The Self-Driving Startup Nobody Saw Coming | E2289

We pioneered end to-end learning when it was widely dismissed. — Self-driving in a way that economically scales the world is not solved. — Our partnership is not up to 25,000 is over 25,000 or in other words a minimum of 25,000. — Our volume is like double the cars Tesla builds a year. And that's just one of our partners. And if you're a manufacturer selling a car that doesn't have this, I think your demand is really going to fall off a cliff. — Every car is being intelligently driven by a machine that never blinks. You know, you'll pay for your own private chauffeur that that's in your car. — Has Uber tried to buy you? — W is not for sale for anybody.

— Hello everybody and welcome back to Twist. My name is Alex and today we're going deep on one of my absolute favorite topics in the world. And no, it's not about open clock. No, today we're talking about self-driving cars. We're bringing back the CEO of a company that we had on the show back in late 2024 when Wave, a UK-based self-driving startup, was doing incredibly interesting things, working hard to bring this technology to market. Since then, quite a lot has happened. We're going to dive into what Wave has done recently, how close it is to changing your life and my life. So, please join me in welcoming back to the show its co-founder and CEO, Alex Kendall. Alex, how you doing? — Awesome. Hey, Alex. — It's so good to have you back. So late 2024 feels like 29 years ago in AI terms. Has the self-driving world been progressing as quickly as the kind of general AI landscape? — Well, you know, if I go back to when we started in 2017, um, one of our very first blog posts was about a world model that we put together back then. It was, I don't know, not in today's standards, it was like a 20,000 parameter world model. Uh, and we were all excited at the time of end to end AI. Hey, it was going to actually um you know allow us to really truly scale autonomy and that pitch has stayed the same for the last decade. But uh it feels like the whole industry is really uh getting behind what we're doing now because um this has been a contrarian approach for so many years. But in the last uh you know last few months we've brought in investment from um Nvidia uh Qualcomm ARM AMD all the big chip companies and then Uber Nissan Mercedes Salantis uh Microsoft and it just feels like the industry is now believing that this once contrarian approach has the legs to go scale things for the industry. Uh that's it's a big privilege. — Yeah. So you guys wrote we pioneered end to-end learning when it was widely dismissed. We built world models years

before they became fashionable. We prioritized generalization across many environments over driverless optimization in single domain etc etc early and it seems to be correct but since we had you on uh you guys have released I think two new world models uh Gaia 2 and three so I know it's a little bit basic but could you tell folks who are behind what a world model is in this context and then I'm really curious what improved between the generations of the world models that wave uses to power self-driving. A world model is a I mean it's a at the basic core principle it's a it's a model that can understand uh the state of the world a given action you take on the world and how the world evolves and so what that lets you do is I mean first of all it's a really powerful representation learning method it lets you uh you know learn a representation of the world that actually cares about what matters so if you're driving a car you don't care about the clouds in the sky or the cars going the other way behind you care about the road lines the curbs traffic signals in front of you and anything that might intersect with you. And so by learning um how to predict the world, you actually uh cause your machine learning model to represent what actually matters in the world in an unsupervised way. So it's firstly, it's a really powerful representation learning method. And then secondly, uh it gives you benefits of it can be a simulator. It can actually allow you to simulate what's happening in the world um to to learn or to validate or to actually control what's in front of you. And as far as I understand it, I'm going to put this in super basic idiot terms, but it's kind of a video game for your self-driving technology to play it. It creates a world with obstacles, traffic, weather, locations, rules, like which side of the road do you drive on, and then you can create essentially an infinite number of uh testing variants, and then you can put your driver into this world, this generated world, and essentially do infinite miles in a virtual setting. that would take um lots more time and money in the real world to do without the safety implications. — Yeah, that's right. I mean, we have an analogy in our own minds, right? In our hypoc campus, we have world models that actually, you know, when we daydream or sleep, you know, we replay experiences, you know, a gazillion times to actually uh reinforce how we, you know, how we act, how we learn to swing a tennis racket or, you know, do any motor task that we have. So, we do the same thing, but it's a lot more than that, right? It's a it's a representation a really rich representation of the world. Um but yes, one of the best uses is a simulator and we know in robotics and self-driving. Um it's not like uh you know a chatbot or something where you got large scale text on the internet, but getting the data and in particular getting the safety critical data and then proving a system is safe is the hardest problem. And it's an arms race in our industry between learning a driving policy and learning a simulator. uh if you have one, you've solved the other and you solved the problem. But the arms race between them, we find that um for simulation, end toend learning is not only the best approach in the world for learning driving policies, but it's also the best approach at learning to simulate because what building a world model with an end toend uh deep learning model, what that allows you to do is it allows you to use data to model very complex and diverse scenes. Uh it lets you learn very rich dynamics. So to answer your question, what's evolved? I mean yes of course we've scaled up the parameter count uh the data sets it's now at frontier scale for uh for the robotics industry but our world model learns from this is the advantage we have in self-driving is we have hundreds of pabytes of data across um uh everything from internet scale data to dash cams to the automakers that we partner with. We've got over a dozen different companies now sharing data with us that we aggregate at scale and to train this this world model. Um so what's changed? So we've scaled up data and compute uh parameter count but then we've also improved a number of things algorithmically. Uh so it's not only uh video but also understands radar and LAR. Uh it understands multiple sensors. So um a typical self-driving car might have a dozen or so cameras, might have you know five, six, seven or how many other radars. So it can understand all of these. — Um and then on top of that it's controllable. So we can actually, you know, prompt or control it or reimulate something we've seen in the real world or adversarily test something and try and um, you know, try and make our car learn or make mistakes in the world model so we can learn from that. — And when you talk about different sensors and different self-driving cars and what they have equipped to them to me there is a buffet of options you can have in your car. I presume that

uh you know your AI driver can work with what it's offered. So if it has LAR and not radar, radar not LAR, visual blah blah. Uh it can take in I presume any type of information and use that to make its decisions. Is there a minimum level of ingestion required here? — Yeah, that's a great question. I think the sensor debate is often a very heated one in the industry, but really probably not the there's more nuance to it than what might be seen. But at the core of what we do at Wave, we want to be the intelligence layer across any vehicle anywhere. And there's going to be some products that benefit from being camera only, some with radar, some with LAR. And so we support them all. Um, now this is, I think, very natural to do with our approach because our model trains on very diverse data, sensors in different locations, different types. And we can learn to understand which signals to represent and also which signals we can rely on and what a sensor architecture can or can't see because you can do that through a world model. I mentioned it's a really powerful representation. When you learn to predict the future, if your sensor can't see part of the scene, it can't predict the future in that way. So, you can learn this very naturally. — So, you're not just simulating, show me with rain, show me with snow. You're also simulating, okay, I'm in a smaller car with this sensor array in this weather environment. So, you can get super granular then inside your world models. — Exactly, Alex. If you're in fog with camera only, you might struggle. If you got a radar, you might you can predict different things. But um to answer your question, yes, there is a there is a minimum bar of safety you need for a hands-off, eyes off or driverless system. Um now you can achieve all levels with a camera only system if you're really good enough, but it might be faster and more efficient to get there with some radar or other sensing modalities. So what we find in the industry today is that um most of our partners who are building say robo taxis uh it's better to work with camera radar LAR but crucially these are not bespoke you know custom spinning lightars on the vehicle these are automotive grade mass market lowcost uh sensing devices so there's the there's I guess there's a difference there. — Absolutely. Now you've had two new world models come out. You've also raised a an enormous amount of money recently um$1. 2 2 $1. 5 billion depending on kind of how you count in tranches and so forth. Uh my read of that following your

technological progress is that people are very impressed and you have cracked self-driving. I feel like we've gotten to the point where we can say we've figured it out. Is that fair or am I a little bit ahead of the uh the curve here in that pronouncement?

— Big cloud providers may offer you cheap compute, but you'll end up paying the difference in engineering costs and hiring extra developers. You don't want to waste time configuring virtual networks. None of us do. or your access policies. You want your team building your product. So, it's time to look at Render. Render is the all-in-one cloud platform for developers that allows you to deploy, scale, and secure your apps and agents with zero ops. Most cloud platforms ask you to split your focus between product and infrastructure or they force you into platform constraints. You know you're going to outrow in six months. But just connect your GitHub repo to Render and you are live. Web services, cron jobs, the whole stack in one platform. It's time to find out why 5 million developers are already using Render. 5 million. Go to render. com/twist and apply for the render startup program. You'll get anywhere from $500 to $100,000 in free credits depending on your stage and who your backers are. That's render. com/twist. — Oh man. uh self-driving is I think not only the hardest problem but it's going to be a continue to open problem for some time. I think the key thing to realize and despite if you live in Silicon Valley or Shanghai, despite what you see on the roads every day, um self-driving in a way that economically scales the world is not solved. And I think that what we bring is an approach that has demonstrated a path to that solution. And now we're entering an you know a integration and product deployment phase. So what we're going to see with this capital is um I mentioned you know our mission is to bring intelligence to any vehicle anywhere and so we're going to see that start to be deployed uh this year in supervised robo taxi trials starting at London, Tokyo and 10 other cities on Uber and from next year uh in consumer vehicles uh you know we're supported by partners like uh Nissan, Mercedes and Stalantis. Take Nissan for example. Last year they announced they're bringing us into their consumer vehicle lineup. Um uh then earlier this year we announced a robo taxi uh because what we find is automakers they want to work with the same partner across L2, L3 and L4. It really helps speed and efficiency and you can leverage data and integration — and the wave system can do uh different steps of the L1 2 3 4 5 ladder. So you can approach this kind of like uh whatever they need you can offer. — Exactly. And then two weeks ago, Nissan announced that they are going to bring this technology, bring our approach to 90% of their vehicles. You know, they build about 3 million cars a year. So, this is — that's 2. 7 million. — Yeah, this is a an enormous um volume. It's like double the cars Tesla builds a year. And that's just one of our partners. And so um you know, we're really excited about this. and this this business model I mentioned how we had a contrarian technical strategy but there's also a contrarian business model because um — there's three ways to bring autonomy to market right you can build your own cars that's what Tesla's doing but then you're limited to just your own brand — um you could build your own fleet city by city that's what Whimo is doing but it's a very expensive high capex endeavor or what we're doing is we're licensing this to any fleet or automaker and that's I think the largest uh business model that's why we've chosen it it's only possible because we built a flexible and generalizable um like AI driver. And so I think this is also interesting how it's uh enabling a different business model that might not you know might not be appreciated at first thought. — We're going to get to that in just a second. But my question of have we cracked self-driving? You answered in a very interesting way and I was being slightly puckish by asking it in that way. But I was curious, you know, with all the technological progress we've made, are we there? And then you said no because we haven't sorted out the economics of bringing this to the world yet. Those are different points. So I guess the question, Alex, is has wave gotten so good at generalized self-driving now that we're only left with the economic and manufacturing questions for bringing self-driving to mass market cars in the next 18 months? Or are there still technical is there still science risk, I suppose, or are we only talking about market risk? I — I'll give a nuance answer here. And I think um what I try to you know appreciate in self-driving is firstly um to let our results do the talking and not sort of add undue hype and and try to bring a bit of uh technical realism. I think these principles have served way of well over the years. But I think um so I think we're through the scientific risk. Uh certainly so let's start with different levels of autonomy. So for um for hands-off driving uh I think that we've now shown that like last year we drove in 500 cities around the world u Tesla system scale. So Wave and Tesla built this the end to end stack and we've both shown that this scales globally. Um we've got the level of performance needed for a delightful product. People are willing to pay for it. You can see the amazing Tesla announcing a one and a half billion of revenue a year with this. uh clearly there's product market fit with that kind of uh product and the technology is performant to do that. Now what is it going to take to get this from uh hands off to eyes off or drive or drive list that basically the same level of safety um for L3 or L4 a pointto-point system um there is a gap in performance from the systems that say Tesla ourselves have today to get to general purpose driverless you know what Whimo has demonstrated in the geoence areas they operate in to do that at a global scale a way that economically scales with mass market hardware no geoence to be able to do that. There is a gap there. But what I'm seeing is in front of us a very clear path to go do it. And so I'd argue we've moved on from the scientific risk and now it's engineering execution risk and like product integration and deployment uh risk ahead of us. Namely, uh what we need to do is we need to integrate this into vehicles that have the right infrastructure for these products. We've got programs underway with some of the biggest manufacturers. So that's underway from an engineering perspective. we need to scale up um the AI model to reach that level of performance and I think that's a very predictable scaling curve a little bit like the what we saw in the LLM scaling journeys but that's a case of data compute um uh some algorithmic innovation along the way but I think that's a predictable curve that we need to go run up and then third of course um to be able to validate it again that's an engineer activity we know how to do it it's a case of now um scaling the validation activities across the domain to prove that this is safer than a safe and competent human driver before we launch. We work through those three things. Then of course getting regulatory sign off will allow the launch of these products. Even on the regulation piece, the amazing thing is that we've seen regulators put regulation in place ahead of the products being ready. And I think that's um quite amazing to see. You know, of course, the US in some states they allow it, some they don't, but there's a market there for it. outside the US, the UN uh two months ago. So we co-chair the industry committee for UN autonomy regulations and um the UN just legal put in place a legal pathway for L3 and L4 driving uh which and that covers basically every country except the US and China. So there is a now a legal path to getting this deployed as well. So all in all I think we moved from science risk and now it's an engineering and deployment risk. So to get us from hands off to eyes off the path from here to there uh in the AI sense in the technical sense is solvable. We know how to do that. It's data compute and algorithmic innovation. Uh and if you're curious what we mean by that just go read a paper from a major LLM lab talking about their latest model and how they change the back functions of it to see how it make it better. — There are some differences from LLMs though, right? like we there's the um challenge of the real-time embodied inference you've got to do on board the vehicle that's much more constrained. There's a safety critical challenge. Um there's the different modalities. You got much larger dimensional data. Um and then you've got to build a system that's safety aware and uncertainty aware because if you put out a you can't hallucinate for a self-driving car. No, there's a lot of differences there, right? — Not no for sure. Uh but this actually brings me a question that I wanted to ask about the business model here because I love taking the third approach working with manufacturers who are already good at making lots of cars or working with demand providers like Uber who already have a lot of people. It just makes a lot of sense to me to take the technology to where there's already aggregated pools of demand. That makes good sense. But let's look forward a couple years. I'm going to go buy a new Nissan. Uh I have the option to get Wave built in. I click all the boxes. I would like L4, please. I don't want to even touch this steering wheel. put it away. I just want to sit in the back and sleep because I'm a terrible driver. Let's be honest. Um, how do I, the consumer, pay for that? Do I pay a a fee to Nissan for the technology? Let's say it's a 5K add-on. Making up numbers here, not holding you to it. Or do I pay them some and then you guys some? Because to me, the compute side of this can't be entirely local to the card. There probably is some data exchange, some inference costs. So, to me, there's it seems like it'd be something that I should pay you for on a regular basis. So, it's good and it gets improved, but I'm not sure that's the plan. — Yeah, I think the journey the industry is on a journey of figuring that out. So, um uh for consumer vehicles, you'll pay the uh the manufacturer who will then pass through economics to wave, but — there's different models that are being played out. Some manufacturers are looking to uh bundle this with a car and actually include it for free with all the cars they sell for a given model. Some are uh you know, it's like a seat belt. It's a feature that you should expect. Others are looking to have a onetime fee. recurring subscription. Some a bit of both. Some may be a free trial and then after a trial then you subscribe to it. Of course, famously Tesla charges $100 a month for these features. Others have got um lower levels of subscription. So I think there's a bit of a price exploration that's going to be done in the industry. But I think uh it's likely that we will see the industry move to a subscription model because as you say all the intelligence will run on the edge on the car but there are going to be um you know there's going to be a improving performance over time with over their updates and of course for L3 or L4 driving there'll be some ongoing insurance costs and things like this for the manufacturer to bear. So, all in all, I do expect we will get to a subscription model for vehicles and you'll um you know, you'll pay for your own private chauffeur that that's in your car.

— AI tools are making it easier than ever to run your own business, even as a solo founder. But you still need a beautiful attention-grabbing website to help your new company stand out in a very crowded field. And you don't want AI slop. Nope. You want to use Squarespace. That's the easiest and fastest way to turn your idea into a real business because the team at Squarespace cares deeply about design and functionality. And a plain looking or generic or AI slop website, man, that's going to be a red flag for your customers, for your investors, and people who want to come work for you and join your team. But Squarespace will take all the guesswork out of designing your first website with the blueprint AI builder, which has been finely tuned to make beautiful websites. Squarespace isn't just going to help you make a new web page, either. They're going to be your all-in-one platform for launching your business. They're also going to help you set up your email. They're going to handle invoicing, paperwork, all your needs. Go to squarespace. com/twist for a free trial. And when you're ready to launch, go to squarespace. com/twist for 10% off your first website or domain purchase. — You just said something incredibly interesting that sounded very boring, but actually I think it's very important. Uh, insurance costs held by the OEM, in this case, the car manufacturer. I thought that was also a pretty unsettled question as to who's liable for Okay, let's just be honest. A lot of people in the world, a lot of cars, even self-driving cars are going to hit people sometimes, fewer. It's going to be safer, but it's going to happen. So, do you expect that in our example, Nissan would hold the liability there for selling the system, the consumer for using the system or you for coding it? Oh, look this depends on uh both the level of autonomy, the regulatory environment and of course the commercial contracting between all the parties that bring together the product. So there's a lot of factors at play but at a very high level a hands-off system um uh if it's implemented correctly uh as the driver you should remain liable and then an eyes off or a driver system the manufacturer or operator will remain reliable with uh with some insured and contracted liability that flows through to the various parts of the ecosystem. So it really depends. Are you going to build that financial back backing that infrastructure we need to handle the insurance element of this or is that going to be handled by like Chub or like Berkshire Hathway? — We're staying focused. We're not going to build an insurance product, Alex, but uh if that changes I I'll let you know. — I mean I kind of feel like it should be like an add-on offering to what you're selling. All right. Uh let's talk about how big this market is. Um a lot of cars sold every year. Enormous market. Key economic engine ironically. um for the world. How many cars do you think in what percentage of man newly manufactured cars in 5 years do you think are are made with uh the capacity to work with either wave or a similar product uh in 5 years? — So today there are about 100 million vehicles produced each year you know um uh call it 50 60 million uh consumer cars. Um today I think the number for advanced ADAS is what the industry calls it is about 15%. But most of this is like highway lane keep assist or some very rudimentary systems. So the penetration of um call it uh you know outside of China the full self-driving experience is just really Tesla and that's a very small fraction of the market. Uh so this is going to go from like nothing to everything over the next few years. What we're seeing is that today luxury manufacturers are bringing in um the right level of compute on the cars Nvidia or Qualcomm or something like that GPU and surround sensing and uh we're seeing more volume manufacturers like Nissan just announced that uh you know they're going to bring this kind of technology to the vehicles from uh from financial year 2027. So over you said 5 years. I think by 5 years we're going to see uh this level of hardware in a very um significant portion of the market and we'll see this improving experience over time. We'll start to see the introduction in the uh more premium end also of eyesoft technology um uh and and even driverless consumer driverless technology and I think we'll continue to see that flow down but the steady state is that every vehicle is going to be capable of that. I mean when you can for you know a very low monthly subscription get a eyes off driving experience I think this is going to completely change things uh and actually you know yes robo taxis are also transformational but when you think

about the scale there's less than 10,000 robo taxis in the world today but 100 million new cars a year and so the scale of impact you can have through consumer vehicles is enormous um I think the advantage that waves brings because we work on both robo taxis and consumer cars means that um Firstly, the data we get from consumer cars will give us what we need to build general purpose robo taxis. Uh secondly, um the manufacturing relationships with the OEMs is really important because an OEM really wants to focus on volume and the only way they can have a business case to work on a robo taxi is if they can have a single partner that work across um across the spectrum of autonomy. So uh you know for these reasons I think um this will give us the ability to have uh native relationships where it's a vehicle built for as a robo taxi with us just as a software integration. It gives us this high margin software business coming across the spectrum of autonomy gives us the data gives us the global supply chain and geography scale. Um, and so I think for all these reasons, it's a it's a very important opportunity for us um that often gone goes unnoticed. But we're going to see this complete transformation of the consumer vehicle market with our AI in the coming years. And to answer your question in five years, I think uh you know what if you're selling if you're a manufacturer selling a car that doesn't have this, I think your demand is really going to fall off a cliff. — Yeah. Apart from probably the most basic you know like um Tata Nano style cars like whatever is like you know very simple probably not everything else — but actually even regulatory requirements require every car to be sold today to have active braking systems and over time autonomy will be so important for road safety that even the most basic cars like you say I think will still have this technology. Um otherwise it's it's a moral imperative because of road safety. — I can't wait. It's gonna be so when my kids can just like walk out of our house and walk down or across the street and I know that every car is being intelligently driven by a machine that never blinks. It's going to be so much better than the Yahoos who drive around my house currently at like 800 miles an hour at night. It's like it's residential, dude. Break it down. All right, so 100 million cars a year. Going back to the Tesla example, 100 bucks a month is 1,200 bucks a year. Call it a,000 for safety. uh 100 million times a,000 is hundred billion. So clearly we're talking about a staggeringly large market. Do you need more capital to unlock it or does the recent billion dollar plus raise give Wave enough runway to get all the way into production with an OEM and early volume? — Oh, we're in an awesome financial position. We've got over two billion of capital right now. Amazing set of shareholders I mentioned earlier. um and all the capital we need to go get this deployed and bring the business to a free cash flow positive and escape velocity. So um you know the indust these contracts we're signing are decadel long relationships with automakers and so um I think the great thing is we can give them the confidence of the security that you know we may not need to raise to get to that escape velocity. Of course, I wouldn't rule out any further raises because there's always uh opportunities to accelerate and grow into other verticals over time, but for now, yes, we've got everything we need to go run at this opportunity and uh you feel the energy in our team. Uh now it's it's such a privilege now get to go and build and deploy these products. — All right, one last question before I let you go. This has been tremendous. I love learning things. Um I was talking to Wabby. They've done something interesting. They started in the world of self-driving trucks, you know, 18-wheelers, and they've been moving towards cars. Now, today we've been talking about cars in various formats, be they interrupt a taxi fleet or AVOM.

Do you think that the technology that Wave has built with its world models is transferable to large commercial trucks and other forms of Earth movers and, you know, construction equipment down the road? or is that an entirely different data set and therefore a different training question? — Oh, 100% it is. Uh I've got a lot of data points I can share on this actually, but um we started with the hardest application, consumer vehicles, because it would force us to build the most scalable technology. Consumer vehicles is the hardest because you've got to run on hundreds of dollars of hardware. You've got to work literally everywhere. Um and you've got to deal with I mean Nissan has 60 different car lines. You got to deal with really diverse set of products. Um and so — 60 — that's a lot — and even starting to learn in London right it's like one of the hardest uh environments to drive in so you know we've tackled the hardest problem first is in our history to really build something that's scalable but this is a stack that will work with any robotics application um we've done some proof of concepts in areas like uh sidewalk delivery um trucking mining uh warehouse logistics I mean all of these kind of applications. What we find is with a small amount of data uh put into our foundation model, we can learn behaviors in these domains as well. Um even our simulator, our GIA, we can um adapt Gaia to these domains. So, it's a small amount of data, but then the driving policy, the reinforcement learning, and the simulation stack, they all transfer with uh with data. When you think of

middle-aged guys who remain really vibrant, feral, healthy, couple of names come to mind, right? David Beckham and Jason Caliganis. That's right. No coincidence that both of us are drinking I am a delicious daily ultimate essentials which has 16 different supplements all in one delicious drink. This thing's got 92 nutrient-rich ingredients including all the vitamins, all the minerals and adaptogens pre and postbiotics. Since I've started drinking this, I feel more alert. I feel less fatigued in the afternoon and hey, calms my digestion down as well. This stuff tastes great, makes you feel great, and I'm a huge fan. So, start feeling like your best self everyday with I am8. Go to I'm8health. com/twist and use the code twist to get a free welcome kit. Five free travel satchets and 10% off your order. These statements have not been evaluated by the FDA. This product is not intended to diagnose, treat, cure, or prevent any diseases. What I would say though is that 5 years ago, we had a we had an end to learning demo. Lots of people are getting excited about end to end AI for driving. Now we had that 5 years ago. Um what we spent the last 5 years building is learning how to make this safety qualified compliant for the automotive industry. uh and to what it takes to actually make this um you know safe and uh validatable and actually you know runnable in an embedded environment that's an enormous amount of product work and to be able to do that in Germany Tokyo Detroit and all the major automotive centers um like that's really where the challenge is and so I think this expertise is going to scale very nicely and automotive will be the best launch pad for us because we want to become the intelligence layer across every robotic vertical there is automotive first — if you have the world models and you have the simulation experience and you can get your hands on the data. What can you not automate that has wheels? Is there any limit to this or is it just a question of data and then investing the time to bring it to market? — Yeah, I think that's right. I um and the other contrarian view I have a lot of people getting excited about manipulation robotics today, but I think mobility is going to become so far before manipulation. It was interesting when I used to go to robotics conferences in my PhD, they used to divide the hall up into like mobility go this way, manipulation go this way. They're two very different communities. They're going to be the same AI over time, but in mobility, there's a tech stack and platforms and an automotive, you got millions of cars being built. They're so far ahead of manipulation. And so I think we're going to see the, you know, Nvidia compute, the sensors, the software defined vehicles. you know, you can go put that on pick your any vehicle you run from like luggage carousels in airports to um totes in a in a warehouse to some room in your um uh you know in your house and so I think we can scale mobility quite well and then manipulation look there needs to be platforms at scale data and I think uh um I think yes we'll be able to adapt in a few shot setting with the data we get from mobility um but I think manipulation will probably come second. Wow. Well, that's uh an incredibly bullish thing to leave on. I'm really excited about it. And I do want to just want to say thank you uh for agreeing that we can't have Whimo, Wabby, and Wave all starting with the same letter and you're going to work on getting some other letters introduced to the self-driving world. Alex, an absolute treat. Um where can people find your company on the internet? And then also, is there a job you're hiring for that you want to shout out into the void in case someone listening is the right candidate for you? — Yeah, thanks Alex. Uh so I mean we're on all the social platforms of the internet. to search Wave W A Y V. Um if you want to come for a ride with us, uh we have our fleets um you know in London, Tokyo, Bay Area, you'll be able to call it on the Uber app soon. Um so come check it out. Uh come give the technology a go, see what it's like or buy uh one of our cars from next year uh with partners like Nissan. So that's how you can really get get — that's such a flex. Buy one of our cars next year. — Yeah. — How does it feel to finally be here, man? You've been working on this for a long time. Well, it's been a decade, but still still not there yet. So, let's uh — I but you can almost taste it like you're starting to like you're I'm starting to slowly reach for my my credit card, you know, and that's a different feeling than a few years ago. It's exciting. Feels good. — It's No, it's awesome. I mean tell you what's been the most incredible experience is I've been living on a plane for the last uh you know last year flying around Germany, Japan and the US uh and being able to sell this technology when the market has shifted from not even like giving me a meeting to now loving it like that's the biggest privilege. Um but in terms of uh growth, absolutely we're hiring, we're growing. There's so much demand from the automotive sector. Every car manufacturer wants this tech. Um what I think we've built at Wave is unique at the intersection of um uh Frontier embodied AI and automotive. Bringing together these cultures, typically chalk and cheese, we've built a company that has both. And what this means is that if you work on Frontier AI and want to see your work deployed in a consumer product at millions of unit scale in the near term, we're the place to do it. Or if you want to work on an automotive and production grade technology, but with Frontier AI, again, this is the culture. And so um that's the environment we built together. And of course, across the full stack, um machine learning, data, software, all the way through to um product and application and validation for sure. and then interesting roles in operations, public policy, uh, and all of the enabling functions to unlock this future. So, um, yeah, if you're interested, come ride the wave. — That $2 billion won't spend itself. Come help, Alex. All right. Thanks, man. And we'll have you on a lot sooner than a year and a half because that was way too long. So, I'll talk to you in Q3. Thanks a lot, Alex. — See you next time, Alex.

— We're going to sit down with one of the most interesting companies in the world. Just raised a bunch of money. has an interesting take on how to bring self-driving not just to trucking but also to cars. So, please join me in welcoming to the show. It's Wabby founder and CEO Raquel Ratson. Raquel, how you doing? — I'm doing fantastic, Alex and really a pleasure to be here with you today. — Oh, an absolute treat. Now, I want to start with world models because when I was learning about self-driving way back in the day, no one talked about them.

But when you founded Wabby, some of the first publications you did as a company were discussing the Wabby driver and Wabby world. essentially putting world models at the very core of your company. So for folks out there who are a little bit behind uh what are world models and particularly why have you selected them as one of the core technologies at Wabi? — Yeah. So when building Wabi you know we identified that they were you know two very big important kind of pieces of technology that uh um you know were going to be fundamental in terms of bringing self-driving a scalable uh solution to self-driving. on one side was can you build autonomy systems that can truly generalize um and have you know humanlike capabilities of reasoning and the second big piece was about yeah you know in the era of AI uh data is as important as uh you know the model itself right so can we build representations of the world that can enable us to build simulation systems that are as realistic as the real world so that we can expose the system with no consequences to all the safety critical situations etc. Right? And that kind of drove that uh innovation required to bring this uh you know these two pieces uh to market — is a world model in the context of self-driving a very high-end specific video game for your AI to drive around in and to be stress tested. Is that a reasonable way to think of it? I think it's important to uh maybe make the distinction is about what are the things or the characteristics that you need a world model to uh to have in the context of self-driving or physical AI. it can be generalized a little bit which I think will help with you know some of the viewers and listeners today uh which is that it's not just about creating interactive worlds where what is interacting is actually the self-driving vehicle or the robot in the physical AI case but also it's very important to and those have to be you know super realistic right but it's very important that you also have controllability of what they are generating and that's that has actually been uh or is one of the big differentiations in terms of uh building world models for physical AI versus for you know creating pretty uh you know I would say pretty movies or uh you know cool video games etc. — Sure. I I didn't mean to imply that the world models are a video game. But from the perspective of the AI model that's doing the self-driving, they're put into virtual situations, I presume in sequence many thousands, millions of times, and they're forced to kind of react to the environment that is created for them. So from maybe from the AI models perspective, it might feel video game-ish. I'm just trying to give people um something to stand on to understand. — Yeah. So, so there are you know alternate representations of the world that where the self-driving vehicle interacts with that world and you know the key there is that you want to uh create those world models so that they truly represents represent all the things that might happen when you're driving in the physical world for self driving right and yeah so and you are you know the self driving vehicle is acting on them as if it was a video game for the self-driving vehicle yes correct — and the reason why this matters going back to your point about data being so important is that If you have a world model that is a good representation of the physical world, you can stress test your your driving systems, the wobbi driver as you put it. Um, and therefore you can take a quicker approach to market because you've already understood the world versus just mapping a single city. And that seems to be the distinction point between certain self-driving technologies, world models or highde mapping. Is that fair? So I will say that those are I guess two different maybe debates that we can have. Uh one is about how do you train and test the autonomy system and world models are an absolute key in order to allow you to you know in parallel in the cloud you know test uh you know systems at the scale and train the systems to do the right thing right and it can bypass you know many years or you know or centuries of experimentation in the real world right and that's you know that's big and then there is the debate about well what is the information that the autonomy system should have in order to make the right decisions. — I see. Okay. — And that's where HD maps uh and we you know happy to talk about you know uh all the beauty and uh you know uh behind you know high definitionition maps etc. Does use of a world model reduce the need for um oncar sensors or mapping or is it more of a underlying framework that takes mapping and sensors essentially to the next level of safety and reliability? — Yeah. So I will say that um you know regardless of the autonomy system that you uh deliver or the you're trying to build uh world models really uh enable you to train and test that autonomy system to the next level. Now what that means is that it's going to cut down significantly two things uh which is the time to market — right uh it's going to increase the safety of that system right it's going to increase your understanding of the safety of your system which is you know tremendously important um and it's also going to cut down if your world model is very efficient your spend that otherwise you will do by integrating out you know the thousands of engineers that you need over time you know for delivering your technology Um so that's one side but they are very useful regardless of whether you use high definition maps or whether you use different sensors and I think the you know in the debate of the camera only versus multiple sensors maps versus not — to me it's a question about safety versus bomb cost and uh together with co cost of u you know creating high definitionition maps. So what we have for example done is create a way to build high definitionition maps that is super efficient and super robust. So it's not anymore a debate about is it scalable where yes is this scalable and provides you with an additional layer of safety. So it's a no-brainer that you should use that because you know you have a safer product. At the same time the self-driving vehicle if those maps are wrong or not up to date uh it has the ability to react and drive regardless right so you know is um I will say that you know we should debate less about high definitionition maps versus not is about do you have technology that can build those maps really in an scalable manner and you need AI for that right to build those maps in ancal manner and if the answer is yes of course you should use them because then you're going to be safer — going back something you said bomb cost is boom um uh bill of materials essentially like the hardware cost. Okay, cool. I just wanted to for folks out there who thought we changed subject to war very quickly. Not that kind of bomb cost very — very different. Uh now on the on the generalization point, you guys started off with self-driving trucks on

highways. Then you expanded into surface streets and now uh with your latest series announcement and the Uber deal which we will get to in a second uh moving into robo taxis. Does the original world model foundation of the company make it easier for you guys to expand from like one segment of the road world into surface streets and then into I presume residential as well. I'm just trying to understand if the world model itself has accelerated your ability to move from one major area of automation into others. Yeah, 100% uh that I think it's worth mentioning that uh you know the physical AI platform that we built from day one that is composed of the world model the simulator together with the autonomy system was built from day one for uh being utilized for multiple uh you know physical AI use cases. Uh so we had uh you know in mind from day one that can we build that really next generation you know generizable technology that will enable Wabby to actually capture many of these you know multi- trillion dollar markets — right and it has been fundamental both the type of autonomy system that we have which is verifiable end to end technology and I'm happy to go into what that means and why it's very different from the traditional AV 1. 0 know or the what uh has become more traditional now AV 2. 0 know right uh but yeah has been a massive accelerator and what is very exciting about the technology that we have is that for the first time it's not anymore a compromise between this use case and that use case you don't need to fork uh you know build to teams fork the stack uh into you know robot taxes versus trucks uh on the contrary is the same brain and the same uh you know simulator on world model that actually does both use cases The same as for humans, we don't change our brain every time that we actually drive a different vehicle for the first time. This technology enable us to do so. So you actually is additive. You accelerate each program with the other program. — Yeah. — Which is a totally different, you know, mindset compared to what it was in the past. — On one hand, I absolutely agree with you that we humans use one brain for all of our driving needs, no matter what car type, road condition, weather, etc. And so having a single in intelligent mind to handle driving for machines makes a lot of sense to me. On the other hand, human brains are not very specialized. And so is there a place in the future for specialized driving systems that are better at say trucking than driving cars in a city or does the single brain get so smart that we don't need to really differentiate between use case when I guess literally rubber meets the road? — Yeah. So in the case of self driving, you know, the brain is aware of what is driving, which is important, right? Because you don't want to have the same style driving an 18-wheeler, right? Uh 80,000 pounds, you know, cargo truck versus a robot taxi. But that's an example where we don't need to be like super specialized uh in terms of technology. Now when you go to other types of skills uh that are more different u then is where uh maybe specialization makes sense. Uh but a lot of the core characteristics of perceiving and understanding the world in 4D uh not 3D 4D which is you know we live in a 3D world that changes over time. — Yep. — Those capabilities and reason and action uh those are core and common to everything. So at the end of 2025 uh it seems that Anthropic and OpenAI released a couple of AI models especially in the coding context that really changed how people felt about AI, how they used it and it has led to a flowering of new products, features, capabilities. It's been a really tremendous last six months I would say in AI generally. It feels like we've had that same explosion of of capability in self-driving in the last two or three years and especially I would say in the

last year. So Raquel, I'm curious, has there has something fundamentally changed in the AI models and uh intelligence more generally that has impacted Wabby and your competitors in a similar way or am I overanalogizing um general AI versus the more specific stuff that you guys are using? — Yeah. So, so and it's very interesting to see and you know I've been fortunate to be working at the forefront of innovation in EI for 27 years now. Okay. Okay, so I'm going to give you the 277 years view of what has happened and I'm itching myself here on you know life but um the you know what is what has been very interesting is that for the physical world and in particularly for self driving there is three things that are converging at the same time like call it tectonic plates that you need because it's more than just AI — on one side is uh you know is the hardware and the OEMs the platforms that they platform was ready so that you can truly build a scalable safe product. This is the time where all the investment over the last decade uh by both you know trucking OEMs and passenger can OEMs this is actually converging and it's ready now. So that's a big piece of the puzzle in terms of why now deployment and scale you know 26 plus is the year for this or the set of years for this on the you know one other piece that is important as well is the regulatory frameworks are evolving — uh in order to really en enable this uh this deployment. When you look at the consumers of this technology both uh in the robot taxi side humans want to use uh you know self-driving technology was a question mark before whether people will trust and what we see with Whimo deployments is that yes people understand actually this technology is making roads safer and in many ways this is a better product that if it's a human driving right — I don't want to get you off topic here so get your third point in a second but I've been blown away by how quickly normies have taken up oop whimo like it I thought it was going to take them much more time to get comfortable with it but no it was — it's experience and then seeing is believing that's I think uh um you know in many ways and uh you know for humans uh is fascinating I will say and the last bit sorry the last I guess the fourth uh fourth and you know for tracking is a no-brainer right search shortage the cost of human drivers the pervasive safety issues uh you know etc are make a very clear case of why everybody wants to adopt this technology if you build the product that is important for them right — uh or that will solve the their pain points and then the last bit is what you asked me the question sorry to go around you know in a circle right but uh also there is you know massive changes in terms of what AI can do today and what we see really is these next generation companies that you know second mover advant in many ways — of you know the AV 1. 0 know maybe you can deploy scaling is you know it's extremely complex etc right small uh with this next generation of AI technology is so much more powerful and you can truly build through reasoning as I was saying capabilities to really generalize from almost no example and that changes the equation totally in terms of the product that you can build the ability to really solve all the long trail and how quickly you can expand geographically and across use cases, right? As we were talking about before, — so market preparedness and demand, having the right regulatory structures in place, willingness of people on the consumer side to uptake this, obvious market fit on the trucking side and improvements to AI together are really driving the success. — Make everything like now is the moment. But for physical AI, you need more than just the IPS piece. It's all of these things together that are ready now. And is it makes this a extremely exciting time for self-driving and for, you know, it's going to change really the way because transportation is at the middle of everything, right? Uh it's going to change the way that this world works. — Yeah. And I think it's going to change it for the better. Now, one thing we we've talked about is uh the cost of all of this. And one thing that I was really impressed to see reading through coverage of your recent series C was how assets light and efficient your company is, — which contrasted a little bit with the amount of money that you raised, Raquel. And normally when I hear asset light, highly efficient, I don't think this is the company that needs between 750 million and a billion dollars to to, you know, get to the next step of its progress. So what am I

missing there? And uh very politely apart from the fact that you could why did you raise so much money? — I have a brilliant question. Yeah. And many people have asked me the this question. It's like you don't need you know that amount of money. Why did you raise so much money? So, so the uh and it's not just because we could um when you think about um you know the future for a company like Wabi you know we raised actually over a billion dollars um and in this last round and what uh what that means is that uh you know we are the most stable company in the market and that means that uh and that was you know it is important um I thought uh for really being able to um make you know both the right bets the right investments and and think about not just about what we need for the next two years but how is this market going to play — if there is any uh you know delays or anything that happen in the ecosystem whether it's in adoption whether it's you know in certain scaling etc by you know some of our partners uh you know being fully robust to anything and being able to go all in terms of not just our trucking you know leadership positioning and scale and deployment right which is now is the time but also be able to um you know go into the additional vertical that we are adding now robot taxes right without you know compromising or thinking that you know we can actually do that because we are so capital efficient that a billion is infinite money for us Right. So it's you know and uh and it really sets us in a very different place than anybody else in the industry where uh there is going to be uh or there is you know a lot of pressure in a quarterly basis for them to actually show progress to continue their journey — right versus for us — from day one everything was about building for the scale moment and — which has semi back to our point. Yeah, — correct. The strategy has been absolutely spot on, right? Uh and in terms of you know we invested heavily on uh foundational technology, right? And at the beginning was all about building this technology that didn't exist that really you know you invest more you take maybe a bit longer to go to onroad for the first time but when you do suddenly you are placed in a very different position than everybody else right and now it's about that next level of investment for the widespread adoption of this technology. So that's why the billion dollars uh you know why to do this. I really appreciate that in-depth answer, but at no point did you say investing in building lots of rolling hardware and so I'm taking it that you're still going to stay very focused on the autonomy layer and leave the car and truck manufacturing to other people. — Correct. We continue to be a technology provider. We are not an OEM and this is very important for us which is you know we don't believe that uh you know retrofitting or suddenly becoming an OEM is a path for you know for us and we don't believe that this is a safe path as well to market. — Yeah. We believe that partnering with folks that really have, you know, uh, excelled at this over the last century, uh, is actually the right path to really bring that safe self-driving technology. And again, we are not, um, you know, since we are, you know, we have that stability and we can really think long term, right? Uh, we are not pressured to do things compromising that are just shortterm. uh let me show you progress on the short term but that's not really the part that anybody wants for the future. — Yeah, back to your point about you know good foundational technology slower to road but also better long term. So that — and I would say Alex maybe one thing that um you know people didn't necessarily or criticize Wabi in the past was about uh why not to start with quite a lot of you know uh I will say operations and you know commercial operations and what we focus really is build a product uh and uh you know get ready a product that really um solves the pain points and really addresses what the customers want. You know, you mentioned before surface streets and I just want to maybe um add one uh one note there which is you know the industry went with this h have to have uh which is model right which is you have hubs close to the highway and then you drive autonomously between the hubs and then a human will do the end of the trip in both sides — right and that was a reason the reason that they did this is that oh for trucks with technology it's too difficult to drive in surface streets you know surface streets and we want to roll out this to market you know as soon as possible and then simplify the atomic problem and when you end up with that approach is that this is not the product the customers want nobody wants to pay for that dreage which in the economics actually can be depending on your length of hole massive like6 uh you know dollars to point8 per mile which just basically breaks the whole thing. — Yeah. and nobody wants this product, right? So instead, you know, we invested really through this next generation AI technology building for the first time truly technology can drive in generalized streets. Now we can go to the end customer, their door and then suddenly you have a better product. Now that product, roll out your product and that's the phase that we are right now. — Okay. So actually let's I really want to get to the Uber thing in a second, but let's just stay on this. What cuz I couldn't actually chase this down before our chat to my to level of confidence. Where is Wabby today in the commercialization of its self-driving technology in the trucking space? Um are there lots of trucks on the roads that you guys are powering today? Are is there one? I just I couldn't quite figure out where you are now. So Raquel

tell — Yeah. Fantastic question. So there's definitely more than one track. So, so we have you know since 2023 uh since 2023 we've been doing a commercial operations um with you know some of the best uh of the top you know shippers carriers uh in North America. Uh we have a massive partnership with Uber Freight for billions of miles of deployment on the Uber freight network uh which really is really nicely you know sitting in between supply and demand. Um, we have, you know, a decent sized fleet of Celium vehicles. Um, I will say — decent. And is that double digits, triple digits? — It's double digits. — Double digits. Um, um, of trucks and where we are is, you know, our commercialization true commercialization path is really through the OEM. Mhm. — Um and I want to make sure that uh I represent our partner with what they feel comfortable or they have say publicly. Okay. But uh as they say uh last year they are quarters away from that uh the Volv Volvo is our OEM partner. Yes. Uh for those that don't know um their fully redundant uh fully validated platform last year was quarters away. That can give you a sense — very yes soon. very soon. — So it's very soon right and 2027 they have also uh say publicly so that you know that will be hundreds of trucks — which is you know pretty a very nice number already for a 27 department right so that's where you know if you want to know where is so that's where the power is now I can see two ways to charge for this just in the case of trucking your current OEM partner you could sell them selling the system, be it the hardware, software, whatever you want to call it, and then let them have it or you could offer it

effectively as a service. And what I'm not sure about is for world model trained AI drivers, how compute intensive the actual operation of driving a truck is. Is that very heavy? Is it remote? Is it local? And is that a thing you could charge for on a recurring basis as a business to your OEM partners, for example? Yeah. So, um I can tell you that Wabish technology both the world model and the autonomy system is super efficient and you can see that by how advanced we are in terms of the technology right about you know uh driver is launched with the IM etc. — Yeah. — Um and prior to this round it's also public how much money we have raised right. So if you put all this together, you can see how efficient we are actually are compared to also other world models companies that just do world models, right? There's a lot of secrets also in how we do this. — Yeah, she she's bragging right now. That was a brag. — I'm not bragging. Yes, I just I think this is important because it's at the core of Wabi is all about sustainable efficient solutions through next generation technology. That's really at the core of our DNA. We are innovators. We have been you know for the last you know I said before two more than two decades in terms of building this technology but sets us apart from yes saying the more it's more philosophy of yes bigger data centers more uh you know more data and then just you know uh expand everything in the cloud. Uh to your point about how efficient is this technology — uh but as it relates to the business model going back to your question. Um so we are a you know it's driver as a service both on the trucking side as well as on the robot taxi side for Wabi we are a technology provider we don't plan to own and operate neither trucks nor robot taxis and that's where our partnerships our customers are tremendously important for us right Uber plays a fundamental role on that go to market right for robot taxes and you know it's very obvious that they are the market so that and they're very incentivized to grow and continue growing that market. So, so that's uh you know that's very exciting right and the second bit uh for trucking uh so it depends on the OEM also like who um who will operate uh those trucks uh and uh this is also publicly known that uh Volvo plans to also operate some of the self-driving vehicles so through building a transportation as a service uh I would say business unit which is uh Volvo autonomous solutions and that's different than some of the other OEMs. For us, it's transparent whether you know doesn't matter whether is through the OEM or is direct to customer uh is the same business model as it relates to what it will just depend who pays us directly whether it's the OEM or whether is you know say a Walmart for example. — But is it a recurring fee or is it a one-time payment? — Yeah, so it's per mile. So it's recurring fee — per mile. Got it. Okay, cool. That's what I was trying to just chase down to make sure that — Yes, it's a very variable uh there. I mean you can do a blend of things like this, right? It's a bit more, you know, uh sophisticated, but uh the big piece is always the uh permal basis. Yeah, — that makes great sense because that means the more they're using it, the more value they're getting, the more money you make. So it seems very aligned. — It incentivize everybody to be on the same page. — Yeah. Yeah. Or perhaps driving in the same direction. Um sorry, that was terrible. Uh okay. Uh before I let you go, I have to ask more about the Uber uh robo taxi deal. Um, you guys said, and I'm pulling through my notes here to find the quote, um, up to 25,000 robo taxis with Uber, I believe. So, Uber has a lot of partners on the self-driving side, uh, including Neuro and Lucid. And that deal was demand network, Uber, uh, Lucid Cars, Neuro, self-driving tech. You guys have announced, you know, your technology, their network, but not, as far as I know, an OEM. Uh so who are you going to work with on the uh making cars side for that partnership? — Yeah. So let me maybe address the Uber partnership a little bit and how Wabby plays a role in the Uber ecosystem. So what is very interesting is that our partnership is not up to 25,000 is over 25,000 or in other words a minimum of 25,000. — Ah okay. So greater than or equal to not up to — I would say — so that already tells you a little bit about the scale of the of the partnership and um and in the ecosystem it's the same as tracking right the in the ecosystem for us since we are the technology provider you know Uber plays the you know the market component right and then there is the OEM to your point uh um that uh will provide the redundant platform where we vertically integrate with and that's again we believe that's the safe path uh safe and scalable and only scalable path to market. Um we haven't announced yet the OEM uh what — would you like to do that today on the show? — I know that you would love that. Um I would what I can tell you is uh that um you know there is is we love again the coming at the right time so second mover advantage of the ocean has been boiled. There is a few OEMs that have that redundant platform ready now and it's very exciting you know uh how excited the ecosystem is about partnering with us and we are very excited about partnering with them. So more details to come another but uh — uh we are yeah as I said very excited about uh you know our entry into robot taxes and in a swift and really exciting manner. Yeah, I I'm [clears throat] really I think Wabby and Wave are the two most exciting companies in the self-driving world today. I think uh apart from the headlines that Whimo grabbed. So, I'm very optimistic to learn more as the year goes on. Uh Raquel, one last question before I let you go. — You worked for Uber's ATG. You told me before the show for four years. Your company has a partnership with Uber Freight and you now have a partnership with Uber's uh taxi service side of things for robo taxis. Um

has Uber tried to buy you? Cuz it feels like you guys are like best friends who live together like why you know like at some point why can't you formalize it? — I would say that through the years since the inception of the company many people have tried to buy uh to buy Wabi. Uh what I can tell you is that you know my goal here is really to build a physically I powerhouse that is transforming the world. So you know web is not for sale for anybody. Um — all right D you need to add a zero to that offer. Try again. — No I I'm optimistic again. We don't have time to get to it today, but I want to have you back on to talk about physical AI in general and how to take the WBY program to everything from delivery bots to possibly even robots inside of factories because I can see the world, — humanoid, you name it. Yeah. Everywhere. — To me, there's a big generalization of the world model approach and solving autonomy to actually bringing it inside of I don't know why it wouldn't work inside of buildings once you built out the right systems for that. So, there's a lot of stuff coming down the road. Um, am I gonna be able to buy a self-driving car like L4, L5 in the next three years, do you think? — years. Um, a level four, level five. — Yeah. — Um, that's uh that would be hard. — Okay. Well, in that case, — experiencing robot taxes at scale. Yes. um uh personally on vehicles on that time frame is harder. — Okay. Well, can you go uh back to work and get on that for me because I would like to uh to buy one because I hate driving and I keep — and I can tell you maybe I can tell you why you say that. I think it's um it's important which is for a decade people thought that you know um level two will go first then it will be level three and then there will be level four — and it makes total sense because it's okay it's just adding plus one to it right as humans okay that makes sense — but what we've seen and what I learned through my career as well is what we've seen with way But that is not the fastest path and it's not even clear that that's actually a path. — So you don't want to go from L0 no help to L1 lane assist to L2 L3 L4. You want to do you skip or do you go backwards? — You need to either you build level four technology. — Yeah. — Or you build level two technology and that's you know separates us from some other end toend companies. — Yeah. — Okay. And I think this is very important to understand and that's you know at the core of why I say three years is difficult because I truly believe because it's a totally different safety problem that you need to solve that it's not just about I drive well I don't have many interventions that's you know a metric that matters for level two whatever that is not a level four metric and and what people don't necessarily realize is that this There's a gigantic difference between — a level two plus product that is performant to a level four system where there is no more human — and you need to go for a level you need to be a level four native technology and that's what we have done. — Raquel, thank you so much for coming on. An absolute treat. Um when you do announce your future OEM provider, please come back on the show and tell me all about it because I want to know the timeline to get that more than 225,000 robo taxis onto the market and the streets. Um thank you so much. Uh what's the website if people want to go and learn more? So wid. aii please come and you know check us out and we are massively expanding as well. So and it's you know the most exciting innovative techn you know company in physical AI and you know is it's an amazing place to work and partner with and we yeah looking forward to tell more and more our story and but more importantly for people to actually really see our deployment in the real world everywhere. Well, as we say here in the States, keep on trucking. Thanks for coming. — Thank you. — Thanks for watching This Week in Startups. If you liked this episode, check out more. If you're a startup founder, Founder University cohort 13 kicks off this fall. It's a 12-week program that provides guidance on building your product, launching to real customers, and pitching to investors. Top startups receive $25,000 or $125,000 in investment. Apply now at founder. university/twist. already have traction. The Launch Accelerator invests $125,000 and connects you with 500 plus investors to help you raise your next round. Apply at launch accelerator. co. If you're an accredited investor looking to gain access to quality deal flow, apply for Jason's angel syndicate at the syndicate. com. We find two to three deals a month. And check out this week in AI, Jason's expertson roundt with top AI founders and operators every week. Find it this week in aai. ai. Check out the Twist Ticker, our daily newsletter at thisweekendstartups. com/ticker. Thanks again to our sponsors for making today's show possible. Follow the show on Instagram. x. com. This week in Startups publishes three days a week, Monday, Wednesday, and Friday at 5:00 p. m. Central time. You can submit an audio or video file question at thisweekin. com.