Lindon Gao on Building Dyna Robotics and Finding a Path Through the Fog

Hello, everyone! I'm Xiaojun. Today our guest is a serial entrepreneur based in the US — Lindon, founder and CEO of DYNA Robotics. He's quite different from Physical Intelligence (PI), who was on our show before. DYNA has taken a path they call the "product research" route. The global robotics industry is still far from reaching a point of convergence. It's an incredibly long race, where you can't see the finish line, and sometimes you can't even see what's around you. So how does each founder make decisions along the way, how do they set their roadmap, and how do they find themselves again when they're lost? Lindon has his own answers to all of these. If you enjoy our show, please share, tip, and like it — give us the full triple combo! Looking forward to 2026, let's grow together with AI. Rarely is there an entrepreneur who Rarely is there an entrepreneur who can get rejected hundreds of times and still not give up on starting a company, right?

Back then I told my team, "We're cockroaches. You guys are cockroaches. We can't be killed." Back then, during a period when we were strategically lost, the disagreement was actually the biggest. It was about the team — our identity as a team took a huge hit. One group of people said, "We need to focus on commercialization," and another group said, "We need to do research." Do you think "smelling the money" and idealism were in conflict? There was no conflict. I'm a very idealistic person. I feel like I got into robotics not to make money. I think "smelling the money" is a path to survival. My dream is the stars and the sea, but my path has to be down-to-earth. Hello Lindon, first say hi to our audience. Hi, I'm Lindon. Coming to Shanghai this time, do you have any different feelings? How often do you come to China? Maybe once or twice a year. We have a team in Shanghai, so I come over sometimes.

My impression of Shanghai is that it's developing really fast. The robotics industry is super competitive. Everyone is talking about robots. How does it feel different from Silicon Valley? In the robotics industry, in terms of how it feels, I actually think both sides are pretty restless. There are tons of releases every day, every company is putting out something new. And every company wants to raise more money. I think, basically, everyone is pretty similar. But fundamentally, I think Silicon Valley and China have somewhat different views on the long-term strategy for robotics. Basically, there's the brain camp and the body camp. The brain camp is basically the US — they believe the brain is the only problem, and the body will sort itself out. China believes the supply chain is the only problem, and the brain will work itself out in the long run. So these two camps are actually quite clearly visible between China and the US. But I have indeed noticed that China has recently also started making moves, for example, at the model level.

So are you in the brain camp or the body camp? We're in the full-stack camp. The full-stack camp. How about, before we talk about DYNA, let's first talk about you as a serial entrepreneur — some of your previous ventures, your upbringing and entrepreneurship. I was born in China, and I moved to the US when I was 11. Actually, the environment I grew up in was pretty entrepreneurial, very much startup-minded. The people around me all ran little businesses on the side. So actually, from very early on, I was already doing small businesses. When I was 14, I started my first business. Many people would call that entrepreneurship, but at the time I felt it was just, in English, a hustle — just making some quick money. At 14, was that middle school or high school? Middle school. Middle school. Back then I was basically just playing video games, and then I felt like there was good money in games. So I started selling accounts, selling in-game gear, and then I built a platform, and made some money from it.

What kind of platform? It was basically a a buy-and-sell thing — a platform where you could go to buy, and you could go to sell. Or I'd go to someone else's platform, or to other people's forums. Actually, sometimes you could buy and sell on forums too. So it was that kind of trading setup. How much was your first pot of gold? Over ten thousand, I'd say. Over ten thousand US dollars? OK, and then? How did that end? the business gradually faded, and once it faded, I sold it — I sold the website. Actually, during that whole period of doing business, it wasn't like one stretch where I was only doing this one thing. It was actually a lot of scattered stuff. So like, today I'd sell game stuff, tomorrow I'd sell software, and the day after I'd sell shoes. Actually, back then, in the environment I grew up in, this was really common in that scene.

The people around me were all doing the same kind of thing. Maybe it was the environment I grew up in — like, my elementary and middle school environment was a really different kind of environment. So the people around me all liked making money. So I was influenced by that atmosphere, and I went along with them, following the trends, and as I learned along the way, I figured out how to make money. Is making money hard for you? Making money isn't hard. Actually, making money is, ... For me it's really... I wouldn't say it's super easy — it's more that I really enjoy... Actually, not that either — I don't love money itself, what I love is the reward that comes from getting something right. At the same time, how was your schoolwork? I wasn't particularly focused on schoolwork. , I might not be a very typical kind of entrepreneur. How did your second venture start?

Back then it was also a trading mindset. In my sophomore year of college, I actually started doing some jewelry trading. At the time it was basically that I found jewelry inventory in various places, and they didn't really want it anymore. They were originally planning to melt the jewelry down and recycle the metal. So I said to them at the time, "Can I buy some of your inventory and sell it on eBay?" And back in my sophomore year it sold really and it gradually turned into a business. Did you sell this company? This company actually later became a self-managing business — I didn't have to spend much time on it; it could make money on its own. And I handed it to someone in my family to manage it for me. This business ran for nearly 10-plus years. Is it still around? Not anymore.

Now I just feel it's too much of a waste — I don't want to spend even an hour on it. Back then I'd spend maybe three or four hours a week and make in a year a hundred fifty to two hundred grand. I thought that was a pretty good deal. But now I feel it's not really worth the time. So it ended. Ended — I ended it myself. So how did you get into jewelry? first through family connections, and also through friends, and also through people in the circle — basically, through family and friends I got to know China's jewelry supply chain, and through some friends in the US I got to know the demand for jewelry. Actually, I turned this jewelry 2C business into a 2B business, because at the time, selling jewelry itself, I felt, was pretty exhausting — packaging every day, then customer feedback... I was still in college at the time, and I had an internship too, so I barely had any time. Later I turned it into a 2B business.

This 2B business was basically me selling to jewelry designers. Actually, the US has a lot of designers, and those designers, back then, had to find their own supply chain or their own factory in the US to get things made, and every run was really expensive, the polishing was really bad, and mass production was really poor too. And mass production was really poor too. So at the time, I started an import-export business, basically connecting jewelry with designers, and back then I would supply them with goods and take a commission. So it was a natural step for you to go to business school, right? When you were choosing a school or a major, did you have other options? And later you went into investment banking. I think my educational path was very free-form. My parents didn't really control me much; they were pretty hands-off with me. So I kept choosing whatever I felt I liked. Actually, at the very beginning I studied medicine, at NYU School of Medicine.

Their med school is really good. Yeah, back then... Actually, studying medicine was my parents' wish. But after two years of studying it, I just didn't want to continue. At that point... So I switched to a finance major. Then why didn't you keep going down the investment banking path? going into investment banking back then was kind of a bandwagon thing. NYU's business school, you know, is also considered really good, so everyone wanted to go into investment banking. When I had just enrolled — when I had just switched to finance — I actually didn't know what investment banking was. But I knew everyone wanted to go. So since I knew everyone wanted in, I wanted to compete with them. So I went, and then... I found it a pretty good experience. How many years were you in investment banking? A bit over two years — two years. And then you started your own company? From leaving investment banking to your current venture, was there anything in between? No, I went straight from investment banking into starting a company, and then I got into robotics.

My first startup was sort of half robotics. Before DYNA, there was another venture. , it's called Caper. It's actually pretty interesting. At Caper, our core business was cashierless retail. Cashierless retail, back in 2017, was really hot, because Amazon had launched a product called Amazon Go. With Amazon Go, anyone could walk into a store that had no cashiers, you could just grab any item and walk right out, and it would check you out automatically. The concept itself was really new, really cool, and it made everyone reimagine what the shopping experience could look like. So at that time... Actually, domestically and abroad, there were probably... too many to count. Back then it was probably as hot as the robotics track is now. There were maybe dozens — over 50 companies.

We're probably one of the very few survivors in that industry. Actually, back then our product concept was very different from everyone else's, because at the time everyone else wanted to renovate the entire store — they would put all the sensor fusion on the shelves and cameras on the ceiling, to do self checkout. Back then, I actually took this concept to supermarkets. I went and met 150 supermarket owners and asked them, "How about this Amazon Go technology? I'll build it for you." What really surprised me at the time was that not a single one wanted it. Then I asked them, "What if I give it to you for free — would you want it?" They said they still wouldn't want it. I was shocked. So I asked them why, and they said the cost-performance wasn't there — "There's no ROI for me.

You want me to spend hundreds of thousands to save a few cashiers? And my cashiers don't just do checkout anyway." So basically, for them at the time, no matter how you sliced it, it didn't pay off, and on top of that you'd have to do renovations. And even after the renovation, the technology still had to... It still required a fair amount of maintenance. So at that point, I actually realized that the business logic of this product itself probably wouldn't work. So we switched to a different product back then, putting sensor fusion and visual recognition into a shopping cart. It actually solved the same problem, because unmanned retail really comes down to two questions: one is who's paying, and the second is how much they need to pay. actually, with a shopping cart you can solve those two questions just the same, and the technology was much simpler than that. Who pays is whoever's pushing your cart, and how much they pay depends on what products they put in, what the products are. So actually we could directly, through visual recognition and sensor fusion, identify these products right inside the cart.

And that's how it gradually came together back then. Was there a really tough stretch in the middle during that startup journey? . In the very beginning, we were actually extremely lucky. We got into YC on our very first application. And after getting into YC, our mentors were actually really famous too — the renowned Sam Altman and Geoff Ralston, who both later became YC's president and CEO. But after we came out of YC, we actually ranked near the bottom of the entire batch — near the bottom in terms of fundraising. We only raised $300,000 in total. So it was pretty tough back then, because a really common question was, "What do you guys have? What's your background? Why do you think this will work?" And on top of that, the US investment community really doesn't like investing in hardware, because they themselves feel that hardware is inherently very costly and extremely difficult to distribute.

So back then, when we had only raised about $300,000, I actually sat down with my other three co-founders and told them, "I'm sorry, I only raised $300,000. If you're willing, we can move into one house together. We can live together. I can't pay you a salary, but here's the thing — we can cover the rent, we can cover your gas, we can cover your food — but we'll have to cook the food ourselves." And so, back then we — the four co-founders moved in together, plus one or two early employees. Back then we were burning $7,000 a month — just $7,000 a month — for two and a half years, before we finally raised our first real money, the seed round. That first round was about $2 million, I think.

So did that get you past the hard part, once that first round of funding came in? Actually, not really. When we were fundraising back then, there was one point that was extremely hard: our product was very contrarian. Contrarian to the point that we were one of the very few companies in the market — really, really few. When we first started, probably nobody in the market had done it before — no one had even thought of making this kind of product. So when I went out to raise money back then, everyone would say, "How could you possibly — how are you going to compete with Amazon?" And there were so many other companies — look, what they're all doing is Amazon-style, technology similar to Amazon's, and they're all doing so and have raised so much money. A lot of them — some companies — raised over ten million right out of YC. And back then we had so little money, so few people, and we were pushing this shopping cart around pitching everyone, and they'd say, "Your shopping cart is going to be obsolete in the end.".

So fundraising was really tough back then. We met with — back then, we — in order to survive until the seed round, we — I personally cold-emailed over a thousand angel investors, and in the end converted two. Because at that point we were almost dead — there was maybe only ten or twenty thousand left in the bank. And then, those two angel investors back then each gave us about $50,000 or so, and that money kept us alive in the end. Then, when we actually went out to raise the seed round, we met with over a hundred institutional investors, and every single one passed — except one, which was First Round Capital. So surviving back then was really not easy. And then — you just asked — after that round closed, getting to the Series A was also a really difficult process, because hardware products have long development cycles, and on top of that, supermarkets themselves have very long sales cycles.

So by the time we truly deployed it and scaled it to the point where it could spread, it had actually also taken more than a year and a half. And then in 2019, when we went to raise the Series A, we actually wanted to raise a ten-million-dollar-level Series A. There weren't many in the market — back then there weren't many institutional investors who could write a ten-million-dollar check. And again, we met with over sixty investors, and basically all of them passed. And then in the end we just kept looking for new investors, and finally found Lux Capital. So back then they invested about — a little over $10 million in us. And this startup later also — this startup later also — it got sold. After the Series A, when we were raising the Series B, we already had a lot of traction. So during the Series B, we got two term sheets and received an acquisition offer.

We actually turned them down twice at first, and then later decided to sell. When was the most painful moment of that startup? Were you getting through it by reading Meditations? How did you know that? I think I saw it written about in a report somewhere. — I think a startup is really — especially in the lowest moments, the lower you are, the more you need to work on your mind. So I think the whole process is a process of cultivating the mind. Because you can imagine — if I meet ten investors and get rejected ten times, I'd say, okay, fine, I'll go meet the eleventh, then the twelfth. But after you've met a hundred, you start doubting yourself. You start thinking, "Is it me that's not good enough, or is the project no good? A hundred people have already rejected me, after all." So back then, through that experience, my understanding of myself and my understanding of the world actually improved quite a lot. I guess it's about looking at it from a different perspective — every failure, to me, was like a round of reinforcement learning: how do I do better next time?

So it gradually forced me to grow. So during that period, I actually grew the fastest — but it was also the most bitter time. You mentioned you had a profile of investors, a mental profile — did that start back then, after you'd met hundreds of investors over the course of it? Actually, no. I think it was precisely because back then I didn't have a clear mental profile of investors that I got rejected over a hundred times. Back then I was focused only on talking and not on listening. So I really, completely — I never really got on the same wavelength as the investors, so it was very hard to build any kind of resonance. This is something I only realized much later, in hindsight. How many years did that startup last? About six years. That's quite a long time.

The moment you sold it, how did it feel? Did you suddenly feel rich? Actually, the moment it sold, I think my first reaction wasn't "I'm suddenly rich" — it was realizing that we'd done the right thing. Because back then, $350 million was, for the startup environment at the time, pretty good, right? One of the reasons I decided to let it go back then was that I felt I could build something bigger, in a bigger industry. My hope was that what I'd learned over the past six years could better help me do my next thing — something I believed could be built into something even bigger. So back then, that was the main reason I wanted to sell, not really for the money. What year was that? 2021 — the timing was great, right?

after that, AI started to explode. Actually in 2021 it wasn't that obvious yet, right? , it was really in 2022 that it started showing a really strong — a curve. Looking back at your two earlier ventures — one startup — what do you think your learning from it was? Actually, my strongest core takeaway is that you should work on the most essential thing. Over my past several ventures, I took so many detours and did so many non-essential things. So metimes it's a mindset like, for instance, hoping that if I just hire this one person, my problem will be solved. You really can't hire with that mindset. Or thinking, if I land this customer, then I can raise funding. Not necessarily, actually. And on the product side — on the product side, actually, that's where I've learned the most.

Building a product is always about doing something valuable. It sounds simple, but I think a lot of the time it's extremely hard to actually do. For example, back then, taking unmanned retail as the case — the thing is, everyone was holding a hammer looking for a nail. Like, "I have this really cool black-box technology, so I'll put it into an unmanned store to save on the cashiers themselves." But actually, when we were really building the product, we found that saving labor costs wasn't, at its core, the most valuable thing about the product. The most valuable thing turned out to be that while we were saving customers time, when they interacted with our screen, they saw value. For example, I put a snack in the cart, and it recommends another snack to me, or for instance, after I put in this snack, if I buy two more, then — you'd get a coupon.

That's when we realized that the core of our product's value was that we could increase the average basket size, not how much labor we could save. And at the same time, in the process of building the product, we also discovered a really big opportunity in advertising, so the overall product value became really, really high. So that value is something that could never be achieved inside an unmanned retail store, because there you don't have a screen that can interact with you. So those are some of the lessons we learned along the way. How do you think about a more essential question instead of an immediate one, like landing one more customer or raising one more round? — I think landing one more customer, if your product hasn't achieved ROI, then landing one more customer is meaningless. Because landing one more customer, unless you know that in the long run, say, once you scale up, the ROI will work itself out — but if the whole business flywheel itself hasn't been built, there's no way to achieve product-market fit.

A lot of people think product-market fit is just "I signed a customer, and the customer is willing to pay me a certain amount, so let's just do it." But in the robotics industry, I think there are actually a lot of companies that found so-called product-market fit, only to discover later that the customization they'd need to do, or the cycle times customers demand — they fall far short of what customers expect from the essential value of a robot, in terms of product form. So I think about this by always tracing the product back to ROI: does your ROI hold up? If your ROI doesn't hold, especially in hardware and robotics, I think if the ROI doesn't hold, it's not a product, and there's no product-market fit. You also mentioned earlier that your YC mentor was Sam — did he teach you anything?

I think Sam is an extremely capable person, but back then I wasn't capable enough myself to get on the same wavelength with him. So the core idea he told me back then was simply launch, launch, launch, launch, and iterate — launch and iterate. That's what I got from him back then. He was pushing me to do it all the time. He still put in a lot of time at YC back then, right? Not bad — maybe we'd meet twice a week. Twice a week. Besides that, were there any other learnings from him? I think my understanding back then maybe hadn't reached a higher level, hadn't reached his level. Are you quite disciplined? Why do you say that? I think I saw that you get up really early, right? I usually get up pretty early. Five or six. When did this start? I guess it started when I began my startup journey. I think it was fairly early on. That's just how it's been.

Is it intentional, or are you just used to this kind of routine? There's a book called Atomic Habits. The idea is that your atomic habits and that momentum can, over the long run, help you reach goals you originally thought you could never achieve. I think for everyone, in the process of growing, if you have really good habits, you'll slowly, slowly become the person you want to become. For me, I think as an entrepreneur, you're basically a mental athlete all the time. At its core, you're really just an athlete. So you need to have an athlete's discipline. That's really what it is, because if, say, you're eating unhealthily, or you're not exercising, or you're not sleeping there's no way you can focus deeply on something. So from my perspective, if I want to build a company I have to take good care of my own body.

So things like waking up early, exercise, diet — I have very, very high standards for myself on those. Up at five, in bed by ten, right? That explains it — your physique really is like an athlete's. Ten years of building it up. Ten years ago you weren't like this, right? , yeah. After starting companies, I got more and more disciplined. Because I found that once you're disciplined, you don't have to think about it anymore. It stops taking up your mindshare. And when it doesn't take up your mindshare, doing it just becomes natural. It almost becomes an obsession — like, if I don't work out today, I feel off the whole day. What time do you work out? Usually somewhere between six and seven. Up at five, work out at six or seven. And after that? Then I start working. Bed at ten at night. Around ten or eleven. My diet is very regular. I eat very few carbs at night.

I usually have carbs at lunch, and breakfast is pretty consistent — with a proper nutritional balance and all that. I've just kept eating this way all along. I think this is something I learned from my dad. My dad is actually an extremely disciplined person. So I noticed that with his own discipline, his own condition, his own health changed a lot too. So yeah, I learned it from my dad. Does discipline improve the quality of your decisions as a CEO? A hundred percent. I think as a CEO, the most important things — everyone says this, right? Find people, find money, find direction. Most of a CEO's major decisions, finding people and finding money, actually don't take that much brainpower. So the most important thing is finding direction. Finding direction is really about every decision you make. Every one of your decisions needs to be fundamental, needs to be high-quality. And to make fundamental, high-quality decisions, you need to think the problem through completely, think it through very clearly.

So if your head is in a really foggy state — like if I'm eating late-night snacks every night or whatever — if your body just isn't healthy, it's really hard to make a truly wise decision. The year you sold your company, you were exactly thirty, right? Yeah, exactly thirty. Standing at that moment, what did you feel your plan was for the rest of your life? I felt like life had only just begun. And how did this next startup get started? There were a few years in between, right? Yeah, I spent three years at Instacart. During that phase, a really big motivation behind being acquired at the time was that I wanted to learn from other people. From the acquiring company? , right. Because at Instacart back then there were tons of executives from Uber, Facebook, and Google. I'm someone who came up the scrappy, unconventional way. My first three startups were all done that way. Very much not by the usual playbook.

It was like, I think the business essence and product form should look like this, so I'd take my engineering team and we'd just go and get it done. So when it came to fundraising, I was pretty naive. Were you one of those founders who hit a lot of walls while fundraising in the U.S.? Definitely. I think there are very few founders who can get rejected hundreds of times and still not give up on their startup. Back then I told my team, "We're cockroaches." You were cockroaches? "We can't die." But at that time, I think a really big blind spot in my experience was management — product and management. It was about, within a much larger organization, how could I learn more, learn better things from them. So that's why back then I stayed at Instacart for about three years — just to learn.

Did you learn anything? And were you also looking at other projects? What did I learn? I'm not the summarizing type of person. I'm more like — I let something seep into my bones, and it becomes part of my understanding. But it's hard to put everything into words. But I think some things can be said. For example, how to get resources from the CEO — that's a really interesting process. You have to explain why the thing you're doing matters so much, how to frame it, and then how to dig into its underlying motivation, and then align with him on it. I think that's something I learned back then. I'd actually never thought you needed to do it that way. It's kind of similar to fundraising. But back then, the way the CEO and I needed to interact was very, very frequent. So that's probably when I trained this side of myself.

And at the same time, on management and organization, back then my management was pretty messy. I didn't really think through what kind of person I needed to hire for a given role, thinking it through very, very clearly. Instead, it was more like, I feel this person is right, I feel this person can grow, and then I'd — and if the person didn't do I'd try to correct them. After doing management for a while, I realized some people just can't change. So me people, within the limits of their own understanding, unless they can break through their current understanding, there's no way for them to take on a bigger scope. So that's when, slowly, it gradually seeped in, and I actually became more and more adept at management. What's the most people you managed at once? Directly and indirectly, maybe a few hundred. Three hundred or so. So mething like that, yeah.

Then when did you decide you wanted to go start a company again, that you couldn't keep being an executive? Do you think being a VP fit your personality? Not at all. Yeah, because I actually really love — I actually have a sense of reverence for entrepreneurship. I really love entrepreneurship itself. Because I think entrepreneurship is like craftsmanship — like a craftsman. Say I'm making a piece of pottery today — I have to make that pottery, that painting, that piece of writing exceptionally good. And they'll spend ten, twenty, thirty, forty years of effort completing a work they're truly proud of. I think entrepreneurship is actually the same. Starting a company is really about polishing your own — your own work. And your own work is the company. And inside that company, that work, is your technology, your culture, your product, and your management, and so on.

I think it's — And entrepreneurship is very multi-modal, multi-dimensional — a very multi-dimensional state of things. So you can learn so much every single day. So back then, my decision to go start a company was really a very simple thing for me. I just always felt I was suited to — better suited to starting companies. Why robotics? AI was really hot back then too. Large models were really hot too. It was actually pretty interesting. after I sold the company, I looked at AR for a while, and we looked at some other things too, like quantum computing, and also things like nuclear energy. But at that time, my other current co-founder, York Yang, and I — we actually explored AR seriously for about six to eight months, and then we found that some of the underlying technology might not be fully ready yet.

And the other areas, we felt, weren't a particularly good fit for us. Back then, York Yang and I had always wanted to do something — we wanted to do something that could change the world. Because we felt that if we were really going to start a company again, we wanted to do something harder, something that could change the world. And the things that can truly change the world are actually few. You could, like, you could probably count them on your two hands, the things that can change the world. Which ones did you rule out? Which ones did you rule out? Like I said just now, AR, energy, and then — quantum. And then things like biology-type stuff, which really wasn't a great fit for us. And then we'd also looked at AI at the time, so we were thinking about how AI should be done back then.

When we finally got to robotics, we realized that it actually matched our previous experience really, really . And I even saw at the time that the shape of the robotics industry was actually kind of similar to cashierless retail — it had a bit of that feeling of everyone holding a hammer looking for a nail. So at that point I felt that hardware was doable, infra software, AI — all of it was doable. But our team was missing someone who was deeply experienced in embodied intelligence, so that's when we found Jason Ma. How did you find him? we were in Silicon Valley and talked to quite a few people. Talked to about twenty-some people. About twenty-some people. And back then, actually, our picture of our third co-founder was really clear. York Yang and I have worked together for over ten years, so the two of us pretty much know our own strengths, weaknesses, and where we complement each other. What are your strengths and weaknesses? I have tons of weaknesses, and very, very few strengths.

York Yang has tons of strengths and very few weaknesses. What's your strength? I think I can see very far ahead. Like, for this company, from now to ten years out, I have a pretty clear picture of the vision, mapping out every step our strategy needs to take and the problems our strategy might run into. So I think direction should be my strong suit. You have a good sense of direction. Yeah, my team sometimes says I can smell money. Smell money. Why is that a sense of direction? That's the ability to make money. Which you trained from a young age, right? The ability to make money is actually a direction. It's about choosing business opportunities. Choosing business opportunities. Because I think all companies, in the end, need to land something, they all need to make money. So no matter how long the runway period is, in the end you still have to make money.

In the end, you still have to make money. So I can figure out what the final money-making form looks like. And then I— Then I know how to build toward it. So I think in this area, this should be my strength. And your weaknesses? Name a few. I have quite a lot of weaknesses. Let me think about my weaknesses. I think— I think I speak too abstractly. A conclusion I arrive at is something I've thought about for a long time. Things I've thought about for a long time— the underlying chain of logic is very, very long. So usually when I talk with someone, about the direction I want to go, I'll walk them through my chain of logic, taking twenty minutes to explain it all, and then give them the conclusion. And usually they just can't follow. So my ability to break down my own thinking into language— I probably still need to improve that. Do you think smelling money and idealism are in conflict? There's no conflict. I'm a very idealistic person.

Why do you say that? I think I'm building robots not to make money. Smelling money is just a path to survival. My dream is the stars and the sea. But my path has to be down-to-earth. Are you a realistic idealist or an idealistic realist? I'm definitely a realistic idealist. You're not an idealistic realist? It ultimately comes down to idealism. Yeah, I'm a realistic idealist. And your co-founder— York. What are his strengths? What are his weaknesses? He's been your partner for over ten years. He's also a realistic idealist. He has a lot of strengths. His empathy is very strong. His structured thinking is very, very strong. His ability to execute and land things is also very strong. His technical skills are very strong. His management skills are very strong. A rare technical genius who also has the cognition of a management dimension.

This is really, really rare. This type of person— How did you two meet? He was studying in the US at the time, right? We met through one of our— another co-founder named Yilin Huang. We met back at our previous company. So this is your second startup together? And how did Jason join? Jason—back then, we were looking for research— Jason—back then, for research— our research co-founder, we had a certain profile in mind. And we searched based on that profile. What kind of person were you hoping for? First of all, what we wanted most was research taste— very, very strong. Research taste. Essentially, you can see it from their papers and the things they've put out. Actually, the really striking thing about him was that you could see from his first and second year of his PhD, a lot of his papers were relatively fancy stuff. But as he grew, you could see that the further along he got, his papers became simpler and cleaner.

The core concepts became very clear. I think that's really good taste— that he kept evolving throughout this process. So just from the written work, you could see this evolution. I was really impressed at the time. So taste came first. Second, I think there needs to be a certain degree of realism— meaning, not just academic pursuits, but also wanting to actually land something, to make a real line impact on a product. I think that's very, very important. I think that's very, very important. If the CEO wants to land things, and the research co-founder wants— say, "I just want to make a breakthrough in the brain"— a team like that usually has conflicts. I think every team has this kind of conflict. Did you have conflicts? back then we— back then we— there was a period when we were strategically lost, and that's when this conflict was the biggest.

It was the team— our identity as a team was under huge strain. One group of people said we should focus on landing things, and the other group said we should do research. We should do research. And when was this? This was probably around the end of last year. And what happened? We sorted out our strategy, and we found a breakthrough point— a strategic point where both the research and landing teams could attack. You were on the landing side? I don't have a faction. Actually, for me it's always been about the path of building up— how to reach my final product form. Landing things can make money, but in my view, it's not necessarily something that can truly be a long-lasting product form. So if landing something can't achieve a scalable product-market fit, I don't think I would do it.

So actually, back then I leaned more toward the research side. So who was on the landing side? So me engineers on the team. So how did you find that overlap in the end? We found the overlap— we extended the timeline. Actually, I think this is a very common problem. Whether it's robotics companies or Agent companies, they're all very similar: if I go aim high, then I have no chance of landing anything. And if I go land things, then I don't have the capacity to aim high. But often, it's because the ability to land things— whether landing can be scalable— is capped by intelligence. So the most important thing at that moment is you have to lift up intelligence. And then the team might think, "Then I should just go do R&D.".

Not necessarily. The idea is we extend the timeline— meaning, further out in the distance, can this strategic point of mine improve my intelligence while also landing something? So back then I set for the team a further-out goal, and that brought the team together. This "aim high or land things" question we just talked about— is it something many robotics teams run into? A very real discussion, or problem? I think inside every robotics company, this gets discussed over and over. I don't think there's a single robotics company that doesn't discuss this problem. Because it's a very real problem. Did you discuss it from day one, or did you start discussing it more frequently around that point at the end of last year? Actually, from day one, we already had a certain awareness and concern about this. But back then we weren't very clear on what capabilities landing would require or where the boundary of intelligence was.

So at the time, what we thought was— what we— that our team is a product research-driven team. Product research is still a bit different from a traditional research team. A traditional research team is like, okay, "I want to get to AGI, so right now I'll go in a divergent mode to explore all kinds of possibilities." And then, once they find certain possibilities, they might converge later. But I think in a divergent mode, sometimes the things you do aren't necessarily done deeply enough. Because if you haven't landed anything, you've only roughly seen some things— that doesn't actually mean the thing has been done thoroughly enough. Product research itself is a bit different— at the boundary of research, I have some guardrails. Meaning, all the research we need to do has to have a certain value— it has to bring value to the long-term product.

And we have a certain framework for prioritizing research. Like, at this stage, we need to do this thing first, and then at the next stage, do another thing, rather than diverging on everything at once. Because I think the team's time— our team's energy is limited. So we only place bets on the most important things. The research-driven companies you mentioned— like PI and Generalist in Silicon Valley— they fall into this category, right? I don't think I can define them by that kind of direction, because internally they also have their own understanding of their own direction. So that's their— it's what they themselves need. Do you deliberately differentiate yourselves from these research-driven robotics companies? I don't deliberately differentiate ourselves from others. Actually, all along, our company has always been a research-driven company.

It's just that our research direction has product guardrails built into it. So actually, I'm very clear about the future product form. So I'm working backward— to achieve this product form, what research do we need to do? Let's go back to when you just started the company. What did you do in 2024? What did you do in 2025? 2024 is really not worth mentioning, because in 2024, it was just a few months. Just a few months. We only really got started around the end of 2024. So back then, we were basically— we were already happy just getting the hardware assembled. Had you already found all your co-founders by then? we had found all the co-founders by then. And the team was initially built. And we had even found customers back then. Because at that time, we needed to— if I had a certain picture of how the research and product would land, then I needed to talk with customers. So we talked with a whole bunch of customers back then. And then in 2025, the most important thing— the most important thing for us— was that the foundation model at the time just wasn't good enough to land anything.

Because the foundation model back then— first of all, robots were very slow, they did things very slowly, and very messily. But most importantly, they could usually only do one thing successfully, and then if they had to do the same thing again, or take the next step, they wouldn't do it . So the situation back then wasn't good enough to land anything. So at the time, I set for the team a product research goal: we had to push the success rate to 99%, and push the speed to 50% to 60% of a human's. Only then would we have a hope of commercialization. So all of our research back then revolved around this theme. And around that time, everyone— later on, people probably saw DYNA-1 around April or May. DYNA-1 at the time was, for us internally, and even for the industry, you could say a really big breakthrough. Because we were the first company to show a commercializable demo.

Back then we were folding napkins— it could fold for 24 hours straight, with a success rate above 99%. For us back then, we were very proud. Last year at CoRL, a lot of people gathered around your robot. That was the one folding clothes, right? Yeah, folding clothes. First it folded paper towels, and then clothes. Why did you choose this scenario? The folding scenario itself is really interesting. We were also kind of the first team to genuinely want to commercialize it. Because actually, when I— I remember when I was fundraising, a lot of people said, "Why are you doing laundry? How big is the laundry industry? It's just this tiny." "Why don't you go into factories, do logistics, do all kinds of other scenarios?".

Actually, my choice back then to do folding was very contrarian too. Because at the time, I saw in folding the opportunity for commercialization Fundamentally, it's because I, with the model— because actually, throughout the whole process I'm a very— a CEO who's very involved in the day-to-day So actually, I stayed very close to everyone's technical direction— our engineering direction— and actually, in the early days I was also managing our technical team myself So at the time, I actually saw the model's own capabilities— the most important thing was that our deployment scenario needed a very high tolerance for error Because, for example, if I'm making a latte and I accidentally don't hold the latte properly and spill it then I'd be— I'd be creating extra work for someone else and wasting a latte But folding is different With folding, if I fold it wrong, I can just redo it So the loss rate itself— the loss rate and the quality requirements fundamentally aren't especially high So given the technical— ceiling—.

given my understanding of the technical ceiling at the time this kind of scenario was the most helpful for me So I felt it was a— a state of technology-market fit And at the same time, folding itself is a very isolated scenario I could have one person sitting there folding nonstop And at the same time, after talking with people back then— back then we talked with laundromats— hotels and so on— and we found that it was still a very common pain point In every hotel— every laundry room— there are dedicated folders and dedicated attendants doing these tasks So at the time, I actually wasn't thinking about how big the market for this had to be but rather whether I could get my first commercial loop up and running And the problem is hard enough so for my model's— there's still a lot of room for fundamental improvement So that's why we chose this scenario at the time What's technically hard about it?

Technically, it has to do with— for example, it— because folding clothes is a soft-body problem— So ft bodies are somewhat different from rigid bodies A rigid body, like a cup for example— it might only have one state like the cup sitting with its opening facing up and the other states might be it tipping over or flipping upside down Folding is different In folding, every step I do— requires the previous step to have succeeded So and folding itself also— demands a lot in terms of generalization Because back then, if we were folding T-shirts then we also— couldn't just fold one size or just one type of T-shirt So actually, at the time we could simultaneously validate our— quality, our generalization, and our speed Back then, it was from— from folding T-shirts that we derived some generalization recipes in terms of garment diversity and environment diversity we actually arrived at some better conclusions about recipes How did you prepare.

the data for the clothes-folding scenario to help it generalize better? Do you have know-how in this area? Actually, I think at the end of the day when it comes to post-training it's really all a question of recipe So actually, back then we roughly knew— for example, from what we saw at CoRL— this zero-shot T-shirt thing— back then we would actually test out in how many environments we needed to collect T-shirt folding data before we could zero-shot generalize in a new environment that thing Actually, we thought it would take a lot but later we found that maybe just ten-plus environments were actually enough , so T-shirts were the same— the T-shirts themselves, different sizes different styles different fabrics different materials and so on actually also have a certain— they also generalize to a certain extent So at the time, in this direction given the state of pre-training back then we roughly probed out a ceiling for post-training, I guess So can I roughly understand it this way— that based on the— pre-training capabilities at the time you found a direction that could.

maximize— a direction where robots could get the most work done exactly And over the past year, this direction— with the breakthroughs in technology— has it changed? Has it changed at all? Actually, I think the most fundamental change is that the amount of data needed for post-training keeps getting smaller And this actually, for us— it wasn't surprising at all Actually, once we hit the post-training ceiling back then after that we roughly knew, oh— the most important thing now should be pre-training Did you later spend a few more months on deployment? Really going into the actual scenarios— after you had the demo?

We did spend a few months on deployment Which months were you doing deployment? Roughly June to September Was deployment hard? Hard, very hard Let me just take napkin folding as an example Actually, with napkin folding itself, what everyone saw DYNA-1 do— the napkin folding wasn't a complete fold because it was something we, after discussing— after discussing with the client, an— intermediate form Actually, at the time, folding the napkin required five folds and we only did four folds and the last fold was left to the waiters because that last fold was very hard Because after folding it over— because by then the napkin was already very thick and very small So after the final fold it needed to be flattened out If it wasn't flattened it would— it would spread open So between those two layers— there was this issue At the time we only— when we only did four folds when we went to deploy the client started to complain They had accepted it at the time but in actual operation they were very unhappy because it wasted a lot of time That's the first thing Second, we actually found at the time that just folding the napkins wasn't enough.

Because in a day we needed to produce 1,300 napkins This client of ours is very big— a very, very large client in the restaurant industry They have a huge volume of customers every day So they needed 1,300 napkins folded per day And we found that DYNA-1 at that time could only fold 850 in 24 hours Actually, at the time we— our speed at the time wasn't up to standard So on the speed front we did a lot of research and finally met the standard But after meeting the standard we later discovered a very big problem which was that every time we folded a napkin although our model was very robust there were some corner cases that we could resolve but resolving them took a really long time So actually, the most common problem back then was that we— sometimes our robot would grab— accidentally grab two napkins Actually, even human hands sometimes accidentally grab two napkins But after grabbing two napkins it would take us about two minutes to resolve it.

Normally, folding one napkin only takes 20 seconds So every time it grabbed two napkins our output for that day— would drop significantly So if— if 10% of them were double folds back then we definitely wouldn't hit the target that day So at the time we were thinking how to use reinforcement learning and some other approaches to let the model solve this problem on its own So at the time we were very firm that everything needed— everything had to be solved by the model Actually, this thing itself is something pre-training can't help you solve Because accidentally grabbing two napkins— because accidentally grabbing two napkins leading to decreasing output— this is very hard for pre-training to solve So at the time we also discovered this And then later we actually found that after folding at the restaurant, we had to put them in a bin And when putting them into the bin the pattern was very random very inconsistent So after you put them in, you'd— people who didn't know would come over and look and say.

these aren't folded properly just a messy pile of napkins And so at that point we had to start conditioning the model so that— first put it in the first position the second position the third position placing them neatly in order This in itself was also a very difficult research process So at the time we actually— that's when we figured out what the model fundamentally actually needed And with that knowledge about what our pre-training— the things we needed to do— became very clear to us Do you also do your own pre-training? we do What's the rough split of resources you invest between pre-training and post-training?

Back then our investment was half and half Now it's much more pre-training than post-training Is this a change that happened this year? End of last year? OK At the time we did pre-training and post-training together Actually, because back then we were probing how changes made in pre-training itself would affect post-training And we later found that changes in pre-training have a very, very big impact on post-training Actually, most of a model's— the vast majority of its robustness comes from pre-training not from post-training So you have to do your own pre-training Based on this scenario what were the hardest things at the time? For example, hand dexterity? Actually, no Actually, the hardest things were these very tiny corner cases Actually, every time a corner case came up the team back then had a— a -known question, which was whether we should use engineering workarounds to solve the problem But we were always firm that we wouldn't Because we always— deeply believed that if we wanted to truly achieve scalable product-market fit we had to use one model to make this work.

In terms of commercialization in this scenario— did it go ? It's already started scaling up It's already scaling up? How much volume so far? I can't disclose that for now How many clients? More than ten now More than ten? So you've already— mass production is going right? Is your mass production done in Shanghai? Actually, here's the thing Actually, I felt that the deployment in this scenario didn't meet my bar for the product's— my standard of scalable product-market fit So I chose not to continue scaling When did you make this decision? I made this decision roughly last year What signs did you notice that made you feel it hadn't reached the scale you imagined?

Because the problem being solved wasn't fundamental enough I only accomplished the task— I only got this one task done but I didn't complete the entire workflow I didn't finish— the whole job in that scenario , right A very common situation back then was— roughly every half— every half hour to an hour we would finish folding a stack of 100— a stack of 100 napkins After finishing, the robot would just stop because there were no more napkins So at the time, some staff in the store had to set an alarm on their phones and every half hour to an hour they'd come over to check and if it was done folding they'd give it a new stack of napkins and if not, they'd just go back to their own work And we'd often find— clients would give me feedback saying our robot wasn't meeting the standard They'd say that— and I'd ask them why and they'd say, I don't know your robot is too slow Then we'd go look at the data and we'd find that.

our robot had very long idle gaps like sometimes the robot wouldn't do any work for half an hour straight So I'd ask them were you not putting a new stack of napkins in on time And then I realized, it's actually not that the robot is too slow. It's that the robot itself needs a babysitter. So this thing about robots needing a babysitter is a really big product problem. So back then I chose not to scale, because I had realized that we hadn't fully solved the fundamental business problem. So if you go and scale an incomplete product, that product will become a burden for you down the road. If you're a research-oriented lab, their approach might be that the most fundamental problem right now is the brain, so I need to solve the brain problem first, because they think that defining a scenario now — the scenario you define today might not be the best scenario.

What do you think about this? Which is basically investors asking you, isn't this scenario too small? Honestly, I think you have to go back to the essence of the problem. The brain isn't good enough. Where is it not good enough? lack of data. You could say it's lack of data. But in the napkin-folding scenario, is it really a brain problem? I had already done the entire job, all of it. So in this kind of, in the sense of this scenario, the agent itself was no longer the problem. So the more fundamental issue was that the scenario didn't work out because — it was really because that within the scope of the task, there was no good closed loop. So within the scope of the task, the agent is not the constraint, but in the workflow — that is, within the scope of getting the whole thing done — it is constrained by the agent.

That's the problem we saw back then. So this also determined our company's next step: for the robot, is this the most fundamental task? in a laundry room scenario, actually, I think for a robot there's no single fundamental task. But I think the task itself really can't fully reflect the capabilities of an embodied intelligence agent. I think the vast majority — actually any task — is just a sequence of motions. Actually, no matter how you generalize along that dimension, the only thing you can — the biggest impact you can see is that I can take this one task and deploy it into any scenario, and it can do it. It can follow that sequence of actions. Meaning its own generalization ability got stronger. Its generalization ability got stronger. But that doesn't mean this task has truly formed a business closed loop. So no matter how strong the agent is, if the product itself is boxed within the scope of one task, I think there's still no business closed loop.

I can put it this way: I can take any task and zero-shot it — put it into any new environment and it can do the task. That is indeed a leap for the agent, but I don't think it's a business closed loop. And then last year, you felt that a robot that only folds clothes on a tabletop, or folds napkins, didn't have enough scalable value. So what did you do? the first question we asked back then was why is there no business value? And after talking with customers, the biggest reason — we had actually seen it ourselves too — was that we needed a robot babysitter. The robot babysitter thing is a really big problem. So how do we get rid of this robot babysitter? That meant we had to do the whole workflow. So what is this whole workflow? It's actually very simple. In a napkin-folding scenario, it means I have to finish loading and unloading too.

In other words, I don't just need to fold napkins. I also need to go — when I'm out of napkins, I also need to know where my napkin inventory is. I need to go get them. After getting them, I need to return to my table, put the napkins down, set them on the table. Then I need to fold. After folding, I need to take the folded napkins somewhere, put them on some shelf, and then grab a new bin, bring it over, and start folding again. That's the whole workflow. So once we realized this back then, it became very clear where our next step had to be. It was evolving from tasks to workflows. How far can you go now? Can you do the entire workflow — can it all be done now? Meaning a robot, in this laundry room space, can move around freely, and it can know what it should do right now.

Have you adjusted the ratio of pre-training and post-training for this? Actually, this really comes down to — it's a pre-training problem. Because here, if you want to complete an entire workflow, then at every step of the workflow — that is, in every single task — there's a certain success rate and failure rate. So to complete the whole workflow, that's really like 99 times 99 times — it's a compound effect, it's an — it's an exponential process, where you — you — we need to do every task and we need to know how to solve every corner case of each task. And then when you chain them together, it can keep running continuously. The problem itself is very hard. It's really a pre-training problem at its core. Because back when we first started, with napkin stacking, we actually found that we were using language to control the model to place our folded napkins at a certain position in the bin.

But we found that the VLA's language-following ability wasn't very strong at the time. So we actually traced this problem all the way back — traced it back to the pre-training problem. What did you adjust later? Later we adjusted the pre-training pre-training and — A semi-humanoid robot. A semi-humanoid robot, meaning its legs are wheeled. And the upper body is a humanoid one. — did you say from day one that you wouldn't build humanoid robots? No, we never said we definitely wouldn't, or that we definitely would. It's really just that — our attitude toward this is: whatever direction is right, we'll — once we see whether a direction is right or wrong, we just go in that direction.

I feel like you're one of the very few deployment-focused people doing robotics in Silicon Valley, right? Because I feel like most people in Silicon Valley think what they should build is the brain, that they should be pushing the high end, that direction. Maybe people's perception of us is that we're the deployment camp. But I think I'm in the product research camp. Not the deployment camp. Because the deployment camp itself is probably more focused on deployment. But I think we're more about probing the breakthroughs that future robot products will need. So actually, the core of our team — from a cultural standpoint, from an engineering standpoint — all revolves around research. It's just that we — it's just that our product philosophy is: if there's no closed loop, if you can't run a product closed loop, then we actually don't know which way to go next. So we actually use the product closed loop to explore our next research direction — which direction we need to go.

Does this count as DYNA-2? The robot and model about to be released (not yet released at the time of recording; since released). , yes. From DYNA-1 to DYNA-2, what would you say the company's technical accumulation is? A whole lot. It's really hard to sum up in just a few sentences. But I think the core of it can be grouped into three systems: system two, system one, and system zero. Across all three systems, we've had really, really big improvements. First, system two is the embodied reasoning layer. It's basically the model's understanding of the physical world — whether it's the physical properties, or the state of what's happening right now, or even the understanding of semantics — it has a very strong understanding, so it can know what we're doing right now, which stage of the task we're currently at, and what our next step should be.

And based on this understanding, it issues language commands to our system one. And system one, at its core, is really a world action model. After the world action model receives the language, how do we convert language into actions? And then at system zero, it's really about how we take the motion — the trajectory we want to output — and from the motion-control side, how we use better reinforcement learning methods to achieve fast, smooth, and accurate motions, rendered on the hardware. In all of this, have you observed signs of generalizability? On system one and system two, there are really, really strong signs of generalization. How far can it generalize? Actually, the dimensions of generalization are a very hard topic to explain in itself.

Because, actually, the two most intuitive dimensions are very similar to humans. Say I teach you how to do something — I actually only have two ways to teach you. First, I demonstrate how to do the thing. Second, I use language to describe how to do the thing. So in the process of generalization, when you measure the dimensions of generalization, you can really just expand along these two dimensions. So on the demonstration dimension, it's how many demonstrations before you learn it. And at this level, what we've found is that for some very new tasks, with our current pre-training capabilities, it can already learn from just a few demonstrations. And the second is language steering capability, meaning I tell you — I use language to tell you to do something. Can the model do it? In this direction, we've actually also found a very strong language sensitivity increase. As for these details, we'll go into them in depth when we do the release.

Have you already started having commercial customers? For this complete workflow? For the workflow, we now already have commercial customers. But we haven't fully deployed yet. But we've already gotten the whole workflow running in different scenarios. What kind of companies are they? What kind of companies are your customers? Hotel scenarios, I guess. That's what I can imagine. But can its cost be lower than a person's? I think that's a really good question. It's very fundamental. It's a fundamental business problem. Having one robot replace one employee — under the current foundation model constraints, it's actually very hard to reach one person's daily output. Because the robot itself — because of some effects from hardware and motion control, yes, and even the speed of data — it's very hard to match human speed.

So for us, what we've found is that the most important value metric isn't being faster than a person, but working longer than a person. That is, a person works — in the U.S., one person works 8 hours a day. So the goal I set for my team is actually that one robot working 24 hours a day can match what one person does in 8 hours of work. Then you can actually reach parity in output. This — this is the most fundamental thing for us. Have you achieved it now? We've already exceeded it. Meaning what it does in 24 hours is more than a person in 8 hours? Did you decide at the very start of the company to not do humanoid first, and to do neither manufacturing nor home use? Was that your earliest position?

In the early days of the company, we actually wouldn't say we definitely won't do this or won't do that. Instead, we thought — our framework is what can be done, not what can't be done. While exploring what could be done we found a lot of things that couldn't be done Why couldn't we do home at that time Home is a matter of timeline I think because my last company was actually B2B2C So I know how high the bar is for 2C And I think home itself is, for robots, kind of the ultimate form if every robot could generalize to every home it doesn't mean it can do all the industrial stuff It just means that, within its generalization capability, it has reached a very high level But even if a robot can do things in every home, that doesn't mean the product works Because from a product standpoint, the robot's daily output needs to match what you expect when you buy the robot.

And that expectation is very subjective , exactly It's very subjective So actually, everyone's expectation is different Just like some people with their housekeepers— the collaboration isn't very pleasant It's actually not that the housekeeper isn't a general purpose housekeeper It's that, whether it's personality or how fast they work, there can be huge differences So I think consumer is a very subjectivity-driven market And in terms of safety and the work itself, from a corner case standpoint, there are way, way more of them So I think this thing— I think in thirty to fifty years we can achieve generalization to every home But actually pulling it off doesn't necessarily mean there's a market for it So you have to polish this product to the absolute extreme before it can truly, truly become a really big product You said you look far ahead So looking at the endgame,.

humanoid robots in the home— is that the most generalizable product form? Could it be our era's something like the Model T? I can only say that home, among all the scenarios in our world— at least as far as I currently understand— of all scenarios, has the highest demand for generalization But people often only look at generalization as the single dimension I think there's also consumer value , and not just consumer value— there's how fast it does the task, whether it does things the way you like.

The way you like— for example, I like the plates placed this way, not that way I think there are so many product details here It's not something that just one dimension of generalization can fully explain And there are safety issues and so on Within all this, there are just too many dimensions So I haven't gone deep into thinking through every dimension But maybe not— , generalization isn't the only dimension a general-purpose robot doesn't automatically make a good consumer product at home , and a home robot doesn't mean that home robot can go into any industrial setting So what did you think back then was your sweet spot? Which scenario? I think for robots themselves, if general-purpose robots become real, I think the demand for B2B general-purpose robots will be higher than B2B2C Why?

Because human desires are infinite Human desires are infinite, which actually means that people— if labor costs become negligible, their appetite for consumer goods, what they want, will only grow Because if labor costs are no longer there, then consumer products become cheaper They become cheaper, and people will want to buy all kinds of things So in B2B scenarios, if our robots succeed, then businesses will need more robots to produce more, more products of different varieties Actually, right now, US unemployment is 4.5% So actually, businesses should have hired 95.5% of the total US population If one home robot— three people make up a household, then the US might only need a hundred million robots.

But with real B2B, there's more nuance to think through, like blue-collar versus white-collar and so on But the US will definitely need more B2B robots versus B2C robots, because more people need to work jobs— making food, working in hotels, running hotels and so on, right And maybe B2B robots are easier to standardize , yes It doesn't need to solve your emotional problems, or your aesthetic problems, right? , and safety issues, to some extent, are controllable So based on all this, you didn't choose what many people might consider the crown jewel— the home humanoid robot path Was this something you figured out from day one? , actually the 2B scenario, for me, is a very obvious scenario, because commercial deployment is the closest And if you can't even make B2B work, then how could I make B2C work?

? For example, if in a laundromat I can't even fully match what the laundromat staff do, then how could I do laundry in every home, washing everyone's different clothes? So I think in many ways, B2B is actually the smallest unit to get this done And the scope itself is actually controllable within its bounds How big is the hotel market? I think the hotel market— actually, a lot of people think the hotel market might be quite small, because how big can a hotel laundry room be? But in every hotel laundry room, you actually need two to five people, depending on how big the hotel is And for hotels themselves, laundry is especially important For example, if they don't wash then their rooms can't be given to guests So actually, back then I heard about some hotel owners during the pandemic, when they had no staff, who went to the laundry room and did the work themselves So it's a very fundamental problem So yes,.

this problem is very painful for them So once we make the laundry workflow work end to end— actually, the vast majority of laundry rooms are largely the same So actually, thinking about this problem from our perspective, this is our first closed loop, and it's hugely meaningful for research, because it lets us know where our model's intelligence boundaries are And at the same time, pulling this off is also meaningful, because it can generalize across the 100,000 hotels in the US So the TAM itself is actually several billion And this is just the first workflow And I think the laundry room itself is everyone's smallest unit on the way to home If you can't even make the laundry room work, I don't think home can work either So I think this path is the necessary route, and it's very important, and from a commercial standpoint, the tolerance for error is actually very high, and the requirements on timing and cadence.

are also very, very low And the willingness to pay is also very high And this is something that is especially, especially needed in the US , I think the robots in hotels now— my impression of them is still those food delivery robots, right? yes, exactly Those are quite widespread now , they're quite widespread now But I think there's more— But that's just a very small piece of the hotel I think under this piece, we can actually open up more Like, once we've done the laundry room we can actually do other things We can do cleaning, we can cook and prepare food, or do other things So the hotel scenario itself covers a great many scenarios And I think a hotel is.

a standardized version of a home Each room— the differences aren't that big So from this angle, I actually think hotels, as the first commercial deployment scenario, are a very suitable scenario Can you do that housekeeper scenario? I think once the model's intelligence reaches a certain level, it's totally doable Her work is also standardized, repetitive work, right? , yes Of course, there will be many corner cases But it's relatively easier to standardize But I think standardization itself, in the short term, is about deployment Standardization helps with deployment, because the variation isn't that big But for long-term generalization, you actually still need to do a lot of generalization work So for example, for pre-training, the generalization data needed for pre-training— the data also poses certain challenges.

So in the process, you can actually learn how to deploy, and also learn how to generalize, because going from one hotel to another, and to yet another hotel, definitely requires certain generalization patterns And there are different SOPs involved Do you think the scenario I mentioned— the housekeeper cleaning each room— compared to your current laundry scenario, how much harder is it? I think it might be a bit harder Because, — Is the market space bigger? But I think the most important thing isn't a bigger market— it's whether you can truly form a closed loop in the shortest time I think this scenario itself might have a lot, a lot of other corner cases, like how to clean the shower area And if guests leave the room in a really, really messy state, or break something, or take something away— I think there might be more corner cases here But also, the whole task might take more time to complete I think workflows themselves actually come in tiers It's not that every workflow has the same difficulty I think hotel room cleaning itself— I think it could be classified as a harder stage So actually, I think fundamentally, if you haven't closed the simplest loops, I think jumping straight to the hardest thing and making it work—.

it's hard to form a minimum closed loop Actually, the closed loop I keep mentioning is really the MVP itself— minimum viable product You make a minimum viable product, and then you can expand outward So hotels were one of the first scenarios you chose? So rt of, right—one of them There were others? There were others , all commercial scenarios Not entirely commercial scenarios Anyway, not home scenarios , no And not industrial scenarios either Not many industrial scenarios Actually, here's how I see it: the industry tends to position scenarios as "this is industrial," "this is commercial," "this is production," and so on Actually, my understanding of the market has never been divided by scenario I think every scenario has workflows I can do Every one So I've actually never.

divided things by scenario So people often ask me whether we're a commercial-scenario robot, and I think that's not quite the right question I think I categorize by workflow complexity For example, a relatively short-horizon workflow, like folding napkins This is a really simple workflow Loading and unloading — that's basically it, right? So if we push a bit further, if we say this is a level one or level two level of workflow complexity, then level three might be a longer workflow, level four might be a laundry room scenario, and level five might be a hotel cleaning scenario, and so on.

I think it's really all layered. So basically, for me, it's not about dividing by scenario, but by workflow — by workflow difficulty level. And that difficulty level can be unlocked at each stage of intelligence of the model. This is your "laying eggs along the way" strategy. Laying eggs at different levels of intelligence, and the eggs are workflows of different complexity. , and the complexity we have now is level one, level two. I think the entire laundry room workflow can absolutely stand on its own. Can it count as a level of intelligence — a complexity level? It's hard for me to say, I haven't broken it down that finely. But I think it's probably around level two, level three, somewhere in that range. Because fundamentally, the consistency is fairly high, but the work chain is very long, and for some aspects of manipulation, for example, opening the door of a washer, you really need to apply force, and sometimes you accidentally drop things on the floor, and there are tons of all kinds of corner cases, and so on.

I think the complexity is probably between level two and level three. And as the agent's intelligence improves, harder workflows can be completed better. So my view on commercialization has always been: the scenarios I commercialize need to fall within the intelligence range of my agent. And then, the strategic direction I want to pursue next is actually beyond my agent's current range. , so generally speaking, that's how I think about it. Is the purpose of laying eggs along the way to give the company cash flow? It's actually to help me understand my current agent itself, to let me probe what level my agent is at, and at which level I can do which things. And at this level, at this level of the agent, what corner cases will I run into? And can those corner cases guide the direction of my pre-training?

Look, for a laundry scenario like this, you first made a desktop clothes-folding robot, and then right after that you went on to do the entire workflow of the scenario. The time you're investing in this scenario is very long. So when you're landing a scenario like this, how do you also do the "reaching higher" part? Do you still have time for it? I actually think reaching higher and doing scenarios are not — not in conflict. The reason they're not in conflict is that, first of all, making our pre-training better, bigger, and stronger is something we have to do no matter what. How big, how good, how strong is your pre-training now? What we've released should be the largest dataset in the industry, and within this dataset we found a very, very strong scaling law itself. So actually, the effort we put into pre-training — I really don't think it's any less than any other company out there. It's even more than other companies, by a lot, a lot. And when I look back at, say, a robotics research lab — it's not like every single person there is doing pre-training either.

Their teams are split up too: we do pre-training — this team does pre-training. How many people are on the pre-training team? This team does post-training. How many people are on that team? And then, in which technical directions will I do what research — it's actually the same structure. It's just that on the post-training team side, we do post-training organized by scenario, rather than in a very scattered, divergent mode. That's really the only difference. What signs of scaling law did you find in pre-training? We found that, when it comes to using human egocentric data, on a pre-training set of one million hours, we found a very strong scaling law itself. I think for the industry this should be a very important milestone, because over the past one to two years, for the vast majority of cases, the intelligence ceiling — this intelligence ceiling has been really hard to break through, because everyone says data is the biggest bottleneck.

So back then, the first-principles thinking we did was: if we want to reach a GPT-3, GPT-4 robot foundation model moment, we need to take the lowest-quality data and extract from it what can scale, rather than scaling a high-quality dataset. This debate — actually, last year I had a very different conclusion, because when DYNA-1 was built, fundamentally it used the highest-quality data. And back then we found that we ran tons and tons of AB tests, and we found that the highest-quality data itself could train the best model.

So at the time I told the team: you have to go all in on high quality data. And then we just kept grinding at it. But then I talked with a lot of people in the industry working on LLMs — Video Generation models as . Honestly, at the time I was selectively hearing what I wanted to hear. They would say that when they were doing pre-training, they ran some ablation studies, and after taking some data out, they found their pre-training got better. But this — so at the time, using these points, I would tell the team: this is their observation, so high quality data is what matters most. But later on, when we actually started doing workflows, we found that what matters most is whether the model itself — whether we can give it the most diverse, broadest human experience.

And then, once we reach internet scale data, doing ablation studies and distilling the data only becomes meaningful then. Before that premise is even met, doing too much quality optimization — I don't think that's necessarily first-principles thinking, at least when you look at this from the language model perspective. So after we had this shift in thinking on this front, we started to think about: what will the form of data look like three, five, ten years from now? What will the ultimate form of robot data look like? We thought about it over and over, and we couldn't think of anything better than egocentric or exocentric — that is, egocentric (first-person view data) and exocentric (third-person view data). We believe these will become the ultimate form of data, because, for example, if we were to bet on teleop, or bet on UMI, or bet on some form of data that isn't egocentric, I feel like we'd be, to some extent, betting that robot hands won't work.

But we think the industry consensus is that robot hands themselves are an engineering problem, not a research problem. So when we bet on egocentric, there was a certain — we would think about it on a 5-10 year horizon, and in the end we decided this is the largest dataset, and also the dataset of the ultimate form. So at that point we went all in on this model — on egocentric. So now you think more data matters more than higher-quality data? Is all of your data real-robot data? now we also have some simulation data. I think the data mixture — I don't think there's any need to declare whether I'm in the simulation camp or the real-robot camp. I think it all comes down to the training objective — what your goal is, what kind of model you want to train.

Every model actually reflects its founder's and its team's different thinking embedded in it. For example, OpenAI back then was consumer application all the way, so their models back then, their early models, were more consumer friendly. Anthropic very early on decided we're gonna bet on coding agents. For Google, multimodality is something that's really important to them. So I think for robotics teams, each team probably also has some divergence on what capabilities they want pre-training to achieve. So I don't think there's a right or wrong. Do you use all of them? We use all of them. Whichever works we use. And after trying them all out, they all work ? They all work . Why does China divide founders into the real-robot camp and the simulation camp? Why do you think this debate exists? I don't know why this debate exists. I didn't even know China had a real-robot camp and a simulation camp.

This technical path itself isn't — isn't the essence. Because it's just one of the — maybe a stage-specific — right, it's a — it's a means to an end. why bet on the means? I don't think there's any need to bet on the means. I think the most important thing is still — what we call important is really: what kind of dataset is the ultimate form for us? That's what we need to bet on long term, because if we get even this one thing wrong — the industry actually shifts very fast. For example, early last year it was the real-robot camp, the teleop camp. Then by the end of last year, people's understanding suddenly all shifted to UMI. And this year, people might start talking about egocentric and simulation. I think — this path itself will keep changing all the time. So the most important thing is that we, as founders, figure out what we believe is the most first-principles, most essential thing, and then we place our bet on that most essential thing.

That's really all it takes. I don't think any of it is right or wrong. And that most essential thing is: what effect should pre-training achieve, what goal should it reach? So what goal do you want pre-training to reach? For us, actually, there are two goals pre-training should achieve: one is to decrease demonstrations — so decreasing the number of demonstrations means I can deploy faster. And the other is language following capability. Language following capability, for us, is really the fundamental link for chaining workflows together. So this is actually our current understanding. But I think as the market keeps evolving and changing, our understanding might change too. But I think the most important thing is: your understanding can change, but that understanding still has the product as its guardrail. What we ultimately have to do is definitely to land it in the real world, no matter how long — no matter how long this ramp-up period lasts.

I think we're still in a ramp-up period right now. But no matter how long this ramp-up period takes, we need to see the product's ultimate form, and then we can keep moving in that direction. So what is AGI for robots? What do you think the endgame of the robotics industry looks like? How many years will it take to get there? I think it's really hard to say how many years it'll take, because I think as the amount of data grows, and based on the things we're seeing in our own lab, I honestly think we still don't fully know where the ceiling is. So I think this is a really exciting era. At every checkpoint of technological breakthrough, we can look forward to where the next possibility lies, especially in embodiment. We're actually not very clear on that yet. So I think it could be very close, or it could be very far. And I think the ultimate form of AGI — at least from the robotics perspective — is that you give it a very vague instruction, and it can get the whole thing done.

That's actually very similar to the ultimate form of coding agents or some general purpose agents. They'd be very similar. How many years it might take, I don't know. Yeah, I think it's really hard to say for sure. I think, as— I actually think, if you look at this through the lens of language models, actually, we— starting from, starting from the point when ChatGPT came out, that was about four years ago. From then to now, I think for the industry, it's already at a point where— no one could have imagined that four years later we'd reach the state coding agents are in today. I think a lot of technological progress, especially in the field of AI, actually, a lot of the time it's a state that looks very, very flat, but then suddenly, one day, there's a sudden jump. There's a sudden jump. So actually, I think we— the jump we've seen ourselves is on the egocentric side, finding the scaling law itself.

And what that actually means is the ceiling on our data right now is very, very high. So we want to go all the way to the end of that data ceiling and see, at that end, just how much there is there. So how does probing the height on this main line of intelligence relate to your approach of laying eggs along the way— to each of those scenarios? I think probing the height and laying eggs along the way are actually not two separate things. They can actually be one thing. Because probing the height itself means you— at every point of intelligence, at every dimension— for example, right now we— for example, from a single task— that's our understanding from last year to this year: a task as the smallest unit— a scope— has now expanded to a workflow as the smallest unit of scope. So next, as language capabilities get stronger, demonstrations become fewer and fewer, and generalization gets stronger and stronger, will the next stage be that we can directly enter a language-based— a language-based state?

Meaning, you just use— you directly use language to tell the robot what it needs to do, and that's actually enough? Still within this workflow, right? And actually, I think maybe even— it could, it could simultaneously do very— a great many different kinds of workflows, or robots could start collaborating with each other. So actually, I think this might be— because the scaling law itself is very hard to pin down in terms of what the impact of my scaling law looks like. But I think, with the framework I just described, you can roughly think about which stage of the scaling law we're at, and how our robots can be empowered with more intelligence to do more things. I think this, for us, is the more fundamental thing. On your road to AGI, how many scenarios will you lay eggs in along the way? Have you projected that? Actually, I think fully doing one scenario is basically equivalent to doing all scenarios. Why is that? Because a single scenario itself, what it requires— whether it's the corner cases, the quality requirements, or the speed— actually, if we can do this one thing thoroughly, we'll roughly know where the current technological bottlenecks are.

So actually, for me, doing one workflow versus doing 100 workflows— the difference is actually just the data for 100 workflows. So for us, the core of our product is actually a system: data goes into this system, and the system outputs a workflow. So actually, when we— when I've polished the first workflow into shape, the differentiation for many of my other workflows is actually not that big. It's true that, for example, within each workflow, there are all kinds of different SOPs and all kinds of different corner cases. But I think the fundamental thinking framework for solving these problems is already there. So I think, once these fundamentals— once the underlying logic and essence are figured out, everything else is actually just— what's left is just a data problem.

Look at the large model industry now: for example, people building Agent applications or AI applications are having a really hard time this year, because people found that one new skill of the large model can just swallow up an entire application scenario. Do you think in the big robotics industry this kind of situation will happen? For example, for teams building the large-model brain, some capability leap of theirs could make a lot of the scenario-specific optimizations stop working, become useless. Would that happen? it would. I think doing one workflow itself actually doesn't conflict much with what you're saying. For example, Anthropic— they built a coding agent. Does that mean they're only good at coding agents? Actually, no. It's because they've put a tremendous amount of work into pre-training as and they've also done enough work on post-training to land the coding agent. So they can, at the same time— they can use the same framework to now land agents in other scenarios, like Excel, or other application scenarios.

Actually, for us, doing one workflow through doesn't mean we stop doing pre-training, or that our pre-training is only pre-training for this one workflow. Rather, our pre-training has always been the broadest— under the broadest generalization. And our post-training is our only place for laying eggs. Actually, it's on post-training that we see whether we can do this workflow through, and then probe further: in the process of doing this workflow, what bottlenecks there are, which bottlenecks need us to keep guiding our pre-training— it's a path that guides the direction. , so our pre-training dataset should be the broadest, the broadest and most generalized, and as we continue to scale— when we scale to, say, 10 million hours— our pre-training is still very broad. But by then, the impact on our post-training will vary in size.

We don't know. But what we at least know is, the basic thinking framework is: at the previous generation's checkpoint, with the previous generation's amount of data, our agent was like this, the bottlenecks were like this. In the next generation, maybe these bottlenecks will go away, and some bottlenecks may still remain. But in the process of probing the whole workflow, we'll actually roughly know which ones can be solved by pre-training and which ones cannot be solved by pre-training. Which ones can't be solved by pre-training? Actually, there are many. For example, like we just mentioned, corner cases— like I said, if I grab— grab two napkins at once, it causes my overall output to drop significantly. I think this is something pre-training cannot solve. Then for example, when placing the napkin at a spot inside the bin, whether it can be placed neatly— the placing itself is entirely something pre-training can solve through language. We said the endgame of robotics is hard to predict.

So which moment do you think it's at now? I think discussing this isn't particularly, particularly important. What I think matters is the moment when robots can truly bring value to humanity, and when robots can truly achieve scalable product-market fit itself. And I think we're actually not far from that. I think, from the perspective of our own product and strategic direction, I'd say yeah, one to two years. One to two years. It might achieve PMF in a certain scenario. I think achieving PMF within a scenario is already something that can become— it's already something that can be landed, that can already take shape. That is, once we've done the workflow through, I can do— do 10 workflows at the same time, 20 workflows. I actually think these— these scenarios can be made to work. So actually, I think the ChatGPT moment is that anyone can— it's useful to anyone.

GPT's biggest— its aha moment is that it's useful to anyone. Everyone has their own subjective questions, everyone has their own subjective thinking, but GPT works universally for all of them. So the question really is whether robots need to— whether robots can reach that level of generality. Now, most people probably understand general-purpose robots as physical AGI already being solved. But from my point of view, I think this is a product definition question. What by product definition is, I actually think robots don't need intelligence— they don't need to reach generality to be usable by everyone. , that's my view on product. Maybe I'm not explaining it very clearly. General enough to be used by everyone.

For example, what form could that possibly take? I think I can only say— I can only say this much: actually, it's this is part of the strategic thinking for our next few steps. But I think the robot itself is actually both an agent and a product form, not just an agent form. Not just an agent form— meaning you think intelligence alone isn't enough, right? It also needs to be a product. , yes. It doesn't necessarily need intelligence to reach an infinite ceiling, but it needs, based on the intelligence of that time, to be made into a product. , yes. Can I understand it that way? , yes. Actually, I think a better— a very similar example is actually ChatGPT came out in November 2022. GPT-3 actually came out in June 2020. From GPT-3 to GPT— ChatGPT is actually GPT-3 versus GPT-3.5. So actually, at the model level, whether in terms of model size or data volume, there was actually a change. But you could say that back in the GPT-3 era, the product form that came out was an API.

So at the time, dozens of developers had access to it, and there were also some companies, like, like Jasper, that successfully turned this API into a product of their own. But I think the essential change from GPT-3 to ChatGPT was a product change: it went from an API form to a browser chat form, and only then did it achieve a very— a huge leap in users. Actually, I'm not entirely sure whether, if it had been released in the form of GPT, ChatGPT could have reached such a large user base itself. But I think the product form really can't be ignored. When intelligence reaches a certain state, its product—look at these past few years— even the form of agents keeps changing. Its shell is always changing to match the intelligence level of that time. , yes. So we've actually put very, very, very much effort into the agent itself, and all of the agent— whether it's our egocentric scaling law, or our research, like our own whole body control, or our workflow— from language and language stability to reasoning, to the embodied reasoning level— we're paying special attention to all of it, and we've made very big breakthroughs in every dimension.

But fundamentally, all of them— everything we're doing is oriented toward a product endgame, that's what guides it. So actually, in terms of our sense of direction, we're a team with a very strong— a very strong sense of direction. The product endgame— you still don't think it's a humanoid robot, right? Then what is it like? Actually, you might think— you might think the product endgame is a robot form. Actually, I don't think it's a robot form, but rather a product for how to use robots. So — What do you mean? It's not convenient to say right now. Later on, next time we can— we can talk about it in detail in the next interview.

You won't be launching this product this year, right? , not yet. OK, so what do you mean by a form that interacts with people— a form for how people use robots? I think it's actually about every person— how every person uses robots, or rather, making every person feel that this robot product is a form that's useful to them— not just the hardware, but the product experience. But didn't you say you're doing B2B? , yes. Are you also going to do B2C? No, just B2B. Can B2B reach everyone's hands? I think it can. And it won't take too long. So you can look forward to our next release. The next one? Not this one. Not this one B2B2C I don't think we need to speculate now Do you think the final form is also not a humanoid robot? Human-robot interaction From a hardware standpoint, I think whether it's humanoid or not, actually the most important response is really an ROI question It's like, if I add these two legs can they bring me higher value and does that value far exceed the cost of making these two legs You said you do product research So when it comes to research on the robot brain what do you think makes you different from other companies, like Physical Intelligence Generalist or Google DeepMind.

There's no fundamental difference We're all basically just trying to figure out how to generalize model capabilities how to crack physical AGI how to make language steerability capabilities better or how facial demonstration can be done better Fundamentally, there's no difference at all So you also work on the robot brain, right? We've been doing it all along We started from day one Then why does the industry, when they talk about DYNA instinctively think you focus more on the robot body more on the use-case scenarios? Do you think that's a misunderstanding? it's a huge misunderstanding I think it's probably because we didn't spend enough time on PR at the beginning Because when we launched DYNA-1 we actually framed it around real-world deployment as the core message Like, why did we build DYNA-1?

Because we wanted real-world deployment But at our core, we're actually a very heavily research-focused team Because if we weren't heavily research-driven we couldn't have built DYNA-1 either That was a fundamental breakthrough So actually, from our perspective we've always believed we've always felt that intelligence is the bottleneck for real-world deployment and that we constantly need to improve the intelligence to truly achieve deployment But it's just that while we're improving the intelligence we also turn deployment into a closed loop for validating the intelligence Where do you spend more of your time on deployment or on the brain? More on research? My time, actually, my time is spent more on strategy and team As for allocating— I can't say my time because which project my time is allocated to doesn't necessarily reflect the company's biggest I think you can look at it from the company's trade-off perspective Eighty to ninety percent of the company's ninety percent of capacity is all all on the intelligence And we, we've made the whole deployment engineer infrastructure simple enough.

So actually for us, deployment isn't as hard as everyone imagines It's not that heavy , actually because, we've now reached the point where any data collector can train a model This is, fundamentally, this comes from a breakthrough we made in pre-training which is that our pre-training has very strong generalization So we don't need that many demonstrations to learn Second is a breakthrough we made in infrastructure Our infrastructure is good enough so our data collectors can all do it they can all train models When you see all of, all of our deployment videos and our demo videos they were all made by the data collection team not by our engineering and research team You also just mentioned that at the end of last year your team had a debate about whether research matters more or engineering deployment matters more And you said you found an overlap where both could matter at the same time What kind of intersection did you find?

Back then it was about tackling workflows At that time, we were actually in terms of strategic thinking we realized that the essential task itself wasn't enough for deployment not enough to become a commercial closed loop So back then we found that deployment itself wasn't enough either It wasn't— it wasn't the right time for deployment And back then our infrastructure wasn't very good either So at that time we chose a strategic model which was, we wanted to— internally we called it "one employee" "One employee" meant that if we truly wanted to reach the value an employee can produce in a day we needed to handle loading, unloading, the task and the entire workflow end to end So during that period we actually united the whole team around one strategic point which was, I— our need for the intelligence had to improve so we could complete the entire main workflow.

At the same time, our deployment team had to make the infrastructure better so that we could deploy better And on post-training technique we also had to get better and faster in order to deploy a workflow itself So that's what united the team back then to push toward this point Is this one of the most important decisions you've made in the past two years? at the time it was kind of a battle for our team It was a very important strategic milestone OK Then let's go through from 2024 to 2025 to 2026 the key strategic milestones what were they respectively?

I think maybe let's skip 2024 It was just the beginning In 2025, the strategic milestone was really about using the smallest unit a single task as the unit how to achieve a closed loop of deployment So what we talked about back then was really mainly a lot of work around DYNA-1 At the time we were actually in post-training which things affect the quality of a task's execution our speed and our language controllability There was a lot of thinking going into that And then in 2026, we actually moved forward with the "one employee" strategy as our guiding strategy And at this point, we knew which dimensions our pre-training model needed to generalize toward and what our pre-training direction looked like how we could go crack.

scaling law with the most abundant data At the same time, we needed to do on a workflow That actually means reasoning and language scalability capabilities need to be strengthened And at the same time, on the robot body, we needed to lay out plans because our first-generation robot couldn't move So for the second-generation robot on whole body control the hardware layout also needed focus So this is our strategic framework for 2026 Beyond that, it's not convenient to share yet Not convenient to share But we still have— we've seen quite far ahead Is what comes next very different from the current scenario? It's probably not an extension of this scenario right? What we see is the whole market What we see is how we can let anyone use this robot Let anyone use this robot, oh So that's the next step, right? exactly So what kind of tasks does embodied intelligence need to do to achieve PMF?

Do tasks that can achieve ROI That is, if they if within their own within their own scope of work they can produce more output than a human then they can achieve PMF In your scenario, you should be able to optimize and reach that In the scenarios we've chosen we can basically reach it in all of them If we can't reach it then it can't become a commercial closed loop in the first place People in the robotics industry say isn't it too early to talk about commercial closed loops? Why pursue a commercial closed loop this early You've only been in business for a year or two Actually, I think a commercial closed loop itself isn't really a— the most important revenue isn't the most important thing What matters is that we think through every stage of the commercial closed loop, every single problem and get them all worked out For example, with self driving itself if you don't drive down every street in San Francisco you won't know where your bottleneck actually is So actually for us it's really about this.

which is, we need to run through all the scenarios all the corner cases of the scenarios we want to target Then we roughly know we can roughly categorize the problems into problems solvable by generalization and problems not solvable by generalization And then, on that front, get the deployment flywheel running Because actually, I think it's not just robotics At my last company, Caper when we were deploying after the product came out even when the product itself could already be rolled out to any supermarket, we still spent several years before we truly got deployment working Because deployment itself is a deployment flywheel And the deployment flywheel isn't just about— people might think once my intelligence is solved I can deploy at scale It's really not like that at all When you're deploying your hardware will have issues your software will have issues your model will have issues your motion control algorithm will have issues And sometimes it's after your model has been running for a while one day you suddenly find my throughput was supposed to be folding 1,300 napkins a day.

but today I only folded 800 napkins and you don't know why So actually, within the deployment flywheel you need to, for example, build an observability stack you need to build a support min stack So here actually, a lot of people probably haven't realized that the deployment stage is actually a very, very difficult stage So in— And the deployment stage itself might also take two or three years to truly work through So once you've finished the intelligence itself and then you go run the deployment closed loop you'll find you're two or three years behind Because actually so really, this fundamentally comes from maybe it also draws on my previous experience That's how I knew OK we probably need to figure out these deployment details now You said you can smell money In the robotics industry have you smelled the money?

I have Where? In all the workflows I'm about to do This is the early stage, the early-stage possibility of a commercial closed loop And what about after the early stage? After that, it's the next step of our strategy That's the possibility of the entire market Why do you emphasize closed loops so much? Because only with a closed loop do you get feedback And only with feedback do you know your next step It's because, you know I think crossing the river by feeling the stones is inherently very difficult If you don't know— when you don't have a closed loop you don't have feedback and you don't know where your problems are it's very hard to know that my research needs to go in this direction or that I need to go in that direction But what a closed loop itself gives you is on my path within the smallest—.

the shortest amount of time I can roughly figure out my next two or three steps So actually for me, it's a process of exploring the road ahead You had a China team from day one right? that's right Why did you make that decision? Was it an important decision? It was a very important decision Because actually, I think with intelligence without a supply chain, there's no way to build a world-class robot Your team in China— how far along is it now? In China we have a team of about forty to fifty people dedicated to hardware R&D and supply chain So we've actually got the whole chain working end to end So we've now already entered the initial mass production stage Initial mass production, OK Hundreds of units, right? Hundreds of units Is it the workflow robot or the previous generation? The workflow robot OK For the previous generation, I chose not to mass-produce it Because I felt the previous generation fundamentally wouldn't work out People in the industry say that you have been pretty low-key since the middle of last year keeping a relatively low profile Why is that?

What have you mainly been working on this past year? I think there are a few reasons First, our research taste— our bar is very, very high Actually, the vast majority of the industry Like some of the things PI has put out, we actually finished a few months earlier. Like on the language steering side of things, we actually had it back in August or September, last August or September. But back then we actually had a very high bar. Because we felt like, this thing wasn't at 99% yet, so we didn't want to release it. But later we realized, if our own taste bar for research is too high, we'd end up releasing nothing, and that's not good either. So we've actually changed our view on that to some extent. Second, I think when the strategy isn't clear, putting out what we see as interim work, I don't think that really aligns with our core philosophy.

When we put something out, it's not for the sake of PR. I think we do it because we feel this thing is truly done at its core, and there's a very strong logic behind it, whether it's product, research, or engineering. Only when we think it's worth releasing, worth seeing, do we choose to release it. And I think the strategic goal we set for this workflow is actually a very distant goal, one that takes a very long time, and requires the team to come together to achieve it. So midway, we chose not to release things, because I felt that telling partial stories just for PR — it's better to just tell the whole story. That way people can better understand why we're doing this.

I think robotics might be an even longer race than large language models. What do you think matters most in this long race? Because it's still far from convergence — the whole industry is far from converging. Everyone is still exploring different things. Everyone probably feels the market hasn't converged. Actually, I think the most important thing is just having a direction — what you think the end state is going to look like, and then you walk toward it. That's really it. For you, which matters more, intelligence or product? Intelligence matters more. Because I think without intelligence, there is no product. So intelligence is part of the product. Intelligence is the most important part of the product. Only with intelligence can you truly go and productize. But I think the product also shapes intelligence itself — like how intelligent does it need to be to become a product? So the two are definitely related, but between the two, I think intelligence is definitely the most important.

And that's actually the center of our team. I think the most important thing in this long race is product and intelligence, product, intelligence, and real-world deployment. I think, even in the course of this long race, I think — I think you can see the China camp and the US camp — which is what we talked about at the beginning. The US side thinks the brain matters most; the supply chain — supply chain — eventually everyone will have similar hardware, those problems will all be solved. While the China side thinks the brain might not matter that much long term; the supply chain matters most, because the supply chain is extremely hard. I think both matter. But if you really asked me to look at this over ten years, I think the supply chain might matter more. , because the supply chain itself is what you — it's what the customer ultimately — what the customer ultimately gets is hardware. So in the end, your mass production, your product form — because actually, the brain determines the floor of your product, and the body determines the ceiling.

So I think long term, it may turn into a war over cost and quality — that's something ten years out. So if the supply chain matters most, how do you see this current wave of robotics teams in China? I think everyone actually has their own playbook, and they're all worth learning from. Like Unitree, for example. I think Unitree is actually a really excellent company, especially on the hardware side. Their hardware R&D — they've made their System ID, the sim-to-real gap, so, so small. At its core, that's a really major breakthrough. A really major breakthrough. So I think they do have an edge in that regard. What about AgiBot (Zhiyuan)?

I think AgiBot's playbook is probably — they do have their own embodiments, they have their own hardware. I think they're probably also still figuring out which direction to go long term. But they have plenty of funding, and I think that funding can sustain the organization for a very long time. That's actually part of the strategy itself. So because they have more funding, their margin for error is higher. So they can — their strategy can spread across a broader product line, and then on the embodiment side they can place their bets in a more divergent way. So that in itself is also their strategy, and I think it's worth learning from and respecting. How do you view your competition with, say, robotics model companies, and how do you view your competition with Chinese robotics companies? Actually, I don't think the industry is in a state of competition. I think if you look at the shape of the whole industry, it's actually a very — sure, there might be some element of competition in it, but I think everyone is also very open — people are very willing to say what they're working on, what our direction is, and when we've done some research, we'll share what our research results are.

I think at its core, this is a really good dynamic for an industry. So I think only when it comes to actual market share — once there's real market share, I think that's when real competitive dynamics will form. now, I think everyone is in a state of mutual respect, or mutual learning. When do you think rapid convergence will happen? Talking to robotics teams these days, I feel like everyone has a different approach, and convergence still feels pretty far off. You're right about that, but at every stage there's actually some degree of convergence. for example, after we built DYNA-1, everyone actually converged somewhat on robot success rates and execution speed — like, we actually need to build a useful robot itself.

And also, whether it's UMI, egocentric data, or world models, the industry is gradually converging in those directions too. So these things might change in the long run, but I think we're already gradually in a process of converging. We roughly know which directions are right, which ones are wrong, which things matter, and which don't. So I think maybe in the next one to two years, there's a chance we'll see a more converged pattern. People will probably move more toward fairly similar things — whether in model architecture or data capabilities — things will start to converge to some degree. How do you see the relationship between the robot body and the brain? I think both are very important.

The brain is the robot's floor, and I think the body is the robot's ceiling. That's — that's actually the opposite of what many people think. Others might think the brain is the ceiling and the body is the floor. The brain determines whether something can be done at all. The hardware determines how it can actually be done. That's actually my view. Because whether something can or can't be done is determined by the brain. And how it can be done — that's actually, to some extent, determined by the hardware and motion control. Like, say I fold 1,000 — I can fold 1,000 napkins a day. Actually, if within the capabilities of data, motion control, and hardware, if it's possible, if it allows it, my movements can go faster, my movements can be more precise. So I think both are extremely important. And I think the relationship between brain and body is really, really tight — very hard to pull apart and analyze, because sometimes, because the body itself has video — meaning your camera, your perception, and your motion control.

And the data you collect — if you look at it from the teleop data side, it's like — take motion control itself: data we collect from motion control — with two different motion control algorithms, you can train two very different models. One might be super smooth, really elegant, and the other might be a really terrible robot. So I find it hard to imagine truly doing B2B deployment scenarios if you don't own the robot body itself. It's very hard to hit their quality, speed, and efficiency. So I think both are — So do you think the brain should drive the body, or the body should drive the brain? I think the brain drives the body. The brain — this is actually a common dynamic in our team: the hardware spec has always been defined by the research team.

For example, how many FPS does our camera need? Do I need global shutter or rolling shutter? And the camera angle — how wide does it need to be? All of this is defined by the research team. So I think the body needs to be defined from a research perspective. So if the brain makes a leap forward, would this generation of hardware stop working? Would it have to entirely replace the existing hardware setup? Would that kind of situation happen? I don't think it's that the brain's improvement makes the hardware stop working — it's that after the brain improves, the hardware may hit its ceiling. Like, your hardware can no longer do what my brain wants to do. Like my brain — to use the simplest example, this is the simplest one: say originally, my brain can play against a world-class table tennis player, but my body can't keep up.

Then I'm limited by the hardware. So I'm wondering — is going full-stack now too early, since the brain problem isn't solved yet? I don't think it's too early, because at its core, it's really a co-develop process. It's about what kind of hardware my embodiment itself needs to pair with. And as the brain evolves — after it evolves, my hardware will definitely keep iterating continuously. The hardware body needs at least three revisions before you can truly enter mass production, before you can truly enter real scenarios. And I think right now, the supply chain hasn't fully converged yet. So hardware itself still has a huge number of problems. If you wait three years to start doing the hardware, then you've lost three years of iteration. So by then, sure, the brain problem is solved, and then you go define the hardware — yes, at that point you might define better hardware, but that doesn't mean your hardware can be deployed at that point, can be mass-produced.

Will you switch the gripper for a dexterous hand? In which generation? Does it matter for your scenarios? For now, it doesn't matter. But I think what matters most for the dexterous hand — I think the time to switch is when dexterous hands are ready. now, dexterous hands aren't ready enough for real deployment. Is there a problem that really troubles you right now as CEO? The troubling problem is — it's a matter of time periods. Each period has its own troubling problems. For example, at one point it might be a people problem at another point it's a fundraising problem and at another point it's a strategy problem But I don't think there's been any long-running, persistent main thread of trouble If there were a long-running persistent thread, then that itself would be the problem Maybe it's too big and I can't solve it Lately, I've been thinking about how to get more data That's probably what every robotics company is thinking about right?

But I think our ceiling might be a bit higher than others' because we can train on more data So I think data sourcing might be one of the problems on my mind Are there any good sources for buying data right now? No, we can use them all From DYNA-1 to DYNA-2, did your technical roadmap change? Fundamental changes — it was a very big change Whether in terms of model architecture, our pre-training dataset, or from the reasoning angle, even the hardware changed I feel we've been completely reborn into a new company Were there any struggles along the way? actually, the biggest struggle was that back then we were somewhat confused about the path we were on Basically, we couldn't see the next strategic milestone or where to push next Which month was this last year?

It was actually last year — end of last year, the one we talked about at year-end, right Why were you confused back then? At the time, we felt the deployment — the tasks themselves couldn't be deployed in a way that formed a closed commercial loop And we needed to decide — should we bet on deployment, or bet on research? And back then, picking both — you can only strike at one point So that was actually the result of convergence at the time So we hoped this workflow scenario could both satisfy the commercial closed loop and also raise our agent's — raise the ceiling of intelligence Do you think that up to now you've achieved that goal? I think we have And what you're going to do next is something you can't talk about yet?

, yes But it's already moved beyond this workflow scenario? , yes Do you have a lot of cash on hand? Quite a lot Can it support robotics model R&D? Absolutely Robotics model R&D — how much money do you think it takes? I think the vast majority of the money goes to talent, compute, and data I think in the next phase, when we massively scale up our data — in that phase — that's when we'll need more capital But for us, from the last phase to this one, we haven't burned much money So this can be considered DYNA 1.0, right? The complete laundry workflow — can we put it that way? And then for 2.0, you might use a lot more data to build a bigger model and define a new scenario or paradigm So rt of You can put it that way that's right In 1.0, what do you think you accumulated through that process?

What did you validate? Just to review this whole phase, up to when you launch the workflow robot this year — Actually, it's about which direction the agent needs to go That's what we figured out when DYNA 1.0 took shape The core question was how far our agent really needs to evolve, how much our pre-training needs to improve, and what kind of generalization and language results we want out of pre-training before we can stand up a workflow I think that's our biggest takeaway Does 2.0 break away from 1.0's narrative, or is it an extension of it? You mean the next step in our strategy? I think it's breaking free of the scenario's constraints Breaking free of the scenario's constraints — to tell a bigger story? Not telling a bigger story — but going after a bigger market It's building for a general market Because the flywheel is actually already built The whole system is already in place All that's left is feeding it more data and it can learn new things So you're saying you've built — based on this one workflow, you've built a system Data can come in and then go out and it can form a workflow that goes out It can be — data comes in, and then through this system it gets converted into business value But that scenario isn't a vertical niche scenario,.

, it isn't I see When will we see this? Next year Next year, huh So it's still a year away I think for each strategic milestone, to truly finish fighting that battle might take half a year to a year I don't know how long it'll take But I think the path is very clear What's DYNA's company culture and atmosphere like? We are all very — I think the team itself is extremely resilient Because based on our past experiences, I think some of the ups and downs in direction don't actually affect us that much Second, the team is a very pure-hearted team We — because York and I — York used to be in rock music And I myself really love the act of starting companies itself So we really.

love working with people who genuinely love doing the work Because the robotics industry itself is a bit restless and fickle And everyone has different ideas and opinions And at any given time, some company is always the one everyone is paying close attention to But for us, if you genuinely love what you do, you won't be swayed by the outside world Instead, we have this picture in our hearts.

and we just walk toward that picture So we are a very pure-hearted team, and also a very passionate team, one that really, really wants to make this thing happen Does your company's internal state feel more like a lab or something else? In terms of internal structure, it's more on the lab side Since you've been an entrepreneur for a long time, do you think the organizational forms of these new companies today have changed compared to companies of the past? I think that's a really good question I actually think organizational form has undergone a qualitative change Actually, I — Before the era of coding agents, most large-scale companies were structured layer by layer, hierarchically So in a big company like Google, for example, there might be 10 to 11 levels, and each level has.

its own set of responsibilities But in the current AI era, organizations — my organization included — have become very, very flat So a very good form of flatness is that we — me as the CEO, or my team's senior executives — can directly steer the direction of execution We can reach directly through the engineers and researchers doing the execution, or through just one layer of manager above them, to get it done directly I think this is fundamentally a very big change And I think it's a good change Things get done much faster So you have fewer than 200 people now Not far off, right?

The US side, maybe — slightly more than that A bit more, right Why is it configured this way? the US team is mainly research-focused, and the China team is mainly hardware-focused Hardware-focused , yes And research is a bit bigger than hardware Actually, my — one of my management philosophies — from my Caper startup plus my Instacart experience — I've found that the management style that suits me better is hiring the very best people to do the most important things So our team's average caliber is very, very high Most of our ICs actually had management experience before, and later chose to be ICs because they genuinely love being ICs So everyone has very rich experience So when you give them a scope, they can break it down themselves and execute it themselves So we've reached that kind of state And with fewer people, management overhead comes down So we don't need to spend too much time managing an organization An organization — if we had, say, layer upon layer of people, then it would be very — like, when you do things,.

everything moves very slowly and the whole work structure isn't nimble enough So my philosophy is we find the strongest people to do this, and that's enough So and everyone can better enter a flow state This flow state is, so to speak, everyone being on the same wavelength We all need to do this one thing So when we all can unite into one heart, we can get this thing deployed faster It actually doesn't have much to do with headcount In fact, I think when there are more people, it's harder to become one heart. I — So we — I want to form one heart in the smallest possible unit and move in this direction How long do you think this small-organization state will last? Is it a phase-specific state or a long-term state?

I think the small-organization state is definitely a phase-specific state But when it comes to growing into a bigger organization, I actually think it might be a bit different from the form it took ten years ago Maybe back then, ten years ago it might have taken 1,000 — this thing, for example — So mething that used to take 1,000 people might now only take one or two hundred people to do So it will definitely get bigger, but it doesn't necessarily need to get huge So fewer than before, but bigger than now After agentic coding exploded this year, has it changed your industry at all? I think there's been a huge change Because everything we do has gotten much faster So when engineering costs come down, you actually have more room to explore things in other directions For example, before, something would take a month to do, and now it only takes a week So at the same time, you can explore four things, four directions So fundamentally, the iteration speed has become very, very fast So when you asked me earlier how many years until we reach AGI,.

I actually don't know either Because engineering efficiency itself keeps improving So I think the most important thing is still a team's iteration speed This is your fourth startup Your fourth startup What feels different this time? I've gradually come to understand what starting a company is really about I think before, entrepreneurship was always about judging my own achievements by external standards Back then, judging by external standards meant how much money I raised, how much I accomplished, how high my valuation was, or how many customers I had, how much revenue, and so on But now I think I'm more in a state of using my own —.

a state of going inward, toward my own heart How do I become my best self? How do I think through this problem clearly? How do I think through the underlying logic clearly, so that I can do deeper, more important things? I've gradually — through a kind of self-, cultivating my own heart and mind, I can live out a truer version of myself I think this is the entrepreneurial state that I prefer When did you start cultivating your mind? I started cultivating my mind when I first started Caper so over ten years ago because I actually think entrepreneurship itself is really, really hard and at every difficult moment it's actually a challenge for you a process of cultivating your mind because everyone looks at the world through their own subjective filter and sometimes when you shift your mindset when you choose to look at the same thing from a different angle your world becomes much bigger.

I can give you an example actually I have a pretty strange condition called ankylosing spondylitis Jay Chou has it too my immune system attacks my own spine right and back then I was eighteen or nineteen when I already had this condition and I really resented the world back then I kept thinking why am I so, so so unlucky and every day I had to live in this pain sometimes it hurt and it hurt even more when I slept for a lot of people, like with pain once they sleep it doesn't hurt but for me it hurt more when I slept and sneezing hurt too or things like that anyway it was really, really miserable and back then, for me the biggest breakthrough wasn't that the illness fundamentally got much better it was that I shifted my mindset I started seeing this illness as.

an experiment like, how can I change my routine how can I change my diet how can I change myself to help fix it to solve this problem so after over ten years of experimenting on my own body I actually discovered certain patterns and I got this condition under control fixed it so so that's when you started being disciplined, right? yes and I think entrepreneurship itself is actually the same I used to always think I'd talk to this investor and we'd be totally out of sync they had no idea what I was talking about and I'd think this investor just doesn't get it I thought they were stupid and I'd just keep saying what I wanted to say but now, after shifting my mindset I've realized investors think on a dimension.

that's very different from mine the world they see is something I can actually learn from so recently I've been talking with some investors and I found the views I couldn't accept before they've actually given me a lot of new inspiration so in the process of shifting your mindset you'll find that the thing for you is no longer as hard as you imagined or as upsetting so when you turn things positive your world keeps getting bigger you told me earlier that you categorize investors what categories do you use? each person they probably care about different things some people care more about financials some care more about technology some care more about the product itself some care more about the endgame some care more about founders, and so on I think these are one of their dimensions of analysis and everyone weights these dimensions differently right do you still, like during your last startup painfully read Meditations?

yes, and I've read many, many other books from similar philosophical systems the West's Meditations and China's Wang Yangming and even the Buddhist system the Christian system the core concepts are actually all very, very similar the core concept is really about self-cultivation how you get along with this world what your relationship with the world is how you see the world how you can shift your mindset how you can live a better life so when you understand yourself better when you understand who you truly are when you better understand the world's underlying logic you'll find that the things that used to drain you internally were never worth the inner friction what would drain you internally?

I'd agonize a lot over how others saw me I think I'm a bit of a people-pleaser so I'd care about whether, if I didn't do something right this person would dislike me or or if I didn't think a problem through clearly whether this person would think I'm dumb or even sometimes I'd like, give a talk and then I'd watch the video afterward and think why did I say it like that and there's a pause right here and later I found a book that put it really interestingly it said the person in this world who cares most about what you look like is yourself that's really it so once I figured this out at some point I pretty much stopped caring I just thought I'm just two minutes of someone's video-watching they finish this video and move on to the next one so it doesn't matter right, exactly for me I felt it was really, really important but in the end it wasn't there's no need to overthink it at all.

you said you're not doing this to prove yourself anymore so why do you start companies now? what's your driving force now? I want to better live as my true self because I think every person when they're first born is actually in a very pure state a state where a mountain is just a mountain this thing is just what it is and I should just do it this way and as you slowly grow up you find that society has this rule this thing you can't do and if I do this thing people will laugh at me and I don't like being laughed at it's kind of like Adam and Eve eating the apple of knowledge and I think living like that you increasingly see a mountain as not a mountain so I want to live in a way where I follow my own heart because I think that's true freedom true freedom is being able to choose to do whatever your heart desires.

without caring about what others think the moment you walked in I asked you a handsome guy starting a company a handsome guy as a CEO what does that feel like, right? so why do you carry so much baggage? you're good-looking and everything seems great but I feel you agonize internally more than others I don't think I'm actually that prone to it my cofounder York actually told me I'm a very insensitive person so I think it's not a problem and now I'm agonizing less and less I care less and less right and about the being handsome thing I think it's just that I've gotten healthier I think everyone is like this the healthier you live.

the fitter you look and people feel that maybe your complexion or your demeanor is better I think it's a state of me getting healthier and healthier has this latest startup brought any different experiences? I think it's very different because this time, for me I really cherish this opportunity to start a company because going back to who I fundamentally am I truly love entrepreneurship as a field so when you love something you keep wanting to know the areas where you want to improve, and compared to others where exactly you can stand and I think robotics is a World Cup I've told many people this I think it's a World Cup because the strongest people from every industry are all in this arena and everyone has their own strategy their own playbook and a lot of what people are doing I deeply admire I deeply respect so in this industry I come in with a sense of awe so I really cherish it and I want to play this game do it and I hope that in the process I can live more expansively and also improve my abilities.

put my craftsmanship hone my craft that's the first thing and I think maybe what's left, I think, is that this time what's very different is that the complexity of the problem is extremely high so it really challenges my personal ability to set direction in a certain sense it's a higher challenge for me how do I, in a very noisy, very difficult, very complex.

environment where the technology hasn't converged yet see a clear path that's a huge challenge for me but I really love it I think it's truly a really interesting thing seeing a clear path does it sometimes feel shrouded in fog? I think it always is this whole industry has always been shrouded in fog but if you can find your own path through the fog that feels very fulfilling to me are there things outsiders get very right about DYNA and things they get very wrong? what they get very wrong is that we're a deployment company actually, from day one we were never building this company with deployment as the core thesis and what they get pretty right is that what we're building is valuable right what outcome for this startup would you consider a success?

what outcome would you consider a failure? I think success is very hard to define but from a company's perspective if we can get every single person using our robots I think that's one mark of success giving everyone the possibility of using our robots everyone yes, every single person and for me personally my own definition of failure is a failure of mind cultivation if I truly believe this thing has failed then I've actually failed it's a state of mind cultivating the mind do you have any rituals? in your daily life, besides discipline mind cultivation is a long-term like, do you meditate or something? mind cultivation and meditation aren't quite the same system but meditation does help with mind cultivation because meditation itself can quiet your mind so you can think properly think things through I think mind cultivation is your view of the world at every juncture at every challenge there's always a way to get through it some ways involve actually changing.

other people and changing the world to get through and some difficulties can't be overcome by changing other people or changing the world those are your own difficulties and if it's your own difficulty you have to transcend it through mind cultivation have you ever gotten through a difficulty by changing yourself? many, many times any in the last two years? yes for example when I was lost on strategy that was actually a huge challenge for me I just couldn't figure it out at the time end of last year? yes, end of last year, I felt I just couldn't figure it out and the team would feel so what are we doing which direction are we heading and then later I let go of some of the baggage in my mind or the noise that was interfering with me and only then could I think with real clarity.

to solve the problem I think a lot of the time solving a problem isn't about how many hours you work because I came from investment banking working 100 to 120 hours a week Actually, that was the norm back then But I found that a lot of the time some things can't be solved just by piling on more time You need to switch to a different mindset to solve them You block out a lot of the noise and let go of a lot of baggage.

What was your baggage? The baggage of having to be a CEO who never makes mistakes I was really afraid of making mistakes back then Because I was, in a sense, a founder who had already succeeded once So the baggage I put on myself was that I couldn't afford to make mistakes Because everyone came for my experience But I felt that under that baggage it actually made a lot of my decisions really unwise Or even decisions made just to avoid making mistakes And I think decisions made just to avoid mistakes are fundamentally not good decisions I saw you said something before: you wanted to prove you weren't just a flash in the pan that's right That was my baggage back then Do you still think that way now? No That was a self-proving kind of thinking , and it's it's a mindset shaped by how the outside world judges you But I feel like how the outside world judges me doesn't matter anymore So it was at the end of last year that you got through that mental state So unds like the end of last year was really important for you The end of last year, early this year, yes When I got through that state, I truly, after I got through it, I felt, from the bottom of my heart,.

grateful. The first thing, the first emotion that hit me and had a huge impact on me was gratitude I was grateful to my teacher at the time Grateful to my team Grateful to my wife and my kids Because in that state, they truly stood behind me My two co-founders really backed me all the way So I felt that as a CEO,.

I had made so many wrong decisions just because I couldn't afford to make mistakes And yet you were still willing to bear with me, and and still, in the same way, support me in getting this thing done So I was incredibly grateful at that time So this state, for me, is now, for me, a really good state to be in What wrong decisions did you make last year? I think it was really about strategy — the strategic direction wasn't very clear And when it came to company culture, I don't think I spent enough time tending to the company culture So at the time I felt the team's cohesion wasn't strong enough And that led to both the projects we needed to bet on and the things we needed to do we made some fairly wrong — not wrong technical choices exactly just ones I felt weren't first-principles technical choices and they wasted some of the team's time Does that have to do with deployment versus research?

That either-or choice? Not particularly It's more about the next step in our strategy Let me put it this way I don't think we had thought through the next step of our strategy and then we just started looking at what others were doing And we felt we had to catch up So we lacked our own judgment of the market and our own first-principles thinking So the team ended up in a kind of drifting state So meone ships something and we'd go look at that So meone ships something else and we'd go look at that too Being pulled around like that was really painful We were constantly in a state of being led by the market's rhythm And then, when we truly found what was most essential to us a direction that felt completely coherent to us the whole team started to converge And it became a very, very stable team with a really, really strong momentum That process itself was a state of everyone being in it together.

I think some barriers some barriers in your own mind actually require other people to help you get through Every time someone hits a hard wall, it's really the ceiling of your own cognition And as for what's above that ceiling, what exactly is up there you actually don't know Because it's not within your cognitive range So when and sometimes I break through on my own And sometimes it's really, really hard and you need other people's reminders and awareness to break through So back then, I had a teacher by my side I had people who loved me deeply my team who cared about me, and and so on, and together we only then did I realize this is the way this path should be walked Because I think thoughts scatter outward very easily How do you pull your thoughts back in?

Have you learned how? I don't think I've reached a level where I can describe it very simply in words how my view and understanding of the world has changed But I think the root of your thoughts may be whether you can offer goodwill to this world I think if you always walk this path from a place of treating people with kindness I think you can go very far Maybe that sounds a bit abstract, haha Rather than proving some kind of ego , truly getting to know yourself truly understanding what kind of person I really am what the state of "seeing the mountain as a mountain" versus "seeing the mountain as not a mountain" looks like And gradually clearing away the fog And gradually discovering what your purest self looks like I think this is a state of mind That changed this year I changed enormously this year.

I went through an inner rebirth Is that very common for founders? Common, — on the entrepreneurial journey, constantly seeing yourself and then making changes I think this is a path Everyone has their own path And that path is the direction they most yearn for For some people, that direction is achievement on the material level And for some, it may be an achievement of the inner self But I think, at the most fundamental level, every founder runs into hard walls And when they hit those walls, how do they solve the problem? Do they solve it by changing the world to fix the problem or by changing themselves to fix the problem? Actually, everyone has their own way of walking it I think, fundamentally speaking, I feel that maybe if it's by changing yourself changing your thoughts and changing the lens through which you filter the world to walk this entrepreneurial path I think it lasts longer Because I think starting a company is a really, really long process And in this long-distance race, if you have no way to.

keep improving yourself and become more and more certain of what you're doing I think it's very hard It's what people call internal friction Your spirit just can't sustain you You don't have enough stamina to finish this marathon And this marathon is different in that you don't know where the finish line is , so you just keep running , so that's also perhaps more — you're absolutely right It's an even harder thing When you don't know how long you'll be running.

And you keep running down this road And you completely, completely can't see the finish line and don't know where it is and don't know what the road ahead looks like So back then, how did you steady your own mindset and keep running? I think this is something every founder in the robotics industry has to face It's a thing . I feel like the moment you start running it's actually not that hard But when you get to the middle of the journey and can't see the finish line and can't see anything around you either that's the most painful part , yes But until the end of last year, you'd only been a founder for a bit over a year Why did it come so fast? Why did it come at that point? Hard times themselves actually don't have an absolute relationship with time It's that everyone, at every stage, has their own hard wall And at every different period too For some hard times, they choose not to talk about them Or some hard times, you feel, aren't actually that serious Take me getting through this one, for example From the outside world's perspective, it actually wasn't wasn't a big deal It was just a very simple.

change of strategic model But at the time, given my mental state back then and my own understanding of the world at that point it was, for me at that time, a really hard wall And so I And getting through this hard wall doesn't actually mean I won't face more hard times in the future So I think everyone is actually going through hard times It just comes down to how they themselves perceive that hard time So suppose this venture of yours turns out worse than your last one How would you feel? Actually, I think it would be the same Keep a level head Fundamentally, I'm here to win I want to play this World Cup And I feel I played with my own style I played as my most genuine self I played my heart out And that alone makes me really happy Because I think in life, actually, you can't always be staring at the finish line Because if you're staring at the finish line, you can never enjoy the process Actually, I think in life we're always on a journey.

Every milestone is just a checkpoint not your goal So if you have no way to enjoy the journey and just keep fixating on the finish line then I think you'll find this journey is pure agony Maybe that's also what I — it was my view before I shifted my mindset I felt this — I kept fixating on the outcome I just had to make this thing happen I just had to do this and that And it actually made me not really enjoy the process So now, my thinking now is completely different Fall into a pit climb out, fall into another pit climb out again That's really just — this is what I think a lot of founders need to go through Last few rapid-fire questions Okay One food you love, from anywhere in the world If I have to pick just one, I'd say Fuzhou banmian noodles and bianrou wontons It's from my hometown One place you love, anywhere in the world I'd say wherever my wife and kids are, I love it How old are your kids?

I have three daughters One, two, and three years old So you married early No, I got married when I was 30 A piece of knowledge — A little-known but must-know piece of knowledge It can be a piece of trivia Does cultivating the mind count? Every guest of ours recommends to our audience one must-read book for life Hopefully one that truly had an impact on your life That would be "Life Has Only One Thing" (by Jin Weichun). The papers that, in your mind, have influenced the course of robotics I'd say Jason's "Eureka" Actually, this paper itself redefined how robots might self-learn in the future And it's also the main paper through which I got to know Jason Based on your current understanding, what's one key, important bet?

Infrastructure, I'd say You guys put a lot of effort into this, right? Infrastructure Very important, because actually it — your model iteration speed and how far your model can scale up — I think both are extremely important Our studio is called the Language Is World Studio When you first heard that name, what were you thinking? From a robotics perspective, it's really about how robots perceive this world — it's really through language and the world So for me, actually, both modalities are very important — the most important modalities for robots The world is vision , the world is vision and audio physical properties and all the sensory inputs And language is, for robots, their understanding of the world and of language turning language into action.