Liang Wenfeng Investor Meeting · Edited Transcript

Transcript preparation

Recording: 5/20 · edited 2026-07-16

Audio file “deepseek 0520.m4a”, approximately 3 hours 44 minutes in total. This transcript was automatically transcribed and then organized with AI; speakers are not distinguished, and square brackets indicate positions in the audio. Some proper nouns and numbers may have recognition errors; refer to the original recording.

[00:00:01]

When the other colleagues and I first started this company, we never began with the thought of how much money I would ultimately make, whether we would enter the capital markets, go public, or anything like that. That was not our original intention.

The first few dozen people never thought that way at all. If they had, they would not have joined. So overall, we approached this with tremendous goodwill toward the world. We believed it would be useful to humanity, something beyond money.

Of course, later, after we realized the opportunity could be very profitable, other temptations appeared. That is a separate matter. But our starting intention, our vision, and the vision we have maintained to this day were not built around maximizing commercial profit. I think that point is crucial.

About twenty years ago, the management thinker I admired most was Jack Welch, the former CEO of GE. Looking back now, much of what he said may no longer be right, but he got one of the most important things right: a company’s most important asset is its vision.

You manage a large company through vision, not through rules and procedures. What is vision? It is not a slogan hanging on a wall. Vision is how you act, not what you say; it is how you actually operate. I have forgotten Jack Welch’s exact words, but that was roughly the idea.

So how do we manage and organize so many people? In fact, we do not have an organization in the conventional sense. We are organized by a vision. We do not really have an organization.

That has advantages and disadvantages. In the future we will try to build on the strengths and avoid the weaknesses, but this is our distinctive character. We are not operating through “what KPI must I achieve?” or formal evaluations. We have only the vision.

The vision is not even written down. Nothing has ever been formally written. It lives in how we do things and in our attitude toward the world. Everyone in the company may understand it differently, and each person’s personal vision may differ, but we are aligned on the broad direction.

I think it is still about holding great goodwill toward the world and wanting to do something worthwhile. That is what organizes us.

I will speak first, and everyone can ask questions afterward. I may explain what follows through the lens of this vision. The vision is real; it is not invented. We truly think this way and act this way. Otherwise, many of our decisions would be impossible to explain.

Why are we so committed to open source? Because the vision itself requires open source. Without that vision, you could not organize people.

For example, Zhipu also open-sources models, but its open source is different from ours. There is a sense that Zhipu is compelled to do it and that it was not its original intention. For us, this is precisely our intention.

From the beginning, we thought very clearly about open source. First comes the vision. Second, we believe that if AI is to succeed commercially, open source is beneficial.

That sounds contradictory and counterintuitive, because historically open source and commercialization have conflicted. But I think AI is different. Historically, a software company’s market might be worth only a few billion dollars a year. If it open-sourced its software, the business might disappear, leaving only tens or hundreds of millions.

But AI is large enough that it may eventually account for, say, ten percent of human society’s GDP. That is an enormous number. No one can monopolize it alone. You have to share it with others, or you will not survive.

This is different from open-sourcing a piece of software in the past, because that software market was not as large. AI is simply too large. If we try to monopolize the benefit, history will certainly abandon us. I think this is primarily an objective law, a view of history.

It is not that refusing to open-source would let me monopolize the market. That does not fit objective reality even in theory. You would encounter enormous resistance, and other forces would find ways to prevent you from reaching that goal.

In that situation, I do not think we should necessarily follow traditional commercial thinking. You need a mechanism that ensures the benefit you can obtain is limited; only then might you succeed. You need restraint. I think restraint is necessary.

If we want to make AI succeed in our hands, the first requirement is restraint. We cannot think that some percentage of all human GDP, or some percentage of China’s GDP, should belong to us. The more you think that way, the less likely you are to succeed.

So from the beginning we felt that restraint was necessary. The more restrained you are, the more likely you are to accomplish something like this. This is a commercial consideration, though of course it is also a macro-level consideration.

I think this is intuitive—at least it is intuitive to me, or at least it is what I genuinely think. We do not have many other advantages. We are not especially capable, we are not richer than others, and our people are not simply better than those at other companies. None of that is true.

Think about it: when we founded the company two years ago, we did not have much money, many GPUs, much name recognition, or much influence. We were simply a group of very ordinary people.

[00:11:49]

I really am part of a group of ordinary people. If you like the story of ordinary people doing extraordinary things rather than geniuses doing extraordinary things, that has a lot to do with our restraint. It comes from the same source as our restraint and our vision.

So, does open source conflict with commercialization? I think that in AI, if you lack restraint, you cannot get started. Open source is part of restraint. Our restraint is expressed not only in open source but in many other ways. Overall, we do not need to keep debating open source or restraint.

The more restrained you are, the easier it may be to succeed—or at least that is what the evidence has shown so far, and what makes sense so far. Otherwise there is no way to explain how we succeeded: we had no special weapons, a very low starting point, very few resources, and a randomly assembled group of ordinary people. I myself am simply a university graduate, not from the very top school.

This restraint is also part of our vision. AI is enormous, and so are its benefits. We are very restrained, but if we can make it work, the eventual benefit will still be enormous. Take only a small share and it is already more than enough. So there is no need right now to think about which part of the benefit to take or how to take it. The benefit is large enough.

You only need a tiny share for it to be sufficient. That is why we have said that we take only a reasonable profit and care about your willingness to use it, rather than the size of the profit. That is different. This is not exactly how we set API prices. For API pricing, we consider a reasonable profit to mean buying a batch of equipment on the market and recovering its cost in ten months. We consider that reasonable.

Under current conditions, considering risk and upfront investment, a server might be amortized over three or five years in accounting. Commercially, however, we think recovering the cost in about ten months is enough. That is the logic behind our current API pricing. Our V3.2 Flash and the other models all aim to recover equipment costs in ten months.

That is our standard. It is not profit maximization. If we wanted to maximize profit, we should set a higher price. In this price range demand is inelastic: even if we doubled the price, token consumption would not change much.

If I doubled the price, total revenue would approach double. Hold on, let me check something. Ah, that is great.

Let me tell you a story about DDCP, one of our models. At first we worried that demand would be too high, so we set the price relatively high. The team was not happy. Later I lowered the price to one quarter, and everyone was very happy.

I think that reflects our real thinking. As I said about the vision, we want this to be useful to people, rather than make the maximum amount of money. We want everyone to be able to use it while we earn a reasonable profit. I think this was also what other people at the company felt: when we lowered the price, many people cheered in the company group chat. Everyone was happy.

That is the purpose of spending so much effort and care to make the model well. The purpose is for it to be very cheap and very effective, so everyone can use it fully. That makes us happy. It is our motivation, our vision, and the consensus that allows the company to come together and do this work.

This is probably unusual. For our competitors, a price cut is certainly not good news, so they would not be cheering. If you cut your price in half, your revenue and ARR fall by half. Yes, that is one way we differ.

We think this is enough. Internally, recovering costs in ten months makes me very satisfied commercially. Externally, we think users are also happy to see this price. It is a win-win for the company and for society, for everyone.

I think that is okay. Someone just wrote on the screen that ten-month payback is too profitable. There is indeed room to lower prices further. There is definitely room. There is also room to optimize the models, so overall there is considerable room for price reductions.

But we can achieve ten-month payback and others cannot. Alibaba or Tencent may not have our optimizations, and their costs may be several times higher. There is still a lot of optimization work here.

Someone just asked why we do not keep lowering prices. The reason is that demand is inelastic. If we lower the price further, demand will not increase, or will increase only slightly. Everyone can already afford the current price and is satisfied; people are not avoiding it because it is too expensive.

So a price cut would not first bring the company more revenue, and it would not create more value for society either, because people are already satisfied with the price. Making it even cheaper would not add much to society’s sense of well-being. Yes, okay—but on the question of pricing, we certainly are not starting from the goal of maximizing company revenue or profit.

[00:25:32]

The highest possible revenue or profit is not our starting point. This is part of our restraint. In the short term, a higher price may bring more revenue; in the long term, it is genuinely hard to say. I think restraint is a strategy.

For me, restraint is a strategy. Sometimes you can give up something in exchange for more of something else. Open sourcing is the same: it can be seen as pressure on us, or as a concession we make.

First, internally, this concession makes us happy. Employees feel a sense of achievement, and it gives us cohesion. It also benefits society: society is happy, and other practitioners and ordinary people are happy. So in the long run, I understand this restraint as something that increases the probability that we can build AGI.

When considering a decision, I have no doubt that AGI will have enormous commercial value. On that basis, my priority is not how to add a little more share or take a little more share. My priority is how to increase the probability that we can succeed.

This restraint may appear in many other areas too. For example, last Spring Festival our user numbers suddenly surged, but we did not try to retain those users, monetize them, or seize the commercial benefits they represented.

We did not try to take users away from anyone or make money from them, but we worked very hard to find ways to serve them well. We never had the idea that we should become the next super app, compete with someone, or become the next ByteDance or Tencent. We had no such thoughts at all.

We could have done that, but we chose not to. My understanding is that this is also part of restraint. You should not think you have to earn money from everything. Once you have users, you may feel you can become the next ByteDance and then swallow up that opportunity.

I think this is commercially viable and possible. If we had spent a large amount of money last year competing with ByteDance for users, that would have been one possible strategy. But we chose a very restrained approach: I will not fight you over this, because there may be a watermelon behind us while what is in front may all be sesame seeds.

I should not grab every sesame seed. Of course, this sesame seed may be fairly large, but I think the AI watermelon behind us may make everything in front seem small. Looking at it now, it may have been right that we did not push hard on the consumer side last year. There really is a bigger watermelon behind us; what was in front really was only a few small sesame seeds.

If I had had a lot of money last year and made this business extremely large, what would I have gained? I would not have gained anything. These are my real thoughts, because I think the AGI opportunity ahead should be enormous. The AGI opportunity ahead will always be enormous.

I do not even need to think about whether I will hold a position in it later, or what my business model will be there. We simply do not need to consider that. As long as the business opportunity is that large, you will certainly find a way. We will pick up the sesame seeds in front of us too, but we will just pick up a few casually. We will not stop, or treat them as important work.

So the consumer-side daily active users last year were probably a small matter. But we picked them up too, and maintained their use at relatively low cost, because they may be useful later. We do not know what these users are useful for now; at present they are purely a cost, but they may become useful in the future.

If something can be picked up casually, we will pick it up along the way. Looking at this year, there may well be an opportunity for ARR revenue in the API or AI business. If this demand can continue to expand, and if it keeps expanding so that we can buy more GPUs, reaching several hundred million dollars in ARR is quite possible.

If AI can generate one billion dollars, my company’s cash flow may basically turn positive, covering our R&D costs and all our other expenses. So this is also possible, but we have not treated it as a priority.

We will do this, but I think it is an important matter. It is not our first priority, nor is it what we truly care about today. The bigger opportunity should still be ahead. The opportunities in front of us include last year’s consumer business and this year’s business market. We need to do them and do them well, but they are not our goal.

Or rather, most people at our company do not think this is an especially important matter. They do not think it is a problem of equal importance to AGI.

Let me say a little more about open source, because many earlier questions were about it. First, I think we will open source, and our strongest models may also be open sourced. I cannot see any benefit to keeping them closed source; I cannot see any inevitable benefit. ByteDance keeps its models closed source—what benefit does that bring? I cannot see one.

Even if you open source a model and tell everyone everything, the barrier remains very high. It is still very difficult for others to get it running. Second, even once they get it running, making the cost very low is also extremely difficult. It is not easy at all.

It is not the case that once I open source something, others can easily achieve the same deployment cost as I do. There is still a great deal of work involved. Although the underlying principles are understood, not every company is willing, or has the will and ability, to organize the people needed to reach that goal.

I am quite used to this too. A company may simply not be good at doing this, because the obstacles are too great. Controlling the cost is difficult, and there are many managerial and physical constraints. This is also an advantage for a startup: if the startup is too small, it lacks the strength to do this; if it is a large company, it is difficult to organize.

[00:38:36]

Every part of this has its difficulties, so this is the sweet spot for a company of our size. If we were larger, we might have no other problems; if we were smaller, our friends would lack the necessary strength.

As for open source, I think we should set a price. At present, we should recognize that it should not put pressure on people. As for the pricing model, I would not charge a very high fee. I might also price it on the basis of recovering the cost in ten months. With a ten-month payback, we would already be able to make independent third-party deployment unprofitable. Third parties cannot do it; they cannot reach this cost, so they certainly cannot do it.

I would use a ten-month payback, which would make independent third-party deployment unprofitable. Third parties cannot do it; they cannot reach this cost, so they certainly cannot.

So open source will not affect my revenue. Of course, if I wanted to make a hundred times the profit, then open source would...

I heard what you said, but the video seems to have dropped, boss.

It may have been because a phone call just came in.

As for open source, I think it has no impact at all on our business model. The premise is that we earn only six times our profit: recovering the cost in ten months corresponds roughly to a sixfold profit. If we earn only six times our profit, open source will not have any impact.

But if you want to earn a hundred times the profit, then open source will indeed affect your ability to earn a hundred times the profit, because a third party can deploy it at perhaps twenty times the cost, which is still lower than yours.

Can this model continue over the long term? I think it can. Under our vision, I think open source can continue over the long term, and that is what we plan to do. You can see it as restraint, and also as something that gives you greater long-term returns.

This strategy gives us more opportunities technologically, and increases the probability that we can achieve AGI. It gives us more room to act.

Think about it: we do not need to work overtime at all, because it is not that difficult. For other companies, it may be very difficult, because you are trying to think about too many things.

Actually, it is not difficult at all. It is really not difficult. At the beginning, from the outside it may look as if we chose a very difficult model: we have to do research and tackle the hardest problems, as though it were a hard mode. But we gave up many things elsewhere, which means we still have considerable strength and can work very easily.

So my actual judgment about open source is that it is sustainable. Open source and commercial payment do not conflict, provided we are operating at a sixfold profit.

A sixfold profit looks very high, but it is not actually high. Given how efficient AI is now, a reasonable profit may simply be about this much. In the future it may fall to, say, four times or three times. I think that is already... it cannot go much lower than that. But there will still be a substantial profit.

Just look at selling APIs. I do not think selling APIs is particularly attractive, but even on this point you can say that there is no conflict. No, I do not see any conflict.

I am also not worried at all that others will deploy our models and compete with us. We actually hope they can deploy them.

We will do our best to help the open-source community and assist everyone in deploying our models. I am not worried that they will compete with me for this business, because the market is large enough. My only concern is that they will fail to deploy it, get some details wrong and achieve poorer results, or end up with higher costs.

Right, there is no conflict here.

Last year, when I was asked about the business-to-business business, a common question was: if we open source our consumer-side model, will that conflict with our own consumer business? We do not have a traffic advantage. Tencent, for example, has a great deal of traffic; if it deploys our open-source model, will it take all the consumer users away, including ours?

But in fact it will not. There are many reasons for this.

The question is whether the open-source model we provide is the same as the model we deploy ourselves. It is the same. We will not open source an inferior model while deploying a better one ourselves. It is the same model.

This also shows that there is no conflict. Throughout last year, on the consumer side, I was basically open source, and we did not see a conflict in the consumer service. We really did not see one. That is the open-source part.

Next is the company’s long-term vision. I think our goal should be AGI. Everyone may define AI differently, but that does not prevent us from making AGI our goal.

From a technical-roadmap perspective, the AGI roadmap is actually quite clear. With the current generation of AI technology, if you can describe a problem very clearly and give it complete context and instructions, it has already surpassed humans. But there is a definition and a prerequisite: you must give it complete context and complete instructions.

That definition is difficult to achieve. For example, when we meet today, we actually have a long and rich context in front of us. Everyone may have decades of context, and AI does not have that. What AI can do now is perform better than people within a limited context.

But it still cannot replace human beings. One thing is still missing: continual learning. People can learn continually. You hire an employee, and they may spend two months becoming familiar with the company’s environment and their work. After those two months, they can get started.

They can do many things. They can understand what you say. For example, if you say, “Bring Wang over,” they know who Wang is. AI does not have the two months of learning that came before.

If you tell AI, “Bring Wang over,” you must tell it who Wang is, what position Wang holds, where Wang is, how to find Wang, and what to watch out for when looking. You have to give AI all the context. AI can work in that situation, but you cannot give it all the context, and it is not realistic.

So AI cannot replace your employees. But if AI had the ability to learn continually, it could spend two months learning at the company just like your employee, and then...

[00:51:02]

Then it could replace people everywhere. So the next step we are missing is learning how to learn.

We can understand AI development as a staircase. The step we took last year was CoT, or chain of thought. We discovered that reasoning through a chain of thought can bring intelligence to a higher level. By thinking for itself, AI can raise its ceiling and do more things.

Then we crossed another step. This year’s step is Agent, because we found that with an Agent approach, it can do many more things and its range of capabilities becomes broader. Its upper limit of intelligence becomes higher.

Why a staircase? Because each later step is built on the foundation of the previous one. Agents use CoT, and CoT also uses the earlier step. The earlier step is the language model, so no step has been wasted.

So the direction of AI development and intelligence is traceable. This year’s step is called Agent, but the Agent step will also be completed. Once it has solved every problem it can possibly solve, it still may not be able to replace your employees, but it will have reached the ceiling of its capabilities.

It is like CoT. Once CoT reaches its ceiling, it has already surpassed the very best humans at math olympiad problems and programming. But it still stops there; its technology has not reached AGI.

So you can see that the direction of AI intelligence is traceable. After Agent, we think the problem to solve should be continual learning: how to let a model keep learning over time, rather than giving it very strong training once and expecting that to be enough. It should be able to learn continuously for a relatively long time, like a person.

This problem is related to completing tasks and so on; they are related and solve the same underlying problem. We are currently standing at the Agent stage, and the next bottleneck we can see is continual learning. The next problem to solve is how to learn continuously. This is visible and relatively clear.

You are stuck at the obstacle in front of you: you have to cross it, and there must be a way to cross it, but it will take time. After continual learning, we may arrive at a singularity. That singularity is when the model can keep learning and can already do everything humans can do.

It will then be able to develop its own versions, conduct its own research, and develop its next version and earlier-generation AI models. It will reach a singularity in which it can iterate on itself.

But this singularity is not really a singularity; it is also a gradual process. This process may be a fairly long transition, not an abrupt change. But habitually, we all think it may be a singularity.

A long time ago, prophets believed there would be a singularity here, but in fact it is not a singularity; it is a continuous process. After this step is complete, I think we will reach embodied intelligence.

This is our speculation, and what we think the timeline should be: first solve learning how to learn, then reach the intelligence singularity, the singularity of self-iteration, and only then reach embodied intelligence. Once embodied intelligence arrives, it enters the real world: it can do housework for you and provide elder care.

We think this is a relatively ideal roadmap, but everyone’s view is different, with no right or wrong answer. We simply think this roadmap is the easiest.

This roadmap is easy because there is very little new work to do at each step. We would not need to work overtime on this roadmap. But if the roadmap were reversed, and embodied intelligence had to come first, doing it ourselves would be exhausting; it would be very hard work. We do not want that roadmap. We want to do things more easily.

If we solve continual learning first, then the self-iterating singularity, and then embodied intelligence, the path will be easy. Later on, you can use earlier technology to help develop later technology. After the singularity, embodied intelligence would not need to be built by people or by us; the model itself could produce it.

So this answers what our long-term goal is. I told him: this is our long-term goal, what we mean by AGI.

Let us return to reality. The most important reality last year was that everyone had to build chatbots and compete for consumer traffic. This year’s reality is that everyone has to compete for To B revenue and take part, because if you do not, you are not even at the table, right?

But we do not think this is important. Within our company, what we truly care about is the AGI roadmap I just described and how to make the next technological breakthrough.

There is something strange, though: the things you want most are precisely the things you cannot get. The things you do not care about as much are instead fairly easy to obtain.

There is a strategic advantage here. We have AGI in mind, and we are building AGI. When we build applications and then build the consumer and To B businesses, we do not need to devote much thought to them. We can do them with very little effort.

I think that working on relatively lower-level technology from a higher technical position creates this kind of dimensionality reduction advantage. At least on the consumer side last year, we saw that clearly. We did not spend much effort on the consumer side; at one point we did not even want to maintain those users, but we could not drive them away.

They really could not be driven away, so they ultimately stayed. But, slightly... This year, To B revenue currently looks fairly promising. I think the figure may also be quite good compared with our peers, I estimate. But we did not devote much effort to it; we simply did not...

[01:02:45]

It was something we did along the way.

When putting internet intelligence into production and moving forward, this AGI step is one I have to take; it is a stage on the path to AGI. So I make all these technologies available to everyone through APIs. I am not doing anything extra.

We are still doing AI; this is a by-product. I only need a few people to maintain the API. We do not even need customer service, sales, or anything else. Users will come on their own.

Or rather, the users considered consumer users, whether on the consumer or To B side, are all by-products of our path toward AGI, intermediate outputs. They do not conflict with my work on AGI. I am not doing this in order to build a consumer or To B business. We are building AGI for the sake of AGI, and this happens to produce something that I can commercialize.

This is different from other companies. Other companies build models in order to do this, to serve consumer users or To B users. For us, that is not the original intention. Our original intention is still to pursue AGI.

I think this is a form of dimensionality reduction advantage to some extent. AGI is a larger vision, and this vision can bring together more excellent people; it has stronger cohesive power. So I have an organizational advantage, and I use that advantage to... It is a kind of dimensionality reduction advantage.

But if you are a commercial company whose vision is to serve consumer users well, that is a different story. You have other advantages in products, user service, and traffic, but you do not have a technological advantage.

The favorable situation now is that model technology is the most important thing. You need to make the model good; everything else... Right, and this explains the path of our development. We really chose AGI, and I never thought we should acquire a large number of users.

When we suddenly became popular last Spring Festival, that was not in our script. We had never thought about it. We simply wanted to make the technology good. But I then realized that, compared with organizations that are fully commercialized and product-oriented, our organization had additional advantages in talent and organization.

It was also quite remarkable. At that time, everyone on the consumer side was fighting desperately for users, but in the end users were led away by someone who was not competing for them. This does show that the logic I described earlier makes sense: there really was an advantage in talent and organization.

This talent advantage does not mean that our people are smarter than theirs. It is about how we organize and motivate these people, and how they collaborate. That is where the advantage lies. Bringing smart people together does not mean they will naturally collaborate or naturally be passionate about pursuing and completing a goal. You need a vision.

Our earlier experience taught me that the AGI vision is very powerful.

OK, that is the question. The next question is: what is the importance of core interests?

I said earlier that we should be very restrained in many areas. But what are our core interests? In fact, there is only one: our greatest core interest is maintaining the stability of the team. This is our greatest core interest, and can even be considered our only core interest.

As long as I can maintain the stability of the team, we will definitely succeed—we will definitely achieve AGI. It is that simple. As long as everyone stays and we can continue working, there is basically no risk. We may encounter setbacks sooner or later; if we encounter setbacks but everyone stays, we can continue again.

Money is certainly not a problem, and resources are not a problem. Other factors are easy to obtain. For us, there is only one core interest and one thing we cannot compromise on: we must maintain the stability of the team.

This is also a very major challenge we face, or what I think is our greatest risk. Of course, this risk was substantially reduced by our recent round of financing, because the options people received were still quite substantial, and the amounts were still fairly large.

In terms of team stability, as long as the most important and longest-serving employees remain stable, the others are unlikely to leave. Even if they have slightly fewer options or slightly lower income, they will not leave. They are not here solely for money; everyone hopes to work in an environment that can achieve AGI.

So it remains attractive to talented people. Historically, our employee turnover has been relatively low; compared with our peers, our talent turnover has always been relatively low. But this remains our greatest challenge, our only challenge, you could say.

Everything else is a matter of time. At most, other issues would make us six months or a year late, but they would not mean we could not achieve it. We certainly do not lack money or resources; in fact, we do not lack any of these things.

So many of the things we do now are intended to maintain the stability of the team. Apart from this, I think we can give up everything else; we can exercise restraint about everything else. We have always been very restrained and unwilling to become an opponent of any large or small internet company.

I hope I can empower them, or assist everyone in doing this, and help everyone do it. This is also part of the commercial meaning we discussed earlier. The premise is that everyone should not concern themselves...

[01:15:12]

Under this premise, we are very willing to assist and help anyone, even our competitors, including Alibaba, Zhipu, and Moonshot AI, to do better. We do not lose anything by doing so. We are open source anyway, and open source means explaining the boundaries as clearly as possible and how to do it, hoping that you can reproduce it. If you cannot reproduce it, tell me and I will tell you how.

That is already part of open source. It will not change just because you are a competitor. Of course, if you are a partner, I will do more. But at the level of major interests, there is no conflict.

When dealing with the outside world, our attitude is that we only pursue the main AGI line. This is what I just mentioned—GPT, CoT, Agent, and so on. We focus only on the main line. The AI field is broad, and there are many things we feel are not on this main line. For example, 3D and video generation may not have much to do with the main line of intelligence, so we will not do them.

There are others, such as world models. I do not think they currently have much to do with the upper limit of intelligence either, so we will not do them. But we are happy to help others who do them. Whether we have the time is a separate matter; there is no conflict of interests.

We also hope these AI technologies can be used in all kinds of production environments, raise society’s productivity, and help every industry improve its productivity. We are highly motivated to do this.

Whether I have the time, whether we have enough people, or whether our colleagues are personally interested is another matter. But there is no conflict of interests; we hope to achieve this goal. We also believe it has no conflict with business: I will still receive the interests I am supposed to receive.

I think that by maintaining this attitude, we have not actually lost anything. We have not received less because we open sourced, because of our goodwill, or because we helped others. For example, we still have many consumer users from last year, and they remain quite stable.

Our To B business this year also looks fairly promising to me. Our goodwill has not harmed my commercial interests at all; it has had no negative impact whatsoever, and may even have helped. This seems counterintuitive, but that is how it is. Or let us think about it the other way: if I acted against this principle, would I inevitably get more? No.

There was a question about how to understand the claim that world models are unrelated to raising the upper limit of AI intelligence. I was talking about the current stage; this is our judgment. We have a view, and this is our own AI roadmap, not the only possible roadmap.

Based on our understanding and judgment, the most important thing right now is to train AI well. Training AI well does not require a world model, and does not even require multimodality. If you narrow the scope of AI training and omit multimodality, there are simply some tasks you cannot do; that does not affect the validity of the algorithm.

Multimodality will ultimately still be necessary. What matters now is training; the next problem is continual learning; after that, the problem is getting it to ask its own questions. There is no world model or video generation in this roadmap.

When video generation first appeared, it became very popular, as if it were something every AI company had to do. If you did not do it, you were somehow not an AI company. That puzzled me, because if you think carefully, it has little to do with the roadmap for intelligence.

In fact, you saw that after Sora first appeared, everyone worked on video generation—large companies and small companies alike. But later the small companies all cut it. It has nothing to do with the upper limit of intelligence.

Commercially, however, it is a good business; it is a good business commercially. But that has nothing to do with intelligence. We will not do something because it is a good business. We will do it only because it is on the roadmap for intelligence.

Video generation is relatively clear, so I use it as an example. World models are less clear, because many things can be called a world model.

Our judgment is that world models and intelligence are not the most important things at this stage. The most important things are AI training and then solving continual learning after AI training. This is our company’s judgment; naturally, each company’s judgment is different.

What I said earlier is that, for our company, the most important issue is the stability of our people. From another perspective, what are we lacking? What is the gap between us and the United States? In fact, there is only one gap: resources.

We do not have that many GPUs; the number of GPUs we have is still relatively small. We currently have roughly twenty thousand H-equivalent GPUs. Most of them have only just arrived in the past month or two, and many machines may still not have arrived.

Our total compute capacity was relatively small last year. This year we are expanding it very aggressively. We currently have roughly twenty thousand H-equivalent GPUs, and in the coming months we will buy a large number of machines, basically all from NVIDIA.

How many GPUs do we need? More is definitely better. Within what we can afford, more GPUs are certainly better; there is no doubt about that. So our strategy now is to buy as many GPUs as we can at a reasonable price.

After this financing is used up, however many GPUs I can buy, I will buy. The speed of spending is not part of the plan; as long as the price is reasonable, I will buy as many as are available. If I spend all the money within six months, I think that is a good thing. If I spend it all within six months, that would be wonderful—ideal.

In reality, spending all that money is very difficult. We cannot buy that many GPUs; they are hard to buy and very expensive. We cannot simply pay an extremely high price—we also have to make sure the price is reasonable.

[01:26:41]

If I could spend all this money within six months, that might be the ideal outcome. Turning the money into NVIDIA GPUs is certainly better than leaving it in a bank. In the bank, I may currently earn about two percent, apparently. But buying NVIDIA GPUs has a ten-month social cost.

So the more we can buy, the more we will buy. Once we buy the GPUs, we will have plenty of room afterward. By providing services or doing something else, we can always generate cash flow. With cash flow, we can survive; we do not need to keep a great deal of money on our books.

So we are only worried that we will not be able to buy enough GPUs. If we could turn all the money into GPUs, we would do so without hesitation. We are also willing to pay a certain premium to turn the money into GPUs, because it is such a good deal.

Even after paying the premium, it would still be difficult to achieve this goal. Objectively speaking, if I could spend twenty billion this year, that would mean our procurement department had performed exceptionally well.

Our main gap with the United States is resources; the gap in people is not very large. There is almost no gap in people, because they are the same pool of people, and many may be Chinese. When Chinese people go abroad, some stay in China, some stay overseas, and some move overseas; it is not that the smart people all go abroad. That is not the case.

It is fairly random. The smartest people may not be slightly more than half overseas; slightly less than half may remain in China. China does not lack talent, and our base is large, with so many people joining every year.

Talent is not the bottleneck; resources are the greatest bottleneck. Resources first affect talent development. Because we have less compute, we have fewer opportunities to run experiments, so our talent overall lags behind the United States. The talent gap is essentially also a compute gap.

We cannot really afford to train the largest models available today. Even if we spent fifty billion, we still could not train them. Even if we could assemble the hardware, we could not afford to operate it. The largest models today activate roughly eight hundred B; in China, we are still at a scale of dozens of B, with the largest domestic models perhaps activating only dozens of B. That is an order-of-magnitude gap.

If I wanted to train a model as large as the AI model, I would need roughly fifty thousand GB300s, or two hundred thousand Huawei 950 cards. That is only for training and does not include research. So our greatest gap with the United States is resources.

Our current resources, including what we will have this year and over the next few months, including the large resources arriving soon, are only enough for us to run more experiments at the scale of ten B activated parameters. I still have many experiments to run and many things to clarify at the scale of dozens of B activated parameters. We are still quite far from being able to train an eight-hundred-B model; there is a long way to go, and we do not have that many GPUs.

So the difference between us and the United States, in my view, is a difference in resources. We may consider all the differences we see—including differences in talent, model capabilities, and applications—to be differences caused by compute resources.

Compute resources are constrained because GPUs themselves are unavailable domestically, and because our capital investment is lower than that of the United States. We invest much less capital, while the share represented by talent compensation is small. They may offer salaries of one hundred million dollars, but when you calculate it, talent compensation still represents a small share; the bulk is compute.

This problem is basically unsolvable for now, because Huawei’s production capacity is also limited. If I want to train an eight-hundred-B model, I need two hundred thousand of Huawei’s latest cards. That is only for training, without considering research.

So we will not even consider competing with the United States at such a large scale right now. Our current losing position is at the scale we can afford to train and operate, the scale of dozens of B activated parameters; we must first do that well. When we have more resources, we can then take it to one hundred and fifty B, one hundred and fifty-six B, or two hundred and fifty B activated parameters.

So we currently have a gap with the United States that appears difficult to close. You can force yourself to train such a large model, but you cannot conduct sufficient research. Before training it, you cannot conduct sufficient research.

Another question many people care about is where the ultimate gap between all the large-model competitors will appear. Where will the final differences between large models show up? I think there may ultimately not be much difference.

The ultimate differences should be in three areas: cost, time, and user experience. Apart from those, there may not be much difference.

Cost is easy to understand: what cost can you offer the same service, at the same quality? Take BYD’s batteries as an example. Given the same level of technology, can other companies offer them at this price? I think that is difficult. It is not easy to achieve, and it is certainly a barrier.

So cost is definitely a difference. I think cost may be the first difference. The second is time: when can you achieve it? A few months earlier or a few months later makes a difference.

The third may be experience. User experience is still somewhat different. There will be some user stickiness and user barriers, but they may not be fundamental. Fundamentally, it is still cost first and time second: do you make it first or later, and do you make the same thing faster or slower?

Once the long-term commercialization path and product lines become richer, how will we set prices? We think the most worthwhile thing to do and spend effort on is still AGI. Right now we need to push AGI further and further forward—to raise the lower bound of intelligence.

At this stage, that should be more worthwhile, or have a greater return, than building more product lines and considering more commercialization paths. I think this will be true for some time in the future, and may have been true at any time in the past. That is, if we...

[01:39:46]

Have we spent a great deal of time thinking about what benefit it would bring to enrich our product lines?

What was the commercialization plan discussed six months ago? Surely it was to run advertising, build e-commerce, embed e-commerce in the product, and integrate it heavily with local services. It would certainly have been useless, because things change too quickly. If you spend a lot of time considering commercialization and commercialization paths, or product lines, at the stage you were at earlier or the stage we are at now, the product lifecycle is very short.

I do not think the time has come. This is our judgment, or at least our experience so far supports it. Over the past three years, whenever you came to me with commercialization paths and product lines, it was a waste of time, because you could not predict or foresee the future. Very little can be foreseen.

Because I am looking at the time, I do not know whether everyone wants to take a break. Should we eat something first, or continue talking? No one has objected, so I will continue.

We conduct world-leading AGI research and commercialize it at the appropriate time. I think we have always been commercializing; I believe we have always been commercializing, just without making commercialization the goal. AGI is our goal, but we have always been commercializing, which is why we have consumer users and To B revenue.

Historically, this strategy has been successful. I think the time when we completely shift to commercialization is still very far away. The priority remains extending the technology and building the next generation of technology to solve more of today’s problems. In the future that I can see, at any point, focusing on products would still be too early.

So this is also part of our restraint. I hope other people will pursue these commercial opportunities. We hope society as a whole and all our partners will use this AI and share the benefits together, rather than my trying to keep everything for myself. That is impossible. We also do not have the energy or enough people in our organization to do all of this.

As for our partners, this financing was carefully selected. As for the specific proposal for how to work together, I think the first consideration is that our interests are aligned: these are the people whose interests are most aligned with ours, who are least hostile toward us, or who most want us to succeed. Not everyone wants us to succeed, because we have harmed the interests of many others.

The processes and decision-making mechanisms for major company strategy, technology, and business decisions: our company as a whole is built on consensus. I am not saying that I decide everything alone; I seek consensus. My authority and influence within the company are built on consensus.

For example, when I want to do something, I first look at what consensus everyone has and whether people want to do it. I may offer some guidance or have a preference, but the effect of that guidance is limited—very limited. It is still built on consensus. This decision-making mechanism is a mechanism for seeking consensus. It is not that I can push something through; it must have consensus before I can move it forward, and only then will I do so.

At present, most of our energy is basically focused on DeepSeek. Let us take a five-minute break. I will write quickly. You can turn on your microphones and give me some feedback. There is no problem on our side; why don’t we take a five-minute break first?

The final gap between different models in their overall performance should be a comprehensive one. When comparing model performance, you have to compare them at the same cost for the comparison to be meaningful. It is like comparing two cars: you compare cars in the same price range.

The difference between a good model and a poor model should not lie in one particular part; it should be an overall difference.

Is Anthropic overtaking OpenAI now something that will last? I do not think it will last; it is certainly an extreme situation. OpenAI and Google will most likely continue to take turns improving in the future. In fact, Anthropic’s advantage in Code Agent is not that large right now; it is not as if Anthropic is crushing OpenAI.

Perhaps half of the people in our company usually think OpenAI is better. Anthropic does have a first-mover advantage, but that advantage should disappear quickly; it is not an advantage it can hold for the long term. All three companies are very capable, and among the three, its efficiency is the highest: its costs and the money it burns should be the lowest.

As the global AI division of labor takes shape, a role Chinese companies are quite likely to play is to have the largest production capacity. In common-sense terms, our capacity is the largest, including for chips; our chip production may be the largest, and we have the most electricity. So Chinese AI may ultimately become one of the three pillars.

Chinese people will make these products the cheapest, and then improve their performance. After all, for many foreign products today, there is not much difference between products made in China and those made in the United States. AI may be the same in the future, but AI made in China may be cheaper. This cheapness may be a systematic cost advantage, just as China provides cheaper services in other industries.

When I do something, my habitual way of thinking is: what will produce the greatest return at this moment? If I think building products will produce the greatest return now, I build products. If I think achieving AGI first will produce the greatest return now, I work on AGI first. Clearly, I do not think building products now offers the greatest return.

If your issue is adapting to domestic chips, domestic chips currently have a historic opportunity. In the past, adapting to domestic chips faced a difficult problem: the ecosystem was poor. You could buy the cards, but could not use them because they did not have NVIDIA’s ecosystem. That is why NVIDIA’s moat is so strong.

But this is changing. NVIDIA CUDA’s moat is rapidly being dismantled, perhaps for three reasons. First, AI now exists, and once AI exists, it is much easier to build this ecosystem than before, because AI can write code.

[01:56:36]

I can use AI to build the ecosystem and reproduce an ecosystem exactly like NVIDIA’s. First, there is AI; second, there are some new technologies. For example, our company has developed a technology called TileLang, which is a high-level language.

Using this high-level language to write CUDA operators, we can very quickly reproduce NVIDIA’s entire ecosystem. Combined with AI, there does not seem to be any obstacle. It is not finished yet and the work is not complete, but this technical route appears to have no real obstacle.

Another point is that CUDA evolved from gaming cards, so in many places the design and configuration of gaming cards are inherited from one another. CUDA is compatible with gaming cards. In the past, AI computing was a very small field, smaller than the gaming-card market, so this made sense.

But now the market for compute cards is larger than the gaming-card market, so there is no reason for the two still to be coupled. The trend now is that they will no longer be coupled. Dedicated chips, whether from Huawei or NVIDIA itself, will all be dedicated chips in the future, rather than the things we had before.

Against this background, the role of NVIDIA’s original ecosystem will be greatly reduced. For a dedicated chip, there is little connection with CUDA; it is not bound to CUDA. Or rather, when the chip is designed, people have already considered how to build its ecosystem.

It is a little complicated, but anyway, domestic AI-chip substitution now has a historic opportunity. We believe that within the next year we will see something confirmed: the ecosystem for domestic chips has no fundamental problem. People previously thought there was a problem, that they could not be used or were difficult to use, but within the next year I think we can reverse that perception, or facts will reverse it.

There is no problem with the hardware or ecosystem of domestic AI chips. The only problem is insufficient production capacity. Adapting to domestic chips poses no obstacle; NVIDIA cannot stop it. In a normal commercial environment, if I could buy NVIDIA cards, domestic substitution would be relatively difficult. But when NVIDIA cards are unavailable, everyone is forced to use domestic chips.

Against this background, adapting domestic cards poses no obstacle at all. I think domestic cards can build an ecosystem as good as, or even better than, NVIDIA’s, but it will still take time.

We currently cooperate mainly with Huawei. Huawei is doing its own adaptation, while we will participate in this ecosystem and become deeply involved with Huawei. Huawei’s problem is still insufficient capacity. Huawei’s capacity allocated to us is roughly sixteen thousand cards; the large internet companies may get hundreds of thousands. We have more than ten thousand, and I think this ratio is somewhat... But this may simply be all the capacity Huawei has.

So we cannot expect to train the larger model later on Huawei hardware, or train models with hundreds of B activated parameters, perhaps this year. Next year or the year after, there may be an opportunity.

For Huawei-card adaptation, our main task is to make the high-level-language compiler good and make TileLang good. Once TileLang is good, the problem may resolve itself. This is complicated to explain, but we are working on it; it will be much clearer once it is finished.

You can understand it this way: when V3 was trained, it still used NVIDIA cards, but it no longer used NVIDIA’s ecosystem. V3 used NVIDIA cards without using NVIDIA’s ecosystem. Instead, we first wrote a high-level compiler called TileLang, and then used the TileLang ecosystem to complete everything else. We already almost no longer depended on NVIDIA’s ecosystem.

As long as I reproduce this whole system and process on Huawei cards, it will be complete. I think this may be a historic mission: completely reversing the perception that domestic-card ecosystems are poor. The timing and conditions are now basically all in place; only time is missing. Within a year, many people should see this or understand that the problem has been solved, leaving only the production-capacity problem.

I am fairly optimistic about domestic compute. I think NVIDIA is digging its own grave here. Huawei’s supernodes, the Huawei 950 supernodes, can completely substitute for NVIDIA’s GB200 and GB300 in performance and price. The price will certainly be higher, but only to a limited extent. Fifty percent more or one hundred percent more would be fine; even two hundred percent more would be fine.

For example, if it is one hundred percent more expensive, I think it can already be considered an equivalent substitute on price. It can also substitute on tasks: every task GB300 can do, Huawei’s supernode can do, with the same latency and so on. The only cost is that four Huawei cards equal one NVIDIA card, while also being two years behind.

Four Huawei 950 cards for one GB300 can be understood. Being two years behind means being two years behind in time. The Huawei 950 supernode is a product for this year’s Q3 or Q4; NVIDIA GB200 was a product from Q3 two years ago. That is a two-year gap.

NVIDIA may already have a new generation this year in Q3. So our chip gap with the United States will, in my view, no longer exist in the ecosystem, but in chips it is fourfold plus two years.

Another question is whether we will vertically integrate upstream. I hope not. The question is whether we will vertically integrate into upstream applications. We hope we will not need to. I hope other people will do this. I do not want to consume everything; I only want to take one piece, the piece we are best at, or the piece we ourselves consider most central, and then with the user...

[02:08:18]

The part directly related to users may be something many of our industry partners care about more than we do. That should be their significance; I should not be the one interpreting it.

Will we build large clusters ourselves in the future? I think we definitely will; we have always been doing this, and all our clusters are self-built. But whether we will develop chips ourselves depends on how large the returns are—it depends on how large the returns are.

Tesla, cultivation... Suppose you operate a power plant. You do not necessarily need to manufacture the generator, right? Someone else can manufacture the power equipment. If its price is reasonable, why would you make it yourself?

So I hope we do not need to make chips. I hope we can buy chips at a reasonable price, so we do not need to make them. I think this may well be the case. The margins on NVIDIA’s chips recently may not be... although...

We hope to do only one piece. I think AI is a very large field; it does not require me to... I only do one piece. If we focus, and I believe the business opportunity and profit here are already large enough, then if the AI era produces many trillion-dollar companies, I think we will be one of them.

We have always done one small part of it; there is no question that we are already one of those companies. I do not need to... I am not brimming with confidence that we must do everything. I think the most likely outcome is that things will be difficult to change, and if you really want to do something else, you will still have more ideas. That is... this is our attitude.

For example, at least in the To B and To C businesses, based on what we can see now, when it comes to truly wanting to build a To C closed loop, we may actually do it better; when it comes to truly wanting to build a To B closed loop, we may also turn out to do it better. The more you want...

There is more to say about To B. The ceiling for the To B business should still be demand. Against the backdrop of this generation of AGI and AI technology, To B demand should be limited. It will grow quickly, but it is not infinite; ultimately it is constrained by demand, not compute.

Given current technology, how large revenue becomes ultimately still depends on demand. Think about it this way: I can recover my cost in ten months. If there is somewhere to deploy the capacity, I will buy more, but there are not that many things to do. Yes, demand should keep growing, and as technology continues to make breakthroughs, this demand will grow.

We have always been working on multimodal capabilities. They are important for products, and important for consumer products. But for the upper limit of intelligence, multimodality is a component; it is not the main line itself.

But as a component, multimodality is something we will certainly build, and we are already building it. We should release relevant models: our V4 and V4’s subsequent versions will support native multimodality. But for multimodality and intelligence, it is a component; we do not treat it as intelligence itself. Its role in the main line is like search.

Search is also a component, and multimodality can be one too; that is how we understand it. We believe in Scaling: the larger the scale, the better the performance, and the more capabilities it can unlock. What prevents us from Scaling is compute. It is not that we do not want to Scale; we simply do not have enough compute to do so.

We have not reached the upper limit, wherever it is. We are not training models this large because I think this size is enough; it is simply that I happen to have this amount of resources. I calculate what size model I can accept and train based on my resources. That is how the size is determined; it is not that this model is enough.

At present, the returns from models are still very clear. We have not yet had the chance to hit the Scaling wall; we are still far from it. When Silicon Valley says Scaling has reached its limit, that applies to Silicon Valley. For Chinese people, we are still far from that; we have not Scaled anywhere near that level.

Scaling includes data Scaling, model-size Scaling, and training cost. We are still far from exploring the upper limit of Scaling; we simply do not have enough compute. But we will also spare no effort to push toward that upper limit.

After we finish this financing, we will have more compute and may train larger models. We need to buy it within the available time, and the shorter the time, the better. If we could spend all our effort within six months, that would be best; in reality, it cannot be done.

I think the core capability of the next-generation model must be continual learning; only then can it be called a next-generation model. Before that, what we can do is lower the cost, improve the results, and increase the speed. But for a major breakthrough, it should have continual learning.

Another question is why we seem to care so much about the model’s computational efficiency. I have indeed found that not everyone cares much about model efficiency. A commercial company has no incentive to pursue model efficiency, because if the model is a little more efficient...

So it does not have much incentive to pursue very high model efficiency. Several startups will not do this themselves. You do not hear them say that low cost is what they pursue, because that does not serve their interests. If the cost is low, how will you make money? If the cost is low, you cannot charge much.

By contrast, this is part of our vision. Or rather, our vision and our colleagues care about cost because our colleagues are ordinary people; they know that using this costs money. They can empathize: other people have to pay to use it, and if it is a little cheaper, more people can afford it.

So many of our colleagues still hope we can reduce the cost further. But from a commercial perspective, you would not think this way. Commercially speaking...

[02:21:31]

From a cost or commercial perspective, this is not the top priority.

For either a startup or a large company, the cost of this service is not actually high. But we hope it remains lightweight and affordable, especially against the background of scarce compute in China. It should be affordable and usable on domestic cards.

I think low cost is first of all a result. Our model architecture has indeed consistently moved toward lower cost, which is related to our vision. We also have many algorithmic methods that can reduce the cost further.

Another reason to reduce cost is that the lower the cost, the larger a model I can train and afford. With the same limited compute, if my computational efficiency is higher, I can afford a larger model.

A large company may not think this way. For a large company, resources can be added, and the problem can be solved by adding resources. But we prioritize cost efficiency.

The value of data-based models—the scope of data is quite broad—data should be almost half of the model. Why do I think that if I wanted that, or if AI could account for twenty percent of GDP and I wanted to take five percent of it, that would be absolutely impossible? I would certainly be defeated by someone else. If someone else said they only wanted one percent, they would certainly defeat me.

If my goal were to take five percent of AI, or of all humanity’s GDP, the calculation would theoretically still work. Look at OpenAI: its calculation seems to work out, at least in theory. But it has a problem: it would be defeated by someone willing to take only one percent.

Because another person might say, “I can do this well, but I only need one percent of global GDP,” and that person would defeat it. Then another person might appear and say, “I only need zero point one percent,” and defeat the person in front.

From a macro perspective, regardless of where the percentage is taken from, there is no difference; they are all the same. The person who wants more will be defeated by the person who wants less. You do not even have to actually take more: if your vision is to take more, you will be defeated by someone whose vision is to take less.

In fact, no one has received any money; it is only a vision. If your vision is to take more, you lose first and face greater difficulties. That is simply how this world works.

OpenAI initially believed it could truly monopolize the world, but in reality it will encounter many challengers. It will face challenges and therefore will not have things easy. The United States has faced challenges before, and in the future it may face challenges from China, because Chinese people are willing to take less and can therefore provide the service more cheaply.

In China, there will also be people willing to take a little less. But eventually there will be a balance. If you take too little, the company’s commercial logic stops working and it cannot survive. So if you take too little, you cannot survive; if you take too much, you will be defeated by someone who takes less.

So for us, we are not trying to take the most money or set prices to maximize returns. We only earn a reasonable return. That is one explanation.

I believe in this. I am not looking for reasons to justify it, because there is no need to look for reasons. This is how I naturally do things, and there must be a reason for doing it this way. The reason may not be conventional, but I think the company itself is unconventional.

Our company’s management actually has two lines: one from top to bottom and one from bottom to top. From the bottom up, each person decides what they want to do and does it; nobody manages them, and there are no KPIs.

From the top down, there is the formal line. When we collectively need to do something, everyone in the company has to cooperate. For example, when we release V4, there has to be a division of labor and everyone has to do one part. That is top-down; we call this the formal line.

Generally, we hope the formal line does not take up half of an employee’s time. It should not exceed half. The other half is unassigned: they can do whatever they want. This is a space for research, allowing them to explore what they think is important without prior requirements.

As long as the company can support it, or the company’s compute can support it, or they need very little compute, they do not need to coordinate at all. This is our current organizational method.

Some people think we are top-down, and some think we are bottom-up. I think both are right. My standard is that the formal work should ideally not exceed half.

We generally do not work much overtime. There are two reasons. First, research requires a relatively relaxed environment. If you push people too hard, they cannot do research. Since they need to have their own interest and think about these questions in their daily lives, exploration is possible only in a relatively relaxed environment. This is a need of research culture.

Second, we are highly focused. Being highly focused means there are few things we need to do. I do not have that much work, so I do not need to work overtime.

This is consistent with the restraint discussed earlier. Because I exercise restraint, I often decide not to do things. When there are fewer things to do, each person has less work. You can see that many of our products are incomplete, and we have not gone back to fill in the gaps.

This is also part of our culture.

OK, there are many questions, and I have scanned most of them. Does anyone have any other questions? Please ask.

Investors, please feel free to turn on your microphones and speak. I should mention that Liang Wenfeng has shared quite a lot of sensitive information. Please do not pass on any figures or details, including the number of cards. Also, please do not record the screen and share it externally. Thank you very much. If you have questions, please feel free to speak.

[02:34:59]

Brother Yang, could you share more about the timeline for when continual learning might bring a breakthrough? If continual learning is achieved, what other architectural or algorithmic innovations and other key factors would still be needed?

There are few people, and research is needed. The whole world is studying this problem. For investors, what they see most now is AGENT; but for those of us doing research, what we see more now is learning and how to solve the learning problem.

Learning may not actually be a technology; it is a problem. Solving it may involve many technologies, not one technology. It is not one thing but many things. Or rather, AGI is composed of many things: it needs a model and many other things as well.

It is also an engineering and algorithm problem. This is a specialized question, but there are many approaches and much research.

Mr. Liang, thank you. Thank you very much for this opportunity today. I would first like to respond to and express my gratitude for what you said at the beginning. It moved me deeply and gave us a great deal of inspiration. You said that the team carries the greatest goodwill and hopes to advance this industry, society, and the development of human intelligence, making a small contribution.

And it carries this sense of mission and vision. I think this is very close to the corporate culture of the companies we serve: “cultivate oneself and help others.” I deeply understand why you lead the team in open sourcing.

I imagine it as creating a small bird paradise around a banyan tree, benefiting all things without competing. In that way it is accepted by everyone, all things coexist, and ultimately it is everywhere. So through this investment, we also want to express our recognition, support, and respect for this mission and vision.

At the same time, we hope to contribute in the areas we are good at as the industry develops in the future. We would also like to continue asking for your advice and discussing this with you. For example, as the ecosystem is built together in the future, after open sourcing, how many partners, talented people, and teams in the industry can reproduce some of our current open-source models and results well?

As you hope to develop this ecosystem further, in what areas do you think more high-quality people are needed to connect with our models and reproduce them? Or is GPU compute currently somewhat scarce? In the future, will everyone take the form of a model matrix?

For example, as we make the foundation models better and better, partners and teams in different industries can build vertical-industry models or application models within the model matrix. How is this developing now? Step by step over the next two or three years, what kind of ecosystem do you think it will grow into?

That is the first thing I wanted to ask you about. Second, you also shared many observations with partners about AI hardware. Global AI companies may each invest on the scale of hundreds of billions of dollars, while China currently appears to have some weaknesses in hardware and compute.

How long do you think it will take to solve this and support the development of our AI, so that compute and hardware weaknesses do not hold AGI back? Do you think this is something Chinese people are destined to accomplish, and that we can certainly build it, with only the time and capital investment remaining as questions?

At the same time, perhaps there are two sides. On one side, progress in models will raise model intelligence, causing the hardware and compute consumed by a single task or certain individual intelligent systems to gradually decline. They will no longer need such large computations, because progress in models will make them more clever. I do not know whether I understand this correctly.

On the other side, advances in hardware technology will make compute more powerful and energy efficient. Could this be a path where the two sides move toward each other? At this point, if we use the current 960 or H200 to make a hundred-billion-dollar compute investment, you mentioned amortizing it over three years. Do you think its actual lifecycle, given technological iteration, is four or five years?

Or more directly: if a compute center now builds a ten-thousand-card cluster using today’s cards, could it become relatively less advanced compute after three years? Could there be a situation where it is under construction and insufficient now, but three years later becomes relatively lower-quality compute with surplus capacity? I do not know whether such a phenomenon will occur. Those are my two questions. Thank you.

Thank you. The first question concerns the ecosystem.

We now think that perhaps every company faces the problem of insufficient talent. But I think this shortage is temporary. In the early stages of every industry’s development, there is never enough talent.

This was true when websites were first being built: there were few people building websites, and talent was scarce. Later, when the internet needed backend services, talent was also scarce. But these shortages were resolved very quickly, in two or three years, because large numbers of people were trained.

The shortage of AI talent is also temporary, and we have already seen it ease substantially. There is actually no shortage of AI people; every company will quickly train them, and training people is fast. So overall, whether in the ecosystem, model companies, or elsewhere, the AI industry does not lack talent.

A talent shortage is certainly a short-term phenomenon. Historically, there has never been a long-term shortage of a particular type of person. I remember that more than ten years ago people said there was a serious shortage of pilots. Pilot training takes a long time, but that shortage was also resolved quickly.

So there is no need to worry about a lack of talent. Also, there are currently somewhat too many companies building models in China—still too many. The United States may have only three, while China has too many companies building foundation models.

Ultimately, we will not need so many people building foundation models; the field will inevitably converge. Resources are therefore also quite dispersed and, to some extent, wasted. It is just that...

[02:44:00]

Every company has to do the same thing, but in the United States only three companies need to do it, so resources are concentrated in those three. In China, resources are spread out, so each company gets less. I think this will definitely converge. It will, but it takes a process; ultimately it will converge.

We do not need so many companies, because people may currently think the profit margin from doing this is extremely high, so everyone has to do it themselves. But when they discover that the profit margin may not be so high, they may stop. Recently it certainly has not been that high. I do not believe it is that high, because that would not fit the objective rules.

This means we are at a particular stage: if there were an extremely high profit margin, that would certainly not fit the objective rules. We should earn a reasonable profit. That is the current state of the industry, and I think it will definitely converge.

Among the companies building large models, no single company should monopolize everything and say, “I want to take all the profits.” That certainly will not work. If every company takes only a reasonable profit, we do not need so many people building large models. In the end, three or four companies competing in China would be enough; competition would be strong enough, and prices would certainly be low enough for a price war.

There may not need to be two large companies and two small companies building large models; that may already be enough.

As for the ecosystem, I do not have many specific thoughts. We hope to support more people, but we do not have that much energy. We have the intention and there will be no conflict of interest, but whether we actually do it is another matter.

At least there is no conflict of interest here; we hope for cooperation in which everyone wins. First, I absolutely do not believe large-model companies can take most of the profits. That is impossible, because there are so many large-model companies and the gap between them does not need to be very large.

There are only two differences: time and cost. So it is not that one company will have enormous profits; I do not think there will be enormous profits. Those who control costs well will earn a little more, and those who control costs poorly will earn a little less. That is all.

Did I answer it? Has the first question been answered?

Everyone can... you can believe that many people will certainly be able to... in the future it will actually... everyone’s data applications will go through this cycle of iteration...

Can everyone hear me now? Thank you. Yes, thank you for your answer. I also deeply understand and respect your strategic position on the industry ecosystem. For example, regarding data, I believe all model companies can already access public data through various channels. The method should not be a problem; it is only a matter of time and cost.

Later, when we truly reach AGI, perhaps the vision or ideal state is that models can iterate and learn by themselves, training themselves. In this area, do you think current data can be used as simulated data, or is the quality of real data still the highest?

If real data is still required, would that mean AI intelligence remains limited by humanity’s past, because it depends on real data that humans have actually produced? Or can the ceiling be broken through simulated data, synthetic data, created data, and other means, allowing the model’s capabilities to surpass all the real data humanity has produced in the past...

I think it can surpass it. I think there are at least two ways to surpass it. Take Go: AlphaGo played a move that humans had never seen before. So it certainly surpassed humans within a certain scope.

But it may also have a ceiling and limitations. We cannot see those limitations at present. We believe, broadly speaking, that it can go beyond the knowledge that humans already have and that we can already articulate.

For this area, will it ultimately rely on real data or simulated data? Will it work?

It cannot be just one thing; there will be many methods.

Good, thank you. I have also taken up your time, and I would like to continue asking about the AI Infra question from earlier.

What was the second question? I am a little...

All right, I will briefly repeat it. I wanted to ask about AI Infra. In the future, people will certainly invest hundreds of billions of dollars in compute. In this area, we believe Chinese people are destined to achieve what is needed in hardware, and that one day there may be highly efficient compute. But in practice, this may still be a constraint for now.

In the future, will the two sides move downward toward each other? On one side, as model capabilities iterate, models will shift from brute-force computation to clever computation, so the compute requirements and consumption per model or task will gradually decline at the margin. On the other side, improvements in hardware such as cards will make iteration faster and faster and improve the efficiency of each card.

What will this look like in the process? If we build a ten-thousand-card cluster and buy H200 or 960 now, could it become relatively lower-quality compute or a relatively obsolete component after two or three years?

NVIDIA cards can basically be depreciated over five years. Huawei cards can be depreciated over at most three years. Huawei 950 is still quite good to use this year; I think it will still be usable next year, but after that it may genuinely consume too much power.

Huawei cards will certainly have a somewhat shorter lifecycle, because they are already two years behind NVIDIA. But I do not think the gap is that large.

If B200 is available now, I think it is worth buying however many you can get. For Tencent, for example, if Alibaba can buy them, then look at the quantity; if they are available at a reasonable price, they are certainly worth it. But when calculating the cost, the assumption is that they cannot be bought.

Understood. Our compute is behind; that is a fact. This fact can be addressed in three ways. First, we accept being behind in models. We can only use models smaller than theirs; how much smaller is a training question. We have to accept some lag in the model and a smaller model size.

This lag has an advantage: being behind means you have more time. With a certain amount of technology, you can then use clever methods...

[02:53:59]

Thank you.

So our gap with the United States may be that we are 12 months behind, perhaps 12 to 18 months behind, or 6 to 12 months behind. Put simply, we are two years behind the United States, while using one-twentieth of the compute to achieve the same thing.

The current narrative is that we are one to two years behind but use only one-twentieth of its compute. In the future, we need to rewrite that narrative: use some fraction of its compute while shortening the time, bringing it down to 6 months or 3 months. I think that is a goal.

We may even surpass them in certain areas. But with an order-of-magnitude gap in total compute, surpassing them comprehensively is unrealistic. In certain key areas, with trade-offs, it may be possible for us to surpass them in some places.

Understood, thank you, Mr. Liang. We are full of confidence and will work hard together. I will leave time for the other partners. Thank you.

Thank you for the sharing. I have two quick technical questions. In the discussion of the technical roadmap, you mentioned that the core problem we need to solve at this stage is continual learning, which is also a hot area of research overseas, called Recursive Improvement. I would like to ask: from a technical perspective, what is the biggest difficulty right now? From your point of view, when can it be solved? That is the first question.

Second, you just mentioned solving continual learning first and then pursuing intelligence. I would also like to understand the technical root of this idea. Does it mean that after solving continual learning, DeepSeek will also work on general intelligence? Those are my two questions.

The technical question is a little difficult to explain... The difficulty is that we have not yet found a method that really works. No one in the world has found a good method yet; everyone is exploring. So we are still in an exploratory stage, and we do not know who will find the next method for solving the problem.

We are still in the exploration stage. We have many ideas, and many ideas that currently look promising, but none has worked through to completion yet. Right, that is the first point.

The second point is that internally we place great importance on this narrative: when we train our next model, we hope it can help with our own development. It can improve DeepSeek’s efficiency. Our model’s first priority is to improve DeepSeek’s own work efficiency, so that when we develop the next model, it can provide more help.

Or more simply, the first goal of the models we build is not for everyone else to find them easy to use, but for us to find them easy to use. First, they must be useful to us. Once they are useful to us, I can develop the next model faster.

Many people internally think this way: first it must be useful to us, first it is for our own use. This is the fastest way to achieve AGI. If it works well for us, that may mean it works well for others too, but first we must ensure that it works well for us.

The narrative is a little strange, but many people really do think this way. It is not that I want users to find it easy to use; I want it to help us more, so that we can achieve AGI much faster. That is the logic of this narrative: to help us achieve AGI.

But first, it helps us achieve it. We need this help to achieve AGI. It is now very clear that we really need artificial intelligence to help us achieve AGI. Although it is not working autonomously yet and still works together with people, it is already very useful.

The second question is about what you just mentioned: solve continual learning first, then enter general intelligence. Is that your expectation for the future? I want to understand the technical root behind it. Why solve continual learning first and then enter general intelligence? What is your subsequent understanding of this?

Because solving continual learning can greatly accelerate our R&D progress. If I solve continual learning first, the general-intelligence problem will no longer be difficult. With AI assistance, if AI can learn continually, its capabilities should be extremely strong.

The current Agent’s capabilities are limited because it cannot learn continually; it cannot learn continually effectively. If we can finish continual learning first, AI’s capabilities will be extremely strong, and it can greatly improve the efficiency of our own research.

Once continual learning is built, general intelligence may become easy; using it to do the work will be easy. That is why I say this is an outcome we very much hope to see. It would save us effort and make things easy. Otherwise, building general intelligence manually now is tiring and difficult; it is data-intensive and labor-intensive, with a poor cost-benefit ratio.

Thank you for sharing.

A small question: please look at the online questions in the chat group. How much longer do you think AGI will take? Can domestic hardware catch up during that time?

In the chat window of the Zoom meeting.

Okay, I have seen this.

Huawei 950: Huawei is currently giving us sixteen thousand cards, which should be public information. That is probably an order of magnitude fewer than the large internet companies. Huawei can only give us this many because the price is not cheap either.

The large internet companies have greater ambitions and greater need. For us, we can buy some non-compliant cards. So our purpose in buying Huawei 950 is still to help Huawei build its ecosystem well.

Sixteen thousand Huawei 950 cards are equivalent to only four thousand B-series cards. So it is not a large quantity, and its significance is not very great. It is not enough to train a next-generation model; it is only enough to train our current-generation model. But it can let Huawei first make this part work, regarding Huawei 950.

And then, how much longer will AGI take? Can domestic hardware catch up in that time?

I think that in the AI undertaking, or in AI generally, China should be able to reach roughly the level of other countries within one or two years, or perhaps match foreign models this year. With the current approach and paradigm, this is not very difficult, so it should be possible this year. But it is not AGI yet.

[03:05:48]

I think it at least needs to be able to learn continually.

Can domestic hardware catch up in this time? I think the domestic side may need a few years. China first needs to solve the ecosystem problem, because the ecosystem is a matter of confidence. After solving the ecosystem problem and then the production-capacity problem, I think it should be possible to solve them gradually.

I do not really believe that five years from now we will still be stuck on production capacity. We are certainly stuck on production capacity now; this year, next year, and the year after, I think we may still be stuck there. But five years from now, perhaps not. I remain fairly optimistic.

The second question concerns thinking about the future organizational structure and the scale of our workforce. First, our previous structure was highly decentralized, because we did not have an organizational structure. But as our people increase further, we will certainly need to make some changes.

For now, all I can say is that many changes will be needed, but it is difficult to express them all at once. In the end, we should still have different departments. Some departments will need a relatively rigorous hierarchy, while others may retain a relatively loose and flat structure.

As the number of people increases, we will make this adjustment. We probably need to make it immediately, because I am already making it. If we did not make this adjustment, many things could not move forward. There are indeed many departments that should have an organizational structure.

And there is another question: was CV earlier than which version? I think the GCV4 version we released online is still relatively rough, and many capabilities still need time.

For me, a comfortable release cadence is generally one version every two or three months. The last release may have been at the end of April, so the next one may be at the end of June—roughly that kind of schedule. If nothing unexpected happens, each version should still be better than the last.

At the scale of 50B activated parameters, I feel that ultimately there will not be much difference between us and the current wave of open-source models. In inference speed and results, I feel there may not be much difference.

But compared with its large-size model, the model it has not made public, the gap should still be quite large. The gap is that we probably cannot reach that number of activated parameters; it would definitely require a larger model, perhaps 150B.

For 150B, judging from our current training progress, optimistically we could begin training it at the end of this year, or at least next year in... The gap is still quite large. Right, this is the gap with OCE.

Hello. I have a question. You often say that the process of achieving AGI is gradual rather than involving an abrupt change. Can I understand this as a process with no critical point?

There is no critical point, but it is nonlinear. A narrative we currently believe is that AI can accelerate AI research—AI can accelerate AI research. In other words, it is not linear, because you can use AI to accelerate your own research, so later it may become nonlinear.

Understood. So as for your current view, I understand that the earlier conclusion may be that continued scaling of language models is enough, sufficient to reach this state.

All I can say is that I have not seen an upper limit to language-model scaling. We have not seen an upper limit to our current level of intelligence, or to the level of intelligence produced in the United States.

Understood. I am very curious because you said that for the United States, a model with 800B activated parameters could be trained but could not really be used, so it would be difficult to make it available to everyone because it is indeed somewhat expensive.

One thing I have wondered about is that humans have had language ability for nearly one hundred thousand years, while evolution may have taken three point seven billion years before that. But in training AI, perhaps the order can be reversed as you said. Ultimately, though, it may still enter what might be called a world model—or perhaps not a world model, but a physical model or the embodied part—after this upper limit.

Yes, I think embodiment will definitely be needed; ultimately it will be embodied. For our company, the natural endpoint may be embodiment. For a normal person, their needs are not computers, right? Normal people eat, drink, play, dress, travel, and live; they do not need computers.

What they need is... they still need embodied intelligence to solve concrete human-labor needs. If the goal is to solve the need for human labor, embodiment is unavoidable.

Understood. So speaking in stages, suppose we reach something similar—not necessarily a critical point—where it can self-evolve, evolve itself fairly well, or approach the rise of SV. At that point, I am curious where we would hope to deploy this state of AI first...

It may be different from now. We hope it can directly... Suppose there is no embodiment. What do we mean by AGI, or what do we hope AGI can do? It can help me iterate the next model, just as it can help me iterate the next model.

Then if embodiment exists, we also hope it can iterate the next embodied system and build the next robot.

I am still curious because, looking at previous DeepSeek interviews and so on, taste and intuition are very important in choosing important directions and research directions, rather than simple engineering optimization. If AI can self-evolve later, will taste and intuition still matter, or what will matter instead?

AI does not currently lack taste or intuition; it lacks continual learning. AI’s taste and intuition are fine. If you ask it to write an article, I think its taste and intuition are not a problem.

There are a few more questions from earlier. Let me look. I saw several questions on the screen, but I cannot see them here now.

Mr. Liang, I have left a question on the screen; let me read it again. I wanted to ask about continuous learning. You also mentioned, as have many researchers, that it is an unsolved research problem. And the coding agent, especially catching up with MILES, is...

[03:17:47]

Scaling to the Office level and MIS level is a relatively definite goal. For an unsolved research goal and a relatively definite Scaling goal, how do you think research resources should be allocated, especially researcher talent, to achieve the best balance and outcome?

The model Office is a relatively definite goal, but for model MIS, I do not think it is certain; I can only call it a goal. Model Office, I think, is relatively definite.

CoT does not consume resources. Any research by a friend does not consume cards, or consumes very few cards; what it needs is ideas. It does not consume talent resources either, because no one needs to work on it continuously. It is not a project; it requires many people to think about the question.

So we do not need to allocate resources to it, because it does not need resources. Training models, building and releasing models, and efficiency experiments for models require resources. As I said, the consumption of resources in people and cards is not high.

That is why we call it “lottery hunting.” The barrier is low, and anyone can try, but who will discover something is something I may not know—perhaps it depends on talent, or perhaps on something else.

So we do not need to allocate resources here. The difference between us and other companies is that we spend time discussing and thinking about the problem, and treat it as an important matter.

Within the company, it is an important problem and one we spend time thinking about, but it does not require a great deal of resources to work on.

Now let me look at another question: the hallucination problem in large models affects user experience quite significantly. There is a method that can solve hallucinations, but it is a long-term proposition. Hallucinations can be addressed and improved through better post-training; it is a problem that can be solved and improved.

It is just that no one has devoted a great deal of effort to it. Or, for me, hallucination is a problem, but we may classify it as a product problem. We will solve it, but it is not a priority.

There was also a question about data labeling. Our situation with data labeling is related to our capital investment. Given the structure of our capital investment, we cannot support the cost of labeling that much high-quality data, because the cost is very high.

The cost of data labeling in the United States is not much different from the cost in China. China has no cost advantage in labeling data, especially high-end data; there is no cost advantage. This makes it difficult for us to invest in labeling data the way the United States does. This path is difficult in China because data labeling is simply too expensive. Whether we outsource it or label it ourselves, it is painful.

So we are basically using a two-legged approach. It is not that we cannot label data at all; some data has a low cost and some has a high cost. We label the low-cost data first.

So you can also understand that half of the people in our company are currently labeling data. Half of the core researchers, the most important people, are labeling data. We are concentrating on data labeling. At this stage, solving the AI problem depends on labeling data. You only need to look at it—it is a matter of data.

Mr. Liang, thank you. What you said really resonates with us; it was excellent. Thank you. I have several questions.

First, you just said that Chinese models are certainly more efficient than American models. You also mentioned other areas where China may be stronger in the future. In what areas do you think we will be stronger than the United States, in intelligence or otherwise?

I think we may be able to do better in many aspects of experience. Without even discussing our own experience, I think our own product experience is quite good. I do not think our user experience will necessarily be worse than America’s.

In products and product capabilities, we may not necessarily be worse than the United States. Costs should also be lower than in the United States, so China may well remain competitive.

In other respects, if there is a structural advantage, I do not think there may be one. But in cost and products, I do think we have certain structural advantages.

Cost is easy to understand: because they do not have to do it, they do not develop this capability. They certainly do not value it as much as we do. We can treat it as something very important, but for them it is not important.

The same applies to products. Native products from many companies still have reasonable product capabilities. So I think these two areas may offer structural advantages.

Good. My second question: you just mentioned post-training. Our relative cost of investment is high, while companies such as Anthropic and OpenAI invest huge sums. After this financing, do you think we will increase our investment in post-training?

The gap is mainly in labeling high-quality data, and then in AI research. We will certainly increase investment, but labeling high-quality data is typically not a matter of capital investment.

The bottleneck in labeling high-quality data is time, in my view; it simply requires time. OpenAI and companies overseas, including Anthropic, started earlier, have more capital, and have more cards.

Under these circumstances, China may be considered to have started only in the past six months, so I think we need more time. This is not closely related to capital investment, because even without more capital, the original capital would have been enough to expand at the fastest possible speed.

But this speed has a ceiling. The bottleneck is not that we can immediately get more people, nor that we are stuck on money or cards. It is simply that we are in a process of rapid expansion.

So we think the high-quality-data problem should be handled fairly well within a year; I think that is something China can expect. The situation may not be as bleak as it looks, but it does require time.

Thank you. The third question is that we have seen Anthropic use its own models to build its own products, launching many vertical applications in finance, law...

[03:29:17]

And perhaps moving into healthcare in the future. Do you think we will consider vertical applications at some stage as well?

I have not yet thought clearly about what our domestic business model will be in the future, or what the smoothest path will be. We have not reached that stage. Domestic and overseas conditions may not be the same, and it is still difficult to judge what the situation in China will look like then.

Given our current situation in China, I think the most reasonable approach is to devote all our effort to general Agents. Other Agents should have lower priority, including finance and medical Agents. We should first work on Coding, because Coding Agents can do many things, as can many other vertical Agents. At this stage, we think Coding Agents should still be the most important.

You explained that very clearly, thank you. We have another question. We also admire you for building DeepSeek through such a pure research approach.

But the industry has now entered the capital markets and the path of capitalization. You are also a very responsible person, both toward your colleagues and toward investors. How will you balance the direction of pure research and pure AGI work with the capital markets? You will certainly continue into the capital markets and eventually have a public company. How do you think about that balance?

I now think it should be possible to do all of it. Suppose we have several hundred million dollars in To B revenue this year, in addition to our consumer users. That already gives us a commercial foundation. If our To B revenue grows next year and demand expands further, the company will not be far from net profit; it may already be profitable.

We may no longer be in a phase of simply burning money, so I feel we will have considerable room to operate. In the worst case, selling APIs might be enough to support a listed company. If there were no further technological progress and our technology froze here, we could devote all our effort to selling APIs and making these services good. I think that would still be enough.

So I am confident. It really is not that difficult. We are in a highly leveraged position and in a very rapidly developing field, so it may simply not be that difficult. We hope to have bigger dreams, but we also have results we can rely on as a floor.

Thank you. That was very helpful. Those are all the questions; thank you.

Thank you for sharing. I would like to follow up on something mentioned earlier. I think DeepSeek’s biggest difference from other companies is that our organization, or organizational form, is different.

But organizational form is related to our goals, and we may also need to consider the boundaries and efficiency of the organization itself. From a macro perspective, is there a good organization we could learn from? Historically, perhaps Bell Labs, or some other form? Or do we think there is no ideal model and that we need to explore gradually ourselves?

Perhaps in a new era, we can only explore it ourselves. Our organizational form is certainly different from that of several American companies, right? Those three companies are different from one another too, but at least they explore from the perspective of commercial companies.

I mainly wanted to ask more about the organization.

First, we have no model to imitate. We make every decision based on actual circumstances and facts, working out what we should do. So it is a product of its era, or a response to reality; it is not the result of imitation.

That is, under these circumstances, this may indeed be the optimal solution, or at least we believe this is the path we have chosen. We have certainly thought through and selected every step. In any case, this is the result we chose. It is not that we saw someone choose it and then chose it ourselves; we analyzed the advantages and disadvantages and made the choice.

The future may be the same. We are not imitating anyone. I think we are still different from Bell Labs, because Bell Labs explicitly did not need commercialization, since it was a... But we explicitly do need commercialization. Ultimately, we have to survive. We are a company, after all, and the government will not give us a single cent.

So we can have an extremely ambitious mission, but at the end of the day we are a company and need to consider how to stay alive. To B is therefore certainly important to us, because we may have to rely on it to survive in the future. It is just not important now because it is currently a cost line.

So I think this is still different from Bell Labs. We are fundamentally a company. Historically, many companies have also pursued things beyond profit, but pursuing things beyond profit does not mean they are not companies.

Many companies have done great things because they pursued something beyond profit. That pursuit ultimately did not harm commercialization; instead, it enabled better commercialization. We are fundamentally a company. We simply make choices about which money to earn, when to earn it, how much to earn, and how to earn it.

Mr. Liang, I have two small questions and would like to ask you quickly. You just mentioned MILOS, which may not be a very definite goal, but you will certainly move in that direction. You also mentioned that activated parameters might be, for example...

[03:41:02]

The next generation may be around 150 to 250 B. In that case, do you think 150 to 250 B would benchmark against O4.7, or might it benchmark against something else? That is the first question.

The second question concerns inference. You mentioned that we now use some programming languages other than CUDA across inference. I understand that when we were based on NVIDIA’s ecosystem, we may have used a lot of PTX. Now that we use more things like the TileLang you mentioned, will this substantially reduce our inference efficiency, or reduce efficiency in the short term?

How do you view the efficiency loss caused by this change in programming language? Or is it ultimately something that can complement the system and improve efficiency?

It improves efficiency; it improves efficiency substantially.

OK, so there is actually no negative impact?

Right, it improves efficiency substantially. That is why this is an opportunity. In the past, we could not do without the CUDA ecosystem. Now we can abandon that ecosystem and use a simpler method, TileLang. It is a high-level language, and writing the program is fast; the amount of code is small, so I can rewrite it.

Understood. So both of these are actually major opportunities brought by AI, as you mentioned, rather than shortcomings that need to be fixed in the short term.

Yes, it is a major opportunity arising from technological development. It is not an AI opportunity, because we also have a project in which we use AI to write TileLang.

So will that be faster?

All TileLang is currently written by people, but it is already much faster than writing CUDA used to be.

Understood. So even at the execution-efficiency level close to the hardware, you think there is no impact?

A loss of 1% to 2% is acceptable, in my view.

OK, understood. Good, thank you. That was very clear.

Does anyone else have a question?

No problem. Let us stop here for today.

All right, thank you.

Thank you all very much for your time. Thank you, Mr. Liang. Thank you.

Bye-bye.

Bye-bye.