AI, and we started aggressively partnering on cloud, on cars, on infrastructure, on insurance, and working with the best companies around us so that we could focus on the critical-path problem. By always focusing on the hardest problem first, we’re not going to stumble into some glass ceiling that we hit down the road. I think that was the mindset.

Retrospectively, it’s become important because I think it’s a mistake to focus on one vehicle form factor, because it’s really unclear what will have product-market fit. Is it a normal SUV like today? Is it a bidirectional vehicle where you sit facing each other? Is it a 2-seater electric vehicle? Are there other applications that really take off? I think that’s unclear.

At the end of the day, if you focus on one form factor, you’re also going to get a small part of the market. We’ve tried to focus on how we build a generalizable driver that can address everything and then follow where the market is most advanced at any one stage.

We started off with retrofit when there wasn’t really anything for us to integrate into. Then, during COVID, grocery and fleet started skyrocketing in profits, and so we started partnering with that sector. Then things changed. Automotive companies, which would never speak to us and said, “This is unsafe. It would never work,” all of a sudden started talking to us and building software-defined vehicles so we could integrate with them.

Now, today, we’ve got the excitement of robotaxis. We’ve followed the market while being open to integrating into anything. The end state, of course, is consumer vehicles, delivery, robotaxis, trucking, and non-automotive forms of robotics. We want to license our embodied AI and make all of them intelligent, safe, and possible.

21. Stack ranking business opportunities

Where you sit right now as CEO, do you have a stack ranking of those different modalities and which one is most top of mind for you?

Yes. Today, it’s all about consumer vehicles and robotaxis. It’s about point-to-point driving in urban and highway situations with mass-market consumer-vehicle hardware. This is an aligned product that we can now commercialize with great partners. This is the real focus for us today, and it can generate an extraordinary business.

You look at the scale, whether it’s data, revenue, number of vehicles, or exposure: about 100 million cars are produced each year. People forget that. Often, they only think of robotaxis, but there are fewer than 10,000 robotaxis in the world. Getting this natively integrated into consumer vehicles opens up tens of millions of vehicle opportunities, which is going to far outscale robotaxis in the short and medium term.

Before long, of course, every vehicle will be capable of driverless operation, which is clearly the steady state we’re going to reach, given the safety, convenience, and product benefits that it can bring.

What geographies do you plan to launch in first?

We’re really looking to make sure the system is global, but we’re focused today on Europe, including the UK, as well as Japan and North America. We have fleets and offices in London, Stuttgart, Tokyo, and the Bay Area. We’ve driven around most cities throughout Europe and North America, and we’re growing in Japan. Those are the markets we’re starting with, given the customers we work with.

Do you have a sense of where you will go mass market first?

It’s those markets. In general, we’re seeing faster movement in North America and Japan, closely followed by Europe. Of course, China is unfortunately very hard for us to work in, and vice versa. I won’t get into geopolitics, but it’s a bit of a shame that’s where the world’s at.

22. The AI 500 road show results

In China, this technology is moving so quickly, and the rate of innovation is very impressive. I think it’s worth looking at if you haven’t, from an automotive-technology perspective. It’s another proof point that when this technology is deployed, consumers really love it. Alongside price and safety, it’s now a top-three reason why people buy a car in China.

Wow. There’s autonomy. You had amazing results at the AI 500 roadshow last year. I don’t know if you want to talk about it, but you took one single AI model and deployed it all over the place, with many of those geographies not having any prior training data. Were you surprised that it worked so well?

No, because that was the core thesis: by building a general-purpose AI model, we could drive anywhere. But to actually see it was huge validation.

We drove around the world. There were some very cool anecdotes. We drove in places like around the Arc de Triomphe in Paris. Have you driven around there?

I have, yeah. Well, not driven, but I’ve gone there.

It’s chaos, right? We went north of the Arctic Circle in Finland and Sweden and tested driving in 22-hour darkness, snow, and things like that.

We were even in Tokyo during a typhoon, when local trains and buses shut down. That’s actually how we were giving a bunch of media drives. We had a day of journalist drives with no disengagements—perfect autonomy the whole day—in literally the heaviest rain I’ve ever seen.

Our cameras are situated just behind the windshield, including the front camera. The windshield wiper was constantly wiping water off it, but it drove reliably around central Tokyo, which was quite remarkable. There were fewer pedestrians on the road that day, but it was still busy traffic.

We’ve seen all of these things, and it demonstrated that, as the first company to do something of that scale, we truly can generalize to any vehicle anywhere, given that over half of these cities had no prior training data. It was a true zero-shot generalization test at scale.

23. Is this superhuman performance?

That almost sounds like superhuman-level performance.

I’d be careful with that word. You and I, if we jump into a new city, can get a rental car and drive there. But to claim superhuman—I mean, there are some great reports that I’ve read, including one from Swiss Re, where they measured the impact of an autonomous system, Waymo’s operation, over its operational history and found that it had much fewer accidents and safety events than equivalent human drivers.

I think there’s some robust evidence that’s really showing autonomy can be superhuman.

24. Introduction

We have work to do to demonstrate that, but we're making good progress there. Sadly, over 99% of accidents are caused by human error. Even just making a self-driving experience that can never be distracted, drunk, or impaired, that can see 360° all at once and make decisions over 10 times a second, presents an opportunity to drive accidents down to near zero. That's what we want to see, but I think you're going to see scale deployment really show that. We want to go do that in the coming years.

25. Scenarios that still challenge the model

Are there situations that still make you nervous at this moment? What would be the kind of configuration that would put you on the edge of what the model can do?

I think it's very hard to give general statements to that, because where the models today tend to struggle is when you have a culmination of many failure modes. If it's dark, you have high-speed traffic, and you have adversarial behavior from someone else driving on the wrong side of the road or not yielding, maybe combined with bad weather, that's when you have multiple confounding factors. Then maybe you have a sensor failure. When you get multiple confounding factors, that's when you tend to find challenges and failure modes.

We need to keep addressing that with better system resiliency or redundancy. We're building that with some next-generation vehicle architectures. We need more data to improve generalization and risk assessment, and we need to make sure we deploy responsibly in areas where we're confident we meet the required bar for safety. We'll do that responsibly with our partners as we grow.

I think the other nice thing about being deployed in consumer vehicles and robotaxis, starting with supervised trials, is that we can incrementally grow this with communities and regulators to the end state where every vehicle is autonomous at a global scale.

26. The trolley problem in self-driving

I remember there was a moment in ethics in AI, maybe when self-driving was hot, when people were asking questions—maybe toy questions—about whether a car should prioritize the safety of the passenger versus a pedestrian. End-to-end, you're kind of sidestepping those questions, and maybe you're not inserting yourself. Is that true, or do you need to push the data in certain ways to get certain types of ethical behavior that you want?

This is the trolley problem. We've got two bad decisions, and you have to choose bad decision A or bad decision B. There are many ways of framing it. The practical reality is that, as an industry—and I think I speak on behalf of most AV companies—we need to design a system so there's always a third good option.

It's called a minimal risk maneuver. If you're in a bad scenario and you see risk, there's always a third option you can take, which is to minimize risk. That typically involves reducing kinetic energy, maintaining a predictable course of action, and stopping or pulling over the vehicle. We always need to prepare to do that if any systems fail. There needs to be a fail-operational ability to do that. We design these systems to always have a third good choice, rather than being stuck between two bad choices.

27. Minimal risk maneuvers explained

The second point is that so many things have to go wrong to get into that two-bad-choice state. The probability of those kinds of scenarios is so small. Even if you are in that position, the probability that you can accurately sense those two things in those scenarios makes this more of a theoretical thought exercise than an actual practical decision the industry needs to take. We focus—and the industry focuses—on building these minimal risk maneuvers that take us away from those situations.

28. Training data and crash scenarios

Having said that, if you do get into that unlikely situation, then of course the behavior will depend on how you bias it from your training data and the system that you design. We need to be thoughtful with the principles and the knowledge that we distill into the system, and with the behavior that we generate. There are some good principles around managing risk, making the system predictable, and, of course, maximizing safety.

I would imagine that in your end-to-end training, you're not seeing a lot of crashes. If there are crashes, you're probably generating them in the dream data somehow. I think that the types of crashes that you're feeding into your system would give it a certain point of view.

I'm really proud of our safety record, actually. We've had no incidents since we started operating in 2018, so we have a really fantastic safety record around the world.

There are a few things that are important here. Firstly, we don't just learn from the data of our vehicles driving. We have enormous data partnerships that give us quite general-purpose knowledge. Unfortunately, there are still a lot of road accidents today, and we partner with dashcam providers and car companies that get data from their consumer vehicles. Sadly, there are still a lot of quite frightening scenarios that they experience. The good thing is that we can use that data to learn from it.

Secondly, we have our world model, GAIA, which can simulate scenarios that are too rare or unsafe to see in the real world and make sure that we're robust to them through synthetic data and simulation, too. Ultimately, with those strategies, we need to build robustness to those events.

We still see some pretty bizarre stuff on the roads around the world, including some pretty unsafe behavior. We can resimulate, augment, or modify these scenes to build resiliency to them. It's a constant game of fleet learning and data iteration to make sure we're grinding up performance over time.

I remember when Tesla launched what I believe they said was end-to-end dream full self-driving. You really felt it, I think, when that version came out, at least in my Tesla. One of the things that was really notable about it was that it stopped at every stop sign in San Francisco, which San Francisco drivers don't really do. It's kind of annoying if you're really following that rule to a T.

It made me think that the average of everybody's driving is probably not what you want. How do you deal with that in your end-to-end training?

It's a great question. I was just in Japan last week, and there everyone drives about 20% above the speed limit. It opens up an interesting question around policy, regulation, and liability. If you're an eyes-off or driverless system where the system's taking liability for driving, it should meet the road rules. But if you're a consumer, a lot of products today in a driver-assistance setting allow you to set the speed above the speed limit.

It's a really interesting question. I hope that self-driving improves road safety to the point where we can actually increase speed limits because it's safe enough to do so. But we're probably some distance away from being able to have those kinds of arguments.

29. World models versus policy learning

Today, the first thing to realize is that when you're training a world model, you want to see as diverse data as possible. You want to see good driving and bad driving. You want to really understand the full spectrum of dynamic events in the world.

Then, when it comes to actual policy learning of how you drive, there's a data distribution that matches whichever customer's preferences. The OEM driving style and behavior, as well as the rules and policies of the country they're operating in, become really important. We can separate those two data distributions. The first one should be as diverse and as large as possible. The second one could be highly curated and small, whether it's RLHF-style feedback or expert demonstrations. There are many ways to do it to make sure that you drive in a safe and considered way for that application. You can separate these data distributions.

30. Transitioning from CTO to CEO

Actually, one thing I wanted to ask you, somewhat as a fellow entrepreneur, is that you started, I think, as CTO with another co-founder who then left, and you became CEO. I wonder what that experience was like. I guess it was earlier in the journey, but if you want to share about that.

Yeah, it was early on. At that stage, I'm sure you remember, everyone does a bit of everything. As we were getting started, for a number of reasons that unfortunately made sense for my co-founder, he left the company to pursue other things. We're still in good touch today.

The transition was my first real experience of change-management communications for the company. I think we were about 25 or 30 people at the time and had just raised Series A. It has matched my experience over the last decade, where every year has been a completely new challenge.

I started the code repo and was coding a lot of the first demos. Then I started working on hiring the team, finding commercial partners, and growing investments. Now I'm managing customers and managing through an executive team, where I'm so thrilled to be working with people who are truly at the top of their game. They're industry legends, whether it's in science, engineering, product, finance, or commercial. Across the board, we've got truly world-class leaders.

I think it's probably a few things: the importance of building complementary teams, working with people that you can learn from, and also getting used to outgrowing a skill pretty quickly and having a new challenge. As soon as I get good at something, it's no longer the relevant skill for me to be working on because it's usually me plugging a gap for the organization and then growing the organization to absorb it. Then there's a new challenge.

I enjoy that, though. I think that's part of the goal. It's the whole thing—it's part of the adventure of building a company. I think the next year for me is now this new phase, where I'm starting to live on a plane and work with global customers. We'll be 10 years in, and it'll be our first product launch.

31. Advice to your younger self

We're going to start to have major public road exposure through that, while keeping up strong and consistent execution velocity. This is a completely new challenge that I'm really quite excited for. Let's see what comes after that.

If you could send a 30-second message of advice back to yourself when you were just starting out, what would you put in there?

Oh, man. I feel like if I built Wayve again, I don't know.

If you built things again, how much faster do you think you could build them the second time around?

I would totally say it's like 2 or 3×, right?

Yeah.

I took so many shortcuts, did things that I was hesitant to do, and made mistakes that I think I could just take the tactical playbook from.

You had to distill it to a higher level because you only have a 30-second window—not just to give the loss-function recipe and relationships and partnerships, but just really tactical things.

I always thought the interesting question was: I feel like we'd be a bit useless if we went back in time 200 years with today's knowledge. What would you do? I don't know what you've learned through Weights & Biases, but you've built on top of stuff that just didn't exist 200 years ago. We're standing on the shoulders of giants.

There's a bit about timing. In many ways, if I developed our product 3 times as fast, we may have been ahead of the industry, when automotive didn't have the right infrastructure or products to actually accept our product. I think timing is everything. In many ways, self-driving has been a question of staying alive and keeping the company capitalized while making progress on our next-generation approach.

If I were to distill it down, I've always had conviction in this approach. It's not like I'd say, "Keep going, Alex," because I've been kind of resilient toward that one. I've been fine with how bold and brave I've been in building this.

Honestly, the biggest thing has been—two things actually genuinely come to mind. One is talent growth at Wayve. I've seen so many great, fantastic, brilliant people and the growth of the expertise we have in the company. I think moving faster and harder, and starting out of a PhD and growing from London compared to the scale of the amazing people we can work with today, there's a great opportunity there that I could have leaned more into.

So, do you mean like try to get the experts earlier?

32. Algorithms versus infrastructure

Yeah, yeah, I think so. You know, we started off with a generalist mindset really trying to figure things out and had to build a lot ourselves, but I think maybe it's a mixed feeling on that one. The bigger thing that comes to mind for me is that I over-indexed on algorithmic innovation compared to infrastructure. The big learning I have with deep learning is that it's 1% algorithms and 99% infrastructure. Whether it's the cleanliness of your data, the reliability and iteration speed of your learning loop—we've been very happy adopters of Weights & Biases over the years, and the acceleration you get with it—or introspection, measurement, and evaluation tools, I think we've been on a journey there and underinvested in it.

I think going harder and faster there, rather than pursuing shiny, sexy algorithmic innovation, is something I wish we'd done earlier. The thing is, we were held back by not having a robust platform for many years. As soon as we had a robust robot platform, it was like giving our AI, which was blind, glasses. All of a sudden, it could see and drive in a remarkable way.

33. Evaluation and simulation challenges

Focusing on that earlier and harder, and building that maturity, was quite key to bringing up performance. We were too late in that. In simulation and evaluation, I think both large language models and embodied AI present a real challenge in evaluating the performance of models. It's an open-ended problem. It's not like you can have a clear metric around that today.

Developing these techniques has been a journey. We've consistently tried different things, from generative models to procedural game engines, to NeRF and Gaussian splatting, to now generative world models. Throughout those times, we haven't been afraid to forego sunk costs. We've spent millions of dollars and months of time building something, and then just deleted it and moved on because we found something better.

That culture and that journey have been important. Today, I think we have something on the cusp of a major breakthrough there that will really transform the industry. But that's another example where getting that right earlier would have helped. It's chicken and egg: if you figure out how to measure it, then you can just optimize against that target. Getting that right is something we should have moved faster on.

34. Taking Bill Gates for fish and chips

Okay, some quick questions to end with. When you took Bill Gates for fish and chips in your self-driving car, were you nervous that something would happen? What did you talk about?

Yeah, that was one of the first times we'd gotten the system reliable enough. Actually, I think there were a few interventions on that drive, so I think Bill was still excited at the time.

We talked about how, when he was commercializing Microsoft and Windows, he had to get a bunch of design wins with PC OEMs. In a similar way, we've had to get design wins with the biggest car companies in the world. There are some interestingly similar dynamics there that he was interested in.

He was in London for a day when we went to get fish and chips. One of his team members went in to get the fish and chips, handed it through the window, got in the car, and Bill just put it down and sat there. He wasn't going to eat it. I said, "Come on, Bill, we've got to eat it. We can't not eat the fish and chips."

He said, "All right, then." He grabbed a piece of fish in his hand, brought it up, broke it in half, and just started eating this battered piece of fish in the back seat of the car with me, which I loved. He clearly has a soft spot for some greasy fish and chips.

35. Parties at the first prototype house

Yeah, nice. Nice. All right, what about when you built the first prototype in your rented house and said you threw one of the biggest parties of your life? What was that like?

We invited everyone we knew in Cambridge to the house. The living room was where we worked with a bunch of computers. The small bedroom was our boardroom. We had another bedroom full of servers, and it was constantly 40°C because of the 60 GPUs that were running there.

Then we just turned the house into a bit of a party house. We had a few musicians on the team, and we were jamming. That was a really fun vibe. Those days are very special, and I really enjoyed the lasting memories of a team that was really in the trenches together during that time.

But that was quite a big one.

I don't know about you, but I feel like I'm always chasing that feeling of the early days, when everyone's cranking together. Such a good feeling.

I just visited our Japanese office last week, which is about 20–30 people. It really feels like that. That's super exciting.

I think there are pros and cons to operating at a different level of scale. Today, we're about 1,000 people globally, but the culture you set at that scale really drives through to where we are today. I think it does require different skills operating at that scale, but the impact you can generate is just tremendous.

36. Wrap-up

Holding on to disruptive innovation and not falling for the innovator's dilemma—these principles are still things we really fight for, and I think they still hold true today at Wayve.

Well, you've had some real success. It's been a joy watching you guys grow.

Chris Sacca. Who said it? Chris Sacca. Thanks so much for listening to this episode of Greymatter Decent. Please stay tuned for future episodes.