S12E3: The Road to Driverless Vehicles with Tiancheng Lou, Co-Founder of Pony AI

When I started my journey, when people talking about AI, people mostly expect AI to mimic human driver's behavior. Like AI try to support our daily work. Most of the things human can do, AI can do by itself. So the next stage after, AI be able to do innovation. This is the next milestone that can achieve and that's huge impact AI can bring to the real world.

Welcome to The Room Podcast. I'm Claudia Laurie, the founder.

And I'm Madison McIlwain, the funder. We are on a mission to open the door to the room where it happens, sharing the stories of some of the world's most iconic builders.

Authentic storytelling is core to our community. It all started as two best friends who met in the heart of San Francisco as product managers in 2018. When I left Uber to start my company, Madison was our first yes. It was these conversations that birthed The Room Podcast.

And nearly 5 years later, our guests have built businesses with a combined enterprise value of over $100 billion. You'll recognize conversations with your favorite technology founders, from Perplexity to Vercel to Zillow to Cloudflare, all across our 110-episode catalog. The best way for you to get to know our community, both online and offline, is by subscribing to the podcast and our newsletter, which you can find at theroompodcast.com. Perkins Coie supports the most innovative entrepreneurs and investors in fast-moving and high-growth sectors, addressing their myriad of legal needs. But the firm doesn't just provide end-to-end legal and business counseling to its startup clients. It also facilitates introductions to key advisors and sources of capital. Perkins Coie's interactive website, Startup Percolator, offers access to programs, resources, and rich dynamic content. Designed to assist entrepreneurs on their startup journey. To learn more, go to startuppercolator.com and perkinscoie.com.

This podcast is brought to you by Mercury, the banking platform businesses like The Room Podcast use to simplify their finances. It's time banking did more than hold your money. Now it can. With Mercury, you can pay bills in seconds, close the books faster, and even send invoices. Not only does Mercury do away with a patchwork of tools, it eliminates guesswork, giving you complete and accurate visibility into your business's finances, all from one account. Apply in minutes at mercury.com. Welcome back to The Room Podcast. Today, we're excited to take you inside the room with Tiancheng Lou, co-founder and CTO of Pony AI, one of the global leaders in autonomous vehicle technology. A two-time champion of Google Code Jam and the youngest T10 engineer in Baidu's self-driving division, Tianqing's journey to autonomy started early. After earning his PhD from Tsinghua University, he worked at Google X before co-founding Pony AI in 2016, launching what would become a trailblazer in both the robotaxi and autonomous trucking space. Since then, Pony has raised over $1.3 billion from investors including Toyota, Sequoia Capital China, and Redpoint. In late 2024, the company went public, and today it's operating fully driverless robotaxi services in major cities across China and expanding globally. In this episode, we dive into the early days of Pony, the technical breakthroughs behind its virtual driver, and how the team adapts autonomy across different types of vehicles, road conditions, and regulatory environments. We also talk about fundraising in deep tech, lessons from millions of real-world kilometers driven, and Tianxing's vision for where the AV industry is headed in the next 5 years. This conversation gave us a rare look under the hood of one of the world's most ambitious AI and robotics companies, and we're so excited to share it with you. Let's open the door. Tianxing, thank you so much for joining us on The Room podcast today. It's a delight to have you on.

Yeah, my pleasure.

We love to start at the beginning with our guests. Before we dive into the founding journey of Pony.ai, tell us, where did you How did you grow up and how did that shape your view of the world?

Oh, I grew up in a traditional Chinese family. For me, I would say, maybe we can save time for later, for the view of the world, I would say I start to view the world as virtual. It's not a real world in middle school times. Later, I spent a long time to verify, to disprove, but unfortunately I realized it's not real. So yeah, this is a long story, but we can save time for discussion later.

We'll definitely dig into that shortly. In that vein of an early identity, did you always think you were going to become a founder?

You can't be ready. Once you have some thoughts to build something by yourself, then let's get— just kick off, get it started. The afterward will be ready. That's the way I work on things.

Absolutely. I think from my personal experience, if I had known what it would be like to start a company, maybe I wouldn't have started it. But once you're in it and you're doing it, there's no turning back. You had a really impressive career in many ways prior to starting Pony.ai. You are a 2-time champion of Google Code Jam. Tell us, how has your competitive programming background influenced your approach to tech and autonomous driving tech?

I'm very active in this programming competition community for at least 20 years. So until today, I still active doing some programming competition for relax. I mean, I'm trying to call this way. There are many different ways to do things. So then the different ways can result in many very different results. That's one thing we learn. Also, during the competition, you have to do everything right, 100% right, in order to get a good result for that. That's also help me to just say, think of how to get things done in 100% way.

Absolutely. I think certainly a coding competition is a great way to show how strong of an executor you are. How did you find out about these competitions? How did you get into that?

Oh, having heard the word IOI, International Olympiad in Informatics, that's a high school competition. I start preparing for that. I try to be qualified as one of the team members in the national team. So that will be my start of this journey. Yeah, actually more than 20 years ago. Then I start learning different algorithm, practice for coding, and meet different friends and keep improving and learn how to do 100% correct. That's what I think I learn.

That's incredible. I think we've had past podcast guests win other IOI competitions, so it's definitely a consistent thread here.

Oh, really? They win or just got a gold medal? This is, I should say, different.

Was it Johnny from Perplexity?

Johnny Fu?

Oh, fun. Do you guys know each other in San Francisco? Have you reconnected?

Not that much, but I definitely heard each other's name.

Amazing. Johnny was on our podcast last season.

Okay, good to know.

Well, you went to Tsinghua University for undergrad and then again to get your PhD. Tsinghua is really the most competitive and prestigious university in China. Tell us, how does your university experience inform your passion for autonomous tech?

First and foremost, I would say I learned a lot. For example, the mathematics, also the algorithm, computer, and helping build a solid foundation for non-learning of all the things, especially during the— also that's undergraduate. And for PhD, I start learn how to realize the impact of different area, try to think what are most important things in the world and then contribute on that one. So I would say for the undergraduate student, say, I just your boss, your supervisor give you a problem and try to solve. I think you are good problem solver. But for PhD, you have to try to convince other people why some of the areas are important. That's one thing. But I have a very fruitful experience. I spent 8 years in Tsinghua, very challenging life there. Also, I meet some of the good friends. One thing I really want to mention that some of good friends there also join the company Pony, who become good friends and later we have different industry tracks. Finally, we converge this new company. Yeah, that's also another biggest asset for my experience while undergraduate.

It's always incredible when you might go to school with someone, part ways, and then merge back together with your company or some other part of the journey. So certainly I think one big takeaway from talking to all of our guests is how important it is to build a network for the long term, even if it's just staying in touch with folks every couple of years. So after Tsinghua, you started your career at Google X. Then you became the youngest T10 engineer at Baidu's autonomous driving division. How did you get into the autonomous vehicle space? Tell me about those years before Pony.

I started my career at Google, it's not Google X. I would say Google is a large company. Google X is, at that time, is a small team. And I would say in the middle of my career, they formed Alphabet and spun off the Google part and X. So I started working on the social things. At that time, social network is a trending topic, more than 10 years ago. So I would say I watched the Waymo vehicles, I tried Waymo vehicles, and then I realized, okay, it is possible technically to get this done. So I decided to join. That's the thinking process for me at that point, I would say more than 10 years ago. So at that time, most of technology are not mature enough for today. I would say many people believe that to be true. Especially for the ones they tried Waymo vehicles without the driver. So at that time, they still have a driver to take care of the vehicle. They're still testing very limited area. So vehicle very, is large, very expensive. But that moment I realized the potential of 3 different things. First, that is the sensors. At that moment, the LiDAR, the sensors are very early stage, but I realized the potential of this LiDAR. I should say without those sensors, it's really hard to get this done. So the next one is the chips. At that time, we were like, okay, let's give a growth of computing resource for the vehicle, unblock the potential of this vehicle. The next one also is the learning algorithm, I call it this way. Today, yeah, people realize a lot of different learning is booming. So a lot of things we should learn so we can build the potential, build the future of the cars. Then, okay, it is possible. Then I start working on this. I talk to the team and say, okay, can I join? Then I join the team and start working on those things. But I say today, you may say it's obvious to say that it's possible. I say that's about 12 years ago, and try to put it this way. So this, some of the friend may think, okay, so, uh, told me, so are you crazy? I'm crazy, I work on different things. I should say, no, no, no, that may take some time, but it's possible. Let's try to get this done. I want to only contribute to accelerate the process. That's the right thinking and why I joined the team.

12 years later, it's very easy for us to see self-driving cars on the road and think that was obviously where the future was going. I was at Uber over 4 years ago and worked there as they decided to, kind of part ways with their autonomous driving group. And I think even 4 years ago, it was really unclear for Uber to get a sense of what the future of self-driving was. So I can imagine having that insight 12 years ago, you were certainly one of the first to really believe in this space. Now that it is a little bit more obvious that self-driving is the future, what advice would you give younger listeners to the podcast on how to break into the autonomous tech space?

It's, I will say today, I'm not sure 100% sure the right timing to start something from scratch. I shouldn't try to put this way. Maybe back to a little bit, try to help the, yeah, younger or younger listener to understand what the real, what was happening there. Say, since you mentioned that about 4 years ago, people, it's hard for some of the outside people, non-tech people realize it's possible because the real progress is hard to visible from outside. You don't really understand the industry before joining that. Big progress being made over the past 10 years. Every year for some of the key metrics, we try to improve a lot. I'm trying to think this way, but from outside, let's say if you just took a ride or took a word, took a talk, something like that, it is really hard for you to really get a sense at what is going there. You may— that may give you a wrong impression, say that the industry is not uncertain, is not making progress. So that's what caused wrong impression. Let me try to put it this way. So if you really want to understand the industry, try to join us, actually do some effort to help you really understand. Don't start at a very far distance. That may trick you. That may give you a wrong impression about those sort of things. But on the other hand, I would say if you really want to start something, something technically, just try to be patient. So if you expect something can be done in months or in one year, sorry, this probably is not the right direction. So for this industry, all the companies should invest a significant amount of time and effort to at least get something done. But the impact of the things getting done will be huge. That's the nature of this industry. So yeah, probably I would say most of the technical guys here have some awareness of those things, but hope my suggestion work can help in this case.

That's a really helpful framework for our listeners. Let's dive a little bit deeper into Pony and the founding journey really. In 2016, you co-founded Pony.ai. Tell us a little bit more about the aha moment. Moment when you decided to start your own company, and who was along for the ride in those early days?

Definitely my partner, the CEO of Pony AI, James Peng. Later we become peers, I should say, peers at Baidu, then we become co-founders at Pony. So I talked to James Peng, we talk a lot on this part. I will say it's important to have the founders of company have a very good consensus on what's going to happen in the next couple years, and not something we're not talking about next month. Or next 3 months, next quarter. We're talking about a few years later, even 10 years later, what is going to happen. And realize, okay, this is good, this very good consensus between us. Also another key thing I put here that James has a very good technical background. So it's important for, I would say, the founding team to really have a good sense on technology. Otherwise you may be very surprised after key decision there, maybe have a wrong alignment of the excitation. Of the development of the technology, also other thing, product. So yeah, so I would say find a good partner will be the changing moment for me, decided to just think of how really find a good company.

It's super helpful to think through what were the things you needed to have in order to have conviction to start Pony. And just to summarize, it feels like you had the team chemistry component with your co-founder, and then you also had these three parts of the puzzle, you mentioned everything from the sensors to the chips needed to be in the vehicles that would have actually been the technological advancement that would have made this popular and possible, really. Looking at that landscape, Claudia mentioned some ways in which the US ecosystem was evolving, but could you draw out for our listeners the difference at the time in 2016 when you're getting started between American innovation in autonomous driving and China's?

Let's just touch a little bit on the ecosystem, I mean upstream or downstream of the autonomous driving, I'm trying to put it this way. So I would say this ecosystem, including autonomous driving technology, evolved over the past few years. I would say it's not by the first day. So it is very matured. We are not waiting for the mature sensors or chips. So all the teams, all the companies working together to think of what will be happening in the future. So we help all the other companies have to build how to understand what's the real requirement, what should the right direction go. And all the companies put together, try to increase the volume and drive down the cost. I would say this was contributed by all the companies, not just us. And for maybe a different way to say that is it's important to have a bunch of companies that work together and join the community and try to mutually benefit from this development. It's not just one by one company or two. This is the same for everywhere in the world globally.

[Speaker:MADISON_MCILWAIN] So there is this kind of ecosystem collaboration happening within China. I think maybe a little bit less here in the States in terms of how people are thinking about their secrets and the magic that's going on behind the driving, if you will. And you guys have created a virtual driver technology. Can you, given that backdrop of kind of coopetition or collaboration, within the ecosystem. What is setting this virtual driver technology apart for you?

First and foremost, I would say definitely the safety record. So maybe one thing to clarify that, so we are working on the fully autonomous driving. Assuming if you ride a Waymo vehicle, there will be no driver in the vehicle. That's what we call L4 autonomous driving.

You call it L4?

Level 4.

What's at Level 1?

Level 1? Oh, I don't really, not really sure, but I would say Level 2 is easy to suggest some driving assist system.

Okay.

Okay.

Gotcha.

Level 3 is a little bit complicated, but L4 is just without driver. Let's make it simple. Just L2 and L4.

Yeah.

Those are the ones our listeners need to know, L2 and L4.

Yeah. Yeah. So for the rest, it's, I know the definition, but they're really hard to like, explain. But for Level 4, so given Level 4, because there's no driver in the vehicle, the safety is very important. Level 4 autoguiding, it's way safer than just a regular human driver. That would have achieved. So set apart and try to answer, to have a better safety record is definitely the most competitive edge or one of the most important competitive edge for different company. So we have very good safety record for us, but to get this done for everyone in the team to set up the system, the sensors, the setup of sensors, the algorithm, all the strategies, all those decisions, it's the goal is to try to achieve a much safer system.

Super helpful and safety is paramount. And especially when you're having this intersection of physical infrastructure, government, and then of course people's daily lives, it, it's absolutely critical. And we are talking about the virtual driver, of course, in vehicles that you and I drive. But then another important part of the virtual driver, the robo-taxis and the robo-trucks. At some point, the roads are going to have virtual drivers and every machinery operated on the road. So maybe could you tease out this different technology for us? What is the difference between the needs of the backend for those types of vehicles versus what we all drive daily?

Maybe I'm going to start with something counterintuitive. First and foremost, I would say truck is much harder than taxi. I should say driving on the highway is much harder than driving locally. For machines, a good driver, we are building virtual driver, is a driver. So assuming a good driver can learn, so can be able to drive different cars in different areas. That's what we are achieving here. But different drivers, they learn everywhere. When driving car, so I driving truck, I should learn everything actually from other driving system. Start from early age for me. So I understand, okay, oh, these are pedestrian, cars, how the physical world are working here. So that's what I learn how to driving. That's a part of learning. So same for the virtual driver. So the, let's say, robot taxi. It's not learn purely leverage robot taxi data experience. It can leverage everything, including the trucks, including others like videos and et cetera. That's why we believe that, so what good driver should be able to drive different cars. But back to the taxi and truck, unfortunately, I would say driving on the highway, it's much, much harder. So back to the idea, Level 2 and Level 4. For Level 4, you have to be much safer. People's expectation that you have much safer than the human driver. But driving on the highway is a simpler game. For hard game, for complicated game like driving locally, it's easier for the AI to be beating the human. But for the simpler ones, it's much harder because human beings also do a very good job. So slightly, maybe another example I could say, when I started my journey, started my company, there's AlphaGo time. I'm sure you remember those things. AlphaGo, one of the best AI Go player beat one of the best human Go player. But Go is a really complicated game. I'm not sure most people have a good mastery on the Go playing, right? So for those complicated things, AI is easier for AI to overcome human. But for simpler things, like many simpler games, it's really hard. Maybe try to think of this way. So driving on the highway, like trucks, the scenario is simpler, but the requirement is much higher. So that's why usually they say robot truck is much harder than robot taxi. But then go back to think. So Good driver should be able to drive different cars. That is the philosophy behind Pony technology.

And having that seamless integration between different types of vehicles is definitely critical to having the entire road ecosystem become autonomous. And as we think about what that future might look like, there's so many different conditions at play, right? Not everywhere is sunny San Francisco where it's relatively risk-averse to be able to drive on the roads. There's usually not much weather, but there's that, there's road obstacles. I was recently in Australia, and let me tell you, driving on the left side of the road. Now that's really hard. How are you thinking about, and how are you building into the models of the virtual driver product, all these different scenario plannings of the unknowns a car might meet on the road?

Here's the easy way to illustrate what was happening here. So the way we train the AI driver, in general, there are two ways to train the AI, or two ways to train people as well. Say, one we call imitation. Just say, I have a small baby, I say, just try to imitate what I'm doing here. Just try to copy me. Let's say this is a way we call imitation. So for the drivers, let's say to train, use the imitation learning to train AI drivers, say collect a lot of data, ask the AI to try to follow my structure, try to follow my pattern, try to learn from what I do, how do I drive. This is one thing, but this was the early stage we call imitation. Recently, most companies switch to that concept of reinforcement learning. The idea is that say we build a virtual world the virtual ones and have AI do, okay, don't ask me how to drive, but these are the virtual world we have tried to practice. This is where you try to imitate by yourself. You try to come up with this idea that by yourself and learn how to drive. Since I touched a little bit on AlphaGo, so AlphaGo, the newer version of AlphaGo is AlphaZero, is using the idea of reinforcement learning. So forget about human Go player, just try yourself and try to beat, just beat others. Then, okay, then get a very tremendous, very surprisingly good results for that. Same for autonomous driving. So we are building a virtual world to try to model everything could happen in the real world and expose those things to the AI. Then let the AI to try to learn from itself and how to improve in this virtual world. That's how what's happening. But that's, let's say, in the virtual world, can you simulate the raining condition? Just put the AI in the raining world, let's say, try to do it by yourself. You should be able to tolerate the rains, the workers, whatever, that's the labor rate, whatever, all these things. And try to get that done. But on the other hand, one thing I'll show you, I did touch here, that it's important for the virtual world, you're very similar to the real ones. Say the AI driver, okay, learned in that virtual world. When we test it on the real world, it's important for AI, okay, oh, it's the same thing. It's just the same as the same pattern as the virtual world, and they're doing well. So this is how the things are training here.

Absolutely. That's a really helpful way to think it through. And there's obviously so much technical complexity that's going into building this product, as we talked about. And you've been at it since 2016. A little bit, we know the future, right? Now it's 2025. We're seeing these cars on the road. But when you were in the process of building, when was that first aha moment for you? When were you able to put the first driverless car on the road?

For Pony, that's about 2022. Over the past 8 years or so, we're gradually increasing the safety record for us. So as 2022 finally reached some of the turning point there, say we got approval from the regulator, say, okay, yeah, good to go. At that moment, it proved that we go much safer than human. So then the regular human drivers, then we have good to go. For 100% of Level 4 autonomous driving company, this should be one of the most important milestone for them, to be able to be get approved to be operate without a driver.

And what has been that feedback now that you're in market with your product, having real users get to be in the backseat? What are you hearing about their experiences?

Oh, the overall, I would say, they appreciate the comfortness of the autonomous vehicle. They're just sensing that. I heard many people just say they just want something a little bit uncommon. Say that I saw some of the passengers, they tend to take autonomous vehicle for a nap. They just call the ride, call the ride, and they go somewhere, whatever. They just make sure the whole trip may take 1 hour, and they took a nap there. So there are a couple of people doing so, but I would say Robotaxi is not designed for those things, but I'm a little surprised they took a nap, but at least this is a good feedback for the comfort of vehicle there.

Oh, I love that. I mean, yeah, I have not napped in the back of a driverless vehicle, but I think I should move that up on my to-do list. Yeah.

Oh, one thing I would say, people like consistency. When you call the Robotaxi, you have a consistent experience. I'm not saying inconsistent to imply too bad or things, but I would say you don't have to handle this uncertainty that the driverless behavior is different for the tech drivers, but like, yeah.

I think from my experience as a rider in autonomous vehicles, I get really motion sick and really car sick in a lot of vehicles. And when a driver is driving an electric vehicle, let's say, and it's really easy to accelerate and decelerate, the jerkiness of the movement gets me sick, gets a lot of people sick. And so from a comfort perspective and comfort principle, having the programmatic acceleration, deceleration certainly at least helps the entire ecosystem of riders that get carsick.

Yeah, sure. Somehow, some kind of new standard could be a new standard of the driving. When carsick is harder, you can just think about carsick, motion sick. Yeah, people can try use a cell phone on the car. That's much harder. Also can use the laptop and try to start write something. It's much easier for them to get sickness there. So I would say that would be higher standard for the driving.

Obviously you're building something that is incredibly technically complex and involves many components, as we've discussed a few of what those are. And you're the CTO of the company. You mentioned your co-founder is also technical, so I'm sure you're both able to think through these problems in depth together. But how have you been thinking about the prioritization of development across Pony.ai's different business units, which we've discussed, or Robotaxi, Robotruck, and then of course your own vehicles?

For me, probably the fact about the competition. Competition is result-driven, right? When we talk about competition, we talk about result, right? No one talking about what's happening in the middle, right? But for us, I will say for the company, we're also by result-driven. Let's say our goal is to improve the safety record to this tier. We are here. Then we identify which component have biggest gap. This is very common in the technical companies. People do iterative process. Let's say a goal is there. We are here. Let's think what will be the next step, and we'll just go step by step. We'll never do the waterfall way, let's say. We'll just do planning and discussion, say what is the trajectory, because no one will really know what's going to happen there. Again, so this is very common in the tech company, I'm trying to put it this way, but in the traditional industry, some other industry, it's really hard for the people there to really get a sense of iterative process, because that probably implies that you don't have a clear plan to your destination. It's clear, but you have to do it iteratively. But given that the company have a very strong intention, also capability to identify the gaps from here to the goal. Hope that I can cover, answer some of Pablo's question.

Yeah, no, absolutely. I think especially as we have more of our listener base becoming growth stage founders and they're thinking through the trade-offs of prioritization of their own time and their business units, it's really helpful to hear how you've been thinking about it and been able to scale so successfully.

Oh, one thing I will say regarding the cost. In old days, the vehicle very expensive. Let's say that. In the past 2 years or 3, we significantly drop down the cost of the vehicle, but drop down the cost and not just say by buying cheaper ones, because the cheaper ones tend to have inconsistent quality. So we have to upgrade the software, the tech stack to compensate those drawbacks. In early days, because we never made happen to, we call the driverless, means no driver. To drive out so that we keep adding things. But in the past 2 or 3 years, we're thinking how to get rid of expensive things. So we are, yeah, I would say the company team is working towards the goal. It's not just the clear say that we keep the pace, which is just the mission-driven and the goal-driven and try to figure out what to do next. So reduce, reduce and controlling the cost will be the priority for us in the past 2 or 3 years. I'm trying to go this way.

Well, speaking of cost, everyone when facing an entrepreneurial journey has to think about how they want to finance the venture and you you all have gone down the path of venture capital, although recently you did IPO as well. So I want to give a shout out to that. Congratulations on the IPO in November of last year.

Yeah, one day before Thanksgiving, not a big Thanksgiving, but yeah.

Yes. Yes.

No, it's such an incredible milestone. I mean, that we're all aspiring to. So again, huge congratulations. And on the milestones to get there, you've raised a cumulative of $1.3 billion, and that has been with notable investors such as Toyota, but Sequoia Capital China, Redpoint China, and many more. So take us back all the way to that 2016 moment when you're just getting started. Who was the first person to say yes to investing in Pony?

Uh, Sequoia China.

Can you take us into that room? What was that conversation like?

Oh, wow. For me, I was— it's surprising. I tend to remember some old things. Uh, at that time, as the head, the Neil Shen, officer of Sequoia China, just visited US and we have a dinner there. And that guy was surprising. He asked some of the very wide range kind of question. I saw him ask something like concept, like a technology plan for that, but that is not the case. Just random questions, cover all the areas. Like for example, I like soccer, like video games, all different things. And then later, afterward he decided, okay, this is right to go. Say, actually, you like to imagine, say yes. But for me, and a couple years later after Neil say, why you doing something Why you're asking such big amount of random questions there? So his feedback, I say that's the way to just making friend or talk to someone because they say the feedback, the reaction on the random questions for him value a lot when you're trying to handle a company. The key thing here, I say there will be a lot of surprise for the startup. So the capability to handle those surprise are important. If some people struggle with some new things, that may be not good for a founder, good for people there. Also very decisive for something. It's important for the founder to have decided, to make decisions when there should be a decision there. So also this is one thing he want to touch up. So yeah, I should say maybe Neo crushing on some of the soft things and for me random things that are very interesting. There's, yeah, they cover a lot of different things. Yeah. Also the virtual world, et cetera. Also the different hobbies, the different things.

Yeah.

Is there any other advice maybe around fundraising or go-to-market that you'd have for first-time founders listening to this conversation?

Success can't be copied, but the failures are. To get rid of most of the common mistakes will be really helpful. I'm trying this way. Don't try to follow the success strategy, but try to get rid of the failure strategy. I'm trying to do this way. So don't try to copy success. Maybe it has this. One thing that I say, try to keep a balance of real commercialization, also the potential of the market. The thing I say keep a balance means I say don't try to do commercialization from day one and ignore side on potential market. But on the other hand, it's not good to say that just try to seek for largest market potential, but forget about commercialization. So you need to keep a balance to try to find some sweet point there. Usually there are many common mistakes that some of the founders try to just extremely optimize for one or another, and most of them, so larger percentage of them, you'll say, may not come up with a good result.

I think from a lot of early stage founders that I've spoken to in terms of what were some mistakes they made the first time that they started a company, it's often around either market that they're operating in or business model commercialization plans. So that definitely resonates. I'd love to shift gears a little bit to chat a bit more about the future of Pony.ai and just your upcoming plans and how the ecosystem is evolving. With Pony operating in China, US and expanding globally to other markets. Tell us, how do you adapt your autonomous driving systems to different regulatory environments across markets? And talk to us a little bit about country and city expansion for Pony.

Okay. First of all, I would say the technology did not matter by the different area. I'm trying to put it this way. Technically, it's not a big problem. This is different from Level 2 and the Level 4 system. So we are talking about Level 4. The key things, the key challenge of L4, Level 4, meaning full-autonomous driving, it handles extreme cases. So when I test, when I judge the maturity of Level 4 system, I will use this mastery of the corner cases, extreme cases, it's not the common cases. So one thing maybe surprise you a little bit is that the common driving scenarios in different country can be different, but extreme case are shared. People may tend to have different driving patterns in different city, But I will say the reckless drivers are shared, are same, are exactly the same. It's the generalization is not a huge challenge for our Level 4. So that maybe surprise you a little bit. People tend to say, okay, people different city, for example, in San Francisco, New York, people have different driving pattern. Yeah. But again, as maybe have said, crazy guys are the same, exactly the same. Uh, Sweden. Yeah. So that's the interesting part. So for Level 4, I say what do we learn in different country, they help others as well. So there's no clear difference there. For Pony, I will say, yeah, the scaling is definitely one of the target for us. So yeah, today I would say we try to optimize the cost structure of the vehicle and try to scale up and reach better commercial result. But for the regulatory and technical side, changing different areas, it is not a big issue technically. But when people are talking about at Level 2, because most of people can't use Level 2 system because many of the guys have tried the driver assist system. Very few of us try the dream of vehicles, right? So when I talk about a Level 2, generalization will be a big challenge. So that's why people may have this intention, say, okay, wow, generalization hard, but for Level 4, not the case.

It's not the case. Right. And that's ultimately where kind of the industry is going to be ending up. Well, you mentioned commercialization and scale. You're going to be scaling your fleet and you mentioned car upgrades and vehicle upgrades. How important is the mass production and scaling of autonomous vehicles to Pony, but also to the overall industry?

I would say it's an essential part, essential point on the path. I would try to say it this way. So scaling, it's not just by saying, let's put more money to build more cars. The first step is the safety, as you said, should be clearly verified. Otherwise you have more cars, you have more risk exposure to run into some bad thing. Another thing, the consistent quality of the car should be high. When you have thousands or tens of thousands of vehicles, It's hard to verify just one by one. You want to make sure they are consistent. So this is one thing maybe I haven't touched on the part. So why we are starting to invest a lot on the robot taxi compared with trucks, buses, or sweepers and etc. When we talk about tens of thousands or hundreds of thousands vehicles. So passenger vehicles are one that matured. There are a lot of the car OEMs, the car makers can build tens of thousands, hundreds of thousands cars in the world. This is a mature industry. Many people know how to mass produce the passenger vehicles. So those things are one thing we should leverage on, work with them together to help us to build a large number of cars.

The other side of that piece is you're able to get hundreds of thousands of cars on the road, passenger vehicles. How do we monetize? And Pony charges fares for its fully driverless robotaxi services in 3 of China's largest cities currently. Are we now in the commercialization era of robotaxis where previously we might've been more in the testing era? Model advancement era?

When I talk about 2022, we're already starting charging money. This is the table. So the table that you're able to operate, let's say, let's scale the fleet to charge money. We're already starting charging money. We charge the same as a regular taxi driver. So the daily ride is the same as a taxi driver as well. So that's what we just say, it just acts as normal. People just view us as just a taxi. It should be like that. So when we scale, we are willing to keep this right card. To just have other passenger riders to view us just a regular taxi driver. That's what we're trying to achieve. City-wise, we're pretty happy for this one. I choose for this way. I'm not tend to enlarge a different number city, not only for China, but for the other city as well. We really want to first like pick out now a few city to really have a good coverage of that city for the vehicle. But regarding the scaling, I will say technically there are many other challenges. So we are on the way to be there.

It's incredible what you've been able to accomplish thus far. And I know a statistic that is just astounding is that you have 40 million kilometers in autonomous driving testing that's powering Pony AI's algorithm. How much more are you trying to get? What is the North Star?

The testing miles, 40 million, is mostly for verification. We have to verify that good enough. but the training process is done in the virtual world. If you ask most of the Level 4 companies, they will give you a similar answer. For the Level 4 companies, this is also the reason why most of the Level 4 companies were never talking about the amount of data is huge, because those are not the key things we use to train. Give you a sense of how the training things works. So I would say every day it's able to just generate the virtual miles, like the hundreds of billions miles in the virtual world. This number is crazy. There's no way to collect such in the real world. Such a data from real world.

So there's this difference between training data that is happening under the hood, so to speak, that's helped to power the autonomous driving model itself, model being what is the backbone of the Pony.ai Virtual Driverless. However, there's also accumulated data that you have driven on with the model, and that's the 40 million kilometers number.

I use 40 million data to train the teacher and have a teacher to train the AI driver. Maybe different from those of you, but this is what happening. To train the teacher is important to, I say, not directly train the students. That how it work. So because the teacher once got experienced, okay, the teacher don't know how to drive, but the teacher is have good judgment of what's driving is a good driving. I don't know how to drive good, but I know what is the winning and losing it. Maybe use example back to the AlphaGo, AlphaZero, the story. Say that the engineers that's writing the AlphaZero, some of them may not know how to play Go, did not have a good player for the Go, but they know who will be the winner. They just write a rule for that, the who deciding the winner, and I like the AI to learn by itself.

These other milestones that you're looking for, what are some north star metrics that you and the team are excited about right now?

Better cost structure. The depreciation of the hardware is the cost we've been paying, right? The income, the type of taxing fee, are the money we collected. Eventually we will reach some of the balance between these two, right? Let's forget about the R&D effort. I focus on the car first. This is called a single car model. The same as other, like Waymo and other company, we're talking single car model. When I talk about cost structure means that when finally I want the taxing fees we collect per car, which can exceed the depreciation for the car, well, the car. That gave us a very good point, good position for scaling. So also I will say that's the real purpose for scaling. Before reaching this key point, I will say for me, I'm a little bit hesitant to, hesitant to think of scaling too much. Otherwise, meaning that when I add one car, I'm losing money. It's not a good thing for the company.

And certainly coming from the world of ridesharing, that's been a big flaw with a lot of these business models in the past. I'd love to ask a little bit more about the role of AI in autonomous driving and how that's evolving over the next 5 years. Obviously, there's been plenty of breakthroughs over the past couple of years. We've had the founder of Perplexity on the podcast. Right now, a lot of people are talking about Deepseek and how that's really revolutionizing AI. Tell us, what are your thoughts on the future of AI?

When I started my journey When people talking about AI, people mostly expect AI to mimic human driver's behavior. Like AI try to support our daily work. So say I try to write something and the ChatGPT can try to beautify it a little bit, something like that. Also, I try to make some videos, can have to just think about those things, actually give the idea, you can give a first draft, et cetera. And later I will say the Pony autonomous driving, especially for level 1 driving, The thing we call build an agent, try to be able to do as most of the human things can do. This we call agent. So if we compare agent with something before we call that supporting roles, it's a big upgrade for the role of the impact of AI. This is just the middle point for the AI. Let me try to think of this. Most of the things human can do, AI can do by itself without the assist of human, just with the minor, minor light assist of human. So the next stage after that, that may take maybe a very long time. I'll say AI will be able to just create some new things. Let's talk about innovation now. AI be able to do innovation. This is the next milestone that can achieve and that's huge impact AI can bring to the real world. But today I should say, according to my knowledge, is no clear evidence, no clear trajectory technology can make the things happen. But I will say that may be something could happen in the next few years or so. My expectation would be 10 years down the road. So there should be some AI can do some innovation.

That's incredible. Obviously being in the Chinese ecosystem in autonomous driving, Deepseek is a popular company that has really become a headliner from China. What are your thoughts on Deepseek and how that might be helping move forward the AI progress?

Let's say Deepseek and some other companies. The data exposed to them is very limited, also the same as computing power. But since they have better ways to organize things, they can have a better job. Now this is one good example to remind people that smart people can do a very bad job to help the AI, or let's say tell the AI to do. Yeah, this is the one thing most of teching guys, tech listeners can understand what I mean. No matter imitation learning or reinforcement learning we do. So one thing to do really good about DeepSix that Not only the results, but also the thinking process. They try to get all this done correctly. So back to a little bit, say, I'm a result-driven people, sure. But I say when the result is wrong, they don't really care the process. But if the result is right, having a good process to result in good result is very important. So that's how the DeepSight trained the J-AGENT. The same as they have very clear objective function in that model. That's also one thing I learned, but sorry, later part of the answer may be a little bit technical, but I assume most of them guys, they understand and going there. This is one thing I want to really encourage the younger listener that this is a good example to help people realize. So different execution, different way, different algorithm can change the world. That's not just by the data volume, also the compute.

Yeah, that's like the approach and the structure of how to approach the problem. So as we look to Pony's future, just wanted to ask, you know, about something that I read recently. You're expanding to Luxembourg. What's the next market? Market you're going to be launching in, and what are your goals for global market expansion?

Global expansion is definitely one of the goals for the company, but today I'm not sure it's the right time to share the name of the cities down there. But one thing I will call out, we are seeking partners in different cities. So when we talk about global expansion, we're always seeking partners. We find partners to help us work together to come up with a solution for that one. We won't do it by ourselves, just by our—

It's a really exciting time across the board in AI, autonomous vehicles, and we opened the call Speaking about being in the South Bay and all the energy that people are saying is back in San Francisco, we're curious, Tiancheng, what's next for you personally?

I've been up a lot between different countries, but not only China, US, but also other countries. I would say my home was in the airport. Yeah, a lot of time in the airport.

Luckily, SFO was not too bad of an airport, and they have that new Terminal 1 open finally after I don't know how many years of construction. We wish you good luck on the travel this year. Our closing question, and it's our hero question that we ask all of our guests, who is a woman in your life that has had a profound impact on you and your career?

With career, my family, back to the Nasdaq listing time. Also my bringing family to on-site, and my wife and daughter was there and they took pictures as well. You may see a small kid, and that's my, that's with my daughter. Yeah, the thing was my wife also joined the journey, but they are not taking picture at that moment. I will say for me, because I spend most of time at the airport and didn't travel, it's important to have a good family to really understand what you're doing there. Also, I truly appreciate. But there's one thing, maybe just one more word about this one. For me, I will say because we have a large number of employees in the company, we should learn how to do something to help our employee to convince, to really convince their family that we are doing a good job. We will host a lot of family event to ask family members to join. To be gone to company to test drive the car and to tell the family members, okay, wow, the party, the family, the member of your family doing a very good job. So yeah, usually the engineers, they have a very deep understanding of technology, they understand that we are making progress, but not for their friends and their family. But they need appreciation not only from company but also from their circle, right? We will try to do more and do a lot of things to help their family, their friend, to understand that they are doing a good job. That's why I say I'm doing something.

We love how important your family has been to your journey, and also having that be a part of the company culture is so incredible. And autonomous driving certainly is an innovation sector that you have to see to believe. We're so excited to see more Pony cars and drivers on the road in the coming months and years. And really, just thank you so much, Tiancheng, for joining us here today in the room, or maybe next time we'll have to do this in a car.

Yeah, of course. Yeah, looking forward.

Thank you so much for the time. Really appreciated having you on, Tiancheng.

Okay, thank you.

Thank you so much for joining us at The Room Podcast. If you want more from The Room every week, subscribe to our newsletter at theroompodcast.com/newsletter. We'll be back next week with a new episode and inspirational guest. Tuesday, 10:00 AM Eastern, 7:00 AM Pacific. See you in The Room. Perkins Coie supports the most innovative entrepreneurs and investors in fast-moving and high-growth sectors. Addressing their myriad of legal needs. But the firm doesn't just provide end-to-end legal and business counseling to its startup clients. It also facilitates introductions to key advisors and sources of capital. Perkins Coie's interactive website, Startup Percolator, offers access to programs, resources, and rich dynamic content designed to assist entrepreneurs on their startup journey. To learn more, go to startuppercolator.com. And perkinscoie.com.

This podcast is brought to you by Mercury, the banking platform businesses like The Room Podcast use to simplify their finances. It's time banking did more than hold your money. Now it can. With Mercury, you can pay bills in seconds, close the books faster, and even send invoices. Not only does Mercury do away with a patchwork of tools, it eliminates guesswork, giving you complete and accurate visibility into your business's finances, all from one account. Apply in minutes at mercury.com.