Liang Wenfeng
Restraint on the Path to AGI

The other colleagues at our company, too—when we first started this company, we weren’t thinking about how much money we would ultimately make, or entering the capital markets, or going public, or anything like that. So that was never our original motivation.

Those first few dozen people never thought that way at all. If they had, they wouldn’t have joined. So broadly speaking, we did this with a great deal of goodwill toward the world, believing it was useful to humanity and about something other than money.

Of course, later, once we realized how profitable this could be, other temptations arose, but that’s a separate matter.

But our original motivation and vision when we started, and the vision we’ve held on to ever since, are not about doing this in a way that maximizes commercial returns. I think that’s a fairly important point.

About 20 years ago, the person I most idolized in management was Jack Welch, the former CEO of GE. Looking back now, most of what he said may have turned out to be wrong, but he was right about one thing, the most important thing: what matters most in a company is its vision.

You run a large company by relying not on rules and systems, but on a vision. What is a vision? It’s not a slogan hanging on a wall. It’s not what you say, but what you do—how you actually operate. Anyway, I’ve forgotten Jack Welch’s exact words, but that was roughly the meaning.

So how have we managed and organized so many people? In fact, we have no organizational structure; the vision leads us. We organize around a single vision. We have no organizational structure.

There are advantages and disadvantages to this. Going forward, we’ll look for ways to build on the advantages and avoid the disadvantages, but this is what characterizes us. We don’t operate in terms of what KPIs I have to meet, or being evaluated, or things like that. There is only the vision.

This vision isn’t even documented. We’ve never written it out, never put any of it down in writing. It lives in the way we work and our attitude toward the world. Perhaps everyone at the company understands this vision differently, and each person’s vision may differ, but we agree on the broad direction.

I think, again, it’s about trying to do a little something with a great deal of goodwill toward the world. This is what brought us together.

Now let me speak first, and then you can ask questions when I’m done. What follows will probably revolve around this vision. This vision is real; it isn’t something we made up. We really think this way, and we really act this way. Otherwise, you couldn’t explain a lot of what we do.

Why, with this vision, are we so insistent on open source? Because the vision itself demands open source. Without this vision, we couldn’t bring people together.

For example, Zhipu also does open source, but its approach to open source is different from ours. There’s a sense that Zhipu is doing it reluctantly. They feel it isn’t what they really want to do, whereas for us, it’s exactly what we want to do.

And we’ve been very clear about open source from the beginning. First comes the vision, and second, we believe open source offers advantages in making AI commercially successful.

That sounds a little contradictory and counterintuitive, because historically, open source and commercialization have clashed. But I think AI is different from what came before. Historically, a software company might have a market worth several billion dollars a year, but if it releases its software as open source, that market might disappear, leaving only tens or hundreds of millions of dollars.

But AI is big enough. For example, it could ultimately account for 10 percent of humanity’s GDP. That is, in fact, an enormous number. For one person to monopolize this—that’s impossible. You have to share it with others; otherwise, you certainly won’t survive.

This may be different from open-sourcing a piece of software in the past, because the market for that software wasn't that big. But AI is truly enormous. If we try to monopolize the gains, history will inevitably leave us behind. I think the most important point is that this is an objective law, a way of understanding history.

Not open-sourcing doesn't mean you can monopolize this market. Even in theory, that doesn't square with objective reality. You will inevitably encounter a great deal of resistance, and other approaches will inevitably prevent you from achieving that goal.

In this situation, I don't think we necessarily have to follow conventional business thinking. You only have a chance of succeeding if there is a mechanism that ensures limits on the gains you take for yourself. You need restraint. I think restraint is necessary.

If we want to make AI a success through our own efforts, I think the first thing we need is restraint. We mustn't think that some percentage of global GDP, or some percentage of China's GDP, will all belong to us. The more you think that way, the less likely you are to succeed.

So from the very beginning, we believed we needed restraint. The more restraint we exercise, the better our chances of making this work. That's a business consideration, and of course a big-picture consideration as well.

I think this makes intuitive sense. At least it does to me, and I genuinely believe it. It's not as though we have all that many other advantages. We don't have some extraordinary ability, we don't have more money than others, and we can't say our people are better than those at other companies. That simply isn't the case.

Think about it. When we founded this company 2 years ago, we didn't have much money or many GPU cards. We had no name recognition and no ability to draw people in. We were just a group of very ordinary people.

We really are just a group of ordinary people. If there's a story I like, it's not that a group of geniuses did something extraordinary, but that a group of ordinary people did something extraordinary. This has a great deal to do with our restraint, and it is part of the same thread that runs through our restraint and our vision.

So does open source conflict with commercialization? I think that in AI, you can't get off the ground without restraint.

Open source is part of restraint. Our restraint shows up not just in open source, but in many aspects of what we do. But overall, we don't need to deliberate over open source, or over restraint.

The more restraint you exercise, the easier it may be to succeed. At least that's been borne out so far, and it explains things so far. Otherwise, there's no way to explain why we were able to succeed. We have no special weapons, we started from very little, we have very few resources, and our people are really just a group of ordinary people who happened to come together. I myself am just a university graduate, and I didn't graduate from one of the very top schools.

This restraint is also part of our vision. AI is such an enormous undertaking, and the potential gains are enormous too. We exercise a great deal of restraint. As long as we make it a success, the eventual gains will be huge for everyone. Even a small share will be very substantial. So I don't think there's any need right now to consider which share of those gains we'll take, or how. The gains are big enough.

Even a small share is already enough. So, as I said before, we only seek a reasonable profit, and we focus on the intention, not the size of the profit. Those are two different things. I'm not talking about our API pricing here. For API pricing, our idea of a reasonable profit is roughly being able to buy a batch of equipment on the market and recover its cost in 10 months. I think that's a reasonable profit.

Under current conditions, considering the risks, the upfront investment, and so on, even if we depreciate the servers over 3 or 5 years for accounting purposes, we think recovering the cost in roughly 10 months is enough from a business standpoint. OK, we think that's enough. That's the logic behind our current API pricing. Our V3.2 Flash and the others all recover their equipment costs in 10 months.

That's our benchmark. It isn't actually about maximizing profit. If we wanted to maximize profit, we'd have to set prices higher.

That's because user demand is inelastic in this price range. Even if we raised the price by half or doubled it, token consumption wouldn't change much.

If we doubled the price again, total revenue would nearly double. Hold on, let me cross-check something. Oh, that's great.

Let me tell you a story. We have a model called DDCP. At first, we were worried demand might be too high, so we set the price relatively high, and nobody on the team was particularly happy about it. Later, I lowered the price to a quarter of what it had been, and everyone was delighted.

I think this is how we truly feel. Like the vision I described earlier, We want to make this useful to people, not make as much money as possible for ourselves. We want everyone to be able to use it while we earn a reasonable profit. That's how other people at our company felt this time, too. When we lowered the price, lots of people cheered in the company group chat. Everyone was delighted.

Because that's precisely why we put so much effort and care into making this model good. The goal is to make it very affordable and very effective, so everyone can use it as much as they need. That makes us very happy, and we see it as our motivation, our vision, and the shared conviction that brings our company together to do this work. It's a conviction we share internally.

That's probably somewhat unusual. A price cut certainly isn't good news for our competitors, and they definitely wouldn't cheer. They have revenue and ARR to think about. Cut the price in half, and ARR falls by half too. Yes, that's what makes us different.

We think this is enough. From inside the company, recovering our costs in 10 months is already very satisfying commercially. And outside the company, I think this is a price people are generally happy with and welcome. It's a win-win for everyone—for the company, for society, and for every individual.

I think—OK, someone just posted a comment on the screen saying that recovering the initial investment in 10 months means the profit is too high. There is definitely room to lower prices further. There certainly is. There's also room to optimize the model, so overall there's quite a bit of room for price reductions.

But while we can recover our costs in 10 months at this cost level, others can't. Companies like Alibaba or Tencent don't have our optimizations, so their costs would be several times higher. There's still a lot of optimization work involved here.

Someone just asked why we don't keep lowering prices. It's because demand is inelastic. If we lower them further, demand either won't increase or will increase only very slightly. At this price, everyone can afford it and everyone is satisfied. Nobody is being kept from using it because the price is too high.

So lowering the price would not, for one thing, increase the company's revenue, nor would it create more value for society. Everyone is satisfied with this price. Lowering it further would not add much to society's happiness. Yes, OK, but on the issue we were just discussing—how we set prices—we certainly do not base it on…

Maximizing the company's revenue or profit is not our starting point. This is part of our restraint. In the short term, raising prices a little might increase revenue a little, but in the long term, you really cannot know. I believe restraint is a strategy.

To me, restraint is a strategy. Sometimes you can give up some things to gain more of something else. Not open-sourcing is actually the same. You could see it as pressure on us, or as us forgoing some of our gains.

First, forgoing these gains is a very positive thing within the company. Everyone is happy, employees feel a sense of accomplishment, and it brings us together. It also benefits society, which is very happy about it; others in the industry and the general public will all be happy too. So my understanding is that this kind of restraint can improve our chances of succeeding with AGI in the long run.

When I consider a decision, I have absolutely no doubt that AGI will have enormous commercial value. With that as my premise, my first consideration is not how to gain a little more market share or take a little more for ourselves, but how to improve our chances of success.

This restraint can show up in many other ways too. For example, during the Spring Festival last year, we suddenly gained a huge number of users. But we did not pursue locking them in, monetizing them, or grabbing commercial gains by extracting value from them.

We did not poach users or make money from them, but we worked very hard to find ways to serve them well. We have no intention of building the next super app, competing with someone, or creating the next ByteDance or Tencent.

We could have done that, but we did not. As I see it, that is part of restraint too. Do not try to make money from everything. You get users, think you could build the next ByteDance, and then try to gobble that up too.

I think that could work commercially. Spending a lot of money last year to compete with ByteDance for users would have been one possible approach. But we chose a very restrained approach: not fighting over it. Because there is a watermelon waiting further down the road, and everything in front of us now is probably just sesame seeds.

I should not try to grab every sesame seed. Of course, some of these seeds may be relatively large, but compared with the AI opportunity ahead, I think everything in front of us now is small. Looking back, we may have been right not to focus on the consumer side last year. We can now see that there really is a bigger watermelon ahead, and what is in front of us really is just tiny sesame seeds.

If I had had a huge amount of money last year and pursued this on a massive scale, what advantage would that have brought? There would have been nothing to gain. That is what I genuinely think, because I believe the AGI opportunity ahead will be enormous, and that it will always be enormous.

We do not even need to think about whether we will have a place in it then, or what our business model within it will be. We do not need to think about that at all. When the commercial opportunity is that big, a way will surely emerge. We will pick up some of the sesame seeds in front of us too, but only a few along the way. We will not stop for them or make them a priority.

So I think things like consumer DAU last year may be relatively minor. We picked up those seeds anyway, keeping users engaged at a relatively low cost, because they might be useful later. Right now, we do not know what use these users will be, and they are purely an expense, but they may be useful in the future.

Since it is something we can pick up along the way, we do. Looking at this year, there is very likely an opportunity for ARR from APIs or AI as well. If demand keeps expanding, and that expansion lets us buy more GPUs, reaching hundreds of millions of dollars in ARR is entirely possible.

If we could reach $1 billion with AI, the company would essentially become cash-flow positive and be able to cover R&D and all other expenses. That is possible too, but we have not made it a priority.

We will do this, and I do consider it important. But it is not our top consideration or what we are really focused on today. The bigger opportunity lies ahead. The opportunities in front of us include the consumer side last year and the business side this year. We need to do these things, and do them well, but they are not our goal.

In other words, most people at our company do not see this as enormously important, or as an issue on par with AGI.

We can talk a little more about open source, since many of the earlier questions were about it. First, we will open-source our models, and we will probably open-source our strongest model too. That is because I do not see an advantage to keeping them closed, or any inherent advantage in doing so. ByteDance's models are closed. What advantage does that give them? I do not see one.

Even if you open-source a model and tell others everything, the barrier is still very high. The barrier for others to actually use it is also very high. It is hard to put it into actual use, and second, it is also very hard to operate it at a very low cost. It is not that easy.

Just because I open-source it does not mean someone else can easily match my deployment costs. There is still a great deal of work involved. Even if everyone understands the principles behind that work, not every company has the willingness and ability to organize the people needed to achieve the goal.

I am very familiar with this as well. They probably will not do it well, because there is too much resistance. Controlling costs is very difficult, and there are many managerial and physical constraints. This is also an advantage for startups. If a startup is too small, it lacks the capacity to do this; at a large company, it is difficult to organize.

There are difficulties in all of this. So it falls in the sweet spot—the sweet point—for a company of our size. If we were bigger, perhaps we might not have other problems; if we were smaller, our team would not have enough capacity.

So when it comes to open source, I think we need to set a price. You should understand that, for now, we are not going to squeeze people. Even for models we charge for, we will not charge very high fees. We will probably price them to recoup our investment in 10 months. Pricing for a 10-month payback already puts competitors…

If I price for a 10-month payback, I can make it impossible for third parties deploying independently to turn a profit. Third parties cannot achieve that cost level. They definitely cannot.

So open source doesn't affect my revenue. Of course, if I wanted to make a 100-fold profit, open source would…

We can hear you, but I think your video has frozen, sir.

I think a call just came in.

So I don't think open source has any effect on our business model, provided we're only making a 6-fold profit. Recouping our costs in 10 months works out to roughly a 6-fold profit. If that's all we're making, open source has no effect.

But if you're trying to make a 100-fold profit, open source really does affect your ability to do that. A third party will deploy the model, and even if their costs are 20 times higher, they could still undercut you.

Is this model sustainable over the long term? I think it is. Within this vision of ours, open source is sustainable over the long term, and at least that's what we plan to do. It involves restraint, and you could also see it as leaving you a share of the benefits over the longer term.

This strategy gives us more opportunities at the technological frontier and improves our chances of succeeding with AGI. It gives us more breathing room.

Think about it. We don't even need to work overtime, because it's not that difficult. But it might be very difficult for others, because they have too much on their minds.

It's actually not difficult. Not difficult at all. When we first started, it was a… From the outside, it might look as if we've chosen a very difficult mode—hard mode—because we're doing research and taking on the hardest problems. But in fact, because we've given up so much in other areas, we still have plenty of energy and can do this with ease.

So my assessment is that open source really is sustainable. There's no conflict between open source and charging commercially. Not if we're working on the premise of a 6-fold profit.

A 6-fold profit might sound high, but it actually isn't. With AI as efficient as it is now, that's probably a reasonable level of profit. In the future, it might fall to, say, 4-fold or 3-fold, and I think that's already… I don't think it can go any lower than that. Even then, there's still a substantial profit.

That's just looking at selling API access, though I don't find that business particularly appealing. Still, looking at that business alone, there's no conflict, and I can say I haven't seen any.

And I'm not at all worried about other people deploying our models and competing with us. In fact, I hope they can deploy them.

We help the open-source community as much as possible and support everyone in deploying our models. I'm not worried about someone competing with me for this business, because the market is big enough. My only concern is that they might be unable to deploy the models, or get some detail wrong and end up with worse results or relatively high costs.

Yes, there's no conflict here.

And here's a question I asked last year, quite often when we had the B2B business: on the consumer side, wouldn't going open source conflict with my own consumer business? I might not have an advantage in user traffic, whereas Tencent has a lot of traffic of its own. It could deploy our open-source model, bring in all the consumer users, and take away every one of mine.

But that isn't actually what happens. There are still many reasons for this.

And here's the question: is the open-source model we provide the same as the model we deploy ourselves? It is. We don't release a slightly inferior model as open source and use a better one for our own deployment. They're the same.

That, too, shows there's actually no conflict. Throughout last year, we essentially open-sourced what we used on the consumer side, and we saw no conflict with our consumer service. We really didn't see any conflict. So that covers open source.

Next is the company's long-term vision. I think our goal should be AGI. People may define AI differently, but that doesn't stop us from making AGI our goal.

In terms of the technical path, the roadmap to AGI is actually relatively clear. The current generation of AI technology already surpasses humans if you explain the problem very clearly and provide complete context and instructions. But there's a definition and a condition attached to that: you have to give it complete context and complete instructions.

But it's very difficult to meet that condition. For example, even this meeting we're having today has a long, substantial history of context behind it. Each of us may have decades of context, but AI doesn't. What AI can currently have is the ability to outperform people within a limited context.

But it still can't replace humans. It's missing one more thing: continual learning. People can keep learning, too. When you hire an employee, they spend about two months getting familiar with the company environment and their work. They spend those two months talking with people, and then they're ready to get to work.

They can do many things. They can understand what you mean. For example, if you ask them to get Xiao Wang, they know who Xiao Wang is. But AI doesn't have that context; it hasn't spent the previous two months learning.

If you ask AI to get Xiao Wang, you have to tell it who Xiao Wang is, what position he holds, where he is, how to find him, and what to keep in mind when looking for him. You have to give AI all the context. Under those conditions, AI can do it, but you can't provide all the context, and it isn't practical either.

That's why AI can't replace your employees. But if AI gains the ability to learn continually and spends two months learning at your company, just like an employee…

Then it can replace people out in the world. So the one thing still missing for the next step is learning how to learn.

You can think of AI's development as a staircase. The step we climbed last year was CoT, or chain of thought, because we discovered that chain of thought could take intelligence to a higher level. By letting AI think for itself, we can raise its ceiling and enable it to do more.

Then we climbed another step. This year's step is agents, because we discovered that an agent-based approach lets AI handle more tasks, broadens its range of capabilities, and raises the ceiling of its intelligence.

Why a staircase? Because every subsequent step builds on the foundation of the one before it. Agents need CoT, and CoT needs the preceding step, too. That preceding step is language models. So not a single step we've climbed has been wasted.

So we can trace the trajectory of AI's development and the direction of its intelligence. The step we're climbing this year is agents, but eventually we'll reach the top of that step, too. Even after solving all the problems it can solve, AI still won't be able to replace employees, yet it will already have reached the ceiling of its capabilities.

It's the same with CoT. Once CoT reached its ceiling, it had already surpassed the best humans at solving Olympiad math problems and writing programs. But it's still stuck there. That technology hasn't reached AGI.

So, you see, there's a trajectory we can trace in the development of AI intelligence. I think that after agents, we need to solve continual learning: how to get a model to keep learning over a relatively long period, like a person, rather than relying solely on intensive training.

This is connected to things like completing tasks; they're interrelated and address the same problem. Right now, we're at the agent stage, and we can see the next bottleneck: continual learning. The next problem to solve is how to keep learning continuously. We can see that, and it's relatively clear.

We're stuck at this obstacle right in front of us. We have to get past it, and there is certainly a way to do that, but it will take time. After continual learning, we could reach a singularity. By that, I mean that once a model can learn continuously, it can already do everything a human can do.

It can develop a version of itself, do further research on its own, and then develop the next version and more advanced AI models. That's how we reach a singularity where it can iteratively improve itself.

But this singularity isn't really a singularity; it's a gradual process. It, too, could be a relatively long, gradual change rather than a sudden shift. We just habitually think of it as a singularity.

Long ago, visionaries predicted there would be a singularity here, but in fact it's a continuous process, not a singularity. And I think embodied intelligence comes only after this stage is complete.

This is our hypothesis. The timeline we envision is to first solve the problem of learning to learn, then reach the intelligence singularity—the point where iterative self-improvement becomes possible—and then move on to embodied intelligence. Once we reach embodied intelligence, it can enter the physical world, do your housework, and care for you in old age.

We see this as a fairly ideal roadmap, but everyone has a different perspective, and there's no right or wrong here. We simply think this roadmap is the easiest path.

With this roadmap, there's very little new work to do at each stage. Follow this roadmap, and you don't have to work late. But if the roadmap is reversed—if, for example, you have to achieve embodied intelligence first—you're making things very difficult for yourself. It's hard work. We don't want that kind of roadmap; we'd like an easier path.

If we solve continual learning first, then reach the singularity of iterative self-improvement, and then solve embodied intelligence, the path is very easy. That's because, as we go along, earlier technologies can help develop the later ones. If we tackle embodied intelligence after the singularity, people won't have to do it—we won't have to do it. The model can create it on its own.

So that's the answer to the question of what our long-term goal is. I told him that this is our long-term goal. This is what we mean by AGI.

Let's all come back to reality. Last year, the overriding reality was that everyone had to build chatbots and compete for consumer traffic. This year, the reality is that everyone has to compete for B2B revenue and get involved. If you don't participate, you don't even have a seat at the table, right?

But we don't see that as the important thing. In other words, what we really care about internally is the AGI roadmap I just described and how to achieve the next technological breakthrough.

But there's something very strange about this. The thing you want most is the thing you don't get, while the things you don't care that much about come quite easily.

There's a strategic advantage here. AGI is what we have in mind, and AGI is what we work on. So when we turn to applications, consumer offerings, or enterprise offerings, we don't need to devote that much attention to them. In fact, they take very little effort.

I think when you're operating at a higher technological level and then work on relatively lower-level technology, you can bring that overwhelming higher-level advantage to bear.

At least on the consumer side last year, that really was the case. We didn't put much effort into it. At one point, we didn't even want to retain those users, but even when we tried to drive them away, they wouldn't leave.

They really wouldn't leave even when we tried to drive them away, so in the end they're all still here. But a little... And this year, the growth in our enterprise revenue looks fairly promising so far. I suspect the numbers are relatively good even compared with our peers. That's my estimate. But we haven't put a huge amount of effort into this, and we haven't even...

When we do something like that, we're just doing it along the way.

As we roll out internet intelligence and move forward, this stage of AGI is a step I have to take. I have to pass through it on the road to AGI. So I make all these technologies available to you through APIs, and I haven't done any extra work.

We're still working on AI, and this is a byproduct. We just need a few people to maintain the API. We don't even have a customer support center, and we don't need sales or anything else. Users come on their own.

In other words, what people think of as the consumer user base—both the consumer and enterprise businesses—are byproducts and intermediate outputs on our road to AGI. They don't conflict with my work on AGI. I'm not doing this to build a consumer or enterprise business. We pursue AGI for its own sake, and when this happens to emerge from that work, we take it and commercialize it.

That's different from other companies. They build these models for that purpose—to serve consumer or enterprise users. But that wasn't our original intention. Our original intention is still to pursue AGI.

I think this is, to some extent, bringing a higher-level advantage to bear. AGI is a bigger vision, and that vision brings together more outstanding people and creates stronger cohesion. So we have an organizational advantage, and we use that advantage to... That's what I mean by an overwhelming higher-level advantage.

But if you're a commercial company whose vision is to serve consumer users well, that's a different story. You may have other advantages in products, user service, and traffic, but you don't have a technological advantage.

What's relatively favorable for us right now is that model technology matters most. You have to build a good model, and the rest... Yes, and that explains our trajectory so far. We really chose AGI. We never had any intention of building a large user base.

When we suddenly became famous during Chinese New Year last year, that wasn't in our script. We'd never thought about anything like that; we were just trying to build good technology. But that was when I realized that, compared with fully commercial, product-oriented organizations, we had an additional advantage in talent and organization.

That's fascinating, too. At the time, everyone was fighting tooth and nail over the consumer market, but in the end, the one company that wasn't fighting took it. That also shows there's something to the logic I described earlier, and that we really do have an advantage in talent and organization.

This talent advantage doesn't mean our people are smarter than theirs. It means we have an advantage in how I organize these talented people, motivate them, and get them to work together. Simply bringing smart people together doesn't mean they'll naturally collaborate or enthusiastically pursue a shared goal and see it through. That's why you need a vision.

What those earlier experiences taught me is that the vision of AGI is tremendously powerful.

OK, that's the question. And the next question is about the importance of our core interests.

Earlier, we said we need to exercise a great deal of restraint in many areas. But what are our core interests? In fact, we have only one. Our most important core interest is keeping the team stable. That is our biggest core interest, and you could even call it our only one.

As long as we can keep the team stable, we will succeed. We will achieve AGI. It's that simple. As long as everyone stays and keeps going, I will definitely… Basically, there's no risk. It may just happen a little sooner or later, and we may encounter setbacks. Even if we do, as long as no one leaves, we can pick up and keep going.

Money certainly isn't a problem, and neither are resources. Everything else is easy to obtain. We have just one core interest, one thing we cannot compromise on: we must keep the team stable.

This is also a major challenge we face and, in my view, our biggest risk. Of course, our latest funding round has substantially reduced that risk, because everyone has received a fairly large number of stock options, with a fairly substantial value.

When the employees most important to the team's stability and the longest-serving employees are secure, few others leave. The others won't leave even if they have somewhat fewer options or earn a little less. They didn't come just for the money. They all want to do this work in an environment where they can achieve AGI.

So we remain attractive to talented people. Historically, our staff turnover has been relatively low, and consistently lower than that of our industry peers. But this is still our biggest challenge, and you could say our only challenge.

Everything else is just a matter of time. At most, other factors might delay us by six months or a year; they won't stop us from building it. We certainly aren't short of money or resources. In fact, we aren't short of any of those things.

So much of what we're doing now is aimed at keeping the team stable. Apart from that, I think everything else is something we can do without, something we can exercise restraint on. We've always been very restrained, and we don't want to make enemies of any internet company, large or small.

I want to empower them, or help and support everyone in doing this work. That's also part of the commercial purpose I mentioned earlier. The condition is that no one interferes with…

On that basis, we're very willing to support and help anyone do better, even competitors like Alibaba, Zhipu, and Moonshot AI. We lose nothing by doing so. It's open source to begin with, and even in open source, without spelling out the boundaries as clearly as possible, we want you to reproduce how it's done. If you can't reproduce it, tell me. I'll show you how.

That's part of open source in the first place, and it doesn't change just because you're a competitor. Of course, we'll do more for a partner. But in terms of our broader interests, there's no conflict.

When we engage with others, our position is that we focus only on the main path to AGI. I mean things like GPT, CoT, and agents, which I just mentioned. We stick to that main path. AI is a very broad field, and there are many things we don't see as being on that path. For example, 3D and video generation may have little to do with the main path toward intelligence, so we won't pursue them.

We also won't pursue things like world models, because at this point we don't think they have much to do with the upper limits of intelligence. But if others pursue them, we're happy to help. Whether we have the time is a separate question, but there's no conflict of interest.

We want these AI technologies to be used in a variety of production settings, improving productivity across society and helping every industry become more productive. We're strongly motivated to make that happen.

Whether I have the time and people, and whether our colleagues are personally interested, are separate questions. But there's no conflict of interest, and we want to achieve this goal. I also don't think it conflicts with business at all, or diminishes the returns I should receive.

We've maintained this attitude in the past, and I don't think we've actually lost out on anything because of it. We've never received less because I embraced open source, because we showed goodwill, or because we helped others. For example, the consumer user base we built last year is still fairly large and stable.

I'm also fairly optimistic about the enterprise side this year. Our goodwill has never hurt our commercial interests; it may actually have worked in our favor. It seems counterintuitive, but that's really how it is. Look at it the other way: if I acted against that approach, would I necessarily gain more? No.

One question is how to interpret the statement that world models have nothing to do with raising the upper limits of AI intelligence. What I was describing is our judgment at this stage. We have our own perspective, and this is our own AI roadmap, not the only roadmap.

Based on our understanding and judgment, the most important thing right now is to get AI training right. You don't need world models to do that. You don't even need multimodality. If you narrow the scope of AI training a little and leave out multimodality, there will simply be some tasks it can't perform. That doesn't affect whether the algorithm works.

We'll need to do multimodality eventually. What matters now is training. The next problem to solve is continual learning, and after that, getting AI to ask its own questions. But neither world models nor video generation is on this roadmap.

When video generation first appeared, it was incredibly hot. It felt like something you absolutely had to do, as if you weren't an AI company unless you did it. I found that very strange. In fact, if you just think it through carefully, you can see that it has nothing to do with the roadmap toward intelligence.

In fact, after Sora first came out, everyone got into video generation, big companies and small ones alike. But the small companies all cut it later. It has nothing to do with the upper limits of intelligence.

But commercially, it's a good business. It's a good business, but it has nothing to do with intelligence. We don't do something just because it's a good business. It has to be on the roadmap toward intelligence.

I used video generation as an example because it's relatively clear-cut. The meaning of "world model" isn't so clear, because many things can be called world models.

In our view, world models and intelligence are not yet the most important things at this stage. What matters most is AI training and how to solve continual learning after that training. That is our company's assessment, and of course every company sees things differently.

For the company I just mentioned, the most important issue is workforce stability. Looking at it from another angle, what are we lacking? What is the gap between us and the United States? In fact, the only gap is resources.

We don't have that many cards; our card count is still relatively low. Right now, our compute is roughly equivalent to 20,000 H-series cards, most of which have only just arrived or arrived in the past month or two. There may also be many machines that haven't arrived yet.

Our total compute capacity was relatively small last year, and this year we are expanding it very aggressively. We currently have the equivalent of roughly 20,000 H-series cards, and over the next few months we'll buy a large number of additional machines. They are essentially all NVIDIA.

How many cards do we need? Right now, more is definitely better. There is no question that, within what we can afford, the more cards we have, the better. So our strategy now is to buy as many as we can at reasonable prices.

When it comes to spending all the money from this funding round, I'll buy as many cards as I can. We don't have a planned spending schedule; as long as the price is reasonable, we'll buy everything available. I think spending all the money within six months would be a good thing. If we could spend it all within six months, I'd be thrilled. That would be ideal.

Actually, spending this much money is far from easy. You can't buy that many cards. They're hard to get, and prices are very high. But we can't just pay exorbitant prices, either. We have to make sure the price is reasonable.

If I could spend all this money within six months, that would probably be the ideal outcome. Turning money into NVIDIA cards is clearly better than leaving it in the bank. If you leave it in the bank, I think the interest rate is probably around 2 percent right now. But if you buy NVIDIA cards, that's 10 months of societal costs.

So we will definitely buy as many as we can. Once we have the cards, we have a lot more room to maneuver. Whether by providing services or doing something else, we can always generate cash flow. With cash flow, we can stay afloat, and we don't need to keep a huge pile of cash on our books.

So our only worry is that we won't be able to buy that many cards. If we could turn all our money into cards, we would do it without hesitation. And we're willing to pay a certain premium for that. I mean we're willing to pay a premium to turn that money into cards. It makes so much financial sense.

Even if we pay a premium, this goal is actually hard to achieve. Objectively speaking, if we can spend 20 billion this year, that would be an outstanding performance by our procurement department.

The gap between us and the United States is mainly in resources. The talent gap isn't large. In terms of people, there's almost no difference. They're the same kinds of people, perhaps because they're Chinese, too. Among Chinese people, some stay in China, some remain abroad, and some move overseas. It's not as though only the smart people went abroad.

In fact, it's fairly random. It may not be the case that a little over half of the smartest people went abroad, but a little under half remain in China. There is no shortage of talent here, and the pool is large. So many new people enter it every year.

Talent isn't the bottleneck; resources are the biggest bottleneck. First of all, resources affect talent development. With less compute, we have fewer opportunities to experiment, which creates a gap with the United States in the overall level of talent. Even the talent gap is fundamentally caused by the compute gap.

We actually can't afford to train the largest models today. Even if we spent the entire 50 billion, we still couldn't train one. Even if we managed to put one together, we couldn't run it. The largest models today have roughly 800B active parameters. In China, we're still at tens of billions, and even the largest domestic model probably has tens of billions of active parameters. That's an order-of-magnitude difference in scale.

To train a model the same size as AI's, I would probably need 50,000 GB300 cards or 200,000 Huawei 950 cards.

That's counting only training; it doesn't account for research yet. So the biggest gap between us and the United States is in resources.

The resources we have now, and those we'll have this year over the next few months—even including the large amount of resources arriving soon—are only enough to do more experiments at the scale of 10B active parameters. Even at tens of billions of active parameters, there are still many experiments to run and many things to figure out. We're still quite far from being able to train an 800B-active-parameter model, and it will take a long time to get there. We don't have that many cards.

So I see the difference between us and the United States as a difference in resources. Every difference we see—in talent, in model capabilities, and even in applications—can be traced back to differences in compute resources.

When it comes to compute resources, on the one hand, we can't buy the cards themselves in China. On the other, our capital investment is smaller than America's. We invest far less capital, and talent compensation accounts for a very small share of it. They offer annual salaries of $100 million and the like, but if you do the math, talent compensation is still a small proportion. The bulk goes to compute.

There is essentially no solution to this problem right now, because Huawei's production capacity is limited, too. To train an 800B model, I would need 200,000 of Huawei's latest cards. And again, that's only for training; it doesn't account for research.

So right now, we aren't even considering competing with the United States at that scale. Right now, we're losing. Still, we first need to build good models at a scale we can train and run—that is, tens of billions of active parameters. Then, when we have more resources, we can scale up to 150B, 156B, or 250B active parameters.

So in this area, there is a gap between us and the United States that looks difficult to close for now. We could force our way through training a model that large, but we couldn't do enough research. In other words, we couldn't do enough research before training it.

There's another question that a lot of people are interested in. In all this competition among large models, where will the ultimate differences show up? What will ultimately distinguish large models from one another? I think there may not be much difference in the end.

The ultimate differences will come down to three things: cost, timing, and user experience. Beyond that, there probably won't be much difference.

Cost is easy to understand. At what cost can you provide the same service, at the same quality? Take BYD's batteries as an example: can another company offer the same level of technology at that price? I think that's quite a difficult challenge. It's not something you can do easily, and it's definitely a barrier.

So cost is definitely a differentiator. I would probably put cost first. The second is timing: when can you get it done? Being a few months earlier or a few months later makes a difference.

The third is probably experience. The user experience does differ to some extent. There may also be some user loyalty and barriers to switching, but those may not be fundamental. What matters fundamentally is, first, cost, and second, time. Do you build it first or later? How quickly or slowly can you build the same thing?

How will we set prices in the long run, once there are more paths to commercialization and a broader product lineup? I think the most worthwhile thing to do right now, and the thing most worth putting our effort into, is still AGI. We need to push AGI further forward now. Raise the floor of intelligence and keep pushing it forward.

At this stage, this is probably more cost-effective, or offers a greater return, than building out more product lines and thinking through more paths to commercialization. That will probably remain true for a while, and it was probably true at every point in the past as well. So if we...

What would be the point of spending a lot of time expanding our product offerings?

What commercialization ideas would we have come up with if we'd discussed this half a year ago? Surely it would have been running ads, doing e-commerce, embedding e-commerce in the product, and deeply integrating it with local services. That would clearly have been pointless, because things are changing too fast. With this product, whether earlier on or at the current stage, if you spend a lot of time studying commercialization, paths to market, and product lineups, what you come up with has a very short life cycle.

I don't think it's time for that yet. That's our judgment, and at least all our experience so far supports it. At any point over the past 3 years, talking to me about commercialization paths or product lineups would have been a waste of time. That's because we can't predict or see into the future. There's very little we can foresee.

Looking at the time, should we all take a break? Shall we get something to eat first? Or should we keep talking? If no one has a preference, I'll keep going.

Do world-leading AGI research and commercialize it when the time is right. I think we've been commercializing our work all along. We just haven't made commercialization our goal. Our goal is AGI, but we've continued to commercialize along the way, which is why we have both consumer users and business revenue.

Our experience shows that this strategy has worked. And I think the point at which we shift entirely to commercialization is still a long way off. So the biggest priority is still advancing the technology, building the next generation of technology, and solving more of today's problems. Compared with the returns on those efforts, I personally think it's too early to focus on products at any point in the foreseeable future.

So this is also part of our restraint. We want others to pursue these commercial opportunities. We want society as a whole and all our partners to work together on these opportunities and on how to use AI, and to share in the returns. I'm not trying to keep it all for myself. That's impossible. Besides, we don't have the capacity to do that many things, and we don't have that many people in the organization.

As for partners, we were actually very selective about this round of investment. The specifics of how we'll collaborate are still proposals, but first of all, I think our interests are relatively well aligned. These are the ones whose interests align most closely with ours, who are least hostile toward us, or who most want us to succeed. Not everyone wants us to succeed, because we've hurt a lot of other people's interests.

The processes and decision-making structure behind the company's major strategic, technical, and business decisions. Our entire company is built on consensus. I don't decide everything on my own; I have to find consensus. My authority and influence within the company are based on consensus.

For example, if I want to do something, I first look at where we all agree and whether everyone wants to do it. I can steer things somewhat or signal a preference, but that influence is limited—very limited. It's still based on consensus. This decision-making structure is really a structure for finding consensus. It doesn't mean I can force something through. There has to be consensus for something to move forward, and only then do I pursue it.

Right now, my energy is basically focused on DeepSeek. Let's take a 5-minute break. I'll write this up quickly. Everyone, please turn on your microphones and give us some feedback. Everything is fine on our end. Or let's take a 5-minute break first.

The differences in the final performance of each company's models must come from a combination of factors. For a comparison of model performance to be meaningful, you have to compare them at the same cost. After all, when you compare two cars, you compare cars in the same price range.

The difference between a well-built model and a poorly built one probably doesn't lie in any one particular stage, but across the whole process.

Will Anthropic's current lead over OpenAI last? I don't think it will. It's clearly a temporary extreme. OpenAI and Google will very likely take turns pulling ahead in the future. In fact, Anthropic's current advantage in coding agents isn't that large, and it certainly isn't enough to say it dominates OpenAI.

Probably half the people at our company generally think OpenAI is better. Anthropic does have a first-mover advantage, but that advantage will soon disappear. It's not an advantage it can hold on to for the long term. All three companies are impressive. Among the three, Anthropic is the most efficient, and it has probably incurred the lowest costs and burned the least cash.

As the division of labor in global AI takes shape, the role Chinese companies are most likely to play is still that of the largest producer. By ordinary reasoning, we have the greatest production capacity—including chips, where we'll probably also have the greatest capacity—and the most electricity. So ultimately, our AI is likely to be one of the three major poles.

The Chinese will make this product at the lowest cost. In terms of performance, too, many Chinese-made products today ultimately aren't that different from overseas products, including American ones. AI could follow the same pattern in the future, but Chinese-made AI could be cheaper. That lower cost could be systemic, much as the services China provides in other industries are cheaper.

This is how I'm used to thinking about my work: What can I do right now that will generate the greatest return? If I think building products offers the greatest return right now, I'll build products. If I think achieving AGI first offers the greatest return, I'll work on AGI first. I certainly don't think building products offers the greatest return right now.

As for adapting to domestic accelerator cards, those cards actually have a historic opportunity right now. Previously, adapting to domestic cards faced a major obstacle: the ecosystem wasn't good enough. You could buy the cards but couldn't use them, and they didn't have NVIDIA's ecosystem.

That's why NVIDIA's moat is so strong.

But that's changing. NVIDIA's CUDA moat is rapidly crumbling. There may be three reasons it's crumbling so quickly. One is that we now have AI, and with AI, building this ecosystem has become much easier than before. AI can write code.

We can use AI to build this ecosystem, and we can create an ecosystem exactly like NVIDIA's. First, because we have AI, and second, because there are new technologies. For example, we released a technology called TileLang, which is a high-level language.

If we write CUDA operators in this high-level language, we can rewrite NVIDIA's entire ecosystem very quickly. Add AI to that, and there don't seem to be any obstacles. But it isn't finished yet; we haven't done it all. It's just that we don't see any obstacles along this technical path.

Another point is that CUDA evolved from NVIDIA's gaming cards, so it inherited much of their design and configuration. CUDA is compatible with gaming cards. In the past, that made sense because AI computing was a very small field, smaller than the gaming card market.

But now the market for compute cards is already larger than the gaming card market, so there's no reason to keep the two tied together. The trend now is to separate them going forward. Whether from Huawei or NVIDIA itself, future chips will all be purpose-built, different from what came before.

Against this backdrop, the role of NVIDIA's existing ecosystem has diminished considerably. Purpose-built chips have little to do with CUDA and aren't tied to it. In other words, how to build the ecosystem was considered from the moment these chips were designed.

It's a little complicated, but in any case, there is now a historic opportunity to replace imported AI chips with domestically produced ones. We think that within the next year, we'll see one thing proven: there is nothing wrong with the domestic chip ecosystem. People used to think it had problems, that it was unusable or difficult to use. But within the next year, I think we can overturn that perception—overturn it with facts.

There is nothing wrong with either the hardware or the ecosystem of domestic AI chips. The only problem is insufficient production capacity. There are no obstacles to adapting to domestic cards, and NVIDIA can't stop it either. In a normal commercial environment where you can buy NVIDIA cards, replacing them with domestic alternatives is relatively difficult. But when you can't buy NVIDIA cards, everyone has no choice but to work with domestic chips.

Against this backdrop, there are no obstacles whatsoever to adapting to domestic cards. I don't see any obstacles to building an ecosystem for domestic cards that's as good as NVIDIA's, or even better, but it needs more time.

Right now, we mainly work with Huawei. Huawei does its own adaptation work, but we'll also participate in this ecosystem and be deeply involved on Huawei's side. Huawei's problem is still insufficient production capacity. Huawei's production allocation for us is roughly 16,000 cards. The big internet companies probably get a hundred thousand or more, while we get ten thousand or so. That ratio is relatively… But that's probably also because Huawei only has so much production capacity.

So we can't expect to train our next, larger models—models with hundreds of billions of active parameters—on Huawei hardware this year. But there may be a chance next year or the year after.

In adapting to Huawei cards, our main work is to build a good high-level language compiler and make TileLang work well. If we get TileLang right, the problem can be solved easily. It's relatively complicated to explain, but we're working on it, and it will be much clearer if we explain it once we're done.

Think of it this way: when we trained V3, we used NVIDIA cards, but we were already no longer using NVIDIA's ecosystem.

V3 used NVIDIA cards, but not NVIDIA's ecosystem. We first built a high-level compiler called TileLang, then completed everything else on top of the TileLang ecosystem, so we were already barely dependent on NVIDIA's ecosystem.

If I reproduce this system and process on Huawei cards, that completes it. I think there may be a historic mission here. We can completely overturn the existing perception that the domestic card ecosystem isn't good. Almost all the favorable conditions are now in place; it's just a matter of time. Within a year, many people will have built something on it, or everyone will understand that this problem has been solved. The only issue left will be production capacity.

I'm relatively optimistic about domestic computing power. In this respect, I think NVIDIA is digging its own grave. Huawei's supernodes—the Huawei 950 supernodes—can fully replace NVIDIA's GB200 and GB300 in terms of performance and price. They will certainly cost more, but only to a limited extent. They can cost 50% or 100% more. A 100% premium doesn't matter; even a 200% premium doesn't matter.

For example, if it costs 100% more, I think it's already a viable substitute on price. It can also replace it in terms of workloads. Huawei's supernodes can handle every workload the GB300 can, and things like latency are the same. The only trade-offs are that it takes four Huawei cards to replace one NVIDIA card, and that they're also 2 years behind.

The idea of four cards replacing one is easy to understand. Being 2 years behind means that four Huawei 950 cards can replace one GB300. By 2 years behind, I mean 2 years behind in time. The Huawei 950 supernodes will ship in Q3 or Q4 of this year, while NVIDIA's GB200 shipped in Q3 two years ago, so there's a 2-year gap.

NVIDIA could already have a new generation of products out in Q3 of this year. So going forward, there will no longer be an ecosystem gap between the US and us, but in chips, the gap is a factor of 4 and 2 years.

Another question is whether we'll vertically integrate upstream. I hope not. So the question is whether we'll vertically integrate into upstream applications, and we hope we won't need to. We hope other people will do this. I don't want to take the whole pie. One slice is enough. Just the slice I'm best at, or the one we consider most essential, and then with users…

The part directly related to them. For many of our industry partners, I think they may be more interested in similar things than we are…

It has to be meaningful to other people; I shouldn't be the one interpreting it for them.

Will we build our own large-scale clusters in the future? I think building our own large-scale clusters is definitely necessary. We've been doing this all along, and we've built all our clusters ourselves. But whether we'll develop our own chips in the future depends on how large the returns would be. It depends on how large the returns would be.

Tesla, fostering… this business. For example, if you operate a power plant, you don't necessarily have to manufacture the generators, do you? Someone else can make the power-generation equipment. As long as the price is reasonable, why should you make it yourself?

So I hope we won't have to get into chips. I hope we can buy chips at a reasonable price, and then we won't need to make them ourselves. I think that's quite likely. Recently, the profit margins on NVIDIA chips aren't necessarily… although…

We hope to focus on just one slice. AI is such a huge undertaking that I… don't need to do it all. We'll do just one slice. Stay focused, and I think the business profits here are already large enough. If the AI era produces several trillion-scale companies, we'll be one of them.

We've always worked on just one small piece of it. There's already little doubt that we're one of the players, and I don't need to… I'm not confidently declaring that we're going to take some particular position. The most likely outcome is relatively hard to change, and if we really want to do something different, we'll have more ideas. This is… this is our attitude.

For example, at least in the B2B and B2C businesses, what we're seeing now is that we've actually done better than those who set out to build a closed-loop B2C business. And those who set out to build a closed-loop B2B business may not have done as well as us either. The more you want…

And I can say a little more about B2B. The ceiling for a B2B business will still be determined by demand. With the current generation of AGI and AI technology, B2B demand will be limited. It will grow rapidly, but it isn't infinite. Ultimately, the constraint is demand, not compute.

With today's technology, how much revenue you can generate ultimately depends on demand. Think of it this way: I can recoup the cost in 10 months, so if I had a use for it, I'd certainly buy it all at once. But there just isn't that much work to do. That's right. Demand will keep growing, and as technology continues to make breakthroughs, that demand will grow further.

We're continuing to build out our multimodal capabilities. They're very important for products, especially consumer-facing products. But when it comes to the upper limits of intelligence, multimodality is one component, not the main thrust itself.

But we will definitely pursue multimodality as a component, and we're already doing so. We'll release models for it. V4 and subsequent versions will support native multimodality. But in relation to intelligence, we see multimodality as a component, not intelligence itself. In relation to the main thrust, it's like search.

Search is a component, and so is multimodality. That's how we understand it: as a component. We believe in scaling. The larger the scale, the better the results, and the more capabilities you can unlock. What's holding back our scaling is compute. It's not that we don't want to scale; we can't scale because we don't have that much compute.

We haven't even begun to test where that ceiling is. We train a model of this size not because we think this size is enough, but because this happens to be how much we have in resources. The size comes from calculating what we can afford and train with our resources, not from deciding that this model is sufficient.

As things stand, the returns from models are still very clear. We haven't even had a chance to hit a scaling wall; that's still a long way off for us. When Silicon Valley says scaling has reached its limits, that's Silicon Valley's story. For those of us in China, we're still far from that point. We haven't even been able to scale to that extent.

This scaling includes scaling data, scaling model size, and training costs. Because we don't have that much compute, we're still quite far from being able to explore the upper limits of scaling. But we'll do everything we can to push those limits, too.

Once we've completed the funding round, more compute will let us train larger models. If we can buy it in time, we'll buy it as quickly as possible. Ideally, we'd be able to put all that capacity to use within half a year, but in practice we can't.

I think continual learning has to be a core capability before you can call something a next-generation model. Until then, what we can do is lower costs, improve performance, and make things faster. But a major breakthrough requires continual learning.

And another question is why we seem to place particular emphasis on models' computational efficiency. We've found that not everyone actually values model efficiency. Commercial companies have no incentive to pursue it. If a model becomes a little more efficient…

So there isn't much incentive to push model efficiency very high. Some startups themselves… You've probably never heard them say that low cost is what they're pursuing. That's because it isn't in their interests. If costs go down, where do they make their money? If costs go down, they can't charge for it.

For us, by contrast, this is part of our vision. Or you could say our colleagues care about these costs because of our vision. Our colleagues are all ordinary people, and they know it costs money to use this. They have the empathy to understand that other people have to pay to use it, so making it a little cheaper makes them more willing to adopt it.

So many of our colleagues want to bring costs down further. But from a commercial perspective, that's not really how you think. Commercially…

From a cost or commercial perspective, this isn't the top priority.

Whether you're a startup or a large company, the cost of this service isn't high at all. But we want it to be relatively lightweight and affordable. Especially given the shortage of compute in China, we want it to be manageable and usable on domestically produced cards as well.

I see low cost as an outcome first and foremost. In terms of architecture, our models have certainly been moving consistently toward lower costs, and that has to do with our vision. There are also various algorithmic approaches, so costs can come down further.

Another reason we bring costs down is that lower costs let us afford to train larger models. With the same amount of compute, in a situation where compute is limited, greater computational efficiency lets us support a larger model.

Large companies don't necessarily think this way. They can solve the problem by adding resources. But we consider cost efficiency first.

When it comes to the value of data-driven models, data is a fairly broad category, and it should account for almost half the model. Why do I think it would never work if AI accounted for 20% of GDP and I tried to capture 5% of that? Because someone else would beat me. If someone else said they only wanted to capture 1%, I would definitely lose to them.

If my goal is to capture, through AI, 5% of all humanity's GDP, the math works in theory.

Look at OpenAI. The math seems to add up, and in theory there's no problem. But there's one issue: they'll lose to someone else who only wants to capture 1%.

If someone else does just as well but says they only need to take 1% of global GDP, they'll win. Then if another person comes along and says they only need 0.1%, they'll beat the previous one, too.

At a macro level, wherever you take those few percentage points from, there's no difference; it's all the same. Those who take more lose to those who take less. Even if you aren't actually taking more, if your vision is to take more, you'll lose to someone whose vision is to take less.

In fact, none of them have even made the money yet. It's just a vision. If your vision is to take more, you've already lost from the outset. You'll face greater difficulties. That's how the world works.

OpenAI genuinely thought from the beginning that it could monopolize this world, but in reality it will face a great many challengers.

Once they face competition, things won't be so comfortable. There are the challenges the US faces, and in the future they may face competition from China too, because Chinese providers are willing to offer this service while taking a smaller cut.

In China, too, there will be people willing to take an even smaller cut. But eventually an equilibrium emerges. That's because if you take too little, the company's business model doesn't work and it can't survive. Take too little and you can't stay in business; take too much and you lose to someone who takes less.

So rather than extracting as much profit as possible or setting prices to maximize returns, we earn only a reasonable return. That's one explanation.

I believe in what we're doing. I'm not looking for reasons to justify it. There's no need to look for reasons. This is how we've always done things, and there are certainly reasons for doing so. Those reasons may not be entirely conventional, but I don't think our company itself has ever been conventional.

Our company actually has two lines of management: one top-down and one bottom-up. Bottom-up means everyone decides what they want to do and pursues it independently. No one manages them, and there are no KPIs.

Top-down is the formal side. There are things we need to do collectively, so the entire company has to work together. For example, to release V4, we have to divide up the work so everyone takes on a portion. That's top-down, and we call this top-down work our official work.

Generally, we want this official work to take up less than half of an employee's total time. It shouldn't exceed half. For the other half, people aren't assigned tasks; they do what they want. We let them explore whatever they consider important within their research area, with no requirements set in advance.

If it's something the company can support and our compute resources can accommodate, or if it doesn't require compute and there's little work involved, they don't even need to come to us to coordinate it. That's how we're organized right now.

Some people say we're top-down, and others say we're bottom-up. I think both are right. My rule of thumb is that, ideally, official work shouldn't take up more than half of people's time.

We generally don't work much overtime either. There are two reasons for that. First, research requires a relatively relaxed environment.

If you push people too hard, they can't do research. They need to take an interest themselves and think about these problems in their everyday lives, so they need a relatively relaxed environment to explore. That's what a research culture requires.

Second, we're very focused. Being very focused means there are fewer things we need to do. And when there isn't that much to do, there's no need to work overtime.

This is in line with the restraint I mentioned earlier. Because we exercise restraint, in many cases we simply choose not to do something. When there's less work overall, there's less work for each person. You'll notice that many aspects of our products are incomplete, and we haven't filled those gaps either.

That's also part of our culture.

OK, there were a lot of questions, and I've gone through most of them. If you have any more, please ask.

Investors, please feel free to unmute and speak. Just a reminder: Liang Wenfeng has shared quite a lot of relatively sensitive information, so please do not disclose any figures, details of the situation, or GPU card counts to anyone outside this meeting. Please don't record the screen and share it externally either. Thank you, everyone, and if you have questions, feel free to unmute and speak.

Liang, could you share a bit more about the timeline for a breakthrough in continual learning? Including what architectural and algorithmic innovations, and what other key elements, would be needed to achieve it.

There are few people working on it, and it needs research. People around the world are studying this problem now. In other words, investors are currently paying the most attention to agents, but researchers like us are more focused on learning and how to solve this learning problem.

In fact, learning may be a problem rather than a single technology. There may be multiple technologies that solve this problem. It won't be one technology or one thing, but many things. Or, to put it another way, AGI consists of many things: it needs models, and it needs many other things too.

In fact, this is also an engineering and algorithmic problem. It's a fairly specialized topic, but there are many approaches and a lot of research.

Thank you, Mr. Liang. Thank you so much for giving us this opportunity today. First, I really want to respond to and thank you for your opening remarks. They were deeply moving and gave us a great deal of insight. You said that this team, with the best of intentions, wants to work in this industry to advance society and human intelligence and make whatever contribution it can.

You also spoke of the sense of mission and vision behind that. I think it's very similar to the corporate culture of a company we back: cultivating oneself and helping others flourish. It gave me a deep understanding of why you lead the team in pursuing open source.

To me, it feels as though we're building an ecosystem like a banyan tree, a paradise for little birds. If you benefit all things without competing with them, you are welcomed by everyone, all things coexist, and eventually you are everywhere. So through this investment, too, we will express our recognition, support, and respect for that mission and vision.

At the same time, we'd like to contribute to this industry in the future in areas where we have strengths. I'd like to keep asking about and discussing that as well. For example, when it comes to jointly building the future ecosystem, now that you've gone open source, how many industry partners, individual researchers, and teams can reproduce the models and results you've released reasonably well?

As we move to the next stage of developing this ecosystem, in which areas do you feel we need more top-tier talent to build on our models and reproduce them? Or is the bigger issue right now a relative shortage of GPU compute? Will everything move toward a model-matrix approach in the future?

For example, we keep getting better at building large foundation models, while partners and teams in different industries build vertical industry models or application models within that model matrix. How is this developing at the moment? As it evolves step by step, what kind of ecosystem do you think it will grow into in 2 or 3 years?

That's my first question. Second, you just shared many observations about AI hardware with the other partners. A single global AI giant can invest as much as 100 billion dollars, while China currently seems to have a weakness in hardware compute capacity.

How long do you think it will take to resolve this issue and support our AI development so that weaknesses in compute and hardware don't hold back AGI? Do you see this as a mission the Chinese people will inevitably accomplish—a matter only of time and financial investment—and believe that we can definitely do it?

But at the same time, there may be two sides to this. On the one hand, as models advance and become more intelligent, the hardware and compute resources consumed by a single task or by certain intelligent agents may gradually decrease, so they may no longer need such enormous amounts of computation. As models improve, they can compute more intelligently. I'm not sure whether I've understood that correctly.

On the other hand, advances in hardware technology will make computing more efficient. Are these two sides on a path toward convergence? If you were to invest $100 billion in compute right now using the current 960 or H200, you just said you would depreciate it over 3 years. But in terms of generational shifts in technology, do you see the actual life cycle as 4–5 years?

To put it bluntly, if you build a cluster of 10,000 cards in a computing center using today's cards, won't that compute be relatively uncompetitive in 3 years? At this stage, capacity is still under construction and in short supply, but could we end up with a surplus of relatively low-quality compute 3 years from now? I wonder whether that could happen. Those are my two questions. Thank you.

Thank you. The first question is about the ecosystem.

From what we can see, probably every company is facing a talent shortage right now. But I think this shortage is temporary. Every industry has a shortage of talent in its early stages of development.

Back when people first started building websites, very few people knew how to do it, so there was a severe talent shortage. Later, there was also a shortage of people doing server-side work on the internet. But these shortages were resolved very quickly, in just 2–3 years, because large numbers of people were trained.

The AI talent shortage is temporary too, and we've already seen it ease considerably. There really isn't a shortage of AI talent. Each company can train people quickly, and developing talent doesn't take long. So across the AI industry, whether we're talking about the ecosystem, model companies, or anything else, talent isn't in short supply.

Talent shortages are definitely a short-term phenomenon. There has never been a prolonged shortage of a particular kind of worker in history. I remember people saying over a decade ago that pilots were in very short supply and took a long time to train, but that was soon resolved too.

So there's no need to worry about talent shortages. And right now, there are a few too many companies building models here in China. There are still too many.

The US probably has only three, but China has too many companies working on foundation models.

Ultimately, there's clearly no need for so many people to work on foundation models, and the industry will inevitably consolidate. Resources are therefore spread relatively thin, and some are being wasted. First...

Every company has to do the same work, but in the US, only three companies need to do it, and resources are concentrated in those three. In China, resources are spread very thin, so each company gets less. I think this will definitely consolidate. It certainly will. It just takes a process, but ultimately, consolidation is inevitable.

We don't need that many companies. Right now, everyone probably thinks the profit margins in this business are very high, so they feel they absolutely have to do it themselves. But once they realize the profits might not be that high, they probably won't. Lately, the profits certainly haven't been that high. I don't believe there are such high profits to be made, because that goes against the underlying economic realities.

This means we're going through a phase. If profit margins are extremely high, that necessarily runs counter to the underlying economic realities.

We should be making reasonable profits. So that's the current state of the industry, and I think it will definitely consolidate.

So when it comes to everyone building large models, you clearly can't expect to take all the profits for yourself, let alone have one company monopolize the market. If each company earns only a reasonable profit, there really isn't a need for so many people to build large models. If China eventually has three or four competing companies, that will be enough competition, and certainly enough to ensure price competition.

For large models, perhaps two big companies and two small ones would be fairly sufficient.

I don't have that many thoughts on the ecosystem. We'd like to support more people, but we don't have that much capacity to spare. We have the intention, and there's no conflict of interest, but whether we actually do it is another matter.

But at least there's no conflict of interest here, and we want cooperation and mutual success. First of all, I absolutely do not think large-model companies can capture most of the profits. That's impossible. There are so many large-model companies, and the gaps between them don't need to be that large now.

There are only two differences: one is time, and the other is cost. So the gap isn't enough for any one company to make outsized profits. I don't think it allows for that. Those who control costs well will earn a little more, and those who don't will earn a little less. That's all.

Does that answer your question? Have I finished answering the first question?

Everyone can... You believe that many people will certainly be able to do it in the future, right... In the future, actually... things like everyone's use of data will cycle and iterate...

Can you hear me now? Thank you. Yes, thank you for your answer. I also deeply understand and respect the strategic position within the industry ecosystem that you described. For example, when it comes to data, I imagine all model companies have ways to obtain any data that is currently public. The means themselves wouldn't be the issue; it would just be a matter of time and cost.

Then, if we take the point when we achieve true AGI as an example, the state everyone envisions or considers ideal may be one where the model iteratively improves and learns on its own—in other words, trains itself. In that context, do you think we can use simulated data, or do you think real-world data is still the highest quality?

If real-world data is still necessary, wouldn't that continue to limit AI's intelligence to humanity's past...?

Because, in this respect, it relies on real data that humans actually had in the past? Or can we break through that ceiling using mock data, simulated data, generated data, and so on, allowing the model's capabilities to surpass all of humanity's actual past...?

I think it can go beyond that. I think it can do so in two respects. For example, in Go, AlphaGo played moves that humans had never seen before. So within a certain domain, it clearly surpasses humans.

But that too may have a ceiling, a limit. We just can't see that limit right now. So broadly speaking, we believe it can build on the knowledge humans already possess—the knowledge we can already express in words—and go beyond it.

Then will that depend on real-world data or simulated data going forward? Will it work?

If it doesn't work, there are plenty of other approaches.

All right, thank you. I've already taken up quite a bit of time, but I'd also like to follow up on my earlier question about AI infrastructure.

What was the second question again? I'm a little...

All right. I'll briefly restate it. I'd like to ask about AI infrastructure. Looking ahead—right now, everyone is investing in compute in increments of $100 billion. In this area, we in China can be confident that we will eventually achieve our hardware goals and have highly efficient compute someday, but as we work toward that, hardware may still be a constraint for now.

Then in the future, will both sides drive those requirements down? On one hand, as model capabilities improve through successive iterations, we'll shift from using compute through brute force to using it intelligently, gradually reducing the marginal compute requirements and consumption per model or task. On the other hand, improvements in hardware capabilities, such as accelerator cards, will speed up iteration and make each card more efficient.

So what happens in that process? If we build a 10,000-card cluster now and buy H200s or 960s, will that compute no longer be particularly high-quality in relative terms after 2–3 years? Will those devices become relatively outdated again?

You can generally assume a 5-year depreciation period for NVIDIA cards. For Huawei cards, it would be 3 years at most. I think the Huawei 950 is quite good to use this year and will still be fine next year, but if you keep using it beyond that, I think it will really consume too much electricity.

Huawei cards will certainly have a somewhat shorter life cycle, because they were already 2 years behind NVIDIA to begin with. But I don't think the gap is that large.

However many B200s you can buy right now, I think they're all worth it. For example, from Tencent's perspective—or if Alibaba can buy them—you'd have to look at the quantities available, and if you can get them at a reasonable price, they're certainly worth it. But this isn't really a question of calculating costs. You probably can't buy them.

All right. It's a fact that we're behind in compute. We offset that in three ways. The first is accepting that our models will lag behind. We can only use smaller models than they can. How much smaller is a training question. We have to accept that our models will lag to some extent and be smaller.

There is one advantage to being behind. It means we have more time, and since we have a certain level of technical expertise, we can find clever ways to do things.

Thank you.

So our gap with the US might be 12 months, or 12–18 months, or 6–12 months. In any case, to put it simply, we're 2 years behind the US, but we're doing this with only one-twentieth of its compute.

The narrative is that we're 1–2 years behind but use only one-twentieth of their compute. We need to change that narrative in the future. I think the goal is to keep using a fraction of their compute while narrowing the time gap further, to 6 months, then 3 months.

And in some respects, we may even surpass them. But with an orders-of-magnitude gap still remaining in total compute, it's unrealistic to surpass them across the board. In certain priority areas, though, where we choose to concentrate our efforts, we may be able to pull ahead.

I see. Thank you, Mr. Liang. You're full of confidence. Let's work toward this together. I'll let others have the floor. Thank you.

Thank you for sharing that. I have two quick technical questions. Earlier, when discussing the technical roadmap, you said the key problem to solve at this stage is continual learning, and that this is also a hot research topic abroad, referred to as recursive improvement. I'd like to ask: what is the biggest technical obstacle right now? And in your view, when might it be solved? That's my first question.

My second question: you also said earlier that you would first solve continual learning and then work on intelligence. I'd like to understand the technical reasoning behind that. Do you mean DeepSeek will also pursue general intelligence after solving continual learning? Those are my two questions.

The technical issues are actually a little difficult to explain… The difficulty is that we haven't yet found a method that works really well.

No one in the world has found a good method yet. Everyone is still feeling their way forward. So we're still in the exploratory stage, and we don't know who will discover the next method that solves this problem.

We're still exploring. We have a lot of ideas, and many look promising at this point, but we haven't fully worked them all out yet. Yes, that's the first point.

Second, there's a way of thinking that we place quite a lot of importance on internally right now. When we train the next version of a model, we want it to help with our own development work. It needs to improve DeepSeek's efficiency. Our model's top priority is to make DeepSeek itself more efficient, so it can provide more help when we develop the next version.

To put it more simply, the first goal of the models we build isn't to work well for everyone. It's to work well for us.

They have to be useful to us first. Once they're useful to us, we can develop the next version faster.

Many people internally think this way. It has to be useful to us first, and we have to use it first. And this is the fastest way to achieve AGI. If it works well for us, that suggests it could work well for other people too, but we first need to make sure it works well for us.

This way of framing it is a little unusual, but many people really do think this way. More than making it work well for users, we want it to be a greater help to us. That's how we can achieve AGI much faster. So the logic behind this is that it helps us achieve AGI.

But first, it has to help us get there. We need that help to achieve AGI. It's very clear now that AI needs to help us achieve AGI. It isn't working autonomously yet, only collaborating with people, but it's already very useful.

On my second question, you said earlier that you would first solve continual learning and then move on to general intelligence. Is that your expectation for how things will unfold? I'd like to understand the technical reasoning behind it. Why solve continual learning first and then move on to general intelligence? I'd like to hear more about your thinking on what comes next.

Because solving continual learning can greatly accelerate our R&D progress. If we solve continual learning first, general intelligence becomes much less of a problem. We'll have AI assistance, and if AI can learn continually, its capabilities will be very powerful.

Agents' capabilities are currently limited because they can't learn continually, or can't do so effectively. If we can get continual learning working first, AI will be very powerful and can enormously improve our own research efficiency.

If we build continual learning first, general intelligence will probably become very easy. Using it will make things much easier. That's why I said this is an outcome we'd like to see. It would reduce the effort required and make things easier for us. Otherwise, humans have to build general intelligence themselves, as they do now, and that's fairly arduous, difficult, data-intensive, and labor-intensive work. It isn't very cost-effective either.

Thank you for sharing.

A quick question—please take a look at the chat for the online questions. How much longer will it take to reach AGI? Can domestic hardware catch up within that time?

It's in the Zoom meeting chat.

All right, I've seen this one.

For the Huawei 950, Huawei is supplying us with 16,000 cards right now, and I think it's okay to say that publicly. That's probably an order of magnitude fewer than the big internet companies are getting. Huawei can only supply us with this many, and they're not cheap, either.

The big internet companies probably have greater demand. They need more. In our case, we can buy some cards that don't comply with the regulations. So our purpose in buying the Huawei 950 is still to help Huawei build out this ecosystem properly.

16,000 Huawei 950 cards are equivalent to just 4,000 B-series cards. So it's not a large amount of compute, and it doesn't make much of a difference. It's not enough to train a next-generation model. We can only train current-generation models with it; it's not enough for the next generation.

But we can at least help Huawei get this right first. That's where things stand with the Huawei 950.

Then how much longer will it take to reach AGI? Can domestic hardware catch up in that time?

When it comes to AI as a business, to the work of building AI, I think China could reach a level comparable to what's available abroad in 1–2 years, or perhaps even get to the point of replacing foreign models this year. AI using the current approach, the current paradigm, isn't all that difficult, so we should be able to do it this year. But that still isn't AGI.

At a minimum, I think it needs to be able to learn continuously.

Can domestic hardware catch up in that time? I think domestic hardware will probably need a few years. China first needs to solve the ecosystem problem, because an ecosystem is a matter of trust. Once we solve the ecosystem problem and then address production capacity, I think we can work through it gradually.

I don't really believe we'll still be held back by production capacity 5 years from now. We definitely are right now, and we may still be constrained this year, next year, and the year after. But I don't think that will necessarily be the case in 5 years. I'm relatively optimistic.

The second question is about our thinking on the future organizational structure and headcount plans. To begin with, our previous organization was very decentralized because we didn't have an organizational structure. But as we add more people, some changes will clearly be necessary.

All I can say here for now is that a lot needs to change, and it's difficult to lay it all out at once. Ultimately, we'll need to split into different departments. Some will need a fairly strict hierarchy, while others will probably maintain a relatively loose, flat structure.

We'll make these adjustments as the team grows. We probably need to make them right away—in fact, I'm already doing so. Without these adjustments, we can't move a lot of things forward. There are actually many departments that need an organizational structure.

Then there's another question: which version of CV is ahead of that? The GCV4 version we've released online is still relatively rough, and there are many aspects of its capabilities that need more time.

Generally, a release cadence I'm comfortable with is roughly one version every 2–3 months. The last release was probably at the end of April, so the next will likely be around the end of June. Barring anything unexpected, each version should be better than the last.

At 50B active parameters, I think we'll ultimately be pretty close to the current wave of open-source models. There probably won't be much difference in inference speed or performance.

But there will be a fairly large gap between us and their larger models—the ones they haven't released. We can't close that gap with our current active parameter count. We'd definitely need a larger model, perhaps 150B, for example.

As for 150B, at our current training pace, this year—optimistically, we could start training at the end of this year, and at least by next year, 4… The gap is still relatively large. Yes, that's the gap with OCE.

Hello. I have a question. You've often said that the process of achieving AGI will be gradual, rather than involving a sudden mutation. Can we take that to mean it's a process without a tipping point?

There's no tipping point, but it is nonlinear. The narrative we tend to believe in right now is that AI can accelerate AI research. AI can accelerate AI research. So it isn't linear. Since AI can be used to accelerate research into itself, progress may become increasingly nonlinear further down the line.

I see. So, if I understand your earlier conclusion correctly, are you saying that continuing to scale language models is enough—that it's sufficient to reach that state?

On scaling language models, all I can say is that I haven't seen a ceiling yet. Neither at our current level of intelligence nor at the level reached in the US have we seen a ceiling.

I see. One thing I was very curious about was your point that in the US, they can train a model with 800B active parameters but can't really put it to use, so even after training it, it's difficult to make it available to everyone. Because in practice, it's a bit expensive.

Something I've long been very curious about is that humans only acquired language in the last 100,000 years, after perhaps 3.7 billion years of evolution before that. But you've said we can reverse that order when training AI. Even so, ultimately we'll have to move into what you might call—not necessarily world models, but physical models or embodiment, right? I mean, beyond that ceiling.

Yes, I think we definitely need to move into embodiment, and that embodiment is ultimately where this leads. For our company, the natural endpoint will probably be embodiment across the board. What ordinary people need isn't a computer, right? Ordinary people eat, drink, have fun, and need clothing, food, housing, and transportation. It's not computers they need.

What they need is—so we still need embodied intelligence to address concrete labor needs. If the goal is to reduce the need for human labor, I think embodiment is unavoidable.

I see. Then, speaking in terms of stages, perhaps we don't need to call it a tipping point, but at the point where it can evolve on its own and do so fairly well, or gets close to this upswing with SV, I think AI in that state would first be deployed in the real world…

It will probably be different from now. We want it to—well, assuming there's no embodiment, how do we define AGI, or what do we want AGI to do? Help us iterate on the next version of the model, and then help us iterate on the version after that.

And once it has a physical embodiment, we want it to iterate on the next version of that embodiment and build the next version of the robot.

There's still one thing I'm curious about. In earlier DeepSeek interviews and elsewhere, they've said that judgment and intuition matter when making important choices about direction and research—that it's not just engineering optimization. So once AI can evolve on its own, will that kind of judgment, taste, and intuition still matter? Or what will become important?

AI doesn't lack discernment or intuition right now. What it lacks is the ability to learn continuously. There's nothing wrong with AI's discernment or intuition. Ask it to write something, and as far as I can see, there's no problem with either.

There were a few more questions earlier. Let me take a look at those. Some questions have come up on the screen, but I can't see them from here.

Mr. Liang, I posted a question on the screen, but let me read it out again. Continuous learning, as you and many researchers have said, is still an unsolved research problem. And catching up with that coding agent, particularly MILES, I mean…

Scaling to the Office level or the MIS level is a relatively well-defined goal. How do you think research resources, particularly research talent, should be allocated between unresolved research goals and relatively well-defined scaling goals to achieve the best balance and results?

Model Office is a relatively certain goal, but it's still hard to say that Model MIS is certain. We can only call it a goal. I think Model Office is relatively certain.

CoT doesn't take up resources. For the kind of research my friend Im does, you don't burn through GPUs—you need very few. What you need is ideas. It doesn't tie up talent either, because no one needs to work on it full-time. It's not a project; it's a problem that lots of people need to think about.

So there's no need to allocate resources to it. You need resources when you're training, building, and releasing models, or running experiments on model efficiency. What I was just talking about doesn't consume much in the way of either people or GPUs.

That's why we call it scratching lottery tickets. The barrier to entry is so low that anyone can have a go, but who will uncover what—I don't know whether that comes down to talent or something else.

So we don't need to allocate resources to this. What sets us apart from other companies is that we spend time discussing the problem, thinking about it, and treating it as something important.

Within the company, this is an important problem, one we should spend time thinking about, but we don't need to devote a lot of resources to it.

Let me look at another question below. Hallucinations in large models have a fairly significant impact on the user experience. There are ways to address hallucinations, but it's a long-term task. Better post-training can address them, so this is a problem that can be solved or improved.

It's just that nobody has put a huge amount of effort into it. To me, hallucinations are a problem, but ultimately they're probably a product issue. We'll address them, but they're not a central priority.

There was also a question earlier about data labeling. That has to do with our capital investment. Given the structure of our investment, we can't afford that much high-quality data labeling. It's extremely expensive.

There's not much difference between data labeling costs in the US and China. Labeling data in China doesn't give you a cost advantage, especially when it comes to high-end data. So it's hard to invest in data labeling the way they do in the US. This is a very difficult path in China. Data labeling is just so expensive that it's painful whether you outsource it or do it yourself.

So we're basically taking a two-pronged approach now. It's not that we can't do any labeling at all. Some data labeling is inexpensive and some is expensive, so we start with the lower-cost work.

You could say that half the people at our company are labeling data right now. Half of our core researchers, our most important people, are labeling data. We're focused on data labeling. At this stage, solving AI problems comes down to data labeling. When you look at it, it's all about the data.

Thank you, Mr. Liang. You've offered a very distinctive perspective that really resonated with us. That was wonderful. Thank you. I'd like to ask a few questions.

Here's my first question. Earlier, you said Chinese models are clearly stronger than American models in efficiency. You also said we could surpass the US in a few other areas in the future. Which areas do you mean? Could we surpass the US in intelligence or in some other respect?

I think we could do better than the US in many aspects of the experience. I'm not saying that our own experience, the experience of using our own products, is particularly good. I'm saying that in terms of user experience, we won't necessarily be behind the US.

In products and product capabilities, too, we won't necessarily be behind the US. Our costs will also be lower, so China can still be competitive.

Do we have structural advantages in other areas? Probably not. But I think we have some structural advantages in those two areas: cost and products.

Cost is easy to understand. They don't need to work on it in the first place, so they don't develop that capability. They clearly don't attach as much importance to it as we do. We may see it as very important, but it isn't important to them.

The same goes for products. Product capabilities that have developed organically across many companies are still quite strong. So I think there may be structural advantages in these two areas.

Great. Here's my second question. Earlier, you said post-training is relatively costly for us, and that Anthropic and OpenAI both invest enormous sums in it. Do you plan to increase investment in post-training after this funding round?

The gap is mainly in high-quality data labeling, primarily within AI research. We'll certainly increase investment, but something like high-quality data labeling is typically not a matter of capital investment.

I think the bottleneck in high-quality data labeling is time. It takes time. That's because OpenAI and other companies abroad, including Anthropic, all started earlier and have more capital and more GPUs.

In this context, you could say that companies here in China only got started in the past half year, so I think we need more time. This doesn't have much to do with capital investment. Even without additional investment, the capital we already have is enough to expand as fast as possible.

But there's a limit to that speed. The bottleneck is that you can't instantly get more people—not money or GPUs. Even so, we are in fact expanding rapidly.

So I think we can reasonably expect companies in China to handle the high-quality data problem fairly well within 1 year. The outlook may not be that bleak, but it does genuinely take time.

Thank you. And here's my third question. We've seen Anthropic build its own products using its own models and roll out offerings in many verticals, such as finance and law…

They're even looking to move into healthcare in the future. At what stage would we also consider these kinds of vertical applications?

I haven't yet worked out clearly what the business model in China will look like in the future, or what the smoothest path might be. We haven't reached that stage yet. Conditions in China and overseas aren't necessarily the same, and it's relatively difficult to judge now what the domestic market will look like then.

Given the current situation in China, I think the most sensible approach is to go all in on general-purpose agents. Other agents, including those for finance or medicine, should be lower priorities. We need to focus on coding first. Coding agents can do a lot, and there are also many vertical agents. At this stage, I still think coding agents are the most important.

That was very clear. Thank you. I'd like to ask one more thing. In fact, through our own work with DeepSeek, we've come to greatly respect you. You've consistently pursued DeepSeek with a very pure focus on research.

But this industry has now entered the realm of capital markets and monetization. You're someone who takes your responsibilities to both colleagues and investors very seriously. So how will you balance pure research—a singular focus on pursuing AGI—with the capital markets? In the future, you'll surely need to turn to the capital markets, and you'll surely have public shareholders. How will you strike that balance? How do you think about this?

Right now, I think we can do both. For example, if we generate several hundred million dollars in enterprise revenue this year and also have consumer users, that in itself already gives us a certain business foundation. If we have enterprise revenue next year and demand can grow further, the company won't be far from turning a net profit. It might even already be profitable.

Then we may no longer be at the stage of simply burning money, which would give us considerable room to maneuver going forward. Or, in the worst case, selling API access alone could sustain a publicly listed company. If there were no further technological advances and the technology stopped here, I think it would ultimately be enough to focus all our efforts on selling API access and providing good service.

So I'm still confident, and it really isn't that difficult. This is a field with tremendous leverage that's developing very rapidly, so it may not be that hard. All I can say is that while we're aiming for a bigger dream, we also have business results that can provide a floor beneath us.

Thank you. I appreciate your answers. That's all my questions. Thank you.

Thank you for sharing. I'd like to revisit a few things you mentioned earlier. I think the biggest difference between DeepSeek and other companies is really that our organization is different from others, or that our organizational structure is different.

But organizational structure is also tied to goals, and we may need to consider the organization's own boundaries and efficiency. From a broader perspective, is there a good model we could learn from? Would it be a historical example like Bell Labs? What kind of structure would be relatively ideal? Or should we accept that there's no ideal model and gradually find our own way?

In a new era, we'll probably have no choice but to find our own way. Our organizational structure is clearly different from those of the companies in the US, right? Those three companies also differ from one another, but at least they'll be exploring from the perspective of commercial companies.

What I'd mainly like to revisit is the question of organization.

First of all, we don't have a model to imitate. At every stage, we start from the actual situation, ground ourselves in the facts, and make decisions based on the circumstances to figure out what we should do. So it's a product of the times, or a reflection of real-world circumstances, not the result of imitation.

In other words, this may actually be the optimal solution under these circumstances, or at least it's the path we've chosen as I see it. At every stage, we've thought things through and made choices, and this is the result of those choices. We didn't choose this because we saw someone else doing it. We analyzed the pros and cons and made our choice.

It will probably be the same in the future. We don't imitate anyone. I think we're different from Bell Labs, too, because they clearly didn't need to commercialize their work. Because that was… But we clearly do need to commercialize ours. Ultimately, we have to survive. We're a company, after all, and the government doesn't give us a penny.

So we may have a very ambitious mission, but we're ultimately a company, and we have to think about how to sustain ourselves. The enterprise business is therefore clearly important to us, because we may have to depend on it to survive in the future. It's just not important right now, because at present it's a cost center.

So we're still different, including from Bell Labs. At our core, we're still a company.

Throughout history, many companies have pursued things beyond profit. But having ambitions beyond profit doesn't mean you're not a company.

Many companies became great precisely because they pursued something beyond profit. Ultimately, those ambitions didn't hinder commercialization; they actually helped those companies do it better. We're fundamentally a company. We simply make choices and trade-offs as we think about what kind of money to make, when to make it, how much to make, and what we'll rely on to make it.

Mr. Liang, let me quickly ask two brief questions. First, you said earlier that MILOS might not be a very definite target, but that you'd certainly move in that direction. And the active parameters, for example…

You said the next generation might be around 150–250B. In that case, would 150–250B be targeting O4.7, or could it be targeting something else? That's my first question.

Here's my second question. Earlier, you said that parts of the overall inference pipeline now use compiled languages other than CUDA. Previously, we were based on the NVIDIA ecosystem, so we may have used things like PTX extensively. If we now use more of something like TileLang, which you mentioned earlier, would inference efficiency drop significantly, or decline in the short term?

I'm curious how you view the efficiency losses that come with this change in compiled languages, or whether those losses can be offset over the long term and efficiency improved.

It improves efficiency. Significantly.

OK, so there aren't any negative effects?

That's right, it improves efficiency significantly. So this is an opportunity. Before, we couldn't leave the CUDA ecosystem, but now we can move away from it and use a simpler approach: TileLang. It's a high-level language, so writing programs is very fast, and there's far less code to write. That means I can rewrite it myself.

I see. So these two things are actually fairly significant opportunities brought about by AI, as you described, rather than drawbacks that need to be addressed in the short term?

Yes, it's a major opportunity arising from technological progress. It's not an AI opportunity. We also have a project to write TileLang using AI.

Wouldn't that make it even faster?

Right now, all our TileLang code is written by people, but even so, it's already much faster than writing CUDA was.

I see. So do you think there's no impact on low-level execution efficiency on the hardware either?

I think a 1–2% loss is acceptable.

OK, I see. Understood. Great, thank you. That's very clear.

Does anyone else have any questions?

If there are no more questions, we'll wrap up here for today.

All right, thank you.

Thank you all so much for taking the time. Thank you, Mr. Liang. Thank you.

Goodbye.

Goodbye.