Right? The internet, the worldwide web took all of our digital information and made that searchable and browsable to humans. Now, we need to turn that upside down and make the physical world accessible to AI. The proof is in the pudding already, I feel. You know, like Unitree robots are absolutely everywhere right now. And people have started making money with Unitree. But no one's making money off of Optimus or Figure.

I remember one of my first visits to a robotics OEM in Changzhou. They had a team of just 20 people or so, but they had this huge campus. It's like, "How can you afford this big space?" It's like, "Oh, no, no, no. This was a gift from the government." Right? No one wants to go to war with their own factory. I see a path to global stability and world peace where the big powers are so tightly interlinked with each other that it's hard for them to even imagine going to war.

We have with us Nils. He is founder and CEO of Auki Labs based in Hong Kong. Thank you for doing this interview.

Happy to be here. Good morning.

Good morning. So you are in San Francisco.

I am.

How is the visit going so far?

It's always very nice coming here and tapping into what the robotics ecosystem is doing here. There are some pretty big differences in approach, I find, between what the labs in China are doing and what the labs in San Francisco are doing. So it's always fun to come over, meet the local talent here, see how they're thinking about these problems, and facilitate a bit of exchange between China and the US.

Okay. So in this interview, I wanted to cover your story, what Auki Labs is doing—which is a completely new paradigm in technology, the spatial intelligence and the networking challenges that come with it—and later on, I also wanted to cover the dynamics with China's rise as an empire and building hardware and manufacturing. So starting off with your personal journey, how did you fall in love with technology? What were the drivers growing up that pushed you into that direction, and tell us a bit about your journey that led you to found Auki Labs?

Sure. So I was very fortunate to have software entrepreneur parents. My parents started a tech company actually when I was conceived. My parents took parental leave when I was conceived and used that parental leave to start a software business that ended up doing quite well. So I grew up with computers. I think I had my first computer in my room when I was like five or six years old and was very early on the internet.

I like to say I grew up on the internet. Actually, my English is better than my native Swedish because I spent so much time online. Most of my formative friendships growing up were with people online. And I got into competitive gaming, I got into online culture. Eventually, around 2009 or so, I started my first company together with a co-founder that I had met online. We had never met in person.

The first project that we took on together was helping build out an intracortical neural interface for ASU's sensorimotor lab under Dr. Steven Tillery. We moved to Beijing to do that, to make the money last longer. I had just visited Beijing a couple of months prior, realized that it's a very cool place to hang out. So yeah, we started working on this neural interface.

At that time in my life, I was also falling deeply in love with the concept of meme theory, this concept of how culture and ideas propagate through populations. I went down a very deep rabbit hole related to meme theory and started doing meme theory consultancy. Eventually, the business that I had started with my friend in Beijing became a fully-fledged product company where we were building out real-time analytics tools to track behaviors in gaming populations, social media populations. That's when I really started getting into tech.

Auki got started in a way in 2019, when I wanted to create shared augmented reality overlays for the tabletop game Warhammer. So, very different from what we're doing today. I had just fallen in love with this tabletop game.

So, it's arcade? Is it digital or—

It's a tabletop game, a physical game. You paint little miniatures, you put them on the table, you use measuring tapes to see how far... It's like a very dirty version of chess, right? And this was in the beginning of COVID—actually, I started it just before COVID, and when COVID broke out, tabletop gaming had a big surge. I realized that there were millions of people every month watching these tabletop games online, and I thought it would be a lot better if we could bring some augmented reality overlays to this game to make it easier for new people like me to understand what was going on.

What I didn't know at the time—like, I didn't know anything about spatial computing at the time. I had, I think like many people, a lot of naive ideas about how digital devices think about the physical world. So I didn't know the limitations of the GPS. I didn't know the limitations of—I didn't know concepts like drift and SLAM and things like this. And just fell in love with that problem space.

Essentially, what I needed to make happen is I wanted to have it so that on your phone and on my phone, we could see the same digital overlay. And to do that, our phones need to understand their precise location relative to each other down to under a centimeter. And you can't use the GPS for that. GPS doesn't work indoors at all, and even outdoors it doesn't have that degree of precision. So at first I was very depressed, like, "Oh no, oh no, this thing is almost impossible to build."

But then I was like, "Ah, this thing is almost impossible to build, but we're going to need it." In a future where everyone's wearing AR glasses, this technology needs to exist. In a world where robots are navigating around, we're going to need this. So instead of giving up, I took the opportunity to just sit in my room during COVID and think deeply about these issues, work together with my collaborators. By 2021, we had invented a couple of new techniques for establishing shared coordinate systems.

Okay.

And that allowed us to get some very generous seed funding to build an open-source protocol to allow for app-agnostic, hardware-agnostic exchange of spatial information in a post-GPS era, if you will. That's what we've been working on ever since. When we started building it, we called it a decentralized machine perception network, which is technically true but also a mouthful and people found it a little bit intimidating.

These days, we call it the Real World Web. And the point of the Real World Web is to make physical spaces like this searchable, browsable, navigable to visiting devices. Right? The internet, the worldwide web, took all of our digital information and made that searchable and browsable to humans. Now we need to turn that upside down and make the physical world accessible to AI. The reason I think that's so important is 70% of the world economy is still tied to physical locations and physical labor. So if we want AI to help us with most of the world economy, we have to make the world accessible to AI. LLMs don't see the world. They don't reason in 3D. Even these latest robotics models struggle with long-term spatial context, etc.

So the way we've come to think about robotics as a stack is we think there are six important layers of robotics in no particular order. There's locomotion, of course—the ability to move around the world. Manipulation—the ability to, you know, grab things. But also perception—more specifically, you want spatial semantic perception. You need to be able to tell the difference between a cup and a potted plant, and you need to be able to tell the difference between something that is far away and something that is close. So you need spatial semantic perception.

But once you have spatial semantic perception, you also want to have a memory of where you saw things. And the mundane name for a memory of where things are in space is a map. So you want mapping, actually. And once you have a map, you actually need to be able to return to that map—understand where you are in relation to the map. And the mundane name for that is positioning.

And then finally, you need applications, deployments—like actually how are we going to tie all of these things together into one deployable application. What we work on is... we don't work on locomotion policies. We don't work on manipulation policies. We work on collaborative perception, collaborative mapping, collaborative positioning, and closing the deployment gap. How can we get devices out into the world doing physical AI workloads, be they a phone, a pair of glasses, or a robot?

One of our core beliefs is there's actually plenty of perception tasks in the world that are more valuable than many manipulation tasks in the world. And closing the deployment gap for these perception tasks is, in our belief, the fastest path to a billion ARR in robotics. So we think that a lot of other labs are over-indexing on the importance of manipulation and not thinking enough about how can we get enough robots out there that the revenue starts becoming significant.

Right. Right. Yes. You mentioned the collaborative part of it. Can you give an example of where you can use a collaborative creation of maps and then make it useful for a customer?

Sure. So since our roots were in augmented reality, we started with the assumption that an iPhone is basically a robot with no arms and legs. You know, it's got a camera, it's got an accelerometer, it's got a LiDAR, it's got many of the same sensors and capabilities as a robot. So, let's treat the iPhone as a robot with no arms and legs. It's going to borrow your arms and legs, right? You will carry it around.

And we built, among other things, an augmented reality task manager for retail workers where the store manager, if they find something that needs to get done, can leave a note in physical space with a photo, with a description. And even if I'm not there, I can get guided to that task in augmented reality later. The reason that was important, I learned by interning at grocery stores—actually putting the apron on, going out and working in the store, seeing what the day-to-day is. And I found that a big part of my day was just waiting for the store manager to tell me what to do, looking for the store manager, the store manager looking for me, and vice versa.

In fact, one of Europe's largest retailers with 17,000 employees recently flew out to visit us in Hong Kong. They told us that they estimate across their 17,000 employees that they lose 30 to 60 minutes per employee per day, predominantly in the mornings, due to them not really knowing what to do, where to start. And this brings us to the robots, right? And why collaborative matters. If you can get a robot and a phone into the same coordinate system and reason about space the same way, then the robot can actually populate that AR task manager.

So, what we realized is that, hey, what these guys need right now is not a robot that can do the restocking or whatever. They have humans that will do that faster than a robot will. They need to activate their human workforce better and make the human workforce more productive. They need a robot co-pilot. They need a robot store manager, actually, that tells them what to do. So, thanks to collaborative spatial computing and having all of these devices in the same coordinate system, we can now have the robot generate tasks in the AR task manager, but also vice versa.

You can use AR to tell the robot what to do. Like if you think of a cleaning robot, for example. Today, cleaning robots run on schedules. If someone drops a bottle in an aisle and you want the robot to come clean up just there, it's not easy to tell the robot "come clean right here." But augmented reality allows you to do that actually. If you have a phone that is in the same coordinate system as the robot and you can look at the spill and say, "I want you to come clean this," you can summon the robot.

In fact, one of the coolest tech demos we did just before the end of the year last year was we put on a pair of upcoming smart glasses that are coming out in a month, and we streamed the video from the glasses to a SLAM server on the edge that tracked the 3D location of the glasses in real time. And then we called a humanoid robot to our location. So the humanoid robot could not see me, it didn't know where I was. Or actually, it was my colleague Ashok who was wearing the glasses, and Ashok said, "Hey Unitree, come to me." And the robot, because it's in the same coordinate system, because it has the same sense of space, could go to Ashok because it knew where the glasses were.

So this is what I think is the point of collaborative spatial computing. When you have human devices like your phone and your glasses in the same coordinate system as the robot, it facilitates the communication with the robot. The robot can communicate with us more clearly and we can communicate with the robot more clearly. And we think this is going to be important because robots are going to want to work in the same places as us. It's not all going to be dark stores. We want robots that can work with us in the human environment next to human workers. So how do we communicate with them?

Okay, you have your collaborative mapping... would work with anything which has a camera on it, right? So it could be a stationary camera, it could be a phone, it could be robot or wearable.

Indeed. Right. So you can take a stationary camera like a security camera, you can calibrate it into the coordinate system. We can do that today by having a phone that knows its position show a QR code on the screen. Show that QR code to the camera and calculate in real time the position of the camera so it knows where it is relative to the map. And then give it an arbitrary computer vision task, like look for spills for example.

And if it sees a spill, because we know the precise position of the camera and the precise topography of the store, we can ray-cast and place the spill in 3D space so that the robot knows where to go. So yes, you can work with stationary cameras, with mobile cameras, any. That's the point. That's why we called it a machine perception network. We wanted multiple devices to be able to exchange information about the world with each other. And we said it was decentralized because none of this is like going to our cloud or anything like this. We allow for ad hoc networks to form where robots can shake hands with each other and start exchanging information. Again, now we just call it the Real World Web, but in my heart, you know, it's still a decentralized machine perception network.

Right. Right. Right. And so lot of devices don't have inbuilt compute or not enough inbuilt compute to process things locally, you know, create maybe a cloud of maps and communicate with other devices. So how are you going to tackle that compute challenge?

So that was actually one of the key realizations that made us want to build this. So as I had been working on augmented reality from 2019 to 2021, I became convinced that the glasses were coming. It was very clear that big companies like Apple and Meta were putting a lot of effort into the glasses and that the glasses were going to come. But it also seemed like the biggest barrier to having all-day glasses is just... you can't put that much compute on someone's face and that much battery on someone's face without it becoming prohibitively heavy.

Like the guys at Mantra, whose office we're in now, right? They told me something very important. They said, you know, glasses were invented 800 years ago, but it wasn't until the 1920s or so that people started wearing glasses all day. Because in the 1920s, material science had gotten to a point where you could make glasses that were around 40g. And it turns out that if they're heavier than that, we have 800 years of history that tell you that the glasses will end up in your pocket. So there's no point making 100g AR glasses, something like this. People will take them off.

And that means that okay, if we can't go over 40-45g or so, then we're not going to be putting very much compute on the glasses. We're not going to be putting that much battery in the glasses. So we're going to have to collaborate with compute somewhere else. And unfortunately, you can't collaborate with a cloud 'cause the latency is too great. So in 2021, we formulated this hypothesis that the future of AR glasses is going to be glasses that talk to a hyper-local compute resource—like your Mac Mini at home when you're at home. And when you go to Walmart, it will connect to a computer at Walmart. And when you go to the office, it will connect to a computer at the office. And that the Real World Web has to be this directory of hyper-local compute resources that you can interact with.

In the end of 2024, we showed our first end-to-end demo of a wearable with very, very little compute streaming its sensor data to a hyper-local compute resource that did the SLAM on its behalf, so that we could track the wearable in 3D space in very, very close to real time without doing that compute on the wearable. And that's what we're doing this year with the glasses that are coming out this February. You can stream RGB and IMU from the glasses to a hyper-local compute resource that keeps track of where the glasses are. And because you're only streaming video, you're not doing SLAM on the glasses, you get much, much better battery performance.

Okay. And you, you're hyper-focused on decentralized approach. And with centralized approach there are several challenges including privacy. So how does decentralized fix the privacy challenge or you know, big tech looking into our private lives?

So I think there are two big practical challenges with building a centralized solution. So one is latency, right? For AR glasses at least, but probably for a lot of robotics things also, I think we're going to find that having very low latency is very, very important and there are just limits to what kind of latency you can get to someone's cloud. For AR glasses, it's just not going to happen. It's just not going to happen. And I think the jury is still out on how helpful cloud is going to be for the robots.

The other issue is privacy, as you said. I learned with my previous startup—that was a big data startup where one of my advisors was heading up data for one of the world's largest retailers—that many retailers are so nervous about sharing information about how the products are placed in their store, like their visual merchandising strategy, that they're literally not allowed to store it on a cloud full stop. Right? It has to be on-premise. Cause they're so nervous about, you know, Jeff getting his grubby little hands on that data.

So that made me think that okay, there are two big opportunities here to compete with the likes of Niantic and Google etc., by just biting the bullet and saying data is not going to be how we make money. Data is not going to be how we make money. Instead, we're going to allow for completely sovereign, self-hosted digital twins where all of the data stays on your machines, but you can still give role-based access to any visiting application or visiting piece of hardware that you like. So that the role that we will take is to facilitate those connections rather than being the repository of all that data. Which is kind of how the internet works, right? Like you can self-host a website. You have a choice of where to host your website. It's not just all Facebook or all Wikipedia. And we believe ultimately that that's how the Real World Web needs to work as well. You need to be able to, if you want, entirely self-host the thing and still be discoverable by other parties and work together with other parties.

Okay. General question about robotics. There is a lot going on right now in general purpose robotics and there are kind of two different kinds of companies where in the US most of the companies are pursuing kind of moonshot ideas where very general purpose robots being useful everywhere, and there are Chinese robotics companies—they just want you to use your hardware. So with what kind of approach do you prefer? Are you approaching companies that are kind of closed to, you know, work with you guys?

So, obviously, it's easier to work with the robots that are programmable, right? Because then we can do what we want with that robot. We've been in touch with some of the American hardware companies, but it's hard to collaborate with them because the stance that they are taking is that they're not going to release the robot until their AI is so good that it doesn't need integration with anything else. Now, I don't believe in that approach at all. I think that's going to fail.

I think we're very far away from an embodied general intelligence that can, you know, change your cat's litter box one minute and then re-string your guitar another minute. I think much more likely for at least the next decade, history is going to repeat itself and we're going to have something like an app store for the robots where you purchase specific capabilities for the robot. And I think that many Chinese OEMs have understood this. Like Unitree and Agibot and Booster—a lot of them are embracing this app store approach where, okay, we're going to make our robot programmable, we're going to let people like Auki develop capabilities for the robots, and hopefully we can get a cut.

And I think that's a very healthy approach and the proof is in the pudding already, I feel. You know, like Unitree robots are absolutely everywhere right now and people have started making money with Unitree robots already. But no one's making money off of Optimus or Figure, you know, you can't get your hands on them. So I don't believe in the closed approach that people like Optimus and Figure are taking. You need to open up to the developer community to allow the robots a fair chance of getting deployed actually. I think waiting until you have a robot that is so intelligent that it doesn't need any kind of integration is wishful thinking.

Okay. And in the robotics ecosystem, you mentioned several layers, six layers, right? Where do you think there is like a significant value to be captured, and is there a kind of historical analogy in maybe some other technologies in the past that gives you the idea, you know, this is the area where there is a lot of value to be captured?

So I think right now most of the teams here are so focused on locomotion and manipulation that sometimes they even think those are the only components to robotics. But you know the 1X Neo, right? They are going to charge $500 per month for that robot. My friend in Singapore has a mall security robot company; he gets to charge $5,000 a month for a mall security robot that is only really doing perception tasks. Which again goes back to my point—I think there are many perception tasks in the world that are actually worth good money. And you should focus on closing the deployment gap there instead.

There are on US soil right now, there are tens of thousands of undeployed Chinese robots that haven't gotten deployed because the software was literally just too Chinese. It was hard to deploy and maintain because the software was in Chinese and people had a tough time working with it. So there's a huge opportunity just to make proper English software for these robots that are already on American soil and get them deployed.

And I think the sentiment has kind of been changing in real time. Oliver Hsu wrote this physical AI deployment gap article like two weeks ago, three weeks ago, something like this. And I can already tell that a lot of VCs are starting to change their opinion of what is fundable and what they are looking for in 2026. In 2025, almost every VC I spoke to was looking to fund, you know, world model companies or data collection companies. But now I'm hearing, "Oh, I'm looking for people with a clear path to revenue that can create good flywheels by having robots actually deployed."

And I think there are many, many opportunities to deploy perception—like mobile perception robots—that are scalable to billions really, without having to solve manipulation, which no one knows how far away we are from general purpose manipulation. So that's our big venture scale bet for 2026 is like: hey, let's make it super easy to deploy perception robots. And manipulation is downstream from there.

So to change the topic a little bit... you've been living in China for how long?

Since 2009.

So yeah, almost—yeah, more than 15 years now. So you have kind of a unique perspective of insider/outsider in China. You've been working, you've been observing people, the work culture there. Can you explain what's behind the speed at which China is not just building manufacturing capacity, but also there is a lot of innovation going on there in robotics, and the most striking is the humanoid robotics. So, can you shed some light?

So I think there are many interesting reasons for why China has been doing so well recently. We can't all put it down to just one. But in no particular order, one thing that I think we do talk a lot about in the West, that we are aware of, is the demographic layout of China. They have an aging population. A lot has been said in the West about the potential demographic collapse that China is facing, and they're obviously very well aware that they do have an aging population. As a result, the government has had as a strategic goal for a long time now to automate a lot of labor. So robotics and AI has been a government-level initiative for a long time where they're putting a lot of money and grants behind robotics companies, computer vision companies.

And I think that what we've been blind to in the West is what those sources of funding are and how much they get. I remember like two years ago reading that, you know, "Chinese AI was doomed to fail because there's been this collapse in venture funding." Like look, "Oh, there's no venture money going into Chinese AI." But part of the reason why there's no venture funding going into AI is it's hard to compete with non-dilutive government grants. And you can get a lot of government money in China for developing these things.

I remember one of my first visits to a robotics OEM in Changzhou. They had a team of just 20 people or so, but they had this huge campus. It's like, "How can you afford this big space?" I was like... "Oh, no, no, no. This was a gift from the government." Right? They just want to support us in building this. Then I went to Shanghai and visited one of the world's largest OEMs there. And their campus was so big that we literally Ubered between buildings. And again, massive, massive government subsidies to attract them to build in this part of Shanghai. So, you know, money from the local government, money from the CCP.

So there's a lot of government money that can be had. Just recently there was this computer vision researcher that I wanted to poach away from a lab at a university in Hong Kong and realized that that's going to be very hard to do because he had millions of dollars from the government that he could spend at his discretion for whatever he wanted. Very hard for a private sector company [to compete]. I think he had $20 million that he could spend, right? Just, you know, that's his budget for that year for that robot lab. So whatever he wants to do, he can just do it.

And that's obviously driving a lot. So the demographic—or the looming demographic collapse—has led to very, very heavy government investment. But I think there are some other two other very interesting reasons why China has done so well, and they both kind of have to do with population density. The world's largest freight harbor is Shanghai. And Shanghai moves five times the amount of cargo that Long Beach, the biggest harbor in the US, moves. And the second largest freight harbor, your audience may be surprised to hear, is Shanghai's second harbor, Ningbo, which is a suburb of Shanghai.

And three of the top 10 ports are in what's called the Greater Bay Area, which is the area around Hong Kong and Guangzhou and Shenzhen. So that area is around two-thirds the size of greater Los Angeles. The distance between Hong Kong and Guangzhou is the same distance of Santa Monica to San Bernardino in LA, but there's 87 million people there. And more high-rise buildings than in all of North America and all of Europe combined. In fact, New York is the biggest or the tallest city in the West. They have over 900 buildings taller than 300 ft. Just in Hong Kong, there's over 4,000.

Why does that matter? Well, it matters because what I told you about the GPS. The GPS is actually a line-of-sight technology that doesn't work well in urban canyons like Hong Kong. So there's a lot of research happening for many reasons into computer vision and spatial computing and positioning in the Greater Bay Area. Some of it also for, you know, perhaps darker reasons. Like before OpenAI, the biggest AI company—like purely AI company—in the world was a Chinese company called SenseTime which was founded in Hong Kong. And they are computer vision specialists. In fact, they are so good at computer vision that they could track the Hong Kong protesters even when they were fully masked and, you know, switching clothes and stuff, 'cause they had such good gait analysis among other things. Hong Kong protesters were putting rocks in their shoes to mess up their gait so that they wouldn't be recognized even when fully masked.

Like, SenseTime is just light years ahead of anyone in the West on computer vision. And DJI, the world's largest drone company, was also founded in Hong Kong. And what these companies have in common is that they have to be very good at computer vision. They have to be very good at spatial computing and they have to deal with the realities of this incredibly high urban density that they have.

So I think those are the three main things that are behind China's success. So government investment because of the looming demographic collapse. High population density leading to great supply chain—right, like having these three of the top 10 harbors in one place just makes the supply chain so easy to deal with. And then finally, all this urban density being a forcing function for getting very good at computer vision and spatial computing, but also making things like drone delivery just more viable. Like as a business model, it is more viable to do drone delivery in Shenzhen than in San Francisco because there are more people that you can reach within a battery charge if that makes sense. So I say those are the three things in my opinion that have led to China's recent success.

Okay. What do you think about so many companies pursuing humanoid robots? Like just another day, Chinese ministry announced there are 140 OEMs making humanoid robots and they made about 300 models just in 2025. So how is that going to play out the dynamic of who's going to win?

Yeah. So I've personally met over 40 Chinese humanoid OEMs. There's a lot of them and a lot of it has to do with, again, local governments funding these initiatives. In China, there's a lot of competition between local governments. So, you know, this city wants to outdo that city. So basically every major Chinese city now, of which there are many, are putting money into these kinds of robotics initiatives. And of course a lot of those are going to go to zero. Of course, a lot of them are going to go to zero.

And the CCP is also aware. Have already issued a warning saying like, "Hey, we think that you guys are overspending on humanoids and we want to see more focus on actually deployable robots." But they are at the same time also pushing on the gas and helping facilitate deals to be made. So, one of the largest OEMs did like a $10 million humanoid deal last year, and that's a big humanoid deal. And it turns out that the buyer was the parts manufacturer. So, the company that makes one of the critical components ended up buying $10 million of the finished robot because... and like, I don't know this for a fact fact, but it seems like what's happening is that the government is going in and saying, "Hey, we want this ecosystem to work. So we're going to give you tax write-offs. We're going to give you all kinds of benefits if you just make sure that all of this works. We want to know that we can make the robots."

Yeah, and that's obviously been working very well. It's been working very well for the EVs. You know, when I first moved to China in 2009, the Chinese auto industry was very, very unimpressive. And now it's very clearly one of the best auto industries in the world. Chinese cars were just terrible in 2009 and now they are spectacular. And that has happened again because there's been very strong government push like, "We need to get good at this." And they seem to be betting on the right horse. So in 2023 they had this guideline for the industry that "we're going to build end-to-end supply chain for humanoids by 2027." And they seem to be well, well on way to doing that.

And so betting big on something which is far into the future would need some sort of, you know, first principles vision. Like where is that coming from?

Well I think the CCP is in many ways a very technocratic party. The people are very technically competent. They don't have to win popularity contests the same way that Western politicians do. And even the Hong Kong government—like the first time the Hong Kong government came to visit us to check out what we were doing, they sent a delegation of 20-30 people and more than half of the people that visited wrote code. Right? And these are government officials that program.

This was two, three years ago. I remember it was just a few months after the initial Gaussian Splatting paper had taken off and these government officials had already tried implementing this. They had already deployed it live for a couple of use cases. Like very, very technically sophisticated though. And obviously the universities in China are very good now. Like 10 years ago there were only two Chinese universities in the top 15. Now there are only two that are not Chinese in the top 15. So there's plenty of talent to go around.

I think we often tell ourselves this comforting story that China only copies. And maybe that was true 10, 20 years ago, but it's not true now. You see a lot of innovation in Chinese EVs. You see a lot of innovation in Chinese robotics. And the joke in Mandarin in mainland China is that the robotics race is actually between mainland Chinese people and overseas Chinese people. Like that's where the real competition is. Because when you look at teams like Meta FAIR, at Nvidia or whatever, you know, all of the very well-funded teams here, a lot of them are predominantly Chinese.

Okay. Wow. Wow. For the US companies or for the geopolitical scenario in the US, there is a lot of talk about bringing back manufacturing. A lot of humanoid companies say they are vertically integrated but they buy components from China. Is it fool's errand to think like bringing back manufacturing in the traditional sense? Should it be more collaborative with China?

So I believe what we all should want is a world without military conflict. And I believe that what is motivating China is making sure that they cannot be bullied militarily. I think they have a deep trauma from the 19th century and the, you know, the century of humiliation, and they want to make sure that they can't be bullied again. I believe that what they're going to do this decade or the next is open up the Greater Bay Area to be just as open as Hong Kong. So it becomes a no-brainer for Western companies to come build their hardware there because no one's going to go to war with their factory. No one wants to go to war with their own factory.

So I think it's in China's best interest to avoid war to make sure that everyone is building in China. And I see a path to global stability and world peace where the big powers are so tightly interlinked with each other that it's hard for them to even imagine going to war.

That being said, I also understand the case for why you don't want your factory to be able to blackmail you and hold you hostage. It's important to be able to build things at home so that China can't take their football and go home. And I, you know, I don't know what the right balance is there. I'm hoping for peace. I'm hoping for prosperity. And I believe what China is going to try to do is open up so that we will not re-industrialize in the West. It's in their best interest that we don't. But it's probably in our best interest to strike a balance between, you know, the economic convenience of a globalized world and the safety of knowing that if push comes to shove, we can build things at home.

But it's hard in a democratic capitalist society to make the same long-term strategic bets in the national interest that an autocratic state... like an autocratic state can make long-term plans and execute on them. Right? And one of the unfortunate drawbacks of having a democratic society and a capitalist society where the will of the people is guiding us is that it can be harder to make a long-term plan. And I think we've seen that play out a little bit. That China has been able to move at a pace that we haven't been able to move in the West, in part due to it's just hard to make decisions like that in a fully democratic, fully capitalist society.

Yeah. Hypothetically if you are in power, if you are a president of the USA—I mean president doesn't have all the power but hypothetical power—what would you change the way industries work in the USA to compete with China?

I think the West needs to subsidize manufacturing more. Be willing to subsidize the people that are making bets on actually building. Like a lot of Chinese factories aren't even concerned with profit. Like some of the Chinese robotics OEMs aren't even concerned with making a profit. No. They are focused on hitting the KPIs that will get them the next government grant. So some of the Chinese OEMs are difficult to work with because they're just not that motivated to actually commercialize their technology because they have other KPIs from the government that they want to hit to get the next big grant from the government.

So again, there's a balance to be had there. You can't just make people dependent on government money. That will stifle innovation and stifle the ability for things to be properly productized. Like commercializing something forces you to productize it properly. So I think Western governments have to find some balance where they reduce the risk of trying hard things, building hardware, and there needs to be some safety net and some government support for doing these difficult things, but not so much that we end up not learning how to properly productize things. And China hasn't gotten that fully right themselves either. But they are in some dimensions of this problem space they are outperforming the West now and I think there are things that we can learn from them.

Okay. Okay. And tying it back to AI, how is the adoption of new technology different from western world to China?

I am appreciative but also a bit envious of how quickly decisions are made in Asia these days and especially in the greater Sinosphere. The decision to adopt new technology can happen very, very quickly. And in fact, one of the world's largest OEMs told me over dinner last year that they don't bother sending their robots to the Western market so that there's stock in the Western market because the decision-making process for Western companies is so long that is almost always better to just look for a Chinese customer instead because they might make a decision in days whereas the Western company needs months to make a decision.

And I don't know why that is. And it wasn't always that way. Like when I came to Beijing in 2009, I wouldn't have said like, "Oh, wow. Chinese people are so quick at making decisions." But something has changed and now decisions are made very, very quickly. There's this techno-optimism, an optimism for the future that is making people willing to take bets. Some of it is also of course that you can get government subsidies and write offs and they've made it a safe environment to bet on the future.

Yeah, I'm a bit envious of that seeing how, for example, one of our current customers is a large Asian retailer and their roadmap for this year that we're intersecting with is an order of magnitude more ambitious than anything we've seen any Western retailer do. They're just incredibly ambitious with what they want to get done and cost is not a concern. And part of it is because you know they're getting sponsored by Western companies actually to do it. Like Google is putting a lot of money behind this but even so it's their ambition that is driving it. Like they really want to build a very, very future-proof system and even the most technologically forward Western retailers that we know are nowhere near as ambitious as that, but they tell themselves that they are. They're not aware of how hard the Asian retailers are pushing. So, some of our American and European customers really pride themselves on how innovative they are, but they're only scratching the surface of what their Asian counterparts are executing on right now.

And I think we need to wake up a little bit to the pace that things are happening and we need to get better as especially private companies need to get better at just making decisions. I've had a couple of like literal shouting matches with some privately held Western companies where there's all this red tape for no reason to make a decision. It's like, guys, you're a privately held company. You can make this decision right now. You could... I could put this piece of paper in front of you and you could sign it. There's nothing stopping you from doing it. And in China, they would. We need to learn to do it.

Uh, what's where do you think the world is going with AI? Is it going to boost the economy in a way that surprises many or is it going to be you know net neutral maybe a decade or two out?

The honest answer is I don't know. What I believe is that 70% of the world economy is still tied to physical locations and physical labor and I think that's going to grow not shrink as we learn to program atoms instead of bits as the robots get out there. We will learn that what we really want is to make our material existence more comfortable. So we're going to see more of humanity's GDP being focused on improving the physical world. So I wouldn't be surprised if 20 years from now it's 90% of the world economy is tied to the physical world rather than just 70.

So I don't believe that, you know, OpenAI and Anthropic are going to be the big names in AI. The big names in AI are going to be the physical AI companies. It's a tougher journey. It's not as easy to generalize. There's less data, etc., but it's a much bigger market. It's a much bigger market. And I think that just as LLMs showed us that you can get a co-pilot for almost any white collar job, we're going to get co-pilots for blue collar jobs as well.

I think putting on a pair of glasses that lets an AI see what you see and can guide you through whatever you're doing—whether you're stitching together footballs in Pakistan or disassembling spacecraft in Area 51—having a pair of glasses that helps you do that better, I think is going to become an absolute killer feature in the very near future. Five years from now, I think a lot, a lot of blue collar workers will be wearing these glasses and maybe adoption again will be faster in Asia where there's still more blue collar work. We'll see adoption in Indonesia and Malaysia and places like this where they're still working a lot with their hands and someone who's a Toyota car mechanic today will all of a sudden be able to repair any car because the glasses know how to do it.

If I had to make a couple of guesses, I would say 10 years from now, the average urban Chinese family is going to spend more money on robots than on cars. That's going to happen within 10 years. I think that more of humanity's GDP will be tied up in the physical world in the next 10 years than less. I think that AI is going to move to the edge because that's what the robots and the glasses need. And I think we will actually see that this spend on data centers was a historic miscalculation. The future of AI is on the edge. I believe I think we'll get co-pilots for physical labor and within 10 years, you know, the average urban Chinese family will spend more on robots than on automobiles.

Okay. Okay. And you mentioned something called meme consultancy that you did. So I was fascinated like what is it and can you give an example of...

Sure. Yeah. So in the late 70s this biologist named Richard Dawkins wrote a book called The Selfish Gene where he explained the neo-Darwinian perspective. Darwin gave us survival of the fittest. But the neo-Darwinian perspective on survival of the fittest was trying to figure out where does that drama really play out. The neo-Darwinian perspective was that survival of the fittest is not best understood at the level of the individual versus individual or even tribe versus tribe or species versus species. The true drama actually happens down at the gene level. Survival of the fittest is predominantly a battle between different genes and that we are the unwitting vehicles for this drama.

And in this book, Dawkins had a chapter where he had a thought experiment essentially. What if culture was also subject to natural selection? And he coined the term "meme"—you know, before the internet was a thing really—to describe the equivalent of a gene but for behavior. So a meme is a unit of behavior that can be observed and copied. So, the word "meme" is a meme. The word "word" is a meme. Waving to say hi is a meme. Having buttons on the front of your shirt is a meme. All of these are behaviors that can get copied. Having a roof shaped like this is a meme. It's a behavior that can get copied. And the idea was that certain behaviors will just be more successful through natural selection. Like roofs shaped like this will be better than roofs shaped like this because the roofs that are shaped like this will break when it rains and you're not going to observe them as often. So there, like as a meme, they won't be as successful.

And I was very fascinated with this perspective and realized essentially that you can do a multi-dimensional analysis of memes where a successful meme needs to have, you could say, four things. One, they need to be observable, obviously. Like, I can't make a meme of greeting you by curling my toes in my shoes because you can't see that. Like, it's just not observable. So, duh, things need to be observable.

But that's not enough. After they're observable, it also matters how reproducible they are. Like, let's call that fidelity. Like how easy is it for you to actually copy the thing that you observed? This is very easy as a greeting, but like this is a little bit harder. So fidelity matters. How much will the meme mutate with every person that interacts with it?

And an interesting thing about memes that genes don't have is that they can kind of self-normalize. They can repair themselves in a way that a gene can't because memes travel through your brain and you have all of this interesting scaffolding that can help you repair broken signals. One way to illustrate that is, do you know what a game of telephone is?

Yeah.

Right. All right. So, if we were to play a game of telephone and I seeded the game by telling you, "I'm going to say a number and then I say nyang," because I've said it's a number, you hear "99," but I didn't say 99. The broken signal could get repaired by the scaffolding of knowing that it's a number. Memes can do that. And that's very interesting.

And then finally, the fourth dimension of analysis is what are the incentive structures around this meme? Like what is it that will make you incentivized to propagate this meme? So I started consulting for people like the World Bank and for Sina, who makes Weibo, helping them craft more memetically fit strategies. So for example I helped Sina... so back in 2012 or so there were a bunch of different services called Weibo in China. Okay. So "weibo" is just Chinese for micro blog and they all wanted to become The Weibo. They wanted to just own that name but it's not a legal struggle. It's a cultural struggle like how can you be the default Weibo.

And I was brought on by the company that is now the default Weibo to help them figure out how do we do that? So we developed a strategy together where we would go to advertising agencies etc., and just have coffee with them, pick their brain about things, and seed them with a narrative that we knew would be pretty memetically fit. We would talk to them about the rising Chinese middle class. We shared some interesting stats about how quick it is growing and their spending power etc. And the reason we were doing this is because we knew even though our Weibo was not the biggest Weibo, one thing that we did know is we had the highest level of high school graduates, highest level of university graduates. So it stood to reason that we probably had the most middle class people, or the highest concentration of middle-class people on our platform.

And we started constructing this narrative of essentially if luxury brands like LVMH want to be successful in China, they need to reach this specific audience and the way to reach that audience is through this specific Weibo. And I knew that this had been successful when... actually let's take a step back. We used interesting things that I call like memetic fingerprinting where we want to be able to track if any part of the narrative is coming from us. So we would slightly fudge numbers to make them recognizable. So if we knew, for example, that you know 72% of people had this or that attribute, we might say 71 or something instead just so if that number 71 comes back, we know that this narrative has touched us at one point. Does that make sense? Like we could fingerprint a little bit.

And I knew that we had been very successful with this campaign when one day I was invited to speak at a soirée of sorts and the person who was speaking before me was a person I've never met before working at an advertising agency who was essentially sharing the same story that I had been sharing with these brands with his own slides and with all of these fudged numbers. He had no idea who I was, right? But there he was given an opportunity to speak. What did he want to talk about? And he was telling the story about the rising Chinese middle class and their new spending habits and where can you find them online and actually luxury brands should really look at doing things on Sina Weibo.

Um and yeah over that 2-year time period or so we went from 50 million users to 500 million users and becoming The Weibo. Because we first became the Weibo in English, right? So we first won the battle in English and that allowed us to win the advertising spend with the big luxury brands etc. which just gave a lot more credibility and resources. So that's an example of memetic consultancy. There's something about the memes that activates certain parts of your brain that helps you remember brands or concepts.

We had to make sure that the story was easy to reproduce—like the fidelity part. So we kept iterating on how do we tell the story in a way that it's easy for them to retell it but it also... they need to come across it in the right places. So you know I would wine and dine different bloggers and journalists and get parts of this story out so that there was scaffolding for them. And then finally there had to be some incentive. So we realized that we have to make them the experts; they need to be the heroes. And that was exactly what happened in this scenario where this guy is giving a presentation and he's the hero telling the story because we've given this story to them for them to be the hero to tell the story.

Right, there's a meme about "I made this meme."

Yes. Yes. Right. And yeah, I... that was a very proud moment in my career when I realized, oh, this guy is essentially giving the same presentation we've been giving to people. He wants to share it with the world 'cause it makes him look cool and he's made his own slides and everything, but this is very clearly the narrative structure that we designed, complete with the fudged numbers that we've introduced into the narrative to be able to track was this us or not.

Right. And living in China, you taught yourself Chinese.

I get by. I am not fluent. Uh, which is unfortunate. But I'm working on it. I gave my first VC pitch in Mandarin in December. So, I'm working on it, but still not fluent. I do think that you should get at least some conversational Chinese if you are in robotics. And if you can, get fluent. Because so much of what is happening in robotics is happening in Mandarin and you're missing out if you can't be part of those conversations.

Okay. Okay. Yeah, that was it from my side. Do you want to share? Did I miss anything?

No, thank you. This was a lovely conversation. Thank you so much for having me on.

Yeah, thank you for the interview.

Thank you.

I had fun.