GeekPark

Conversation with the Man behind Hugging Face's Microduck, Eric Pan from ‪Seeed Studio

Sun, 13 Sep 2026 05:35:26 GMT

You've seen the duck.

Microduck — $399, from Hugging Face's Pollen Robotics — sold one unit every four seconds at peak. Ten thousand presold in five days. Over $5 million. Orders are booked into next year, and it left the tech bubble some time ago: your non-tech friends have seen the duck.

What you probably haven't seen is who builds it.

That would be Seeed Studio, in Shenzhen. Before Microduck, Seeed built Reachy Mini — the Hugging Face robot that reached NVIDIA's CES stage and sold out just as fast.

Left: the desktop robot Reachy Mini; right: the robotic duck Microduck. Both were developed by Pollen Robotics. | Image source: Pollen Robotics

Seeed's founder, Eric Pan, was a chipset product engineer at Intel. In 2008 he walked through Huaqiangbei, the electronics market that covers several blocks of central Shenzhen and supplies much of the world's hobbyist hardware, and decided not to leave.

He started Seeed Studio that year, sourcing and building for makers worldwide, and turned it into the junction between global makers and Chinese factories — part curated supply chain, part manufacturer.

Eric Pan, Founder & CEO, Seeed Studio

Eighteen years on, Seeed is one of the best-known open-source hardware manufacturers in the world. Ninety percent of its customers are outside China. It has sat somewhere in the supply chain of most Silicon Valley hardware waves you can name.

We started with one piece of business and ended somewhere broader: how the relationship between Bay Area founders and Shenzhen is changing, where a hardware founder's value now sits, and the rules of the Chinese supply chain that nobody puts in a deck.

One shift Pan has noticed: Silicon Valley teams arriving in China have moved from purchasing to something closer to handing over a child. He's also seeing more failures. His warning to software people entering AI hardware is direct — don't spend your energy rebuilding hardware from scratch. That's a form of respect. Your strength should be outside the hardware. Know what you don't have to build.

Edited for length and clarity.

By Lilyann and Zihua Su · Edited by Jingyu

Additional reporting by Liyuan and Sean, GeekPark (@GeekParkHQ)

01 · The part nobody can copy

GeekPark: Microduck is past ten thousand presales. Did that match your expectations?

Eric Pan: We expected the duck to do well. What we didn't expect was this — that it would break out this far, that everyone would love it, that people entirely outside tech would be paying attention. That caught us off guard.

GeekPark: Looking back, what part of it generalizes?

Eric Pan: A few things, and they run against how embodied AI is usually done.

First, it isn't a functional object. It isn't trying to prove a technology or hit a performance number. The creator is more like Geppetto — he set out to make something interesting, something like his own child, something he loved and put feeling into. That changes the product definition and the form. Bring a humanoid home and you might catch sight of it at night on the way to the bathroom and get a fright. It's too invasive.

A duck feels like a pet. Compared with most machinery it has more spirit in it. Watch the way it moves. We all know it's a pile of servos and a controller, but it moves like a small creature. A silicon-based creature. That's the first thing, and it makes far more people willing to accept it and look closer.

Second, it costs a fraction of a normal embodied robot. A humanoid might be hundreds or thousands dollars. A duck is 399 USD. People realize that for a price they can accept, they can own this very cute thing. That turns them from spectators into participants.

Those are the two key points. The logic they share: I buy a robot, I train it. It used to be a group of academics gathered around one or two machines. Now one person can have a pile of robots to train. The whole sim-to-real cost came down. We build larger robots ourselves, and when we take them to a show or handle them, we worry about safety and about load.

A little duck is harmless. It lowers the cost of taking part in training and the cost of reproducing results. The inner ring — researchers — see the duck and become its distributors. Then the next ring. Then the consumer ring, and everyone wants one. It's a smooth process.

GeekPark: So the first audience was researchers.

Eric Pan: Definitely. Look at how it launched — there wasn't much media coordination. It went up on Hugging Face's own site, and the people who follow Hugging Face are researchers and people working on large models.

GeekPark: $399 is far below other research robots. But there's real engineering in this duck — fifteen motors, a camera, a lidar. How did you get the price down that far? Where did the cost come out?

Eric Pan: There are a few paradigm differences from traditional robots. With many humanoid robots, the first sale usually goes into research, and research procurement is low-frequency. That's high-cost procurement — it needs tendering, and the transaction cost is high.

The duck is D2C, straight to the user, without much of a distributor layer. The pricing was set aggressively. There isn't much margin in this duck — when people try to clone it or hand-build one, their cost comes out well above its retail price. How do we do it? Overall optimization, including using this volume to negotiate upstream, and then a lot of engineering method in between.

GeekPark: What was the sequence behind the Hugging Face partnership? Did Hugging Face acquire Pollen Robotics first, or had Pollen already found you?

Eric Pan: Hugging Face acquired Pollen a few years ago. Our work with Hugging Face came after that — that's when they had more resources and more appetite for open-source embodied robots. That's the starting point. But our relationship with Pollen goes back much further, because they were already our customer. They were using our open-source hardware in their early development, so they knew us long before.

GeekPark: When Hugging Face came to you, was it already a duck?

Eric Pan: It was a duck when it arrived.

GeekPark: And on the rest of the details, was there a natural consensus?

Eric Pan: Largely, yes, because they know our boundaries and our capabilities, and we respect their thinking. It was a smooth fit.

GeekPark: What do "boundaries and capabilities" mean when you work with a customer?

Eric Pan: When we co-create with a customer — overseas or not — the customer focuses on their design, their knowhow, their software. What we want is for them not to just hand us a drawing. We want our customer to be a maker too, someone who can hand-build their own prototype.

They're like a painter. If the brush isn't in your hand, what you can create is limited. With all this open-source material available, you develop a form, 3D print a shell, and through that process you keep feeling your way and revising. At some point you show it to us, and once we can see you've proven these things out and thought them through, our communication cost drops sharply.

I know your intent. You're not handing me a parameter — I take your intent, add our professional judgment, and help you revise it together. That process is ideal and efficient, which is why we've turned down many people who came to us holding only an idea.

But plenty of customers arrive with a prototype, and often that prototype already uses our open-source hardware, which means it's already compatible with our supply chain. They built it out of my building blocks, and my blocks are already in the warehouse. So we're the best and fastest partner to scale it up.

GeekPark: How many suppliers did you connect for Microduck? Was this a set of requirements Hugging Face couldn't have handled directly?

Eric Pan: Going through us was probably a good deal more efficient. Take buying a single cable — who do you buy it from? Cables don't have a clear standard or brand. Do you go to a big supplier or a small one? Your requirements, safety compliance, all of it takes a long negotiation. And that's one cable. Do you want your energy going into that cable, or into product structure, community, core algorithms?

We've bought that cable maybe a thousand times. We're a curated supply chain for robotics — not an especially wide path, but a proven one. Come through us and it will at least work.

Your focus, your value creation, shouldn't be in the supply chain. It should be outside it. We joke that Chinese teams are good at making things cheap and American teams are good at selling things expensive, so American teams should stop spending every day trying to make things cheap.

GeekPark: Microduck's form isn't a phone or a tablet with a mature, process-driven line. The moment a novel form is involved, it gets hard to find a factory that understands what you need.

Eric Pan: And behind it isn't one factory. It's a factory inside a factory inside a factory inside a factory. If I need a cable today, upstream of that cable is plastic, is copper foil, is a production process. It's like a tree.

Everyone talks about the work, about how to do it. From where we sit it's a question of relationships, trust and confidence. Does each link in the chain trust its upstream and downstream? Do the upstream and downstream have confidence this product has potential? Transmitting that information is hard. Before wearables took off, how would a plastics factory have understood the category at all?

GeekPark: Where does Shenzhen's advantage show up?

Eric Pan: Chinese factories have no weak side. Cost, quality, innovation, initiative, sense of responsibility — and they're extremely elastic. The factory here is a larger proposition about engineers and supply chain collaborating. You can move the workers to Vietnam and Chinese engineers can fly over to support it, but that's a surface-level relocation.

Especially for innovative products. If you're a chef, you want to source at a seafood market, not walk along a shoreline picking up random ingredients. The efficiency isn't comparable. Look at the range of China's industrial categories. Say you want to build a robot — robots are complex.

There's electronics, structure, actuation, thermals, optics, acoustics, then packaging, materials, textures. Anywhere you do this, once you expand the form, it isn't just many parts. It's many categories of part. And if one material is missing today, the line stops. It isn't only whether it exists — the capacity has to match yours, and match it precisely.

"If a factory serves only you and leaves its other 90% of capacity idle, the price it gives you might be ten times higher."

Eric Pan: If capacity falls short — you need 100,000 units and the factory can give you 10,000 — the other 90,000 wait. But if a factory serves only you and leaves its other 90% of capacity idle, the price it gives you might be ten times higher. So it's a network, and after all these years it's become a rich, balanced, robust ecosystem.

An American coming to China is plugging into that network, fitting into its support and its gaps. If you want Made in USA, you have to rebuild that ecosystem from scratch. Both are legitimate and both exist, but for most founders, plugging into an existing ecosystem is much faster and much better value.

GeekPark: So it's a floor-level constraint. If one thing doesn't arrive, everything waits.

Eric Pan: If you're building something new, the fastest path is to enter the existing system.

GeekPark: So the relationship between a supplier and a customer, especially early, is more fragile than we imagine.

Eric Pan: Yes. And it's fragile in another way, a reasonable one. You agree with a supplier to build this thing, and one day the supplier tells you they've taken a big order from an important customer and can't serve you anymore, or their lead time has stretched. You can't stop someone else from making money. And this doesn't happen once. It can happen at every link in your BOM. If you're sitting in San Francisco, unfamiliar with all of these suppliers, how do you handle it?

Take Pebble Watch, an early representative of smartwatches. They raised more than ten million dollars on Kickstarter. The founder, Eric Migicovsky, has said that to build the first-generation prototype, he and three members of his hardware team spent roughly six months living near the factory in Shenzhen.

He'd raised more than ten million dollars, he had hard demand, he'd proven the category, and when he talked to suppliers everyone knew the volume was large enough and that investors and major players were involved — and there were still that many problems.

Pebble Watch2 | Image source: Phone Arena

He'd sold ten or twenty thousand watches by then. What about a brand with a smaller number? When they took an order for twenty thousand watches to the watch factories, many would have seen it as a small order. And that's before you get to a smart watch, where I'd have to cut a dedicated mold for you. That's an enormous gap.

GeekPark: How can we help founders? Would it help if teams used more open-source hardware?

Eric Pan: Yes. On method, we try to tell customers things in advance. We produced an Open Manufacturer white paper — when someone comes to China we send it first. Read it, or have your agent read it, and based on your product there may be things you should watch for, problems you didn't know you should be thinking about. Think about them first.

Second, base yourself on open source where you can. Your product doesn't have to be open source, but if you build on reasonably popular open-source hardware you'll avoid many detours — in design, in quality, in delivery, in supply chain.

02 · The hotpot test

GeekPark: If a founder from outside comes to Shenzhen and goes straight to a factory instead of working through someone like Seeed, what detours do they hit?

Eric Pan: The most immediate one is that most factories won't want to deal with them. A factory's thinking is: the bigger the volume the better, the bigger the counterparty the better. If you're Apple, you hand over a card and everyone is instantly circling you. The factory asks who you are, how big your company is.

The answer is: two people. In practice it's usually a little better — the factory is willing to cooperate at first, and then as the work goes on, why can you never get this thing finished, and the trust that had been building collapses.

Both sides are running a trust-matching mechanism. Big factories won't engage with you; small factories don't have mature enough experience. And on the founder's side, your product, your state, the stage you're at — none of it is set up so that you ask for a brick and get a brick to a national standard. That isn't what this is.

It takes both sides investing and committing, and in between you need a mechanism for building trust. What we do is bridge the two ends. We work with small and large factories alike, because we aggregate fragmented demand, strip out the common part and buy it centrally.

When I raise a requirement, I raise it professionally. When I place an order, I come with a professional guarantee — you don't need to worry about getting paid.

But not every product is suited to going through us. Some of you have raised a lot of money and are happy to go direct, or you have an experienced program manager. There are many paths. We're only one of them.

GeekPark: Has this collaboration model become an established convention between Shenzhen and Silicon Valley, or is it still being figured out?

Eric Pan: For us it's a mature model. For Silicon Valley plus Shenzhen as a whole I wouldn't dare say, because our view is limited and biased.

The most natural Silicon Valley attitude is: I've just raised a lot of money, I need to find somewhere to manufacture. People have tried Singapore, Malaysia, India, some inside the US. And then they find the teams that win and take the market are still the ones working closely with China. That's the end result.

Once that result exists, two things can happen. One is Chinese founders going to the US, building a company with a Chinese core and taking the market — Anker and companies like it, going outward.

The other is Silicon Valley teams hiring Chinese people and building much stronger linkage with Shenzhen. Because it isn't a simple buy-sell relationship. It isn't coming to China to source a part, verifying the quality standard, and being done.

It's much closer to tuōgū — a Chinese term for entrusting your child to a guardian before you die. You're handing over your child, or your pet, to a partner you trust, because this is being built from nothing and problems will come from every direction, and everyone has to solve them as a team.

In the previous wave, many teams came and left empty-handed, having spent a lot of money without getting the thing built. They hadn't respected our factory smart. A great deal of frontline manufacturing experience isn't written in books. A worker who looks perfectly ordinary may be the one who can guide you. Whether a highly paid Silicon Valley engineer arrives on the floor with an imperious attitude or an open, at minimum level-headed one — that affects the outcome.

GeekPark: Is it harder to raise now without hardware?

Eric Pan: In the Bay Area, I think so. If you're pure software it's now easy to have no moat and no differentiation — that's been ground down completely in the Bay. Add hardware and you can find new data, new deployment scenarios, a new experience. The complementarity is strong.

GeekPark: What kind of founder makes that attitude shift easily? And what kind can't adapt?

Eric Pan: We have one test: are they willing to eat hotpot with us. Whatever your background, being willing to eat hotpot with us means you're willing to accept an unfamiliar custom or food, or that you already understand something about the culture. You'll be facing all kinds of new ingredients. Do you resist, or are you willing to try?

We've had customers and partners who wouldn't try anything new in China, holding on to their bread. That person won't do this well. Their perception is bounded, and building trust with them is slow — and that slowness can be fatal. That's the simplest test.

Above that test, it inverts: we're trying to use the simplest possible instrument to work out whether a team will go humble. Are they willing to accept what they don't know and try to learn it? Which is the same principle as manufacturing and hardware.

It isn't a buy-sell relationship, it's a partnership. Like a long trek together, heavy packs, into a wilderness with nobody in it. Who do you pick to go with you?

GeekPark: You used the word tuōgū. Is the real relationship deeper than the partnership we imagine? Does most of the business end up being done by Shenzhen?

Eric Pan: It depends on the complexity of the product, or its maturity. If the product is mature and the form is immediately legible — it looks like a phone, you're making a tablet — then it's picking one supplier out of N. But if what you're building has no precedent, everyone has to take risks and run experiments together. That's the wilderness state, and it takes much more understanding and alignment.

Why were we willing to commit to a project like Hugging Face's? Because we could see the project had taste and quality, and that it was distinctive. As opposed to the many people who come to us saying, "I heard that product took off, I want to make a similar robot, how much would you charge." I generally turn those down.

GeekPark: What do you look for in the companies you work with?

Eric Pan: Take Hugging Face. That much influence, building a robot, using data as a flywheel and hardware as the carrier — that's logical. And talking to that team, they're pure makers, it's smooth, and the product is differentiated and inventive.

We look at many direct and indirect indicators beyond our own instinct. If a popular open-source project like Home Assistant comes to us, you know immediately it's worth doing. Or if someone has real influence and has already built a product a proven user base needs, that's easy to verify. But we don't necessarily pick the biggest — and the biggest don't necessarily come to us. We look for the products with more potential.

GeekPark: Why not the biggest?

Eric Pan: The biggest — Luxshare, say, one of Apple's main assemblers — will build a complete team and do it in-house. And much of that is pure consumer product, which also isn't our highest priority.

GeekPark: Why is pure consumer product not a priority?

Eric Pan: Pure consumer product has a lower probability of success and flames out easily. It usually needs its own aggressive team to pull off. We'd rather work on things that are less fashionable but meaningful and influential — exploration in vertical niches, exploration of new forms. That's more worth doing.

GeekPark: How do you decide internally whether something is a consumer product?

Eric Pan: The test for a consumer product is whether it exists to satisfy desire rather than to create value.

What we're good at is realizing a new technology and pushing it out. Our logic for new technology starts with researchers, then businesses, and only then consumers. We'd rather do the business work. The consumer work is already brutally crowded.

03 · What to do when you land

GeekPark: If a software-first team wants to build hardware capability and comes to Shenzhen for the first time, how should they spend a month?

Eric Pan: If you're not from the hardware world, don't spend your energy rebuilding hardware from scratch. That's a form of respect. What we're afraid of is a software engineer who hasn't done hardware — smart, capable — because some of these holes you have to fall into yourself.

So first, walk the boundary. Huaqiangbei, and Shenzhen Robot Valley. Both are worth seeing. The point is to know where the boundary of hardware currently sits. Whatever hardware you see already existing in Shenzhen, or simply already built, you can usually find an existing company to reproduce it or an upstream company to supply it.

After you've looked, get hands-on. Not building hardware — using hardware. Don't start from scratch. Build on what exists. Your strength should be outside the hardware; that's what makes you worth having in Shenzhen.

We're putting together courses on vibe coding a piece of hardware, vibe coding a robotic arm, and you find software has already come this far. We used to separate front end, back end, embedded, microcontroller-level embedded, then FPGA. Now vibe coding cuts through all of it — you can not know C and still control a small embedded device in Python or Rust.

Turn it around and hardware is one link for you. It takes a button that was on your screen and puts it in the physical world, where even a cat can trigger it. In using it, first you build confidence, and second you learn what's already available — which tells you what you don't have to build. If I had one phrase for what we do, it might be vibe hardware.

GeekPark: Seeed runs a maker space called Chaihuo. Are these courses part of it?

Eric Pan: Yes. The courses are open source and free. It's mostly about giving people a way to exchange with each other, and we encourage more people to take them, so a mutual trust mechanism can form. Our philosophy, the one thing that hasn't changed in more than a decade, is "making technology accessible."

Chaihuo Makerspace | Image source: Openthings

Chaihuo's role is before a company starts and in its earliest days. You haven't formed a team yet, so how do you see the range of possibilities, how do you talk to different people? Chaihuo is a starting point. There are events and workshops, and afterwards you may find that with everyone's help you can do it — like a one-person company that hasn't raised anything being able to bang out a prototype at Chaihuo.

Many overseas teams come and stay in that space, and prototyping there is more efficient, because they're not imagining components and buying them from far away. They can go directly to the supplier, to the factory — to design from manufacture. They can see how factories are already doing this, or find something in a similar form that's already proven and in mass production, and ask whether they can borrow and adapt it. The efficiency is much higher.

So the first step is to come to Shenzhen and look. Come for three days, or a week. Then come a second time and stay a month, living and working here alongside everyone, opening up your thinking, meeting people, making friends. Someone will be good at structure, someone else has already been through the electronics problems. Plug into the ecosystem here. Whether you then go back to the Bay Area to start your company or do something else, it'll be easier.

GeekPark: Among founders coming to Shenzhen recently, are there collective tendencies in what they pick?

Eric Pan: The ones that cluster too tightly — chasing whatever just took off — we generally invite to figure it out on their own.

04 · Hardware's value is outside the hardware

GeekPark: Humanoids or small desktop robots like Microduck — which is the easier path to volume?

Eric Pan: This one is simple: the cheaper it is, the easier it ships. In pure numbers, the lower the barrier, the more people take part. That's one factor. The other is the WHY. Many founders in Shenzhen focus too much on the HOW and neglect the WHY.

With a humanoid — what's my reason for buying? I bring home a two-hundred-thousand-yuan humanoid, and then what? Have it fold my clothes? An ordinary person can't make that add up. I can't convince my mother to spend two hundred thousand on a robot to fold her clothes, so who can you convince? Whose problem is that two-hundred-thousand barrier solving? And what problem does a two-thousand-yuan duck solve?

Quite a few. As a high-end figurine, as a toy, as a pet — it's comparable spending. The core is still embodied AI training, of course; other users can then try it as a robot pet they can train. As for other embodied robots, we care about applications. We've been waiting the whole time for one thing: what job can this robot do today?

Take Reachy Mini. The most direct use is service and retail. I put one at a front desk — at minimum it gives my customers a better impression, and for one or two thousand yuan it's a productivity tool if it can queue people, talk to a customer, take a number. Same with robotic arms. An arm used to be a hundred thousand yuan, now it's ten thousand, and at a few thousand it can do a specific job. That's much easier to scale.

GeekPark: What's the key to Microduck still existing in a few years without being copied out of existence?

Eric Pan: The key is that it isn't just hardware. It's already a cultural phenomenon.

And even if it weren't — under the original plan it's a carrier for embodied data. Everyone trains on this pile of things, and because there are so many models on top of it, everyone stays willing to keep using the duck.

GeekPark: So the duck effectively becomes data.

Eric Pan: Right. Which comes back to the point I keep making: hardware's value is outside the hardware.

GeekPark: The duck's software is open source, but the hardware isn't yet, correct?

Eric Pan: Correct, the hardware isn't open source for now. So I don't think this is necessarily the endgame either.

GeekPark: So the ideal endgame is that you build an innovative product, and that product becomes a building block, so the people who come after can assemble better models out of more and more blocks?

Eric Pan: That's our model, though for Pollen or other companies it might not be. The common thread is: don't treat hardware and shipment volume as the only answer.

GeekPark: On the software side people explored a lot — building on things like OpenClaw — and the coding track broke out first. Does hardware end up similar? Everyone has a passion project they want to hand-build, but what scales is still functional?

Eric Pan: They combine — functional hardware plus applications and presentation that have real expression in them. The functional part is already mature. A duck: the main controller is a single-board computer, the legs are servos, the microphone and speaker are already modular. This is integrating them into a new form. It's a combination of functions, high cohesion and low coupling as an engineering matter.

But each person can point it at a different WHY, and the WHY each person defines is where the value and the appeal will be.

05 · Too many shells, not enough ghosts

GeekPark: Since embodied AI got hot, have the people coming to Shenzhen changed? In number, in what they want, in their role?

Eric Pan: It isn't necessarily about embodied AI. It's that after China and the US both tried everything, people found the only place to come is still Shenzhen.

The wave coming now is related to embodied AI, because they were passively migrated here by the large-model track. Some were doing software, then got laid off, or left of their own accord, and needed a new frontier, a new border zone — and software plus hardware is embodied AI. Whether it's a robot, or a robot shaped like a human, isn't fixed, because a robot is a form and a carrier, a configuration of hardware.

How a robot ultimately solves a problem will come from combining with the knowhow of some industry. Say I'm a chef, or a psychologist researching elder care, or a hotel group. At a certain point robots get pulled in by the industry, in reverse. The cost sits in the robot hardware, but the return sits in the industry knowhow.

GeekPark: Comparing two years ago to now, has embodied AI risen noticeably as a share of your orders?

Eric Pan: We don't separate out embodied AI, because at this point everything we make counts as embodied AI. Take the sensors we've built for years. Sensor data used to be collected and then judged, or processed by a person. Now that data goes straight to a large model and becomes a derivative of the model. I don't know whether that counts as embodied AI, but we count it.

We now split our product lines into two halves. The first is Everything to AI — more real-world data is needed. Even the camera on a Reachy Mini is real-world data, and its conversation is corpus.

The second is AI to everything: making AI small, cheap and usable. There's the large end, hundreds or thousands of TOPS of compute, and the small end, an embedded model running on a 30-milliwatt, five-dollar microcontroller. Once these models are built, they can be pushed into existing devices and fixtures and objects.

So everything in our range from here is tied to AI and embodied AI. It's only that our definition of embodied may be broader than most people's.

GeekPark: I think our definitions are the same. I just didn't expect your whole product line to be AI-related now.

Eric Pan: Probably because we don't make things that are only for people to use. We don't do consumer product. In B2B there's a clear trend from human-to-human interfaces, to human-machine interfaces, to agent-to-agent.

For agent-to-agent you take what exists — say a dashboard and a bank of buttons — and use something, an AI camera or a machine finger, to operate it and connect it into a system. We encourage users to open up that thinking. Embodied AI doesn't have to be a humanoid. If we go back to the core definition of a robot, it never said it had to be human-shaped. If I'm a microwave today and I talk with you and can identify what's inside me, I'm a robot too.

GeekPark: Back in 2023 Seeed said it was betting on AI. What did that mean concretely?

Eric Pan: In 2023 large models weren't yet mainstream, and what I was doing was edge-side AI — putting reinforcement learning and classification algorithms into small edge devices.

Where you used to need a server behind a camera, now for a dozen dollars inside the camera you can run a classification algorithm. When large models arrived we went immediately to heterogeneous edge compute. Say you want to use a large model on a camera — that burns a lot of tokens, because it has to read every second. Put a small embedded AI in front, and I only pass the key frames I've identified to the large model. That saves a lot of cost.

GeekPark: So in product terms, more edge computing devices?

Eric Pan: Yes. More compute on one hand, and on the other the data path — MLOps — building out the whole chain from training to deployment. Above that is large-model integration. For example, we use a 4.5B model, and dropped into a specific scenario it can do plenty. I go to a restaurant to order. There's a button on the table, I press it and talk to it. I don't need to discuss the weather with it. I only need it to match and filter the dishes this restaurant has. How big does that model need to be? It can be small.

So in B2B, the opportunity and the room are limitless.

GeekPark: How do you come up with these scenarios? Do B2B customers bring you demand, or do you do your own advance research?

Eric Pan: Both, and it's a loop. Interesting attempts emerge in the community, some take off, people understand what's behind it, and then it turns into something more productized that inspires more people.

We rarely do big product planning. We curate the newest technology in, users run PMF experiments, and what comes out isn't necessarily a serious or large-scale product — but it feeds back to us, or to them, and they influence each other, and eventually one product gets selected out and becomes a hit. This duck wasn't the first duck we tried, either.

GeekPark: Over the past year and a half, AI toys and AI hardware have been hot, but the ones that broke out almost all came from open-source hardware — XiaoZhi AI, Reachy Mini, now this duck. Why?

Eric Pan: It's a logical necessity, because the mass market always lags. It's always a step behind. It gets led there by pilots and early adopters, who are themselves makers and hackers, and they're the first to encounter open-source things.

With something closed, I can see it but I can't touch it. With something open, I become part of it, I can use it early, believe in it, contribute to it, and not worry about being expropriated. So hackers, makers and geeks naturally gather around open-source projects. Meanwhile the large companies have the ability to lead consumers but usually don't dare to innovate. Look at Apple — as strong as it is, it doesn't innovate much now.

An open-source project doesn't have that much earning power, but it can unite the geeks. And when the thing has real value and the friction is low enough, it breaks out.

GeekPark: We've also seen founders building businesses in open-source hardware — M5Stack took off, and now new teams are building similar desktop robots for the maker community. Is that market really that big?

Eric Pan: The maker market is growing fast. When we started in 2008, you had to spend a long time explaining what a maker was. Now you don't have to explain what a maker or open-source hardware is, or why anyone would open source.

Ten-odd years ago most people had never seen a 3D printer. Now plenty of households have one, and you can define anyone who owns and uses a 3D printer as a maker. That's a large number of people, and it's expanding.

Consumers are also changing now that they have AI. Many people never had the opportunity before, but now I can vibe code my way into being a programmer — it's hard to say that person isn't a developer. And from there to maker is a short step. I already have web coding; I want to put my web coding into a device. There's a lot of room in that.

GeekPark: Will open-source hardware become increasingly consumer-grade? Electronic Lego, barrier dropping — does it become consumer electronics?

Eric Pan: The underlying logic is like personal websites, or indie apps. Everyone can make one, and because everyone can, many unserious things appear, or things I made purely to delight myself. Indie hardware becomes the indie game console.

I think we'll see more cyber accessories, more eyewear. Future electronics will be more like personal expression: I don't make much money, but I'm cool, I can make something different. It doesn't have to become a thing everyone buys at high volume. Before, you had to cut a mold, you had to do mass marketing. Now I post on RED, Xiaohongshu, with something 3D printed, and everyone can use it — if you like it, I give you the source and you copy it yourself.

GeekPark: Are the people motivated to build open-source hardware a particular kind of person? Pollen was acquired by Hugging Face, which gives it more incentive. For an ordinary maker, where does the drive come from?

Eric Pan: The main driver is still interest and passion. It's like making short videos, or being a YouTuber, or a Bilibili creator. You're interested in a direction, you share, people get more interested, you keep investing, and it may even become a profession.

If video and news can be done by individuals, hardware built and released by individuals is the same. I made one product, it's interesting, and I only make a hundred sets — or I run a group buy of five hundred each time. Who says our mice all have to look the same?

GeekPark: Do you see a clear trend rising, in China and abroad?

Eric Pan: More and more of it in both. Search Taobao for strange keyboards, or e-paper things. Take the little yellow duck — if you can't buy one today, fine, people start hand-building, and there's a nationwide wave of hand-built ducks. Then after the first pass, well, you have a duck, I don't want a duck, I want a goose, or I have my own IP character, and I start creating.

And why is that person good at creating? Because they studied sculpture, or they were a designer of anime props. Interest is the core of it, not some grand but hollow trend.

GeekPark: Seeed also develops its own hardware. What do you build in-house and what do you buy?

Eric Pan: We don't develop components in-house. If we need a chip or a camera, we integrate the mature option — but we turn it into a module so users can use it easily. Over the past few years we found that users building things out of modules was also unreliable.

Say I want to build something for outdoors and I haven't considered any of that, and afterwards it's hard to change. So we started building open-source devices — as if I've already made it into a box, and you also want a box, so tell me how you want this box changed. Leave the thermal part alone, we've validated it and it needs no new mold, but swap the interface, change the logo, add something. We turn our devices into the user's reference design, then into a solution, with hardware and software already done, that you can customize.

We're building out in both directions so a user's team can be smaller. They don't need a specialist optical or thermal engineer. They need our open-source hardware plus an idea, their algorithm, their market influence, and they can work with us lightly.

It's like software — from doing everything yourself, to IaaS, to PaaS, to SaaS. We're Manufacture as a Service. But the starting point is that you still have to develop on top of our framework.

GeekPark: Is it a shame that you're always working on the frontier, on developer-facing products, and once something matures it moves into a much bigger mass market that may not be yours? Have you considered taking the mass market too?

Eric Pan: Those are two questions. First, once a market gets big, is it still our market? Every company has a range in which it's alive. We built this business around the tech and maker community, so not being able to serve consumers is normal. Or if some enterprise vertical takes off and we don't understand that market, we shouldn't be taking that profit either.

Second, us not taking that market doesn't mean our customers don't take it. My job is to help our customers win it. We're happy to work with small teams to go up against big opponents.

Now that Microduck is out, plenty of competitors will imitate it. Three years from now, when we talk again, Microduck shouldn't have been killed off — it should have played an important role in the process. It shouldn't be that we showed up early to the market and then got wiped out. I don't think it goes that way. Every industry should be like this: helping our users succeed in each industry is enough.

GeekPark: Do you plan to bring Seeed to the front of the stage rather than staying a partner?

Eric Pan: We're the stage, but I don't have to be the actor. We're not hiding behind it either — here we are today, talking about what happens on stage and behind it. But we can't take the actor's job, because the actor has their own audience.

Beyond that, the stage is large and we'll take more opportunities. There's a lot of room in combining hardware and software, and OpenAI and Claude are both pushing toward physical AI. I think that moment arrives soon.

Here in Shenzhen, what troubles us is that there are too many shells without ghosts. The soul is underdeveloped and it's all very hard hardware. But hardware's own value is incomplete, insufficient. It has to be joined with meaningful software before it's a whole solution. And the further that is from Shenzhen, the more valuable it is.

Shenzhen is the rear base. Keep working to make the hammer better. But the nail is somewhere far away.

As I said, our main criterion for judging a partner isn't whether they have hardware expertise. The ability to do hardware well sits outside the hardware.

Editor’s Note: GeekPark is dedicated to telling China’s AI story to the world and fostering a two-way exchange between Silicon Valley and China’s technology ecosystem.

If you have an idea or story you’d like to share, get in touch with the GeekPark Silicon Valley team.

Wang Xingxing's Unitree IPO Afterparty Speech

Wed, 19 Aug 2026 13:47:06 GMT

As Unitree's earliest institutional investors, @GeekParkHQ is in the room at the company's IPO afterparty. Below are founder Wang Xingxing's remarks, edited and translated by GeekPark.

Ten years. It has gone quickly, and it has only been ten years. We believe that with your support and our own work, the next five to ten years will bring more surprises and more in return. Thank you for your trust. Today we formally listed on the Shanghai Stock Exchange, and it went well, but I believe this is only our starting point.

On behalf of Unitree's board and all of our employees, I want to extend the warmest welcome and my sincere thanks to the leaders, guests, and friends who took the time to be here tonight.

What we have achieved would not have been possible without the support of government at every level, and without Spring Festival Gala, which helped turn Chinese robotics into a technology symbol recognized around the world. Over the past few years, the global impression of Chinese robots has been rewritten. Today, most of the robots actually in use across the United States, Europe, and Japan are made in China. That is something to be proud of. What we want, above all, is for robots to be genuinely used in work and in everyday life.

That is looking back. Let me share a little of how we see what comes next.

Tonight, for the first time, I want to put forward what we are calling Physical AI Robot Self-Evolution 1.0 (物理 AI 机器人自进化 1.0). I believe this is the thing most worth doing, now and in the years ahead. Today, across our company and the wider industry, the way we build robot programs and applications is still very primitive. Current AI can do far more.

The idea is simple. We take the world's best large models and give them a set of tools and skills we define. We let the AI automatically search the latest leading research and open-source code, generate its own robot training code, and train itself, including in simulation and with video-generation models. We then move that onto physical robots for further training. The results are scored by the AI itself, with human evaluation in the loop as well. Once that loop is running, it iterates extraordinarily fast. Given today's compute, its speed of programming and improvement already exceeds what a team of hundreds of engineers could do by hand.

Why does this matter? A few reasons.

First, foundation models around the world take a clear step forward every month or two. Inside this loop, we can ride the best models available anywhere to lift our own capability. Put plainly: even in a stretch where my colleagues and I do not get better, the models do, and our self-improving loop keeps advancing.

Second, more data. Every part of the process, the programming, the simulation, the test and evaluation metrics, produces data at scale. More data makes the model stronger. And because our company ships a large amount of hardware, we can deploy across thousands or even tens of thousands of machines, which means the data can grow exponentially in the months ahead.

Third, the skills compound. We add one skill today, another tomorrow, another the day after. The skill base and its ecosystem keep growing.

I see no reason to avoid this. Building a single VLA model, or a single video model, on its own, is no longer enough for where competition and iteration are heading. What we need is a self-planning model architecture. In the early stages the process still needs enough human engineers deeply involved, defining the right metrics, the right pathways, the right algorithms, and the success rates we hold ourselves to. As the model's capability rises, it grows stronger on its own. Our aim is to open the automation era of Physical AI robotics this year and beyond.

I have said a lot, and some of you may feel I am telling a story. Ten years ago, many people who heard me thought exactly that. But I believe that in the next five to ten years, we, and society as a whole, will make this real.

Thank you again to the leaders and guests for your care and support for Unitree. I wish everyone good health and every success. We will keep our heads down and refine the technology, and we look forward to walking even further alongside all of you. Thank you, and let's raise a glass together.

Notta's Founder Ryan Zhang: How Notta Reached $30M ARR in Japan, the Market Almost No One Picks

Tue, 18 Aug 2026 05:58:51 GMT

Most founders building for the world reach for the US first.

Ryan Zhang, the founder of Notta, picked Japan. It is a market almost no one goes into on purpose. The language is hard, user satisfaction has run negative for years, enterprises decide slowly, and the part that matters most is that trust takes years to build.

Against that, Notta reached $30M in ARR, more than 10 million users worldwide, and coverage across 68 of the 225 companies in the Nikkei index. The market is hard to crack. Once you crack it, the difficulty turns into a moat.

"If these difficulties aren't fully out of your control, if you can solve them through product, service, and organizational capability, then they become your natural moat. Once we actually built the language quality, the local growth, the customer service, and the trust in Japan, it got very hard for anyone behind us to copy the whole thing by translating their website into Japanese.""What Japan gave us wasn't only revenue and users. It was a capability to localize deeply."

This year Ryan brought a new product, SpeakON, a piece of AI voice hardware. And he turned from Japan toward the US.

On the AGI Playground 2026 stage, Ryan walked through his thinking from 2020 to today: why he chose the counterintuitive Japanese market, how Notta grew there, how far the localization of an AI product actually has to go, and why, after five years of SaaS, he went back to hardware.

What follows is Ryan's talk at AGI Playground2026, GeekPark's Singapore edition summit.

01 Playing the Long Game in Voice

I started building companies back in 2014. The first was Mobike. After I left Mobike, I stayed in AI voice and never really left. Since 2016, everything I've worked on has been voice.

The theme today is Playing the Long Game in Voice. People talk about long-term thinking all the time. What does it actually mean? To me it's like playing a game. At each stage you build a different capability, and only once that capability is in place can you enter the next stage. When you do, and the capability is there, you can grow faster than before.

The first layer is utility. Does your product solve a problem that hurts enough? Notta started from a very small point inside a workflow and solved one painful problem. After people used it, the reaction was: finally, I don't have to do this myself anymore.

The second layer is distribution. Once the product exists, can users find you in the language and the channels they already know?

The third layer is trust, especially once you're selling to companies. Why would a customer hand you their data, their process, their budget?

The last layer is habit. In the exact moment a need shows up, does the user think of you first? Does the answer just come out as Notta?

Watching Notta grow, I saw utility, distribution, and trust stack up one on top of the other. Today we also launched a new product, SpeakON. It's my attempt at that last layer: can a piece of hardware become a habit, can it hold the moment a user has an idea and turn that into a routine?

The form changes. The underlying task doesn't.

Looking back, the products I've built are pretty different from each other. In 2017, 2018, 2019, I built smart translation devices, all tied to language and voice. Then in 2020 I started Notta, a SaaS plus AI business. Notta now looks a bit like one of those AI OG products. And now there's SpeakON, software plus hardware.

Some people might think I keep switching directions. The way I see it, the form of the product changes, but the underlying Voice AI doesn't.

The translation device solved the problem of not sharing a language. Notta solved the problem that after a meeting, you still have to record, organize, and write up the minutes. With SpeakON, the problem is this: you have an idea and you want to capture it fast. In the past that took a lot of steps, unlock the phone, find the app, type it in. SpeakON strips all of that friction out.

There's another thing. People speak far faster than they type, and when they speak they usually carry more context, because typing is a hassle and you don't want to type much. Voice always had one big problem: it produces a pile of raw material that you then still have to process. For years I've been looking at the same question. Is there a way for voice to not just get recorded, but to turn into a task you can hand off directly?

The product changes, the market changes. The core task underneath hasn't.

02 Why Japan: The Difficulty Itself Is the Moat

Choosing Japan wasn't a mainstream call at the time. For a company that wants to go global, the usual instinct is the US, or Southeast Asia, some market that's easier to enter.

What is Japan usually like? The language is hard. Users are extremely demanding. I don't know how many of you have looked into this, but the user satisfaction of Japanese software products, the NPS, is usually negative. A score of minus 15 already counts as decent, while in the US it's often around 65. Enterprise decisions are very slow. Trust takes a long time to build.

That all sounds like a list of drawbacks. But flip it around. If these difficulties aren't completely out of your control, if you can solve them through product, service, and organizational capability, then they become your natural moat. Not everyone will go into Japan right away, and not everyone can reach the level of trust we reached.

Which means once we actually built the language quality, the local growth, the customer service, and the trust in Japan, it got very hard for anyone behind us to copy the whole logic by translating their site into Japanese.

So what Japan gave us, I think, wasn't only revenue or users. It was a capability to localize deeply.

Choosing Japan wasn't only about TAM, the largest market a product could theoretically reach, or the fact that Japan has more than 100 million people speaking one language. There were a few more practical reasons.

First, the pain point is very clear. Japan is like Germany in that it adopted a parliamentary system early, which produces a huge volume of meetings and client visits. After a visit you spend a lot of time turning speech into minutes. That work is concrete and very frequent. A new employee in Japan has to write down every word of a meeting and hand it to the boss. That's a high-frequency need Japan already had.

Second, Japanese ASR has a real barrier. Plenty of global products say "we support Japanese," but supporting Japanese and actually designing for local Japanese users are two completely different things. Recognition accuracy, how the interface reads, the localized workflow, the way support replies, all of it shapes the real experience.

Third, Japanese users are relatively willing to pay. As long as you save them time and make the work clear enough, they can see the value.

Last, and to me the most important, is trust. Japanese users don't just look at features. They look at who the company is, whether you've told your story in the media, whether they can find you when something breaks, and how fast you respond. All of that is part of your product.

For me, a market with no barrier at all is easy to enter, and easy for anyone else to enter too. The hard parts of Japan happen to be things you can turn into your own capability through long-term investment. That's one of the main reasons we chose it.

03 From a Simple Tool to $30M ARR

Pull Notta's path into a single line.

In 2020 the product was very simple. Just an app. You could record in real time, upload MP3 and MP4 files, and quickly get a written document, the gijiroku, meeting minutes you could use directly. There were no large models then. We could only run NLP models for entity analysis, pulling out numbers, places, and other domains. That was AI from the previous era.

Once the need was validated, the next question became: where do the users come from? We started doing SEO in Japan, building growth strategy and localization, all of it around how Japanese users search.

One of our earliest customers was Yachiyo Engineering, a Japanese firm that does consulting for nuclear plants and water projects. Employees inside the company bought Notta themselves, then started sharing files with colleagues, and that's when the B2B needs showed up. Who can see this? Who pays? Who owns the data? Can the security team approve the budget? The product was starting to shift from B2C toward B2B.

By 2023, what limited Notta's growth was no longer a product feature or an ad budget. It was: how do you run B2B marketing in Japan? Do you have a Japanese legal entity? A local team? Local compliance, procurement processes, customer service handoffs, those became the bottleneck.

When we worked with Japanese enterprises, a customer would hand us a security sheet with something like 100 rows and ask us to fill in every line. What level is this? Where does the data live? Does the company have access control at the door? Are there cameras? Around a hundred rows, and every company's sheet is a little different. You get a sense of how seriously Japan takes trust and compliance.

Through all of this we grew revenue to roughly $30M in ARR. Every stage was painful, and each one demanded a different kind of growth.

SEO wasn't only acquisition. It was a market-sensing system.

Notta didn't start as meeting intelligence. It was a simple voice-to-text tool. Meeting intelligence sounds big, but early users don't care what the thing is called. All they want is: I sat in a meeting for an hour, don't make me spend another hour going back through it to organize the information. The task is simple. Let the user search, edit, process, and share quickly. We spent an enormous amount of time on that experience.

A lot of early-stage founders paint the endgame too perfectly. They treat the product like their own kid and want him to play basketball like Yao Ming, run hurdles like Liu Xiang, and shoot photos like a famous photographer, all at once. You don't have to see the product perfectly at the start. Find the few strengths that matter most and sharpen those. Users don't stay because you take great photos. Maybe they stay because you're tall. Do one thing well first, then think about how to expand.

In Notta's early days, SEO was extremely important. Japanese users were already searching how to write meeting minutes, how to turn speech into text, how to caption a video. We didn't have to teach them these problems exist. We just had to make Notta a clear enough answer. For every task and every search keyword, we built a page telling the user how to get that job done.

We never did programmatic SEO, the kind where you use AI to churn out piles of abstract copy about changing the world. We wrote serious articles about how to actually use AI better. And the search terms kept telling us how users described their own problems, which fed back into how we improved the product. We turned SEO into a market-sensing system. It brought in users and helped the team understand the market at the same time.

No channel is the best channel. The order is what matters.

When people look at growth, they list out SEO, ASC, content, case studies, and ask which channel works best. From what I've felt, the list of channels doesn't matter much. The order does.

Here's an example. Early on we switched pricing from dollars to yen, and conversion went up 30 percent. Just localizing the price lifted paid conversion by 30 percent.

The right order, roughly, is this. First, get the basic localization right and optimize paid conversion. Second, use organic SEO traffic to learn what users are searching for, and keep building pillar pages and scenarios that catch that demand. Third, scale revenue through ASC and search ads. Fourth, do the real localization work: customer case studies, local technical support, enterprise sales. That last part answers a different question, which is why an organization would dare to choose you.

The logic behind it is fairly simple. Solve product uncertainty first, then solve channel uncertainty, then solve enterprise adoption risk.

04 Localization Is Not Translation. It Has to Become an Operational System.

An AI product going global usually hits one question: how far does localization have to go? Translate the website and test the water, or register a company, hire a team, and rent an office on day one? I think both are too extreme.

Some localization has to happen on day one. Japanese UI, onboarding, ASR, pricing, App Store copy and screenshots. These basics are part of the product, not just UI and features. If a user's first impression is that this is an English product translated into Japanese, it feels off. They think you're a foreign product.

Other things are very important and also very expensive. A Japanese legal entity, a local sales team, security documentation, procurement support, customer success, an office, long-term brand building. Hiring in Japan is basically lifetime employment. If you want someone to leave, unless you convince them they'd grow more somewhere else and they choose to go on their own, they have the right to stay with you. So people in Japan are very, very expensive.

Which means you land in Japan step by step. Building a team of dozens right out of the gate is agonizing. Before a need is validated, don't globalize for the sake of globalizing, and don't build a heavy organization all at once. That's painful too.

If I sum up what we did in Japan, it falls into about five layers.

First, product: language quality, UI, onboarding, workflow.

Second, transaction: what currency, how users are used to paying, whether the payment channels are smooth, whether you support local Japanese payment methods.

Third, service: response time, whether you offer an SLA, local-language support, whether you have your own customer success.

Fourth, the legal entity: security, contracts, procurement processes, and so on.

Fifth, branding: can the customer form a long-term memory, and when they think of this kind of solution, do they think of you?

In the end, localization is not a translation team, or a marketing team of a few people refreshing social media every day. To actually stick, you need a real team that can serve your Japanese customers and turn all of this into an operational system, one complete setup running from product growth and sales through operations and management.

05 The Shift From B2C to B2B Was Forced by User Behavior

Notta moved from B2C to B2B not because we suddenly noticed enterprise SaaS gets higher valuations, and not because I had the enterprise strategy figured out in advance. Before this I'd only built consumer internet products. I'd never done B2B SaaS. The shift started from one small user behavior: sharing.

One person uses Notta. Then a team uses Notta. Then the team starts to depend on it. At that point the admin has all kinds of questions. How do we handle accounts? Permissions? Who pays the bill? Where does the data sit? Is there a security certification?

To judge the B2B opportunity, we didn't only count how many enterprise inquiries the sales team got. We looked at where the users' output was flowing.

This doesn't mean giving up B2C. Notta still keeps improving its consumer features, the way Notion and Gamma do. But once you're facing enterprises, you find that while you keep stacking features for consumers, the B2B customer says: "You can't open that feature to my employees. Put a switch in the backend. I don't want them using feature A, or feature B." You end up with a pile of switches, which is nothing like the consumer way of thinking.

After we entered Japan, one thing became obvious. Users aren't only buying software. They're buying the company behind it.

The bottom is still the product, of course. Are the daily recognition results accurate? Is the SLA stable? I remember 2024 and 2025, when we spent a huge amount of time on stability. Back then, especially with Japanese enterprise customers, the moment they found something unstable, the CEO would take the team to bow to the customer.

Go up a level and the customer asks: Do you have a legal entity in Japan? Who runs the Japan business? Who's accountable when something breaks? How do you handle data issues? The security team needs certain materials, can you provide them? Higher still, SOC 2, ISO 27001, everything tied to corporate compliance.

Localized onboarding, CSM, an escalation process, do you have all these SOPs?

Early on these things look expensive and seem unrelated to product innovation. But once you reach enterprise procurement, they are part of the product. Without them, the user vetoes you outright, and you have no way into Japanese B2B accounts.

To me, trust isn't a branding word. Trust is getting a user to actually trust you and hand you their budget.

People often ask when a startup should build its brand. My answer is that brand starts on day one. It's not something you do later once you have money, or by hiring a famous agency to shoot a promo film.

What is a brand? A brand is a personality. A friend once told me a brand is like a mother of two who's starting a stockings business. On day one her brand positioning is: I'm a full-time mom working several jobs at once, with two kids.

A startup is the same. You don't need some grand, lofty brand. You need to know what role you play. I'm building in Japan, I'm going after the local Japanese market, I'm solving one problem for the user, and who's my enemy? Right there, your brand already exists. Your competitor isn't Notion, Gamma, or Otter. Your competitor is a concrete direction. Treat that as your product's personality. Get it right from day one and the taste of your whole brand is already there.

06 Why Go Back to Hardware After Five Years of SaaS

Hardware means manufacturing, supply chain, inventory, logistics, after-sales. Every one of those is far more complicated than software. I remember during Notta I even told a reporter I'd never do hardware again, it's too painful, especially the supply chain.

So why start again? Because model capability is becoming very easy to get, and product forms are looking more and more like general-purpose agents. But from having an idea to calling a model, there's still a lot of friction in between. So we started asking: what is the next interface?

I don't think it's another chatbot, another dialogue box. I think it's this: the moment an idea shows up, a very simple physical action, press once and speak, and it goes and executes for you. The hardware itself isn't the intelligence. It's more like a trigger that helps you finish the task at the right time.

Good generative models are spreading fast. Software features get copied fast too. Internally we can ship a very large feature in about a week now, and that's no longer the team's bottleneck. Whether a user forms a new habit just because you have a new product, that's the hard part. Acquiring software users is still very expensive today, and users already face a flood of apps. The competition is shifting from whose model is stronger to who gets activated first, in the moment the user needs it.

SpeakON: an AI physical button.

SpeakON is very simple. It's an AI physical button. You press it, you speak, the task is done. You say the idea out loud, the software cleans it up and drops it into whatever app you're using. It's also an agent router that can connect to all kinds of tools and products.

For example, you can press the button and say: I have a meeting at 3 today, create a calendar event. SpeakON runs a series of tasks in the background. Or you're in Japan, you don't speak Japanese, and you want to book an omakase in Shibuya. You press and say: book it, this time, this many people. The agent handles it in the background.

SpeakON is not a voice recorder. It's more like a dictation button, an AI button. A recorder's job is to save the sound, generate an audio file or a transcription. That has value, but afterward you still have to listen, check, distill, then copy it into an email, Slack, a note, or some other doc. In other words, once the recording ends, your task isn't finished.

What SpeakON captures isn't a piece of sound to be stored. It's the idea you have right now, so you can record it fast and execute the task fast.

On the interaction side we split it into four layers. Press removes the activation friction: no hunting for an app, no blank page, just press and speak. Speak removes the expression friction: speaking is faster than typing and carries more context. Polish is where AI actually creates value: stripping out repetition and filler, understanding mid-sentence corrections, adjusting the format to the situation. Deliver removes the transfer friction: the content flows into whatever workflow you need.

Hardware brings a few advantages. People's willingness to pay for a physical product upfront is clearer. It's well suited to reviews, gifting, and word of mouth, because there's an object to show. The device stays within reach, which lifts usage frequency. And a device plus software means one purchase that keeps delivering value. SpeakON needs no subscription at all right now. Buy the hardware and you get unlimited dictation, because dictation costs almost nothing now. A user's whole year runs about $10 to $20.

Japan and the US are two completely different tempos.

In Japan, the first thing Notta had to do was earn the right to scale. Customers look at the company's credibility, its local service, procurement, and whether the brand has been built up over time.

In the US, what SpeakON has to do is earn the right to become a habit, by improving efficiency fast and constantly.

One market emphasizes trust. The other emphasizes whether you can iterate on efficiency fast. Underneath, I think it's the same: how do you build trust locally, then scale globally?

07 Four Basic Judgments From Years of Building Companies

These stints of company-building left me with a few basic judgments.

First, choose a situation that forces the team to build long-term capability. Japan is hard, but the difficulty taught us how to localize and how to build trust.

Second, don't build every capability out on day one. Only when the next layer of capability can open the next stage should you invest in it.

Third, let a business model change user behavior, rather than letting a fashionable concept drive it. The shift from B2C to B2B started because content began getting shared and teams began to depend on it.

Fourth, treat interface and trust as strategic assets. Whether your technical capability turns into value that happens in a user's day is a big part of whether the product wins or loses.

On serial entrepreneurship: the next venture should look a bit different from the last. If it's identical, you might just be repeating a domain you already know, without using what you learned to level up. SpeakON differs from Notta in product form, market, and business model. And the methods Japan taught me carry into this new set of judgments and keep working.

Thank you.

Insta360 Founder JK Liu: The Camera Agent Battle Is Beyond the Lens

Mon, 10 Aug 2026 15:37:13 GMT

That's how Insta360 founder JK Liu framed the future of the imaging industry at AGI Playground 2026. Last month, Insta360 unveiled its latest product vision: building a "Cameraman" - an imaging experience defined by simpler operation and better delivery.

"You could also think of it as a Camera Agent," Liu said.

The product will likely take many forms. Even with a light front end, different shooting scenarios still call for different lenses, different focal lengths, and different formats - first-person, third-person, and so on.

As in embodied AI, letting the device itself move freely through general environments remains a real challenge. But a Cameraman may not need a general-purpose brain. When it comes to recording and sharing life, taste that stays on point matters far more than intelligently executing standardized tasks.

Insta360 is already offering cloud-based AI editing directly to users, Liu said. It costs 6 RMB per clip. In the first two months of its limited rollout, the service has produced hundreds of thousands of edits, with an export rate - the imaging equivalent of the "roll rate" in AI video generation - above 50%.

But it's still early. A Camera Agent isn't just a standalone entity or service. What it has to improve isn't only its own intelligence and taste, but a deeper understanding of what people expect, aesthetically, from the physical world.

What follows is the conversation between Insta360 founder JK Liu and GeekPark Founder and CEO Jack Zhang (Zhang Peng).

01 · The Cameraman: the core isn't shooting, it's delivering the result

Jack Zhang: Welcome to AGI Playground, JK. Insta360 recently updated its product mission. Where you used to give consumers a "camera," now you want to give them a "Cameraman." What does that shift mean?

JK Liu: The Cameraman idea didn't arrive all at once. It started back in 2019. We mounted an action camera we were about to launch onto an FPV drone to shoot an ad, and the footage came out beautifully. But capturing shots like that takes real skill to fly and operate. So we worked backward from there: how do we let anyone capture footage like this more easily?

Push that further and you have to see the need behind the need. What people often really want is to stay immersed in the moment, while still coming away with photos and videos that capture and preserve those meaningful experiences. When people go out today, some of them hire a photographer to follow them around. This event has cameramen catching candid shots for us, and we don't have to lift a finger to end up with plenty of good pictures. That's the ceiling of the imaging experience.

Jack Zhang: So from that point on you started taking it apart - what technical means are needed to realize the "cameraman" vision?

JK Liu: To build the future cameraman, a camera needs to do three things well: see, understand, and act.

First, it needs to see more of the world. That’s one reason 360 imaging matters so much to us. A traditional camera captures only a narrow frame, while a 360 camera captures the full scene. That gives both the user and the AI much more context, reduces the risk of missing important moments, and creates a stronger foundation for tracking, reframing, and scene understanding.

Second, it needs spatial intelligence. The camera has to understand depth, motion, orientation, and what’s happening in the environment - not just record pixels, but actually interpret the scene.

Third, it needs to act in real time. That means tracking subjects, stabilizing footage, composing shots, predicting movement, and eventually making smarter filming decisions automatically based on the user’s intent. If it can’t respond instantly, it doesn’t really behave like a cameraman.

So for us, the future cameraman sits at the intersection of sensing, on-device AI, and computational imaging. The goal is for the camera to stop being just a recording tool and start becoming an active creative partner.

Jack Zhang: So the Cameraman is really a direction you'd already thought through internally - you're just sounding the charge now.

JK Liu: Right. We actually talk about it at the company all-hands every year.

Jack Zhang: You just didn't tell us.

JK Liu: We kept piecing the parts together, and now we're going for it. Something clicked when we were talking this morning: if "Cameraman" is expensive to explain, there might be a better word today - Camera Agent. Everyone's talking about agents now, right?

It's probably easier to grasp, without the detour. The core value isn't the process of operating a camera. It's the ability to deliver good photos and videos on its own.

Jack Zhang: We're not buying a gadget, we're using an agent. We've already gotten used to handing tasks to agents in the digital world and getting results back. In the future we'd like imaging to work the same way.

JK Liu: We hope it can work that way in the physical world too.

02 · Light at the front, heavy at the back: the traditional path is moving toward AI

Jack Zhang: Let's start with the imaging line. With the progress in large models, cloud, and on-device compute these past few years, what has changed most about imaging? It's the front-most part of the whole Cameraman. Where does it stand today, and what will the landscape look like?

JK Liu: There really are a lot of new technical paths here. Let me give a few representative examples.

Take denoising. The old approach was traditional CV denoising running from the sensor through the ISP (image signal processor). Starting in 2019, we did end-to-end denoising with AI models, which let even an action camera shoot night scenes well.

Later, the industry started doing tone mapping end-to-end, so in the last few years everyone - us and our peers - made a big leap in color.

But beyond the basics of color, dynamic range, and noise, imaging also has a lot to do with color grading.

That's where you get the things people intuitively think of as "photoshopping" - adding bloom, adding vignettes, layering the image. What we eventually found is that processing this directly through models produces many effects you simply can't achieve on-device.

And that's just for photos. In video, there are far more effects that on-device hardware can't produce at all.

So we think imaging may be going through a paradigm shift in how cameras are built: relying less on the front end. In the future the camera portion at the front could be very light, even very cheap, and it doesn't necessarily have to be premium hardware.

Jack Zhang: The camera is becoming more of a sensor. The optics still matter, but they're no longer the only thing that defines the device

JK Liu: There will always be people who chase traditional craft, of course. But I believe that for everyday users, in most cases, using something light enough and getting a great-looking image matters more.

Jack Zhang: Where does the industry still need to shore things up?

JK Liu: For example, the fine detail you capture with a super-telephoto lens is still very hard for AI to fill in. And cloud generation is asynchronous right now, whereas for a camera, instant capture is a very basic expectation.

So once the models mature, some of this may run back onto the device. On-device compute is currently limited by chip process nodes; once the chips advance, on-device inference and generation should get much stronger.

Either way, one thing hasn't changed: we are seeing more AI-empowered imaging in photography

Jack Zhang: That's the "light at the front, heavy at the back" idea - the optics end still has to meet a standard you can't lower, but the future gains sit further back, tied to AI and to compute.

JK Liu: Right. Take background blur. Traditional optics do this very well, but they basically can't be compressed: how big your sensor is and how wide your aperture is directly determine whether the depth of field and the blur look natural. Simulating that with an algorithm mostly doesn't work. At the same time, AI generation provides another way to achieve a similarly natural result.

Jack Zhang: So if on-device compute keeps climbing, you press the shutter and it's out in a second, and there's actually a generated component inside - is that a race between edge and cloud? Or will something else eventually decide the balance between the two?

JK Liu: I think plenty of industries beyond ours will develop this same branch. Some people are willing to pay less for hardware up front and subscribe over the long run; others would rather buy the service and compute outright, once.

I don't see it as a race, but as two consumption models that will coexist. Some people don't want to pay a lot up front and prefer to pay over time - and for some, this may not even be something they'll use forever, so the cloud is simply cheaper for them. But for high-frequency, must-have users, they may just buy the compute outright.

03 · AI will become a consumer of imaging too - and that's not necessarily a bad thing

Jack Zhang: Would you explore anything at the action layer? When you talk about the Cameraman, have you thought about it having "man"-like ability not just in how it generates and optimizes footage afterward, but at the level of shooting, moving, acting?

JK Liu: The cameraman will have a "brain" of its own - what we call 360-degree awareness of the environment. To capture a good shot or a good video, camera movement is critical: understanding what the subject is, what the scene is, what space or path is traversable, and how to trace that trajectory. That involves the body's own motion, which is fairly complex.

There's also a purely data-level notion of "movement": feed in a wide-angle image and find the good crop.

Jack Zhang: So the Cameraman doesn't necessarily need physical movement - it can be editing and cropping instead?

JK Liu: Right. One is ego motion; the other is focus movement, cropping within the panoramic image you've already captured.

On the whole, moving a drone is still much simpler than moving something on the ground, because above a certain altitude there are far fewer obstacles.

Jack Zhang: So a bipedal Insta360 device walking around the house shooting isn't the right direction to imagine?

JK Liu: In a complex scene like that, traversability is quite a challenge. A robot vacuum already has to solve a lot of problems; an indoor shooting robot has to solve far more than a vacuum does.

Jack Zhang: There's a huge amount of capital and a lot of talented people exploring embodiment right now. For their data needs, the devices of the past aren't necessarily the optimal way to collect anymore. Are there organizations like that talking to you? How do you see it?

JK Liu: Broadly, in this space we think the path for data collection hasn't converged yet. Some collect first-person data, some collect first-person plus third-person, some collect panoramic, and some don't collect data at all - they learn straight from video.

To put it bluntly, it hasn't reached scale yet. There are only a few points of consensus right now - for instance, that you don't use teleoperation in the pretraining stage, because the collection cost and the data volume can't keep up.

Jack Zhang: Could that change the nature of demand? Everything imaging-related we've built so far has been for people. If the age of robots arrives, the demand for sensory data for robots could be even larger than for humans - the way human search volume is falling today while AI search is rising exponentially. If the embodied era arrives, how do you think imaging changes?

JK Liu: I think the impact on our industry's target customers is fairly small, whether from AI or from embodiment. Because imaging serves people's record of their lives, and recording has nothing to do with AI. It's about the beautiful, precious moments of a person's own experience.

The things we shoot on our phones aren't all meant to be shared. Some are just for our own memories, or to put in a digital photo frame at home. People don't tend to generate an image of something that never happened. From the standpoint of recording, I don't think there's much impact.

But beyond recording, imaging has two other kinds of demand.

One is sharing. Here AI can act as an amplifier, bringing out effects you couldn't have captured yourself. The other is creation - using imaging to express an idea, or to shoot a commercial or make a piece of work. There, I believe AI can also play an important role.

04 · The Cameraman needs a dedicated brain - one with taste

Jack Zhang: There's a debate in the industry lately about whether an all-in-one device will eventually "rule them all," and whether a product built for one narrow niche will always have a low ceiling. It rhymes with a question in AI - whether to build a general-purpose agent, or to go all the way on a few well-defined needs. How do you see it?

JK Liu: I think the camera category is fairly special. From first principles, imaging content falls into a few different types, and that alone guarantees the hardware will split into at least several forms.

First, shooting what we ourselves see. For that we'll mostly use a phone, a camera, or a gimbal camera.

Second, POV - first-person view. This kind of content is common during motion, and it demands a wearable camera, so the form becomes glasses, a thumb-sized camera, or an action camera mounted on your head.

Third, the third-person view, which frees your hands. From the Cameraman's standpoint it definitely isn't something you operate; it's independent. So beyond what you hold in your hand, there's what you wear on your body, and there's the thing at some distance from you that moves autonomously and shoots you.

Jack Zhang: So a single all-in-one device won't necessarily realize the whole Cameraman?

JK Liu: There's an important concept in imaging called focal length - how wide and how far a frame can capture. Different focal lengths serve different purposes.

Panoramic is great for shooting everything around you from your own vantage point; you capture it all and let computation and AI handle it afterward.

But from the standpoint of recording, a good portion of your frames has to be about shooting you: the camera up close on your face with the background, farther out on your full body with the background, or farther still with a telephoto on your upper body with the background - each produces a different feeling. And there's a category of beautiful video shots whose common trait is the relative motion between subject and background, which only a third-person view can produce.

So imaging has two broad forms and needs: one is shooting you from another vantage point, the other is shooting other things from your vantage point. Both are important pieces of visual language. For data collection you need both; for applications they're two different kinds of demand. A camera may be like a car: is its essential function getting from point A to point B? Yes - but it still branches into a great many things.

Jack Zhang: So from the brain's perspective - in embodiment everyone's saying you need one general-purpose brain. For imaging, is the future brain one, or many?

JK Liu: Just for the Cameraman's brain, our current read is that it probably won't be an especially complex thing. Could we be absorbed by a larger model that swallows this piece whole? That's possible too.

But there's another line of reasoning from the business side. I believe that whether you're building models or building intelligence, commercially everyone will still pursue their own distinctiveness - the part that only you have.

So by default, it's genuinely possible that one general-purpose brain does everything across every industry well, or that for photography and videography specifically, one brain handles most scenarios quite well.

But from the standpoint of human nature and of business, I think it's very likely everyone will want this: I understand this scenario deeply, so I've trained on more data here, some of it data only I have, and maybe the shooting method itself is something I invented - I turn that into a model or a skill dedicated to this scenario, and I charge a premium for it.

Jack Zhang: When it comes to imaging, what may matter more to us about that brain is better taste - in theory it should surpass mine, if it can deliver that. With what data and what methods can you make sure it reliably delivers higher taste across the board? Is there a clear answer to that today?

JK Liu: That's a great question. The direction we're training on for composition and editing is still common-denominator taste.

The reason our auto-editing - the Cameraman - is relatively deliverable today is, first, that it doesn't act back on the physical world, and second, that it can offer several options at once. Its evaluation standard doesn't converge, so one of them usually lands. The problem is you can't guess the customer's intent 100% of the time.

Internally we have a formula: customer satisfaction equals what you deliver divided by their expectation. But the expectation itself is hard to pin down. As an analogy, say an AI decides what you eat tonight. However well it knows you - the way your wife decides what you're having for dinner - it can't nail exactly what you're in the mood for every single time.

The next direction probably comes back to the customer: we offer several options, learn their preferences from which ones they pick, and feed that back as context or parameters for the next inference, until we get to something personalized for each individual. But that hasn't played out yet, because how often a customer edits videos is nowhere near the volume of data from how often they scroll through them (laughs).

So on taste - on this kind of demand - the hit rate for matching customer needs always seems to have an invisible ceiling.

Jack Zhang: I know you already have a feature that edits videos for users directly in the cloud, and some users are even paying for it. Do you think that share of the service will keep growing as the technology and the taste improve?

JK Liu: I think it's inevitable. You've already spent all that time shooting, so of course you'll want it summed up, whether for yourself or to share with friends - and it takes a real amount of time. You've already spent the money to travel and the time to shoot; from the standpoint of the final step, the demand is clearly there. In the past, editing was simply too costly for customers, so many people skipped that last step.

We currently have some premium editing services, at around 6 RMB per clip. In the roughly two months since launch we've edited hundreds of thousands of clips, and that's still just the limited rollout.

Jack Zhang: How many rolls does it take before they keep one?

JK Liu: Our export rate is above 50% right now. For every two options we give, roughly one gets picked - about that level.

Jack Zhang: One in two gets used.

JK Liu: Yes. But that still depends on where we sit. We judge this to be a must-have, and in the end some people will probably choose to buy the compute outright - after all, at 6 RMB a clip with some chance of having to roll for it, a fair share of people just don't consume that way. It may eventually have nothing to do with buying compute either: as phone compute grows, it might run entirely on the phone, or on a NAS at home.

Jack Zhang: Do you think where you "sit" could change in the future?

JK Liu: Overall, customer behavior and consumption habits are a distribution; it will never be 100% one thing. Not 100% cloud, not 100% buying the compute on the camera, not 100% on the phone or a NAS.

From a company-strategy view, we will consider extending computation from one platform to others. But a big part of it is the maturity and penetration of the platform itself. Charging a customer is a serious matter today; we have to make sure the experience is good. And what can guarantee a consistent experience today is either the cloud or your own hardware. Third parties - whether a NAS or a phone - are still far from our compute target. There's basically no consumer-grade chip or product delivering hundreds of TOPS integrated into a phone or a NAS. But once that trend takes off, our compute service will certainly extend along with it.

Jack Zhang: The impression that stuck with me from this whole conversation is that if we're very confident about AI's intelligence today, we probably shouldn't be over-confident about taste, because taste seems to be personal. So at the level of the Cameraman, if taste is the key, then it's hard.

JK Liu: Let me tell you a joke: every carmaker has great designers, but whether the car looks good really comes down to the taste of the person running the company.

Jack Zhang: In the end it still comes down to the boss. Thank you, JK. Here's hoping your taste stays on point and brings us more great products.

How MiniMax Builds Agents

Mon, 10 Aug 2026 11:03:05 GMT

Model-as-product has, by now, become model-as-agent.

The unique advantage a model company has in building agents is that it can co-design the agent harness and model inference inside a single team. Every decision on the harness side - how the system prompt is written, how tool calling is organized, how errors are handled - directly affects the cache hit rate, scheduling efficiency, and token consumption on the inference side.

MiniMax's past year of agent work started as an internal Feishu (Lark) bot, evolved into a desktop app called MiniMax Code, and finally closed into a training-and-serving loop: the model trains inside its own harness, serves users through that harness, and user behavior becomes the training signal that feeds the next model. That loop is the core logic of how a model company designs agents.

In his talk at AGI Playground 2026, Vincent Wu, who leads Developer Relations at MiniMax, laid out how a model company sees the relationship between model, inference, harness, and agent, and shared what the team has learned building agents over the past year.

Compiled and edited from his conference talk.

01 · The MiniMax agent began as an internal "Agent Intern"

What I want to share today is how we, inside a model lab, think about agents.

The model is our product; that has never changed. The question is: why does a model company need an agent at all, and how should it design its own? We've spent a year on this and moved through several stages.

The earliest was around May or June of last year, when we built a cloud agent (agent.minimax.com), similar to Manus, with its own Linux sandbox and virtual machine that could build websites for users. It used traditional tool calling, but it could already operate the sandbox and run shell scripts. That one was more of a demo.

What really rooted the agent inside the company was something we called the "Agent Intern." It surfaced as a bot inside our everyday office software. From the start, the Agent Intern wasn't just for engineers - everyone at the company used it: engineering, HR, marketing, finance.

Sales faces prospects in different countries every day, and every message has to read differently, so they had the Agent Intern research each person's background, look at what they'd posted before, and generate personalized copy. The designers' use was more direct: they'd vibe-code in Cursor, resolve design-fidelity issues in a code branch, and hand it to the front-end team. Product and support were messier - the Agent Intern helped PMs handle user-complaint emails, pulling the back-end task trace, judging whether the credit consumption was reasonable, suggesting a refund proportion, and then sending the email to execute it. The compute-platform team used it constantly to analyze alerts, since the alert metrics for large-model inference move very fast. The data-analysis team said it replaced the traditional SQL-writing part of their work; HR used it to search schools and screen résumés, a 10x jump in efficiency. Others used it to write code automatically - finish the requirements list, and the agent writes the implementation entirely on its own.

The most interesting change, to me, was in mindset. One engineer put it well: "What we're really doing is having the agent - our intern - create an intern that can do the job." They were no longer an engineer, but something closer to a leader. We call it an intern for now, but we think one day it will be like a digital employee.

These internal practices pushed the next step. We found the cloud agent wasn't enough - a lot of workflows actually live on the local machine. So we turned it into a desktop app, released as MiniMax Code, which can control the user's computer but doesn't require you to be an engineer to use.

At the same time, we found that the M2-series models - especially M2.1 and M2.5 - worked well with third-party open-source harnesses, mainly OpenClaw and Hermes. We worked closely with those communities to tune the models' performance inside them. The reason is that MiniMax's models are generalist: not just good at coding or math, but also at operating a computer and handling managerial tasks, and cost-effective on top of that.

Which brings us to today: our new model, MiniMax M3, and MiniMax Code.

02 · How the agent helps optimize and iterate on the model

Before I get to how the agent helps with training and serving, let me answer a simple question: what does a foundation-model company actually do? It comes down to two things - train models, then serve them.

The inputs to training are compute, data, gyms (training environments), and training algorithms. The inputs to serving are compute plus inference algorithms. Having the model weights and a few GPUs doesn't mean you can serve the model; there's a lot of hard engineering behind inference serving.

The agent fits into this in two ways. The direct way is to have the model help us with R&D - writing better training algorithms and better inference algorithms.

Start with the direct way. When we released M2.7, we shared that the post-training team's workflow breaks roughly into five steps, of which two - running experiments and analyzing them - can be done entirely autonomously by M2.7. The remaining steps aren't fully automated, but all of them are AI-assisted.

More importantly, the harness code used to do this research was itself written by M2.7. Not by humans. M2 wrote the agent's code, and then, through that agent, carried out recursive self-improvement tasks.

Why does this help so much? Before agents, a researcher's time was serial: propose a hypothesis, run the experiment by hand, spend a lot of time debugging - because these experiments are hard to run - get the result, then propose the next hypothesis. Most of that time wasn't spent thinking of new ideas; it was spent debugging.

With a research agent, the tedious-but-hard work of running and debugging gets taken over. Researchers can spend more time proposing new conjectures and debating the right direction with each other. This is a structural change: it's not that you go faster, it's that you can run more ideas at once.

Direct R&D is only half of it. The more interesting part is the indirect way - using the harness to connect training and serving into a closed loop.

The logic is this: real users do real tasks through MiniMax Code; our harness collects these messy, long-horizon workflow traces and environments; and then, through a lot of data engineering, we turn them into repeatable, scorable gyms for reinforcement learning and other post-training.

Gyms are the next key direction for model training. We've now entered the "era of experience," where it's no longer enough for a model to do well on fixed datasets and fixed math problems; it has to perform in real-world work.

We even have a dedicated team for this. We've hired part-time and full-time experts from fields like finance, law, and medicine, whose main job isn't to label data but to use MiniMax Code and the M3 model within their own professional workflows. That usage becomes real-world, high-value training signal that feeds into the next model's RL training.

That forms a flywheel: the model finishes training, serves users through the harness, collects real usage data, turns it into training signal, and trains the next model. Every turn of the wheel, the model gets better, the data quality gets higher, and the next model gets better still.

03 · What an in-house agent does better than a third-party one

M3 is strong at writing kernel-optimization code. Kernel optimization is the heart of inference serving - the code that binds the model weights to the GPU. Our inference team can use M3 to write better kernel code that serves M3 itself. That's another self-improvement loop.

M3 was itself trained inside MiniMax Code, so it naturally runs better in that harness.

But there's a question fewer people discuss: why should a model company build its own agent harness rather than rely solely on third parties?

Three levels of co-design explain it.

Model-chip co-design is already familiar: design the chip and the model together so the model architecture runs better on the chip. Model-harness co-design is easy to grasp too: the model trains inside a harness, so it naturally performs better in that harness.

But the third - inference-harness co-design - is discussed far less. It means that because both the harness and the inference service are in our hands, the harness can in turn help the inference team optimize serving efficiency.

How? More reliable and cleaner system prompts, progressive disclosure of skills, the right tool-calling patterns, good tool-error handling. These all look like harness-side design decisions, but they directly shape the demand patterns on the inference side. More predictable demand patterns mean a higher cache hit rate, more efficient scheduling, and fewer tokens consumed per task.

This is part of the reason Anthropic doesn't let subscription users run its service inside third-party harnesses. Part of it is branding, but another part is engineering - an in-house harness lets them better control the structure of inference demand.

For users, this means cheaper, faster task execution and fewer errors and rate limits. For us, it means serving more users on the same compute. It's a win-win, and it's something third-party harnesses can't do - they don't have control of the inference side, so they can't do inference-harness co-design.

At our scale, these efficiency gains add up. Compared with relying purely on third-party harnesses, we can reach much higher token throughput.

04 · AGI isn't a single model but the whole closed-loop system

Back to the full picture. The agent helps us on two levels. Directly, it helps researchers write better training algorithms and helps the inference team write better kernels. Indirectly, the in-house harness makes inference serving more efficient while turning real user behavior into training signal.

This loop is tightening. It's worth imagining what the system looks like after it runs for six months, or a year. The model is still the core; from training to serving, everything revolves around the model. The model is still the product. But the harness is now an indispensable link.

My conclusion is that AGI may not be a particular model checkpoint, but the whole closed-loop, recursively self-improving system. You can imagine a day when the entire cycle - from training to serving to data collection and back to training - runs with no human in the loop. The model closes the loop through its own harness and keeps improving itself.

For other model companies, this past year points to at least one thing: the agent isn't an accessory to the model. It's the axle that connects training and serving, the lab and the real world. Design that axle well, and the flywheel turns.

Eric Jing: The All-in-One That Matters Isn't Features. It's Context.

Tue, 04 Aug 2026 12:40:52 GMT

At AGI Playground Singapore, Genspark made its stage debut with a new product: GenOffice, a full AI office suite (docs, spreadsheets, slides, PDF). Free, open-source, PC and Mac. One engineer built it in a week.

Genspark has gone from AI search, to general-purpose Super Agents, to a full AI Workspace, and each pivot has landed on an inflection point in the technology. The bet behind all of it: models will commoditize as they compete, so the durable moat isn't the model, it's the context. Pull a user's work data, habits, and history onto one platform, and AI stops being a tool and becomes a work partner. The endgame Jing describes in one line: work does itself.

He's clear-eyed about the odds. 99% of AI startups fail. Genspark is building to be the 1%, and Jing thinks that window is still open.

On August 3, GeekPark founder Jack Zhang sat down with Eric Jing at AGI Playground 2026. The full conversation follows, edited by Founder Park (Founder's community of GeekPark).

01 Users Always Seek Better Tools - Standing Still Is the Real Risk

Jack (Peng) Zhang:

Let’s start with a product you’ve already iterated past - AI search. It was Genspark’s starting point, and you had over 5 million users when you decided AI search wasn’t the endgame for the company, shifting instead to explore new product forms. Can you take us back to that decision moment? What did you see that made you commit to this change?

Eric Jing :

Let me add a bit of context first. I joined Microsoft as a software developer in 2006 and spent most of my career working on search. Over my 20-year career, it all boils down to one thing: building products that connect people with information.

The essence of search is matching people with information. Before founding Genspark, I worked with billions of user queries and learned one key truth: people don’t use search for search’s sake, or just to find web pages - they use search to get things done. People want to complete entire tasks more simply, and spend less time on tedious busywork.

Back in our early search days, we tried all kinds of ways to optimize the experience. Our team even positioned Bing as a “task engine,” trying to help users complete entire workflows end-to-end. But 20 or even 10 years ago, the technology wasn’t mature enough to deliver on that vision.
That all changed when ChatGPT arrived. We suddenly realized this AI technology could actually turn that long-held dream into reality: AI could take over all operational work, and humans could focus on creativity. That was the founding vision for Genspark AI.

From day one, we wanted to build an autonomous Agent that could complete tasks end-to-end. But at the time, base model capabilities weren’t strong enough. As a startup, we couldn’t afford to wait around for perfect technology that might never arrive on our timeline. We had to survive first.

So we had to find the most basic, realistic Product-Market Fit (PMF) with the AI capabilities available at the time. Two years ago, AI search was one of the clearest PMFs in the market - so that’s where Genspark started.
Then Claude Sonnet 3.7 launched. It was the first model that made end-to-end, full-process Agents practically feasible. The moment we saw that signal, we knew it was time to evolve Genspark toward end-to-end task completion.

Outsiders call this a pivot from AI search to AI Agents, but for us it was a natural extension. The technology we’d been waiting for was finally here. We wanted to bring it to market as fast as possible and turn Genspark into a truly AI-native work experience.

Jack (Peng) Zhang:

For you, this was the moment to deliver on the company’s mission. But 5 million users is a massive milestone for any founder. Did you assess the risks of making this transition?

Eric Jing:

It’s not that I saw no risk in transitioning - it’s that I saw far greater risk in standing still and missing the opportunity. Users are always searching for better tools. Even if a new tool feels unfamiliar at first, once people try it and see it’s significantly better than the old way, they’ll migrate.

Our user transition was incredibly smooth, because this was fundamentally a step up in experience. Instead of reading AI summaries and clicking through web pages, users could generate a full presentation or complete tedious document work with a single prompt. For users, it was a pure upgrade. That’s why the transition was seamless for both us and our user base.

Jack (Peng) Zhang:

So in your view, the risk of inaction far outweighs the risk of iterating forward?

Eric Jing:

Exactly. Standing still and missing the opportunity is the far bigger risk.

Jack (Peng) Zhang:

Honestly, the pace of AI product iteration over the past two years feels surreal. We’ve gone from chatbots to AI search, general-purpose Agents, and now AI Workspaces - all in just two years. I’ve been trying to unpack the core drivers. Is it purely model evolution, or are other factors pushing this forward?

Eric Jing:

AI is fundamentally different from every traditional technology we’ve seen over the past 20 to 30 years. I’ve lived through multiple technology shifts - from the personal computer era, to PC internet, to mobile internet. Past technology iterations followed a familiar pattern: talented teams in a given space would iterate and polish the same core technology, making incremental improvements day by day. But AI is an entirely new species. It evolves so fast that every new model release can unlock entirely new possibilities.

It’s like compressing a century of progress like the electric power revolution into just 5 years. That’s why AI is changing so fast it’s hard to keep up - and it’s only accelerating, with no sign of slowing down. That’s the first point: the inherent nature of AI technology is fundamentally different from previous tech generations. The second point is that the entire ecosystem was already primed for AI.

Over the past 20 to 30 years, our daily work has been taken over by all kinds of software and services: email for communication, CRMs for customer data and conversations, online meeting tools for discussion records. A massive amount of digital assets has already accumulated. This work context has been “maturing” for 20 years, just waiting for the AI era to arrive. When mature data context meets fast-evolving AI technology, they create a flywheel effect: the more context you feed a model, the better it gets at solving everyday tasks. This flywheel is what’s driving the digital workspace to evolve so rapidly - users get a noticeably different experience every single day.

02 The True All-in-One for AI Is an All-in-One Context

Jack (Peng) Zhang:

We’ve known each other for years, and I know you’re a firm believer in the “All-in-One” product philosophy. But what I see today is that most people still do things the old way across many scenarios. Meanwhile, you’re building fully AI-native products rebuilt from the ground up - yet you’re still committed to the All-in-One approach. How does that align with your product philosophy?

Eric Jing:

I am a firm believer in All-in-One products. Many of the greatest products in history are fundamentally All-in-One: search is All-in-One, the smartphone is All-in-One, WeChat is All-in-One. It’s a proven path: you build a simple, easy-to-use interface, and keep adding more capabilities into it over time.

In the AI era, I think we’re one of the first companies to put this All-in-One AI concept into practice. And I believe more and more companies will come around to this direction - All-in-One is the future.

But the All-in-One I’m talking about is centered on context, not features. Models are important, of course, but our judgment is that over time, models will gradually become commoditized infrastructure.

Let me show you a slide - this is updated from just two weeks ago. New models are launching constantly: we’ve already seen Kimi K3, DeepSeek V4, and many others. On one hand, every model today is already very capable, each with slightly different strengths for different tasks. On the other hand, the cost gap is enormous - the most expensive models can cost up to 50 times more than the cheapest ones. Most people have no idea which model to use for which task, and the same goes for enterprises.

Three months ago, I attended a closed-door CEO summit hosted by Microsoft. There was an anonymous poll, and the number one pain point for CEOs right now is not knowing which model to bet on - AI is iterating too fast. Another 24 CEOs said that even after spending heavily on AI services, they worry their teams won’t actually use them. The first problem is model selection paralysis; the second is usability.

Behind these numbers is a simple reality: there are too many models on the market today. People don’t have the time to test them all, or the expertise to judge which ones are good. And cost is becoming an increasingly big concern. That’s where companies like ours come in: we sit in the middle, translating raw model capabilities and parameters into simple, powerful, easy-to-use platforms.

Our version of All-in-One has two layers.

The first is model All-in-One: we bring all models together in one place, matching the optimal model to each task. This saves users money while delivering top-tier performance.

The second is context All-in-One. Building a unified interface is easy. But today, knowledge workers’ data is scattered across dozens of apps and services. People are constantly copying and pasting work context from one tool to another. If we can consolidate all that context in one place, the platform can fully understand your work habits. Combined with our broad model selection, it can become your true work partner.

When you need to get something done, a partner that has all your work context at its fingertips can handle it efficiently and cost-effectively. That’s the future we believe in.

We keep adding features and improving capabilities, but the goal is always to capture more and broader horizontal context. The essence of All-in-One is aggregating all the horizontal context from across your work life. With unified context and a unified brain, an All-in-One platform can deliver high-quality results automatically, at low cost and high efficiency. That’s why I’m convinced All-in-One products in the AI era will not only work - they’ll create enormous value for users and enterprises.

Jack (Peng) Zhang:

So the very concept of “All-in-One” is evolving. It used to mean packing different features into one app with one UI. Now the more important part is that all products share the same core context.

Eric Jing:

Exactly. This is a new generation of All-in-One philosophy. The concept itself is evolving. The old definition - one interface with many features stacked inside - still holds true. For AI products, that means a web interface with different sub-Agent capabilities built in.

But as I said earlier, AI technology changes the whole equation. If one person can build a full-featured office product in a week, why make users come to Genspark? Why not bring the service to their doorstep? Why not open the restaurant in their neighborhood, and deliver the food right to their home?
That means letting users get tasks done in the interfaces they already know and use every day - while we capture that work context at the same time.

Jack (Peng) Zhang:

Using technology to create greater value. That’s the new variable this era gives all founders.Now, many founders are hunting for new AI use cases and untapped user value. From your experience, are there any scenarios that you didn’t expect at first, but ended up becoming core sources of user value? Can you walk us through some examples?

Eric Jing:

There are quite a few, actually. Two years ago, we were one of the first companies in the world to launch an AI presentation product. When we first released it, a lot of people pushed back: who makes business presentations in HTML format? Everyone wants PPT files - that’s what looks professional. But from our perspective, HTML is one of the most native file formats for large language models.

Building presentations in HTML lets you generate higher-quality content with better layout and design. Then users told us that for some use cases, HTML versions work perfectly - like sharing information with teammates, where format doesn’t matter but editability does. So we built an online HTML slide editor. Then users asked: the online editor is great, but can I edit these HTML slides in my desktop PPT software? So we built an Office plugin. Then users said the plugin was too expensive and too slow - so we went ahead and built our own native GenOffice from scratch.

We don’t build products in a vacuum. We listen closely to user feedback. Every time we ship a new feature, users come back with new requests, pushing us to make the product better. That’s how we iterated from a basic online AI slide tool to a full-featured desktop office suite. The email client you see today is another example born directly from real user pain points. I use email every day - it’s one of my most-used apps for work communication. But I constantly struggled with the same problem: when I need to find an important email, I can never remember the details to search for it. So we thought: why not build an Agent directly into the email client? We use this email client ourselves every single day.

For example, when we were planning this event, I asked our email Agent to go through all my correspondence with Geek Park, pull together the full schedule, extract the PDFs from attachments, block time on my calendar for each session, and even draft emails to your team in my tone to suggest small tweaks to the agenda. These are small, granular needs, but they’re absolutely core to daily work. If you really dive into scenarios and listen to users’ real pain points, you’ll find enormous opportunities to reimagine traditional software and services. That’s been our core learning along the way.

Our product philosophy at the company is pretty simple, just two rules: First, build products we love and trust ourselves. Don’t build products for other people - build them for us first, and be honest with yourself. We are the first users of our own products. If we don’t love using it every day at work, if it doesn’t solve our own real problems, it’s not ready. Second, we believe there are lots of people in the world just like us, with the same pain points. The more universal and common the pain point you define, the more users it will resonate with. So our approach is: build it for ourselves first, make something we love to use, then trust that more people will feel the same way. That’s our product methodology.

03 AI Value Lives Beyond the Model Layer - The Translation Layer Matters Most

Jack (Peng) Zhang:

Speaking of this, it feels like the definition of “general-purpose Agent” is shifting. At first, everyone talked about Super Agents that could do anything. Now it seems like the priority is doing specific things really, really well.

Eric Jing:

We’ve always maintained an open and respectful attitude toward new technologies. We even try to unlearn all the experience we’ve accumulated over the past 20 years - because new technology is a whole new species, and you can’t judge it by old rules.

Let me show you another set of data. This chart went viral in the industry back in February. Each dot represents 3.2 million people. The gray dots are people on Earth who have never used AI. The green dots are people who have used AI for free. The yellow dots are people who pay $20 a month for AI. And that single dot all the way on the right - just one dot - represents people who use Super Agents and Coding Agents.

That’s the reality I build products for. We want to build a company that stands the test of time. I’m under no illusions: 99% of startups fail. But even with those odds, we want to chase that slim chance of building something great.
When I look at this chart, I think: if we want the vast majority of people on Earth to use AI, we can’t just build cutting-edge experiences that only a tiny fraction of power users can understand. We have to meet people where they are. We can’t make users come to us - we have to go to them.

The old technology playbook was: open a restaurant, and invite people to come dine with you on weekends. But with AI technology, we can open that restaurant right on your street, and deliver the food straight to your door.

That’s our core internal principle:

meet users where they are. Building client-side software used to be incredibly complex, with extremely high barriers to entry. Today, those barriers have dropped dramatically. So why not work with users’ existing habits, deliver service through interfaces they already know, and lower the learning curve?

And as users use the product, we capture their work context, aggregate that data back, and build out that unified brain, that unified context. That’s why I believe being a good “translator” between models and users is so critical.
We have another analogy we use internally. Coding Agents are like supercars - incredibly powerful for developers and researchers, but they’re single-seater vehicles. Only expert users can really get the most out of them.
What we want to build at Genspark is a self-driving car that goes just as fast. You can pick any model you want, it runs in the cloud with unlimited compute, it always stays state-of-the-art - and you never have to worry about “driving” it. It handles everything itself.

If we can deeply understand the strengths and characteristics of every model, we can stand right at the intersection between models and users. That intersection touches consumers and enterprises, software and hardware. Players who stand at that crossroads get access to an enormous amount of context data.

Jack (Peng) Zhang:

That really does build a unique moat.Which leads right into my next question. You’ve grown very fast - your Annual Recurring Revenue (ARR) is already over $250 million and still growing. But there’s a common critique of Agent companies right now: that they’ll end up as nothing more than distribution layers for model providers, just middlemen selling tokens for a small cut. What’s your response to that?

Eric Jing:

First of all, we don’t see ourselves as an “Agent company” - we’re an AI company. Two years ago people called us an AI search company. Today they call us an AI Agent company. Tomorrow they’ll probably call us an AI Workspace company.

In my view, the AI revolution has barely started - we’re maybe 10% of the way through. Product forms will keep evolving. What we want to build is an AI company centered on All-in-One context.

For AI, the most critical thing is continuously compounding context, building that context flywheel: the more context you feed into the unified brain, the smarter and more personalized it gets.

As for the “distribution” argument: I’ve traveled all over the world studying what makes successful companies work. For some core products at companies like Microsoft and Google, distribution channels drive 80% of revenue in certain markets. Distribution itself is a moat. It’s not a dirty word. If we can build enterprise-grade distribution capability, that’s already extremely valuable.

Let me use an analogy to explain the early-stage industry landscape. Large model labs are like fishing companies - they supply the raw material. Coding Agents like Cursor are like high-end sushi restaurants, specializing in one vertical. Platforms like OpenRouter are like wholesale distributors.

And Genspark? We’re the neighborhood chain restaurant that delivers right to your door. Fish is just one of our ingredients. We turn it into different dishes, tailored to your taste.

In other words, we build massive amounts of value-added work on top of models. Nobody calls Apple a “hardware component distributor” - because Apple does enormous integration and innovation on top of those components. It’s the same for Genspark. Inside our All-in-One workspace, we’ve built extensive tooling, orchestration frameworks, and our own Agent-friendly file system. We aspire to be the kind of integrative, innovative product company Apple is.

04 The AI Era Will Give Rise to a Trillion-Dollar Application Company

Jack (Peng) Zhang:

Your global expansion has been really impressive. You’re not just growing in the US - you’re seeing explosive growth in Korea and Japan too. What’s the key decision behind that success? How did you pull it off?

Eric Jing:

The US, South Korea, Japan, France, India, and Brazil are all our core markets today. I’ve traveled all over the world, and you definitely feel the differences between markets. But if you look past the surface, at the core it’s all about human needs - and those are universal. Take New York, for example. We were one of the first companies to advertise heavily in the New York City subway.

We did a very concentrated ad buy there. New York is a unique city: it’s almost hard to believe how often people lose cell signal on the subway. That time with nothing to do but look at ads is an incredibly high-intent scenario. User habits also differ. A lot of New York users work in finance and investing, and they live in Excel. In Japan, people commute by subway and taxi heavily, and they’re power users of PowerPoint. South Korea has a huge design industry, so people use all kinds of design tools constantly. Those are the surface-level differences. But the underlying logic is the same.

I come from a search background, and this is where search taught me a huge lesson: search is an incredible product. When I worked on search, we didn’t care where users came from, who they were, or how old they were. We only cared about their query, and matching that query to the right web page. Do that well, and you can serve millions of people - without ever even knowing who they are. That’s where I learned the principle: great products acknowledge differences, but abstract them away into universal solutions. Users in New York need to build presentations too. Users in Korea need to work with spreadsheets too.

You just adjust the priority and weighting of features based on the unique traits of each market. Every time I visit a new market, I talk to local users and enterprise customers, and we adjust our roadmap priorities accordingly. The end result is a product that acts like a “global citizen” - it works everywhere. That’s how we’ve approached our global expansion, step by step.

Jack (Peng) Zhang:

We’ve known each other a long time, and I remember you saying two years ago that the AI era would give rise to a trillion-dollar company. Do you still believe Genspark has a shot at reaching that scale?

Eric Jing:

99% of startups fail. At Genspark, we operate with that sober reality front of mind, and we do everything we can to be the 1% that survives. I think there’s still a very narrow path to that trillion-dollar goal. We’re aiming for it, and we’ll work as hard as we can to get there. One trend I’ve observed recently: if you take today’s model intelligence level as a baseline, in two years the cost of that same level of intelligence will drop to 1% of what it is today.

People have different takes on the industry, and I fully respect that. But our observation is this: competition among large models is heating up fast, and under that pressure, model supply will trend toward homogenization and commoditization. When models become infrastructure, value shifts to the middle “translation layer” - to product companies, just like Apple in the PC and smartphone eras.

If we can integrate 70 or 80 top-tier, multi-modal models on a single platform, and build up unified user context so we truly understand our users - combining those two strengths while standing at the intersection of all parts of the ecosystem - we can automate most of people’s daily busywork. If we get there one day, our value will be that of an AI company that truly understands its users and solves their problems.

The more work we help you with, the better we get at helping you with the next task. Someday, people might only need to do two things at work: approve, or send back for revision. All the repetitive work will already be done by AI. You just decide “yes, go ahead” or “no, do it differently.” That’s why I believe AI application companies have a real shot at reaching trillion-dollar scale.

Jack (Peng) Zhang:

The past couple of years, VCs have been obsessed with investing in young founders. The logic goes: young people don’t have the baggage of old ways of doing things, so they’re better suited for a new era. But you’re a very successful serial founder, and you don’t fit that traditional “young founder” mold. I’d love to hear your take: what role does age actually play in AI entrepreneurship? And what matters more than age?

Eric Jing:

Before I answer that, let me quickly add one observation about enterprise AI. AI iterates so fast. A few months ago, everyone was talking about “token stacking” - chasing maximum model capability. Now in the US, the trend is “token optimization” - which is really just a fancy way of saying cost reduction and efficiency gains. In enterprise, no company wants to bet their entire stack on a single model. Models change too fast. So cost optimization is now the mainstream conversation in enterprise AI, and open-source models are gaining traction globally. These are the market signals we pick up on, and we adjust our solutions accordingly.

Going forward, I think closed-source and open-source models will converge. Enterprise users want the same thing: strong model capabilities, and reasonable cost. That’s the industry reality right now. AI is moving so fast that it’s easy to get caught up in the hype. But I think it’s important to maintain an open, rational mindset toward the technology. That’s my take on enterprise AI and where model competition is headed. Now, back to the age question. We have a mix of young people and very experienced people at Genspark. I know some cultures place a huge premium on youth. But if you look at Silicon Valley - the founders of OpenAI, Anthropic - none of them fit the classic “young founder” stereotype.

They’re all deeply experienced practitioners. I read a book once that made a point that stuck with me: the brain works like a filter. As you get older, that filter gets stronger. When you learn new things, less of that new information actually makes it through the filter and into your brain. But the more people I meet, the more I realize how much variation there is. There are plenty of highly experienced people who still have incredibly open minds toward new things today - even after they’ve become very successful, they’re still writing code themselves. People like that are incredibly powerful. If someone has deep industry expertise and the energy and open-mindedness of a young person - they’re basically superhuman in the AI era.

That’s why biological age is never a hiring criterion for us. What we value most are mindset, intellectual openness, self-motivation, and bias for action. Those are the core factors we use to evaluate talent. Younger people may have a higher percentage of folks with that open mindset. Among more experienced people, that percentage might be lower. But when you find someone who combines experience with openness, their output is far beyond average. That’s how we approach hiring and talent.

Jack (Peng) Zhang:

Age doesn’t matter - mindset does. Thank you so much for sharing, Eric. I hope you keep finding new horizons and new discoveries along this journey, and enjoy every step of it. We’d love to have you back at AGI next year to show us what you’ve built in the year ahead.

7 Things We Learned From 1,000+ Conversations With AI Founders

Fri, 31 Jul 2026 17:34:25 GMT

We've spent three years talking to the people actually building AI companies: founders, researchers, investors, and product teams across China, SF, Tokyo, and London.

Not panel discussions. Not keynote fluff. Real conversations. The kind where people say what they actually think because there's no audience.

Here's what keeps coming up.

1. AI made individuals 10x more productive. It made organizations exactly 0x better.

This is the tension we hear most often and nobody has solved it.

Employees are writing faster, researching faster, coding faster. But the org chart? The approval chains? The way decisions get made? All of that is frozen in 2015.

One COO told us: "I have 2026 workers trapped inside a 2015 bureaucracy. The friction is killing us."

The next big unlock in enterprise AI isn't a better model. It's redesigning how institutions actually work when machines are participants, not just tools.

This is an organizational design problem disguised as a technology problem. The companies that figure it out first will have an advantage that compounds for a decade.

2. The model got commoditized. Verticalization became the only real play.

A year ago, you could still argue for building horizontally — broad platform, lots of use cases, grow into it.

That window closed.

Every durable AI business we've watched being built did the opposite: they picked one workflow, went uncomfortably deep, and made their product irreplaceable for one specific person with one specific problem.

When intelligence itself is cheap and everywhere, the moat isn't capability. It's how precisely you've embedded yourself into something people do every single day.

The race to the bottom on model quality is over. The race to the top on workflow depth is just starting.

3. The "agent" conversation went from speculative to operational — fast.

Two years ago people debated whether agents were real. That debate is over.

The questions now are all operational:

How do you build an agent that runs reliably for hours without human intervention? How does it recover when something breaks? How does it actually improve over time instead of doing the same thing on repeat?

The mental model that clicked for the most teams: a good agent isn't a chatbot with more features. It's a new hire that owns a recurring process and gets better with every cycle.

Delegation, not generation. That's the real opportunity.

4. The best AI products got quieter — and that's how they won.

Early AI products had an instinct to fill space. Long responses, detailed explanations, constant output. The model was showing off. Product teams didn't know when to intervene.

The products gaining real traction now are radically more restrained.

One founder told us: "We removed 60% of our AI's output. Retention doubled."

There's a fundamental difference between a tool you use and a product that's integrated into how you work. The latter knows when you need it — before you ask. That's a design problem, not a model problem, and the teams who've cracked it are building something with completely different retention curves.

5. Memory is quietly becoming the most powerful moat in AI.

The default assumption in software has always been: context resets. You close the app, it forgets you.

That assumption is breaking down.

Agents that persist across sessions, accumulate preferences, remember past decisions, and build a model of how you work — that's a fundamentally different product from a smart chat interface.

Here's the compounding effect: once your software remembers you, the switching cost grows every single day. Day 1, it's easy to leave. Day 90, it knows your workflow better than your coworkers do.

The companies thinking about memory architecture right now aren't just building better products. They're building something closer to an ongoing relationship.

6. Going global is easy now. Making money globally is the actual hard part.

Generating international interest for a well-built AI product is almost trivially easy now. Distribution is nearly solved.

But converting that interest into reliable, durable revenue across different markets? That's where every team hits a wall.

The founders scaling internationally aren't thinking about positioning anymore. They're debugging payment infrastructure in markets where their revenue model doesn't translate, creator monetization structures that vary wildly by region, and compliance requirements nobody warned them about.

The moat is in the operational plumbing, not the product itself. The best product in the world means nothing if you can't collect payment for it in Southeast Asia.

7. China's AI story shifted, and the new question is more interesting.

Three years ago, the conversation was all benchmarks. How close are Chinese models to GPT? What's the gap? How fast is it closing?

That framing is mostly gone.

The question now isn't "can China catch up?" — it's "which layers of the stack will China own?"

The model layer is converging globally. But the application layer, the distribution layer, the vertical integrations, those are diverging fast along regional lines. Chinese AI companies aren't trying to build a better GPT anymore. They're building the ecosystems around intelligence that are specific to how business works in their markets. @natolambert 's visit to china also mentioned this.

This is a more interesting story than "catch-up." It's a story about strategic divergence — and it has implications for every AI company thinking about where to compete globally.

The thread running through all of this.

AI stopped being a technology story. It became an organizational design story, a product design story, and a go-to-market story - all at once. The companies winning aren't the ones with the smartest models. They're the ones who figured out how to turn intelligence into something people rely on by default.

That's a much harder problem. And a much bigger opportunity.

Pipes, billing, and commerce graphs: How China is playing a different AI game

Fri, 31 Jul 2026 17:29:35 GMT

This article was published on 21st, May, at @GeekParkHQ.

This was not a normal week in AI, even by the standards of a field that stopped having normal weeks sometime in 2023.

Three things happened that, read separately, look like routine product news. Google announced updates at I/O. Anthropic made an acquisition. Alibaba reported quarterly numbers with the usual superlatives. Read together, they suggest something about where the competition is actually headed — it's not where most of the coverage landed. And China is playing the game differently.

The Google story got framed as a model story. It isn't, really.

Yes, Gemini 3.5 Flash outperforms the previous flagship on coding and multimodal benchmarks at a fraction of the cost. That's significant for developers doing the math on inference spend. But the more revealing number from I/O wasn't on any benchmark slide. It was this: planning-related queries on AI Mode are growing 80% faster than AI Mode queries overall, over the past six months.

That's a behavioral signal. Users aren't just asking questions. A meaningful and growing subset is using AI to organize decisions, set up tasks, think through logistics. Whether that trend continues is an open question. But if it does, it changes what the product actually is — from an answer engine to something closer to a working layer.

At the same time, Google shipped Managed Agents in the Gemini API: isolated Linux sandboxes, versionable behavior files, a runtime that developers can hand off to Google's infrastructure instead of managing themselves. It's a quiet move. Combined with the consumer-side numbers, it puts Google in an unusual position — trying to own both where users start and where developers build. That's a harder thing to replicate than a better model.

Anthropic's acquisition of Stainless didn't get much attention, which is probably a function of how unglamorous the product is.

Stainless generates SDKs, CLIs, and MCP servers. It has, by Anthropic's own account, powered every official Anthropic SDK since the early days of the API. The acquisition brings that capability in-house.

The reason this matters is less about any specific feature and more about what it signals. As agents become more central to how people use AI, the connectors between agents and external tools — the generated SDKs, the model-friendly interfaces, the way developers actually operationalize API access — become part of the product in a way they weren't before. Anthropic is making a bet that owning that layer is worth doing. It's a reasonable bet.

The China angle this week is harder to read cleanly, which is partly why it tends to get flattened in English-language coverage.

@AlibabaGroup's quarterly numbers were strong. Cloud Intelligence Group external revenue up 40% year over year. AI-related product revenue growing at triple digits for the eleventh consecutive quarter, now 30% of Cloud external revenue. The customer base for Model Studio grew eightfold year over year. These are commercialization numbers from a company with existing scale in cloud and enterprise — a different starting position than a standalone AI lab.More structurally interesting: Alibaba has connected @Alibaba_Qwen directly to Taobao and Tmall's product catalog, which spans more than 4 billion items. Agent capabilities now cover order management, logistics, and after-sales service. The practical effect is that the starting point of a purchase can now be a conversation that has access to the full transaction graph — something Western AI products are still mostly trying to build toward through partnerships.

Then there's a development that deserves more attention than it's received.

China Telecom has begun trial token plans starting at 9.9 yuan for 10 million tokens. China Mobile has launched a universal token service in Shanghai, where users access multiple AI platforms under a single account and pay through their phone bill. It's easy to read this as a pricing story. It might be more than that. When AI usage gets billed through carrier infrastructure rather than through cloud consoles or direct subscriptions, it changes who controls the default relationship between users and AI services. Carrier billing has historically been a powerful distribution channel in markets where it took hold. Whether that dynamic applies here is genuinely unclear. But the structural logic is worth watching.

@Baidu_Inc introduced a new reported metric this month: DAA, or daily active agents. The number itself should be treated cautiously — the methodology hasn't been disclosed and the figures are difficult to verify independently. But the framing is worth noting.

Token counts measure inference consumed. DAA, as Baidu defines it, measures tasks completed. If that framing spreads, it would shift how the industry talks about AI productivity — away from input volume and toward delivered output. That's a different scorecard, and on it, the current rankings look less settled.

The through-line connecting these stories is less about capability than about position.

The model race produced clear winners and is still ongoing. But a parallel competition has opened up around something harder to close the gap on: who sits between users and completed tasks, and how deeply that position is embedded in existing infrastructure, billing relationships, and developer defaults.

In the US, that competition runs through search, cloud platforms, office suites, and developer ecosystems. In China, it runs through commerce platforms, carrier billing, enterprise workflow software, and tightly integrated service graphs. The companies best positioned in each market are not necessarily the ones with the most advanced models.

Whether that changes the outcome of the broader competition is an open question. It does suggest that reading AI market results through model benchmarks alone is going to miss more and more of the actual story.

Wearing the Future: Hypershell and the New Shape of Chinese Hardware

Fri, 31 Jul 2026 16:09:08 GMT

Walk into some popular mountain parks in China today and you'll see something that didn't exist three years ago: people renting exoskeletons at the base of the trail, the way they used to rent walking sticks. The device weighs about two kilograms, has one button, costs about the same as a decent pair of hiking boots to buy outright. Most of the people renting it have no idea who made it. @HypershellTech made it

And Sun Kuan, its founder, has a specific way of explaining why it exists: "Tools determine how large a person's imagination can be." He said this at @GeekParkHQ's Innovation Festival 2026 (GeekPark IF2026) last December, laying out his founding logic publicly for the first time. The product at the mountain park kiosk is the physical answer to that sentence.

The $120M across its Series B and the new X Series launch in May 2026 are milestones in a longer process. The more interesting question is what got the category here - and what this company reveals about a generation of Chinese hardware founders building in ways that look nothing like what came before.

What the product actually does now

The new X Series settles the question of whether consumer exoskeletons are a real product category. They are.

The flagship X Ultra S: 1000W peak motor output, 22 N·m torque, 30km range per charge, 97.5% gait synchronization across varied terrain, 0.31-second response time - 64.5% faster than the previous generation, TÜV Rheinland verified. Weight 2.5kg. One button.

That last detail is the output of HyperIntuition: Hypershell's self-developed end-to-end AI motion control system, and it signals a genuine architectural shift. Previous consumer exoskeletons ran on rule-based gait models: fixed patterns that couldn't adapt to uneven ground, stairs, or the variation between how different people actually move.

HyperIntuition learns. Instead of working in a separate sequence of steps, HyperIntuition works as one continuous system, processing raw data through the neural network, directly outputting the motor torque required for each joint.

It understands your movement as it happens and provides support instantly. Sun calls it the "external cerebellum" - the part of the system that handles automatic coordination so the conscious mind doesn't have to. "Just like the iPhone used the capacitive screen to achieve single-button interaction," he said at IF 2026, "AI (AI agent) lets the exoskeleton achieve the most simplified interaction threshold - users don't need to do anything other than turning it on and off."

Sun Kuan at the stage of IF2026

The 0.31-second response time and 97.5% synchronization aren't spec-sheet decoration. In a wearable system that assists physical movement, latency and synchronization have consequences in a way they don't in software. A chatbot that gives a bad answer is annoying. An exoskeleton that mistimes assistance on a staircase is a different kind of problem. Third-party verification on these numbers, and a 64.5% improvement over the prior generation, suggests real engineering progress rather than incremental polish.

Where it gets more complicated is the data layer. Every unit in the field generates gait data across terrain types, load conditions, and user profiles. The strategic logic is sound: that data trains better models, better models improve the product, which attracts more users, which generates more data. The problem is that "sound strategic logic" and "actual moat" are not the same thing.

Consumer wearable companies have tried this argument before - Peloton, Whoop, Oura - and found that sensor data was harder to defend than it appeared, because algorithmic improvements were replicable and the data wasn't as proprietary as it looked. The honest question isn't whether Hypershell should be building the data layer. It's whether data is the bottleneck, or whether architecture and model design matter more - and whether 20,000 units generates enough labeled, structured training data to make a meaningful difference.

(20,000 units at time of initial analysis; according to internal data, the number has since grown substantially.)

The SAR program is the more credible piece. 50+ search-and-rescue organizations testing devices in actual emergency deployments - sustained high-load climbing, irregular terrain, fatigue states - produces training data that no consumer dataset replicates and no lab can manufacture. Whether extreme-condition data generalizes to the average hiker is a real question, but the instinct to design field programs with data architecture in mind is sophisticated. Most hardware founders aren't thinking this way yet.

How the category got unlocked

The exoskeleton has been a near-future promise since the 1960s. Understanding why it failed for so long matters, because it explains exactly what Hypershell had to get right.

The hardware problem was fundamental: every additional kilogram of device weight increases human metabolic load by roughly 2%. A 10kg system costs you 20–30% before you've done anything useful. Old exoskeletons ran on fixed gait patterns that couldn't adapt to stairs or individual variation. "Locally effort-saving, but overall exhausting," as Sun put it at GeekPark IF 2026. A device that adds net burden isn't a mobility product. It's a constraint with a marketing budget.

Two things changed at roughly the same time. In 2019, MIT open-sourced quasi-direct-drive (QDD) motor research - power and torque densities exceeding human muscle by 10x at costs that don't require a defense contract. The second shift was AI: learned motion models replacing rigid gait programs. Sun's specific contribution was seeing something others missed. "I found an architecture that uses a single power system to simultaneously assist both legs," he said at GeekPark IF 2026. "At the time this was considered somewhat counterintuitive. But validated through simulation, it became our first-generation product - the world's first single-motor exoskeleton." That insight is what made consumer pricing possible. The M-One Ultra is its direct descendant.

He started the company in 2021, before most of the industry saw these doors opening. That's why nearly a hundred investors turned him down. It's also why the ones who eventually backed him made a good bet.

The decision that revealed the founder

In 2023, Hypershell raised $1.23M on Kickstarter - the first consumer-grade exoskeleton crowdfund, 40% of backers from the US. Then the grey-test scores came back. Nothing above 60 out of 100. Uncomfortable fit. Awkward assistance. Incompatible with backpacks.

Sun killed the product. Scrapped it, rebuilt with a new dual-motor architecture that doubled unit cost, took 18 months instead of 6, meant every second-generation unit shipped at a loss.

"If the experience isn't good enough, from the user's perspective, there is no experience at all."

This matters more than most founding decisions get credit for. The easier path was obvious and available - ship it, collect the cash, iterate later. He didn't take it. Hardware founders who optimize for short-term shipping numbers build a different kind of company than founders who kill their own product when the experience isn't right. The decisions that come later - when to cut a feature, when to delay a launch, when to spend on the right component instead of the cheaper one - tend to follow the same pattern as the first hard decision.

The second generation worked. Tens of thousands units shipped. Top-selling on Amazon in the US and Europe, and JD.com in China. Someone ran a marathon in it and swapped the battery mid-stride. Firefighters in Hong Kong climbed thirty-plus floors post-blaze while preserving enough physical reserve to perform rescue work - the first documented deployment of powered exoskeletons at scale in emergency response.

The competitive landscape

The most important thing to understand about Hypershell's competitive position is that it's not actually competing with the companies most people would name - and the reason goes back to a deliberate choice Sun made at the start.

The obvious first market for a mobility-assistance device is elderly care. Sun didn't go there. Before Hypershell, "exoskeleton" meant "medical aid" - a product you wear because your body is failing you.

A founding brand impression of medical equipment creates a permanently limited ceiling: you can grow the medical market, but you can never become something people want rather than something they need. So Hypershell went to the mountains first - hikers, backcountry skiers, photographers hauling heavy gear into terrain nobody else reaches - and built the category's first mental model around capability rather than accommodation. The downstream effect: elderly users who eventually bought the product posted about family hikes and returning to mountains they hadn't climbed in a decade, not about their mobility limitations. Same device. Different story.

That choice also defined who Hypershell is actually competing against.

Wandercraft is the most-cited comparison, and it's mostly a false one. The French company raised $75M in June 2025, has its Eve personal exoskeleton targeting the US market in 2026 - at $92,000, designed for individuals with severe mobility impairment. Its Atalante X rehabilitation system sells at $228,000 to hospitals. Wandercraft has sold 110 units total. This is a medtech company building toward FDA clearance and insurance reimbursement. It is not competing for the hiker who wants to see the view from a mountain peak. Ekso Bionics and SARCOS are in the same bucket.

The real competitive threat is closer to home: well-capitalized Chinese hardware imitators with the same Shenzhen manufacturing access and global distribution playbook. RoboCT raised $13.7M in March 2026. NAVEE showed up at CES 2026 with its Exo-Fit. Neither is dangerous today. But the pattern of Chinese hardware categories is consistent and worth taking seriously - drones, 3D printers, robot vacuums, e-bikes: once a category is validated, competition arrives fast and price compression follows. Hypershell's answer is HyperIntuition and the data corpus. Whether those hold when a well-funded imitator arrives with similar hardware in 18 months is the real durability question, and it's one the company needs to be thinking about now, not later.

The generation this company belongs to

DJI established the template that Hypershell is executing: neglected category, engineering depth over price, global-first validation, Shenzhen iteration speed. DJI didn't beat the drone market on price. It beat it because its flight stabilization was technically superior to everything else. The first generation of Chinese hardware exports won on cost. This generation wins on capability - and that's a fundamentally different, and more defensible, competitive position.

What's genuinely new about Hypershell's version of the playbook is the AI layer sitting on top of a hardware business. DJI's moat was manufacturing precision and proprietary flight software. Hypershell's potential moat is a motion data corpus that compounds with every unit shipped. The Kickstarter with 40% US backers before a dollar of retail spend is the global-validation move. The M-One Ultra from concept to production in under three years is the Shenzhen-speed move. The SAR program putting units into extreme conditions before retail scale is the brand play that doubles as real-world validation - stress-tested terrain and load profiles that controlled lab testing can't replicate. None of these are accidents, and the fact that they're all happening simultaneously at this stage of the company's development is a signal worth noting.

Sun's own read on the moment goes beyond Hypershell specifically. "We seem to be truly living in a science fiction era," he said at @GeekParkHQ IF 2026, "where things we could only imagine are accelerating into reality. Categories that people once thought impossible, or that existed but didn't have good enough experiences, can now be rethought and redesigned." What he's describing is a structural window - opened by AI and advanced manufacturing — where the gap between what's imaginable and what's buildable has collapsed faster than incumbents can respond. Hypershell got through that window early. The question is whether it can widen the gap before the window fills up.

The unit economics, roughly

The numbers that are public: tens of thousands units shipped, ASP probably around $1,000–1,200 blended across the SKU mix, gross margins 35–50%. That puts cumulative gross profit somewhere in the $7–12M range - real money, but not a business that justifies a $400M valuation on its own. Anyone who looks at those numbers and concludes this is currently a $400M company is confusing a bet on trajectory with a description of present reality.

Base case: 200,000 units at $1,500 ASP, 45% margins within 4–5 years → ~$135M annual gross profit. At 10–12x, that's a $1.3–1.6B business. Not a straight line from $400M entry, but within range if growth compounds and the software layer adds incremental margin.

The scenario that makes $400M look conservative: Ant and Meituan convert CVC positions into real distribution - exoskeleton rental in Alipay ticketing at scale, Meituan delivery worker augmentation contracts. That generates recurring revenue with entirely different margin characteristics. The standard CVC pattern is optionality without commitment, though. Whether this is one or the other is not answerable from the outside - and it's the most important question on the diligence list.

What comes next

Consumer exoskeletons are a rounding error in the overall market today - and the only segment growing at 3x the rate of the others. Hypershell has the entry point and the right brand architecture to expand from it. The outdoor positioning was always a first chapter. The Hong Kong rescue deployment, the HyperLIFT SAR program, the elderly users quietly reappearing in the user base - these are the category expanding outward from aspiration, the only direction a consumer hardware brand can expand from and maintain its pricing power.

The unresolved questions are serious. Can the data and software layer get built before the hardware gets commoditized by imitators? That window is probably two to three years, not five. Can a company whose DNA is Shenzhen manufacturing execution build the organizational capability for platform software - a skill set that has killed more than a few hardware companies that thought they could make the transition? Will the Ant and Meituan relationships have actual teeth?

Here's the honest read: the downside scenario. Hypershell remains a well-run hardware company that dominates consumer exoskeletons without ever building the platform layer - is still a good business. Probably a $1–2B business at scale, which at $400M entry is not an most neutral outcome. The upside scenario, where the data flywheel turns real and the CVC relationships convert, is something larger and harder to price.

What's not in question is whether there's a real company here. There's a real product, in a real category, built by someone who's already shown - on the record, at some cost to himself - that he'll make the right call when it's expensive. In hardware, that's rarer than it sounds. And it's the thing that all the financial modeling in the world can't manufacture.

"Different Organizations Allow Different Things to Emerge": Yang Zhilin's First In-Depth Interview

Thu, 30 Jul 2026 15:44:10 GMT

Editor's note: This conversation, recorded in November 2023 shortly after the launch of Kimi Chat, was the first time Moonshot.AI founder and CEO Yang Zhilin shared his thinking publicly at length. With Kimi K3 drawing global attention to Moonshot, we are publishing English editions of his two long dialogues with @GeekParkHQ founder Jack Zhang, starting with this one.

At the time, Moonshot had one product, a chatbot built on a 100B-plus parameter model, and three clear labels: long context, proprietary and closed-source, consumer-first. Read it as a time capsule. The transcript has been edited and condensed for clarity.

Highlights

The necessary path to AGI is a new kind of organization, not any single technique. OpenAI's success is, at its core, organizational innovation. "Only if you get the organization right can you actually walk the AGI road."
Large model innovation cannot be planned in advance. Mobile-era requirements were deterministic; AGI innovation is post-hoc. You have to try before you know.
Different organizations allow different things to emerge. Google's environment could produce a scientific result, the Transformer, but not ChatGPT. OpenAI invented nothing new, yet an industrial masterpiece emerged from combining three factors: the Transformer, 10^25 FLOPs of compute, and twenty years of internet data.
The Transformer is a new computer. Parameter count is its CPU, context length is its memory. Forty years ago people thought 500K of memory was plenty. The same fallacy is being repeated about context.
The ultimate case for long context is a lifelong AI companion. Trust and complex emotion only show their power over decades, and "an AI that has to reset its context window every day cannot do that."
The super app entry point will most likely be closed-source, because owning the model's evolution creates product differentiation from day one. This is the bet Moonshot has since reversed with the open-weight K series.
An AI-native product is defined by two datasets. Training data determines what the model can do, test data determines whether it is actually usable. Define the datasets and the product is defined.
The classical product manager pointed at one spot on the map and planted a tree. In the AGI era you mark out the whole plot and let the model sweep the field. Not design first, then build; you complete the design through making.
The metric that matters is a Moore's Law of use cases: the number of usable scenarios must double every N months, exponentially, not one scenario and one dataset at a time.
The greatest companies of the next decade will fuse two cultures: Silicon Valley's technical idealism as the drive, and the Chinese emphasis on usefulness and business models as the fuel.

Why "the dark side of the moon"

Jack Zhang: People say your company is a bit mysterious, starting with the name. What is the story behind "Moonshot AI," or in Chinese, "the dark side of the moon"?

Yang Zhilin: It comes from the Pink Floyd album, The Dark Side of the Moon. The founders all love rock music. We used to play in bands. And 2023 happened to be the album's 50th anniversary.When you look at the moon, you only ever see the lit side. You never see the back, and that gives you a strong urge to explore it. Large models feel similar. You want to explore something mysterious and unknown. It is hard, and it connects to the spirit of rock music: keep innovating, keep challenging the existing shape of things, keep imagining what comes next.We settled on the pairing of "dark side of the moon" in Chinese and "Moonshot" in English. It reflects our commitment to AGI. It might also define what kind of people we are.

Jack Zhang: I'm a Pink Floyd fan too. The album has an astronomical name, but it is really about the human subconscious.

Yang Zhilin: Right. The lead track is Brain Damage, about a person experiencing hallucination. Fifty years later, here we are trying to fix hallucination in large models.

Jack Zhang: What did you play in the band?

Yang Zhilin: Drums. The drummer keeps time and gives the whole band a frame to play inside.

Building a new kind of organization is the necessary path to AGI

Jack Zhang: How did you decide to go all in, and specifically to build a company around it?

Yang Zhilin: My understanding shifted enormously over the past few years. At first I thought language models were a tool that could improve results across scenarios. In the second stage, I thought they might be useful for many tasks. Eventually the view became: language modeling may be the only problem AI needs to solve. Everything can be addressed by making the model better, by making next-token prediction better.

In 2018 and 2019 at Google, we started training language models on thousands of chips with Transformers. You would observe so many phenomena along the way, and they kept adding evidence that this path was correct. Just keep walking it, keep finding more efficient ways to scale, and you get remarkable results on problems that used to be very hard: memory, reasoning, common sense, even complex multi-step problems. That experience left a deep mark on me. It was the runway toward founding a company.From 2020 I worked with many institutions to train large models. I was involved in some of the earliest large-scale efforts in China, including Pangu and Wudao. Through that process, I saw the challenges up close. Some were technical. Others were organizational. We found that if you use a traditional organizational structure, training frontier models is very hard to pull off. OpenAI's success is, at its core, also a success of radical organizational innovation.

So you could say I had been searching for one opportunity the whole time: the chance to build a new organization from zero. I believe this is the necessary path to AGI. It matters even more than the technical details we touch every day, because the organization is the deeper layer. Only if you get the organization right can you actually walk the AGI road.

By 2023, both the capital market and the talent market had changed dramatically. The timing was finally right.

Innovation in the large model era cannot be planned in advance

Jack Zhang: What convinced you that the organization is the core problem?

Yang Zhilin: Practice, mostly. Before this year I tried many modes: working inside a large company, working with independent research institutes, other supposedly efficient setups. None of them could produce fundamental organizational innovation.

Here is a simple example. You cannot innovate on large models through planning. In the mobile internet era, I could plan the features I wanted to build. Once a requirement was defined, it could be deterministically produced. You rarely heard of an app that a team suddenly did not know how to build. It was a deterministic event: humans encode the logic, computers execute it.AGI is different. I cannot plan that today we will fulfill a certain requirement to a certain degree, because it cannot be hard-coded or expressed as rules. AGI-style innovation is not front-loaded planning. It is post-hoc. You have to try before you know. You need an underlying machine that does many things in a systematic way.That is a fundamental difference.

Your organization has to match how you do things, so when the underlying logic changes, you need a new organizational form. The internet era produced excellent organizations, superb at things like recommendation-driven products. The new era will likely produce organizations that are superb at AGI. I think that is very likely to happen.

Jack Zhang: Is OpenAI a good template in your eyes? What did they get right, and what might not be optimal?

Yang Zhilin: On results alone, OpenAI made an enormous breakthrough. Without that company, the trajectory of humanity might be different.Going deeper: a good organization needs very high talent density, a shared vision, and the ability to focus efficiently on one goal. They did all of that extremely well.But the most essential point, and the one hardest to see from outside, is this: once you have those preconditions, how do you find a systematic way of doing things? That is the precondition for all the technology, and it is what we most want to iterate on.

Jack Zhang: By systematic, do you mean it can be replicated and scaled?

Yang Zhilin: Yes, but replication in the sense of reusing it across different problems, not copying it to another company. It forms inside one company and is very hard to transplant. But that company can use the same system again and again. Today I can use it to crack long context. Tomorrow, autonomous AI capabilities. The day after, multimodality. It is a reusable system that accumulates into your core asset. Every AGI company should spend serious time polishing this.

Google's organization let Transformer emerge. OpenAI's let ChatGPT emerge.

Jack Zhang: When people discuss OpenAI and organizational design, the debate is often bottom-up versus top-down. How do you frame it?

Yang Zhilin: The top-down frame still applies, especially for large models. Top-down is about leadership vision: can you judge what is right and worth doing, and what you should not do right now? AGI is like a lunar program, a giant system that takes a long time and many tightly coupled people. That kind of top-down design is necessary.Under that frame, what matters is whether the system lets many small units each produce efficiently, and then the top-down frame integrates the output.

Jack Zhang: So in a sense, you are running a big frame within which different directions can produce emergent innovation, because nobody can precisely define the one path to AGI today. The organization has to support emergence.

Yang Zhilin: Exactly. Look at how the Transformer came to be. Google gave those researchers an environment for emergence. Before the Transformer, the pieces already existed: attention, residual networks, LayerNorm, SGD and the training toolchain, learning rate schedules. Everything was prepared. Google provided the environment where people could freely combine them, and the emergence happened.But different environments allow different things to emerge. Google's environment could produce a scientific result. It could not produce a great work of systems engineering like ChatGPT, an epoch-defining product that pushed execution to the extreme while precisely capturing demand. Google's organization and ChatGPT simply do not match.OpenAI invented nothing new.

What emerged from it was an industrial masterpiece. It combined three things on a different dimension from Google: the Transformer architecture, compute centers capable of 10^25 floating point operations, and twenty years of data accumulated by the entire internet, which may be the internet's greatest value. OpenAI saw those three factors and provided the environment for them to combine. Out came a milestone on the way to AGI.That is my point. Different organizations allow different things to emerge. Whatever you want to emerge, that is the direction you should tune the organization toward.

The technical path to AGI is set. The product path is not.

Jack Zhang: A popular line in 2023 goes: the cards are on the table, now bet big. Meaning the technical path to AGI is settled and the game is now about who pours in more resources. Do you agree?

Yang Zhilin: The first principle of AGI is now clear: keep improving lossless compression and you get higher degrees of intelligence, eventually beyond human. There is abundant evidence. Like Newton's laws for classical mechanics, the big direction is basically settled. So that popular line has some validity.What remains is the second layer: under the big principle, how exactly do you do specific things?

For example, how do you build genuinely lossless compression over a long context? That is not simple. Even OpenAI has only taken the first step. Every subsequent step still carries uncertainty.Then, beyond technology, there is the product layer. We are still far from the superintelligent AI of science fiction, and today's products are not necessarily heading in the right direction.Every era has its greatest people and its next-greatest people. The greatest discover the correct first principles. Then a cohort of people, slightly less great but still great, solve the technical, product, and commercial challenges. That second layer is a vast open sea. How to play in it is still full of unknowns.

Transformer is the new computer. Context length is its memory.

Jack Zhang: Long context is Moonshot's specialty. With so many possible directions, why this one?

Yang Zhilin: Everyone should ask themselves what they want AI to do for them, what the human-AI relationship should be. In the ultimate form, one big capability is missing today: a much longer input window.The gap between a long window and a short one is more fundamental than I once thought. One ultimate form of AI is building long-term emotional value with a person, a lifelong companion over nearly unlimited time. Time is a critical dimension. Only over long stretches do trust, complex emotion, and decades-spanning interaction show their power. That is when AI can offer deep value to the human spirit. An AI that has to reset its context window every day cannot do that.

Jack Zhang: Tomorrow it forgets what you did today.

Yang Zhilin: Right. So think of the Transformer as a new computer with two critical dimensions. Parameter count determines computational complexity, like the CPU in the old computer. Context length is the new computer's memory. It determines how much can participate in the computation.Given sufficiently complex computation, the bigger the memory, the bigger the unlocked application space. Look at computing history. Forty or fifty years ago, everyone thought 500K of memory was plenty. Today that is obviously absurd. The same thing will happen with this new computer system. Long context, as the memory of the new computer, is absolutely essential.

Closed source, in service of a super app

Jack Zhang: This wave of model startups includes plenty of open-source players. Moonshot is closed-source with no plans to open up. What is the thinking?

Yang Zhilin: We strongly support open source. Open and closed will be complementary in this field. Open source lets developers try all kinds of applications, with stronger compliance control over data, training, and deployment, and more flexible scenarios.Closed source has its own value. The future's super app entry points, whether in productivity or consumer entertainment, will likely be built around closed models. The two approaches complement rather than conflict. The choice depends on each company's strategy. Ours is to build a super app. That is where all our time goes.

Jack Zhang: If someone wants to build a super app on an open model, buy the engine and modify it, why insist on building the engine end to end?

Yang Zhilin: Building applications on open source is a real opportunity, and the two do not conflict. But if the endgame is a super entry point, it will most likely be closed, because closed source lets you differentiate the product from day one of building the model. When you control the model's long-term evolution, you have full room to create a decisive product advantage.Applications on open models may not become the super entry point, but they can create incremental value: productization, proprietary data, fine-tuned capabilities others do not have. Both paths will exist in the ecosystem.

Also, right now we are in a technology-driven phase. A better foundation model converts into product advantage, so your base capabilities need to stay ahead of the commodity level. In ten or twenty years, when the technology commoditizes, you will need to convert first-mover advantage into more durable moats, like stronger network effects.

Begin with the end in mind: consumer is the only mode that matches AGI

Jack Zhang: So no API business, no helping enterprises deploy models. You are going consumer. The last AI wave produced almost nothing on the consumer side. Everyone ended up doing B2B, which at least reliably generates revenue. Why so firmly consumer this time?

Yang Zhilin: We are not refusing B2B entirely, but the focus and the push are consumer.For a long time, AI technology had no consumer success stories. With the new technical variable, AI can achieve results that were previously impossible, and those results can appear as new applications and new entry points, with exponential revenue growth and fast-growing users.

Midjourney, Character AI, and ChatGPT have all largely proven that AI-native apps have a real shot.And if you are doing AGI, you must choose a business mode that matches it, one that demands extreme innovation efficiency. Only consumer lets you close the loop fast. Only there can the organization form a culture of rapid iteration: updating models, adjusting the organization, and meeting user needs on a daily cadence, everything revolving around data at high speed. Only the consumer side generates that kind of energy, the kind that matches a company whose goal is AGI.This is thinking from the end backward. Co-creating with consumer users is itself doing AGI. It may even be a necessary precondition. AGI cannot be built behind closed doors.

The core is data. Without co-creating with users, you cannot get enough high-quality data, you cannot know what problems the model produces in real use, and you cannot dig deeper into scenarios together with users.

Jack Zhang: So it comes back to first principles about the goal. Without a consumer super app, enough users, and enough data, you cannot actually reach AGI. A company that does not want to build a super app is arguably not a true AGI believer.

What super app means in the AGI era

Jack Zhang: How do you define a super app here? WeChat is one because users do everything on it. Taobao is one by scale and value.

Yang Zhilin: The definition itself is not new. What changes is the value delivered. Only by providing value that could not be provided before does a new entry point appear. In the end you still need many users, high frequency, and large value created in use.But AGI has one property that makes a super app possible: the G, generality. It does not solve one class of problems. The set of problems it can solve keeps growing, so the product boundary keeps expanding. AI penetrates deeper into every part of life, the application's value keeps strengthening, and it fits the definition of a super app more and more. Generality and super app status are compatible.

Jack Zhang: That is a key point. In the mobile era, an app first reached scale, then expanded its service boundary to become general. In the new paradigm, you start with a general productivity engine, and it naturally becomes a super app. Different eras, different genes. Last era was a land grab for scale. This era, the technology engine gives you super app genes from birth.

Yang Zhilin: Well put. One addition: however general the technology, you still start from a subset of scenarios and generalize outward. And the generalization speed can be exponential rather than linear.

AI-native development: define two datasets and you have defined the product

Jack Zhang: Under the old paradigm, product managers, frontend, and backend collaborated toward a defined target, shipped in cycles, ran A/B tests. What does development look like for an AGI-era super app?

Yang Zhilin: Product development changes with the underlying technology. Mobile-era development meant clear requirements mapping to deterministic operations and fully deterministic events, built on the old computer and deterministic code. Deterministic logic gave rise to deterministic graphical interaction. Deterministic GUI plus deterministic systems: that was twenty years of internet product development.

Today the paradigm has shifted. The frontend becomes conversational UI, and more products will adopt it. The backend has been unified, to an extreme degree, into one language model. And it handles more than language. It processes all the world's information. It is encoding and losslessly compressing everything.With both ends settled, most application-layer development no longer touches backend architecture or frontend framework. There may be blends of GUI and conversational UI, but the overall architecture is basically fixed.

Most of what we call development today happens in the middle layer: data. Same interaction, same model class, different data, different product. ChatGPT, GitHub Copilot, Midjourney: essentially the same thing, differing mainly in how the data is defined.This is a massive paradigm shift. What product managers increasingly need to think about is how to build a product out of two datasets. Define the datasets and the product is defined. The training data determines what capabilities the model provides. The test data determines how usable it actually is.Before, there were no AI-native products, only AI features, so this way of working was rare. Many people with strong product sense do not know how to apply it.

Say you want to build something like Character AI, or improve on it. How do you define your two datasets? You need strong data production and processing techniques, ways to acquire data, judgments about which data is effective. We are still making these workflows concrete through exploration. The new development method of AGI probably requires a new organizational form to pull off.

The product manager of the new era

Jack Zhang: Early mobile internet ran on product managers with imagination, people who defined future scenarios by instinct. I once joked with Allen Zhang, the creator of WeChat, and he called himself a classical product manager. I like to call that quality a touch of the divine: they could not fully explain their convictions, but the convictions turned out right. Later, product management became scientific, data-driven, A/B tested. For a super app, do you want more of the divine or more of the scientific?

Yang Zhilin: You look for the balance, but long term, I believe the system will win overwhelmingly and become the mainstream development paradigm. That does not make the divine instinct unimportant. It just needs a strong system underneath. The system should be the main force.

Here is my usual metaphor. Allen Zhang pointed at one spot on a giant map and said, plant a tree here. He turned out to be a god, because the spot was right and the tree grew into a forest. A sharpshooter who hits whatever he aims at.AGI does not work that way. The industrial designer Sori Yanagi liked to say you do not manufacture according to a design, you complete the design through making. Build the thing, and the design is done, rather than designing first and building after.The old way was like scouting forever for the one spot, planting one tree, and celebrating when it survived.

Now, you glance around, say this patch of land looks decent, and the sharpshooter marks out the whole plot. AGI is your main force. It rolls across the entire field and finds every opportunity, every place a tree could grow.With a strong system, you realize that having talent-and-luck-driven judgment pick individual planting spots, out of tens of millions of possible scenarios, is extremely inefficient. The search for product-market fit in the AGI era should be done the AGI way: use its generality, use the user ecosystem, use the system, and push across the whole front at once. That is the biggest difference from the classical product manager.

Jack Zhang: So instinct concentrates at the front, on problem definition and goal selection. Getting there faster and better is the system's job.

Watch the delta, not the balance

Jack Zhang: More concretely, the product people you have hired, what do they share? Reverse-engineer the traits for us.

Yang Zhilin: An open mind and the ability to learn. Together they point to one thing: can this person iterate fast? That is the trait we value most. Not just product managers. Every role, every person, AI or not. Change is too fast now. You basically cannot predict what AI will offer by the end of next year, let alone three, five, ten years out. So everyone needs to learn fast with an open mind, then go execute concretely.For product specifically, you need strong consumer sense. Shipping version one is the easy part, because you have not yet gone through making it continuously better, or the process of defining ever more precisely what product you want. "Better" is deeply abstract. How do you make ChatGPT better? What counts as better? In which direction, and by how much? All hard to define.

A common trap for product managers is defining a pile of features, the old way. That is probably wrong now, because your features are defined through data. That is the AI-native way. And this is not static theory. You learn, then you try. What I said today might be wrong. Fine. Try it, gain something, take another gradient step, deepen your understanding. Out of that process, this era's Allen Zhang will appear.

Jack Zhang: Or not a new Allen Zhang. He was the classic of his era. The next era will produce a completely new figure, whose touch of the divine will look different. That is the most exciting part of progress. And one thing is certain about talent now: watch the delta, not the balance. Maybe ignore someone's history, but look at the increment between their past and their present, their present and their tomorrow. In an uncertain era, a large enough delta means a lot.

Yang Zhilin: Yes. Often, carrying no historical baggage is an advantage.

Jack Zhang: People in the comments are asking whether you are hiring.

Yang Zhilin: We are. Full-time and interns. We are quite open about backgrounds. AGI is a comprehensive undertaking. The technology alone is full-stack: NLP, computer vision, RL, alignment, infrastructure, kernel engineering. Beyond technology, product, operations, commercialization. Ideally, very diverse backgrounds, but one shared vision. We welcome anyone with passion for AGI, for the super app, and for the global market.

The metric that matters: a Moore's Law of use cases

Jack Zhang: Sam Altman has written about Moore's Law for intelligence, and Moore's Law for everything: a Moore's Law relationship between the cost and capability of intelligence. Do you buy it?

Yang Zhilin: Yes, and I think Moore's Law itself is going through a paradigm shift. The original version: transistor count doubles every N months. Then model parameters and compute followed it: FLOPs double every N months.For us, the Moore's Law of intelligence ultimately means this: every N months, the number of usable use cases doubles. It is an extension of scaling laws. Standard scaling laws describe pre-training: add compute, data, and parameters, and watch how training loss changes.But the metric that matters most in the end is the Moore's Law of use cases. How many scenarios reach usability? It has to rise exponentially, not linearly, doubling every N months.

You cannot use the traditional AI method of adding one scenario and one dataset at a time to make it work in that scenario. You will never get exponential growth that way.Measure how many scenarios get unlocked. With that, the search for product-market fit accelerates enormously, and you can try many things at once. Do not plant one tree. Mark out the land, and with a Moore's Law of use cases, you test the entire plot in one pass. A superfast tree-planting machine that knows whether a tree will live or die without planting it. When the scenarios multiply past a certain point, you become a super entry point.

Jack Zhang: The Moore's Law of intelligence and the Moore's Law of use cases should form a double helix. Costs keep falling, capabilities keep unlocking, and scenarios multiply.

The greatest companies of the next decade will fuse two cultures

Jack Zhang: You worked at Meta and Google. Compare Silicon Valley's engineering culture with China's. What is each good at?

Yang Zhilin: Silicon Valley's engineer culture is distinctive. Take Noam Shazeer at Character AI. They built a product with a real degree of product-market fit, and in the early days they barely had dedicated product managers. Engineers with their own ideas, fusing technology with those ideas, taking one step forward on their own. That is worth borrowing, especially in the AI-native paradigm where technology and demand have to move toward each other. Think of constructing those two datasets. Without engineers willing to take that step forward, much of it simply does not get done.

At the bottom layer, we want to absorb the best of both East and West. OpenAI runs on strong technical idealism. "I want to build AGI," business model unclear, heavily funded from the start. The recent effective accelerationism wave is another expression of it. Google and Microsoft were also born of a degree of technical idealism.Chinese culture emphasizes usefulness, thinking within the premise of a business model.

In the next ten years of the AGI era, the greatest companies will probably combine the two. The pragmatic side finds a genuinely good business model, the fuel that keeps you burning. The idealistic side drives you beyond money and usefulness, toward the simple desire to see what the far side of the moon actually looks like.

Jack Zhang: We have always been good at goal-driven usefulness. But making something useful and universal may now require some moonshot spirit: aiming at something high, hit or miss, and moving toward the deep of the universe. Exciting goals are what gather truly excellent people. Thank you, Zhilin.