1. Original intention and vision: a group of ordinary people, doing things with goodwill
I, and other colleagues in our company, first came to start this company because I didn't think of how much money I would make in the end, how I would go on the capital market, go public, and what it would be like. We had no original intention. The first few dozen people had never thought of it that way-if he thought so, he wouldn't have come.
Generally speaking, we do this with a very great goodwill towards the world. We feel that this is useful to humans, and it is something other than money. Of course, later on, after discovering that this matter was very beneficial, there were other temptations. This was another matter. But our original intention, our vision, and the vision we have maintained to this day are not done in a way that maximizes commercial interests. I think this is critical.
About twenty years ago, the person I admired most in management was Jack Welch, the former CEO of GE. Looking back now, most of what he said may be wrong, but he was right on the most important point:The most important thing for a company is its vision.
Managing a large company depends not on your rules and regulations, but on your vision. What is vision? Vision is not a slogan hanging on the wall, vision is how you do it, not how you say it, but how you actually run it. Jack Welch said something I forgot, something like that.
How do we manage so many people, how do we organize? In fact, we have no organization, it is vision-driven, and we rely on a vision to organize it. This has advantages and disadvantages. In the future, we will find ways to leverage our strengths and avoid our weaknesses, but this is our specialty. We don't do it in a "what KPI do I want to achieve" way-there is no assessment, just a vision.
This vision is not even written, nothing has been written. This vision is in the way we do things and in our attitude towards the world. Maybe everyone in our company has a different understanding of this vision, and everyone's vision is different, but it is the same in one general direction: with a very great goodwill towards the world, and then want to do something. We use this to organize it.
Next, I will talk first, and then ask questions again. I will talk about what follows around this vision. This vision is real, not made up. We really think and do it, otherwise you can't explain many things about us.
2. Why insist on open source
Why do we insist on open source? Because this vision itself requires open source. You don't have this vision, you can't organize people. For example, Zhishu is also open source, but the open source of Zhishu is different from our open source. The open source of Zhishu has a feeling of being forced, and they feel that this is not the original intention; but for us, this is our original intention.
When it comes to open source, we thought very clearly from the beginning. The first is the vision; second, we believe that to make AI commercially successful, open source is beneficial.
This sounds a bit contradictory and counterintuitive, because historically, open source and commercialization have been in conflict. But I think AI is different from before. Historically, the market for a software company may be several billion dollars a year. Once it is open source, it will disappear, and there may be only tens of millions or hundreds of millions of dollars left. But AI is big enough. In the end, it may account for 10% of the GDP of human society. That is actually a very large number.You can't monopolize this matter. You must share it with others, otherwise you will definitely not survive.
This is different from making software with open source before. The market for that software is not that big, but AI is too big. If we want to monopolize this interest, then we must be abandoned by history. This is an objective law, this is a historical view. It's not that I can monopolize the market just because I don't open source-this is theoretically inconsistent with objective facts. You will definitely encounter a lot of resistance and there will be other ways to prevent you from achieving this goal.
In this case, traditional business thinking does not have to be followed. You need to have a mechanism to ensure that the benefits you can get are limited so that you can achieve it. Restraint is needed, I think it is needed.
If we want to do AI in our hands, first of all, I think we need restraint. You can't think about how much % of human GDP belongs to me, and how much % of China's GDP belongs to me-the more you think about it, the more impossible it will be. So we felt the need for restraint from the beginning.The more restrained you are, the more likely you are to achieve it. This is a commercial consideration, but of course this is a macro consideration.
I think it's intuitive, or at least my intuitive, or at least I really think so. We don't have many other advantages: we don't have any skills, we don't have more money than others, and we don't say that our people are better than other companies, but we don't. You think, when we founded this company two years ago, we didn't have much money, many cards, no popularity, and no appeal. We were just a group of very ordinary people.
They are really a group of ordinary people. If there must be a favorite narrative, it is: a group of ordinary people did extraordinary things, not a group of geniuses did extraordinary things. This is closely related to our restraint, and is in line with our restraint and our vision.
3. Restraint: API pricing and "ten-month cost return"
Will there be a conflict between open source and commercialization? I think when it comes to AI, if you don't restrain yourself, you won't be able to get up.
Open source is part of restraint. Our restraint is not only manifested in open source, but also in many aspects. Generally speaking, we don't have to specifically think about open source or restraint-the more restrained you are, the easier it may be to achieve it. At least so far, it is confirmed and understandable. Otherwise, there is no way to explain why we can do it: we have no weapons, a very low starting point, and very few resources. Our people are actually a random group of ordinary people. I myself am just a college graduate, not the top school.
This restraint is also part of our vision. AI is too big and has too great benefits. We are very restrained. As long as we can achieve it, the benefits will be very great in the end. If you divide it casually, the benefits will be very large, so now you don't have to consider which part of the benefits to take and how to take it. This benefit is big enough, and if you divide it a little, it will be enough.
So we said before that we can only earn a reasonable profit, and it depends on your wishes, not the size of the profit. This is different. The logic of our API pricing is: a reasonable profit is probably that we go to the market and buy a batch of equipment,Ten months to recover costs。In the current situation, considering that you have risks, upfront investments, etc. -a server, we amortize it financially for three or five years, but commercially, we think it will take about ten months to recover the cost., we think it is enough. This is the logic of our current API pricing. Our V3.2 Flash and others all recover the cost of the equipment in ten months.
This is our standard. It's not really profit-maximizing-if profits are maximized, the price should be set higher. Because in this price range, user demand is inelastic: if I double the price by half, or double it, the consumption of tokens will not make much difference; if I double the price, my total income will nearly double.
Let me tell you a story. Our DDCP is our model. At first, we were worried about too much demand, so we set the price relatively high at first, and everyone in the team was not very happy. Later, I lowered the price again to one-quarter, and everyone was very happy.
I think this is what we really think, the vision I mentioned earlier: we still make this thing useful to people, not that we make the most money; that we can make reasonable profits, and everyone can afford it. This is also the idea of other people in our company-when we lowered the price at that time, many people in the company group cheered and everyone felt very happy. Because this is why we have spent so much energy and effort to make this model well: it can be very cheap, the effect is very good, and it can be fully used by everyone. We just thought it was very happy. This is our motivation, this is our vision, and the consensus that our company can come together to do this.
This should be quite special. Because price cuts are definitely not a good thing for our other competitors, and they must not be cheering-if your income, your ARR are reduced by half, the ARR will be reduced by half. This is a place where we are different.
We thought that was enough. Internally, I am very satisfied with the cost recovery in ten months; externally, we also feel that this price is something that everyone is happy and willing to see, and it is a win-win situation-the company and society, and everyone are win-win.
Someone just left a message on the screen saying that the profit of returning capital in ten months is too high. There is indeed room for price reductions, and there is also room for model optimization, so the overall price reduction space is still relatively large. But this cost-ten months to recover the cost-we can do it ourselves, but other families cannot. Like Alibaba or Tencent, it is not our optimization. Its cost should be several times higher than this. There is still a lot of optimization work here.
Just now, why don't we continue to reduce prices? It's because it's inflexible: if I lower the price again, the demand will not be more, or there will be very little increase. Because this price is affordable for everyone, everyone feels satisfied and will not use it just because the price is high. So if you lower the price, first of all, the company will not have more income and will not have more value to society-everyone is satisfied with this price. No matter how low you are, your happiness in society will not increase much.
When it comes to pricing, we definitely don't want the company to have the highest revenue or the highest profit. This is part of our restraint. In the short term, your price will be higher and your income will be more; but in the long term, it's really hard to say.Because I think restraint is a strategy. For me, restraint is a strategy.--Sometimes you can give up something and trade it for something else. In fact, the same is true for open source, and it can also be regarded as our profit.
First of all, for our company, we are very happy and everyone is very happy. Employees feel a sense of accomplishment and we will have cohesion because of this; Secondly, this profit is beneficial to society, and society is also very happy. Other colleagues and ordinary people will be very happy. So I understand that this kind of restraint can increase the probability of us achieving AGI in the long run.
I have no doubt that AGI will have great commercial value. On this basis, my priority is not how to get more shares, but how to increase the probability that I can do it.
4. Other manifestations of restraint: not robbing users, not picking up all sesame seeds
This restraint is also reflected in many other aspects. For example, there were suddenly a large number of users during the Spring Festival last year, but we did not pursue my pursuit of retaining these users, using these users to monetize them, and seizing these commercial interests. We didn't rob users or make money, but we worked hard to find ways to serve users well. We don't have this idea: I want to make the next super App, who am I going to compete with, I want to make the next byte, the next Tencent. There is no such idea at all.
We could have done this, but we didn't. My understanding is that this is also part of restraint. You don't want to earn everything-it seems that after you have users, you can make the next byte, and then you can eat it. This works commercially: if we spent a lot of money and bytes to grab users last year, it would be a way to play. But what we chose was a very restrained approach. I won't compete with you for this thing--Because there are watermelons in the back, the ones in front may be sesame seeds.
I shouldn't take everything. Of course, this sesame seed may be bigger, but I think compared with the AI behind it, the front is not big. Looking at it now, it may be right that we didn't exert our strength on the C-end last year, because we can see that there are really bigger watermelons behind and there are really just some small sesame seeds in front. If I had a lot of money last year and did this very big thing, what good would it be? You don't get anything. These are my true thoughts, because I feel that the AGI opportunities in the future are very big and always will be very big.
I don't even have to think about what position I will occupy in it or what my business model will be. I don't even have to think about it. As long as there is such a big business opportunity, you must have a way. We can also pick up the sesame seeds in front of us, but we just pick them up casually and will not stop or treat them as an important thing.
So last year's C-end daily activities, I think it might be a small matter. But we also picked it up and maintained user use at a relatively low cost, because it may be useful in the future-although we don't know what use this user has now, it is a pure cost expenditure now, but it may be useful in the future. Since we can get it easily, we will get it easily.
Including this year, it is very likely that there is an opportunity for our ARR revenue in API or AI: if this demand can continue to expand and more graphics cards can be purchased, then it is very possible for ARR to reach hundreds of millions of dollars. If AI can achieve one billion dollars, then basically my company's cash flow will be corrected, and I can cover my R & D expenses and all my expenses. This is possible, but we don't make it a priority. We will do this and we must do it well, but it is not the first priority, nor is it what we really care about today. Greater opportunities should still lie ahead-the opportunities ahead, including last year's C-end and this year's B-end, are not our goals. In other words, most people in our company don't think this is an issue as important as AGI.
5. Talk about open source again: Under six times the profit, there is no conflict
Open source can be said more, because many questions have been asked about open source before. First of all, we will open source, and our strongest models will probably also open source. Because I don't see the benefits of closed sources-bytes are closed-source, what's the benefit? I can't see it.
Even if the model is open source and you tell everything to others, the threshold is still very high. If others want to use it, it will be very difficult; secondly, if they want to use it, it must be done at a very low cost, which is also very difficult, not so easy. It's not that he can easily achieve the same deployment costs as me because I open source. There is still a lot of work to be done here. Although these working principles are understood, not every company is willing or has the will and ability to organize manpower to achieve this goal. It may not be good at doing this because the resistance is too great, it is difficult to control this cost, and there are many management and physical constraints.
This is also the advantage of startups: if they are too small, you don't have the strength to do it; if they are large, they will be difficult to organize. There are difficulties in this matter, so this is a sweet spot for a company of our size, sweet point. If we were bigger, we might not have this problem; if we were smaller, our strength would be insufficient.
As for the pricing, I will not charge a very high fee, but also charge it based on a ten-month return. If I pay back the cost in ten months, I can already... If I pay back the cost in ten months, I can make the independently deployed third party unprofitable-a third party cannot do it, he cannot do it at this cost, and he definitely cannot do it. So open source will not affect my income. Of course, if I want to make a hundred times the profit, then open source will...(Live video is briefly interrupted)
Open source, I don't think it has any impact on our business model. The premise is that we only make six times the profit-ten months to recover the cost, which corresponds to roughly six times the profit. When you only make six times the profit, open source will have no impact. But if you want to make a hundred times profit, then open source will indeed affect your making a hundred times profit, because a third party will deploy it, and it may cost 20 times less than you.
Is this model sustainable in the long term? I think it's okay. Under our vision, open source is long-term sustainable, and we intend to do so. You can argue that restraint is also in the long-term interest. This strategy can give us more technical opportunities, and our probability of making AGI is greater, and we are more relaxed.
You know, we don't have to work overtime at all because it's not that difficult. But for other families, it may be difficult because he thinks too much. In fact, it's not difficult, not difficult at all. It may seem from the outside that we have chosen a very difficult model: to do research and do the most difficult things, which seems to be a hard model. But in fact, we have given up a lot in other places, which makes us still very powerful and we still do it very easily.
So my judgment on open source is that it is ongoing.There is no conflict between open source and commercial payments-provided that there is no conflict at six times profits.Six-fold profits may seem high, but they are not high. Now that AI is so efficient, a reasonable profit may be so much. In the future, it may drop to four or three times, and it is impossible to drop it any further, but it will still make a lot of profits-just by selling APIs. I don't find selling APIs that attractive, but this matter shows that I don't see any conflicts.
I'm not worried about others deploying our model to compete with us. I'm not worried at all. We hope they can deploy it. We try our best to help the open source community and help everyone deploy our models. I'm not worried about him competing with me for this business because the market is big enough. I'm only worried that he won't be able to deploy it-some details are not done correctly, the effect will become worse, or the cost will be higher. There is no conflict here.
When I had the To B business last year, I asked a lot: Our C-terminal-I open source, then does the C-terminal conflict with my own C-terminal? Because I don't have the advantage of traffic, Tencent itself has a lot of traffic. He deployed our open source model and took over all C-end users, so he stole my C-end users. But in fact, it won't. There are still many reasons for this.
Another question: Is the open source model we gave us the same as the model we deployed ourselves? It's the same. We won't open source a poor model or use a better model when deploying it ourselves. It won't, it's the same. This also shows that there is actually no conflict. For the whole year of last year, the C-side was basically open source, and the C-side services didn't see conflicts, really didn't see conflicts. This is the open source part.
6. Our goal is AGI: Technology Roadmap
Below is the company's long-term vision. I think our goal should be AGI. Everyone's definition of AGI may not be the same, but it does not prevent us from taking AGI as our goal.
From the perspective of technical route, AGI's roadmap is relatively clear. With the current generation of AI technology, if you can describe a problem clearly and give it complete context and instructions, it has surpassed human beings. But there is a prerequisite: You give it full context, full instructions-this prerequisite is difficult to achieve. For example, if we are meeting today, we have a long and strong context in front of us. Maybe everyone has a context of decades, which AI does not have. What AI can do now is that it can do better than humans in a limited context.
But it still cannot replace humans. There is still one place missing here: continuous learning. People can continue to learn. If you hire an employee, he may spend two months getting familiar with the company environment and his work. He will be able to get started and do many things. He understands what you say--you say,"Call that little Wang," and he knows who he is. But if it's AI, because it didn't have these two months of study in front of it, you say to it,"Call Xiao Wang over," you have to tell it who Xiao Wang is, what position, where, how to find him, what to pay attention to when looking for him. You have to give AI all the context. In this case, AI can do it, but you can't give it all the context.
So AI cannot replace your employees. But if AI has the ability to continue learning, it can replace the world by studying in the company for two months just like your employees.So we are still missing the next step of "learn to learn".
The development of AI can be understood as a ladder. The ladder we walked last year was CoT, which is the thought chain: we found that through the thought chain, intelligence can reach a higher level, and through its own thinking, AI can do more things. We crossed another ladder. This year's ladder is Agent: With Agent, more things can be done, its range of capabilities is wider, and its intelligence limit is higher.
Why ladders? Because each subsequent step is based on the previous foundation. Agents need to use CoT, and CoT also needs to use the ladder in front-the ladder in front is the language model. So not a single step was taken in vain.
The direction of AI intelligence is traceable. The ladder we walked this year is Agent, but the Agent ladder will also end: after it solves all possible problems, it still cannot replace your employees-it has reached the limit of its capabilities. Just like CoT, after reaching its upper limit, CoT has surpassed the top humans-doing Olympic math questions and writing programs has surpassed the top humans-but it still stops there, and that technology has not been able to reach AGI.
After Agent, we think the problem we should solve is continuous learning: how to make the model learn continuously, rather than saying that you have to give it a strong training-it should be able to do a relatively long period of time like a human being. Continuous learning. This problem is the same thing as completing the task and so on. They are related and solve the same problem. Now that we stand in the place of Agent, the next bottleneck we can see is continuous learning. The next problem to be solved is how to continue learning. This is visible and relatively clear. This is an obstacle stuck in front of you. You must overcome it, and there must be a way to overcome it, but it will take time.
After continuing to learn, we may come to a singularity: when this model can continue to learn, it can already do everything that humans can do. It can develop its own version, research it itself, develop its own next version, and develop a stronger artificial intelligence model. So it will reach a singularity and be able to achieve its own iteration.
But this singularity is not a singularity. It is also a gradual process. It may be a relatively long gradual change rather than a sudden change. But habitually, we all think it might be a singularity-because long ago, prophets thought there would be a singularity. In fact, it is not a singularity, it is a continuous process. Then after this step is over, I feel that it is embodied intelligence.
This is our speculation and our perception of the timetable:First solve "learn to learn", then reach the singularity of intelligence-the singularity that can iterate on itself-and then there is embodied intelligence. After becoming intelligent, it will enter the real world and can do housework for you and support your elderly care.
We think this is an ideal roadmap. Everyone has different opinions and there is no right or wrong, but we feel that this roadmap is the easiest: there are very few new things you have to do at each step. With this roadmap, we don't have to work overtime. But if the roadmap is reversed, such as realizing embodied intelligence first, then it will be very tiring and it will be a very hard job. We don't want this to be a roadmap, we want it to be easier.
Solving continuous learning first, then solving the singularity of self-iteration, and then solving embodied intelligence will make this road easy. Because at the end, you can use the previous technology to help develop the later technology. If you create embodied intelligence after the singularity, you don't need to be done by humans, and you don't need to be done by us. The model can come out by itself.
So this is the answer to what our long-term goal is: This is our long-term goal, which is what we call AGI.
7. Dimension reduction strike: Both C-end and B-end are by-products of AGI
Everyone back to reality. The most important reality last year was that everyone wanted to build Chatbot and grab traffic from the C-end; the reality this year is that everyone wanted to grab To B's revenue and participate in it-because if you don't participate, you won't be at the table at all, right?
But we don't think it's an important thing. Within our company, what we really care about is actually the AGI roadmap I just mentioned and how to make breakthroughs in the next step.
One thing is very strange: you can't get what you want most; it's quite easy to get what you don't care about. There is a strategic advantage here: we think about AGI in our hearts and do AGI. When we make that application and C-side and B-side, we don't need to think too much and only need to spend very little energy.
If you stand at a high level of technology and do a relatively low-level technology, there will be a reduction in dimension. At least on last year's C-side, we saw that this was indeed the case: we didn't spend a lot of effort on the C-side, and we didn't even want to maintain those users for a while, but the users couldn't get rid of them-they really couldn't get rid of them, so in the end they were still there. This year's B-end revenue seems to be growing relatively optimistic and should be relatively good compared with its peers. I estimate. But we didn't spend a lot of energy to do this. We didn't specifically do it at all, just did it incidentally.
AGI This step is something I have to take. It is a step. On the way to AGI, I have to pass this step. I put these technologies to serve everyone through APIs, and I didn't do anything extra. We are still doing AI, which is a by-product. I just need to have a few people maintain this API. I don't even have customer service, and I don't need to sell it. I don't need anything, and the users will come by themselves.
The C-end and the B-end are both by-products of our AGI process. They are an intermediate output and have no conflict with my AGI process. I didn't do it just to make C-end or B-end; we made AGI to make AGI, which happened to produce this thing, so I used it for commercialization.
This is different from other companies. Other companies create this model to serve C-end users and B-end users; but for us, this is not the original intention. Our original intention is to pursue AGI.
I think this is a dimensional-reduction blow to a certain extent. AGI is a larger vision. This vision can gather more outstanding people and has stronger cohesion. So I have an organizational advantage, and I use this advantage-it's a dimensional-reduction attack.
But if you are a commercial company and your vision is to serve C-end users well, that is another story. He has other advantages-in terms of products, user services, traffic-but not in terms of technology. The more favorable situation now is that model technology is the most important thing, and you have to make the model well. This can explain our previous development trajectory: we really chose AGI and didn't want to be a lot of users at all.
When we suddenly became popular during the Spring Festival last year, it was not in our script. We had never thought about that thing at all. We just wanted to do this technology well. But at that time, I discovered that our organization had additional advantages in terms of talent and organization compared to a fully commercial and product-targeted organization.
This is also amazing: at that time, everyone was robbed of the C-end, but it was led away by someone who did not rob it. This really shows that the logic I said earlier is reasonable and does have talent and organizational advantages.
This talent advantage does not mean that my people are smarter than him, but how I organize these talents, how I inspire them, and how to cooperate. This thing has advantages. Because you bring smart people together, it doesn't mean that they will automatically cooperate and achieve a goal with great passion, so you need a vision. The inspiration from our previous experience is:AGI's vision is very powerful.
8. The biggest core interest: team stability
Next question: What are the importance of core interests?
As I said earlier, we must be very restrained in many aspects. But what are our core interests? There is actually only one point:Our biggest core interest is to maintain the stability of the team. This is our greatest core interest, or even our only core interest.
As long as I can maintain the stability of the team, I will definitely be able to make it, I will definitely be able to make AGI, it's that simple. As long as no one leaves and we can continue to do it, I can. There is basically no risk, just early and late; if I encounter setbacks and no one leaves, then I can continue. Money is definitely not a problem, resources are not a problem, and other elements are easy to obtain. For us, there is only one core interest, and only one that cannot be compromised: we must maintain the stability of the team.
This is also a very big challenge we face, and I think it is the biggest risk. Of course, this risk has been greatly relieved with our recent financing-because everyone still gets more options and the amount is still relatively large.
In terms of team stability, as long as some of the most important employees and the oldest employees can be stable, others will not be able to leave. Even if others had fewer options and less income, he wouldn't leave, because he didn't come all for money-everyone wanted to do this in an environment where AGI could be achieved. So it is still attractive to talents. Historically, our talent flow has been relatively small, and compared with our peers, it will always be relatively small. But this is still our biggest challenge, our only challenge, so to speak. Everything else is a matter of time, which will cause us to delay half a year or a year at most, but it will not be said or unable to do it. There is definitely no shortage of money, there is definitely no shortage of resources, there is no shortage of these.
So a lot of what we do now is to maintain the stability of the team. Apart from this, I think we can abandon everything else and refrain from it. We have always been very restrained and are unwilling to become rivals with any big or small Internet factory. I hope I can empower him and help you and help you do this. This is also part of the business significance we talked about earlier. On this premise, we are willing to assist and help anyone-even our competitors, including Ali, Zhipu, and the Dark Side of the Moon-to do better. Because we have nothing to lose: we are originally open source, and open source must make the boundaries as clear as possible and hope you can reproduce them; if you can't reproduce them, you tell me, and I will tell you how to reproduce them. This is part of open source, and it won't happen just because you are a competitor. Of course, if I were a partner, I would do more. But there is no conflict in terms of major interests.
When dealing with the outside world, our attitude is:We only do the main line of AGI. It is the language model, CoT, Agent, etc. I just talked about, just the main line. The field of AI is very extensive, and there are many things we don't think are on this main line-such as 3D and video generation. I think it may not have much to do with the main line of intelligence, and we won't do it. There are also some things, such as world models, that I don't think have much to do with the upper limit of intelligence now, so we won't do them. But when others do it, we are also happy to help-whether we have time is one thing, and there is no conflict of interests.
We also hope that these AI technologies can be used in various production environments, improve the production efficiency of society, and help all walks of life improve production efficiency. We are very motivated to do this. Whether I have the time, whether I have the manpower, or whether our classmates are interested themselves is another matter. But there is no conflict of interests, and we hope to achieve this goal. And we believe that this has no conflict with business, and I still deserve the same benefits.
We had this attitude before, but we didn't get anything less because of it-we didn't get anything less because of my open source, because of our kindness, or because I helped others. For example, for last year's C-end users, we still have a large number of C-end users and are still relatively stable; this year's B-end, I think it is also relatively optimistic. I have not affected business interests because of our goodwill. I have no impact at all. On the contrary, I may have added points. This may seem counterintuitive, but it is. Think about it the other way around: If I violate it, can I get more? No.
There is a question: How to understand that "the world model has nothing to do with the increase in the upper limit of AI intelligence"? I'm talking about at this stage, that's our judgment. This is our own AI roadmap, not the only roadmap. From our understanding and our judgment, the most important thing at present is to do a good job in AI training. To do AI training well, you don't need a world model or even multimodal-you narrow down the scope of AI training. There are no multimodal, but there are some tasks you can't do, but it does not affect the establishment of this algorithm.
Multimodal will eventually be done. The important thing now is training. The next step is to solve the problem of continuous learning, and the next step is to ask questions by itself. But there is no world model and no video generation in this roadmap.
Video generation was very popular when it first came out, as if it was something that had to be done. If you didn't do it, you wouldn't be an AI company. I am very surprised: if you think about it carefully, it has nothing to do with the intelligent roadmap. In fact, you also found that after Sora came out, everyone did it, big companies and small companies did it, but the small companies later cut it. It has nothing to do with the upper limit of intelligence. But in business, it's a good business.We won't do it because it's a good business, we will only do it because it's on the smart roadmap.
Video generation is relatively clear, so take it as an example. The meaning of a world model is not so clear, because many things can be said to be a world model. From our judgment, world models and intelligence are not the most important things at this stage. The most important thing is AI training and how to solve continuous learning after AI training. This is our company's judgment, of course, each company's judgment is different.
9. Computing power: The only gap with the United States lies in resources
What I just said is that the most important issue for the company is the stability of people. From another dimension, what do we lack? What is the gap between us and the United States? In fact, there is only one gap: resources.
We don't have that many cards, and the number of our cards is still relatively small. We now have about 20,000 H equivalent computing power, most of which have just arrived, and have just arrived in the last month or two. There may be still many machines that have not arrived yet. Our total computing power was relatively small last year, but this year we are expanding very aggressively. We will buy a large number of machines in the next few months, basically all of them are Nvidia.
How many cards do we need? The more the better now. Within the range we can afford, the more cards the better, there is no doubt about it. So our current strategy is to buy as many cards as you can at a reasonable price. With this financing, I can buy as many cards as I can. The speed of spending money is not part of the plan-as long as the price is reasonable, I will buy as much as possible. If I spent all my money within half a year, I think it would be a good thing. It would be too happy and ideal.
In fact, it's not easy to spend all this money: you can't buy so many cards, it's difficult to buy, and the price is very high. You can't say you pay a very high price to buy it, but you have to make sure that the price is reasonable.
If I spend all this money within half a year, it may be the most ideal. Because turning money into a Nvidia card is definitely better than putting it in a bank-if you put it in a bank, the deposit may only be two points now; but if you buy a Nvidia card, you will recover the cost in ten months.
So we must buy as much as we can. After buying the card, I have a lot of space. By providing services or something, I can always have cash flow. If I have cash flow, I can live and I don't need to put a lot of money in my account. So we were just worried that we wouldn't be able to get that many cards. If we could turn all our money into cards, we would not hesitate to turn all our money into cards, and we were willing to pay a certain premium here because it was so cost-effective. Even if we pay the premium, it will be difficult for us to achieve this goal. Objectively speaking, if I can spend 20 billion yuan this year, then our purchasing department will perform extremely well.
The gap between us and the United States is mainly in terms of resources, but in terms of people, there is almost no gap-because it is the same group of people, all of whom are China. Some people stay at home and some go abroad. They don't say that smart people go abroad. No. In fact, it is quite random: more than half of the brightest people may go abroad and less than half remain at home. There is no shortage of domestic talents, and we have a large base and are replenished by so many people every year.
Talent is not the bottleneck, resources are the biggest bottleneck. Resources first affect talent training: because we have less computing power and fewer experimental opportunities we can do, our talents are generally behind those of the United States.The gap in talents is essentially due to the gap in computing power.
On the largest model today, we actually can't afford to train. Even if you spend all 50 billion yuan, you can't afford to train it; even if you can pile it up, you can't afford to use it. Currently, the largest model is activated at about 800B; in China, we are still at the scale of tens of B, and the largest model in China may only be activated at tens of B-an order of magnitude away. If I want to train a model the same size as it, I should need 50,000 GB300 cards, or Huawei 9.50 million cards. This is just training and research has not yet been considered. So the biggest gap between us and the United States is in resources.
Our current resources, as well as the resources within this year and in the next few months, including the large resources that will arrive soon, are only enough for us to do more experiments at the scale of ten-B activation. Because at the scale of tens of B activation, there are still many experiments to be done and many things need to be clarified. We should be far away from a model that can train 800B, not so many cards.
So the difference between us and the United States, I think, is the difference in resources. All the differences we see-including differences in talents, model capabilities, and applications-can be considered to be due to differences in computing power resources. Computing power resources, on the one hand, are not available in the country itself; on the other hand, our capital investment is less than that in the United States, and our capital investment is much less. The proportion of talents 'wages in this is very low-look at the salary they offer is 100 million US dollars, but when calculated, the salary of talents still accounts for a small proportion, and the majority is still computing power.
This problem is currently basically unsolved.
So we will not consider competing with the United States on such a large scale now. Now we should do it well on a scale that we can afford to train and use-a scale of tens of B activations-and then wait until we have more resources in the next step to achieve 150B, 156B, or 250B activation scale.
There is a gap between us and the United States that currently seems difficult to bridge. If you insist on training such a large model, you can train, but you can't do enough research-you can't do enough research before training.
10. Final gap: cost, time, experience
There is another question that everyone is more concerned about: Where will the final gap be reflected in all large model competitions? I think there may not be much difference in the end. The final gap should be in three aspects:Cost, time, user experience.Other than that, there may be no gap.
The cost is easier to understand: If you provide the same service and the same quality service, at what cost can you provide it? Can other companies provide the same service, such as BYD's batteries, at this price with the same technical content? I think this is a difficult thing. It is not so easy to do. It must be a barrier. So cost is definitely a difference, and I think cost may rank first.
The second is time: when can you do it? A few months earlier and a few months later, it's different.
The third possibility is experience. The user experience is still somewhat different. There will still be some user stickiness and user barriers among users, but it may not be essential. The essence is still the first cost and the second time: do you do it first or later, and whether you do the same thing faster or slower.
11. Commercialization: Always doing it, but not targeting it
How to price after the long-term commercialization path and product line are further enriched? We feel that the most worthy thing to do and the most worthy of energy now is to do AGI. To push AGI forward now is to push the lower limit of intelligence upward. This should be a more cost-effective and profitable path at this stage than building more product lines and considering more commercial paths. I think this may be the case for some time in the future, as well as any time in the past.
If we spend a lot of time thinking about and enriching our products, what is the use? What was commercialization discussed six months ago? I must be advertising, doing e-commerce, embedding e-commerce in my products, and integrating with local life. It's definitely useless because the changes are too fast. Your product, at this stage, we spend a lot of time doing commercialization considerations and commercialization paths-the product line has a very short life cycle.
I don't think this time has come yet. This is our judgment, and at least the previous experiences support this judgment. In the past three years, it was a waste of time for you to come and talk to me about commercialization paths and product lines. Because you cannot predict or foresee the future, you can foresee very little.
We are doing world-leading AGI research and commercializing it in due course. I think we have been doing commercialization, but we have not targeted commercialization. We aim at AGI, but we have been commercializing it, so we have C-side users and B-side revenue. Judging from historical experience, this strategy has been successful. We should be far away from the time when we completely turn to commercialization. The biggest thing is the extension of technology, to build the next generation of technology and solve more current problems. I personally feel that in the future that you can see now, it is too early for you to focus on products at any time.
This is also part of our restraint. I hope that these business opportunities can be done by others-how to use this AI is done by the whole society and all our partners, and everyone will share the benefits together, rather than I want to keep it all for myself, which is impossible. And we don't have that much energy, and our organization doesn't have that many people to do it.
As for partners, we actually selected this financing carefully. Regarding specific cooperation proposals, first of all, I think that the interests are relatively consistent: those that are most consistent with our interests, those that are least hostile to us, or those that most hope that we can succeed. Not everyone wants us to succeed, because we still harm the interests of many others.
The company's major strategic, technical business decisions and decision-making mechanisms: Our company as a whole is based on consensus. I am not saying that I decide everything alone, but that I seek consensus. My authority and influence within the company are based on consensus. For example, if I want to do something, I will definitely first see what everyone's consensus is and whether everyone wants to do it. It is possible that I will have some guidance or tendencies, but the role of this guidance is limited, very limited, and it is still based on consensus. This decision-making mechanism is actually a mechanism for seeking consensus: it doesn't mean that I can promote anything-it must be consensus, so that I can push it forward, and then I can push it.
At present, the main focus is basically on DeepSeek.
12. Model gaps and global division of labor
The gap in the final effect of each model should be comprehensive. To compare model effects, you must definitely compare them at the same cost for this to make sense-if you compare two cars, you will also compare cars at the same price. The difference between a well-done model and a poorly done model should not be in a specific link, but in the overall context.
Anthropic now surpasses OpenAI, is this long-term? I don't think this is long-term, it must be phased. For OpenAI and Google, the probability will rise alternately in the future. In fact, Anthropic's advantage in Code Agent is not so great now, and it does not crush OpenAI-probably half of our company think OpenAI is better. Anthropic has a first-mover advantage, but this first-mover advantage should be gone soon. It is not an advantage that it can occupy in the long term. These three families are all very powerful. Anthropic is the most efficient among the three, and it should cost and burn the least money.
When global AI dominates the division of labor, China companies are likely to play the most productive role. Commonly speaking, we have the largest production capacity-including chips, chips may have the largest production capacity and the most electricity. China people will make their products the cheapest; in terms of effect, many products made in China are not much different from those made in the United States. The same may be true for AI in the future: AI made in China will be cheaper. This cheapness may be systematically low, just like the cheaper services provided by China in other industries.
When I do things, my habit of thinking is: What will you gain the most from doing at this time? If I think I can make the most benefits from making products now, I will make products; if I think I can achieve the most benefits from AGI now, I will make AGI first. Obviously, I don't think making products now is the most profitable one.
13. Domestic chips and TileLang: Nvidia's moat is disintegrating
The problem of adapting domestic cards: Domestic cards actually have a historic opportunity now. There was a problem in adapting domestic cards before, which was called poor ecology-you couldn't use it after buying it. It did not have the ecology of Nvidia. So Nvidia's moat is very strong.
But this matter is changing.Nvidia CUDA's moat is rapidly disintegrating.There may be three reasons for the rapid disintegration.
First, there is AI now. With AI, it will be much easier for me to build this ecosystem than before, because AI can write code-I can use AI to build this ecosystem and build the same ecosystem as Nvidia.
Second, there are some new technologies. For example, our family has a technology called TileLang, which is a high-level language. Using this high-level language to write CUDA operators can quickly write the entire ecosystem of Nvidia. Combined with AI, there seems to be no obstacles. It has not been completed yet, it has not been completed yet, but there seems to be no obstacles to this technical route.
Third, CUDA-Nvidia evolved from game cards, so it is in line with the design and settings of game cards in many places. CUDA is compatible with game cards. In the past, because AI computing was a very small field and smaller than the market for game cards, this was reasonable. But now that the market for computing cards is bigger than that of game cards, there is no reason why the two need to be coupled. The current trend is that there will be no more coupling in the future. Then dedicated chips-whether it is Huawei or Nvidia itself-will be dedicated chips in the future, and they will no longer be the same as before. In this context, the original role of Nvidia Ecology has been greatly reduced: for a dedicated chip, it has nothing to do with CUDA and is not bound to it. In other words, when this chip was designed, how to establish this ecosystem was already considered.
It's a bit complicated, but anyway: There is now a historic opportunity to replace domestic AI chips. We believe that in the next year, we will see one thing verified:There is no problem with the ecology of domestic chips.I used to think it was a problem and thought it couldn't be used or easy to use, but in the next year, we can reverse this understanding and use facts to reverse it.
There are no problems with the hardware and ecology of domestic AI chips. The only problem is insufficient production capacity. There is no obstacle to adapting domestic cards, and Nvidia cannot stop it. If I could buy Nvidia cards in a normal business environment, it would be difficult to replace them domestically; but when Nvidia cards could not be purchased, everyone had no choice but to make domestic chips. In this context, there are no obstacles to the adaptation of domestic cards. I don't think there are any obstacles to domestic cards to establish an ecosystem that is the same as or even better than Nvidia, but it will take time.
We are mainly cooperating with Huawei now. Huawei adapts itself, but we will participate in this ecosystem and deeply participate in Huawei's ecosystem. For Huawei card adaptation, the main job we do is to make a high-level language compiler and TileLang. After completing TileLang well, the problem may be solved. This matter is more complicated to say, but we are doing it, and it will be much clearer to explain it after we finish it.
You can understand: When training in V3, you still use Nvidia's card, but you no longer use Nvidia's ecology. V3 uses Nvidia's card, but not Nvidia's ecosystem. Instead, we first wrote an advanced compiler called TileLang, and then based on TileLang's ecosystem to complete everything else, so we almost don't rely on Nvidia's ecosystem. As long as I repeat this set of things and this process on the Huawei card, it will be completed.
I think this may be a historic mission: it can completely reverse everyone's previous perception of the bad ecology of domestic cards. Now the time and place are basically complete, but the only difference is time. I think within a year, many people will understand that this problem has been solved, and the rest is the problem of production capacity.
I am quite optimistic about domestic computing power. I feel that at this point, Nvidia is digging its own grave. Huawei's supernode-Huawei 950 supernode-can completely equalize NVIDIA's GB200 and GB300 in terms of performance and price. The price must be expensive, but it is limited: 50%, 100%-it doesn't matter if it is 100%, it doesn't matter if it is 200%. For example, if it is 100% more expensive, I think it can be considered equal in price. The tasks can also be equalized: Huawei's supernodes can do all the tasks that the GB300 can do, and everything is the same with latency.
Nvidia may have a new generation in Q3 this year. So the gap between us and the United States in terms of chips, I think:There will be no more ecological gaps in the future, but in terms of chips, it will be four times two years.
There is also the question: Will we vertically integrate upstream? I hope not. We hope we don't have to do this, I hope it's someone else who does it. I don't want me to eat everything. I just have to eat one piece-the piece that I am best at, the piece that we think is the core, and the piece that is directly related to the user. For many of our industry partners, they are more concerned than we are. That should be someone else's business, not me to eat it all.
Will we build large clusters ourselves in the future? Building large-scale clusters is definitely necessary. We have been doing this ourselves. All our clusters are built by ourselves. But whether we want to develop our own chips in the future depends on how much the benefits are. If you run a power plant, you don't necessarily need to build generators, do you? Power generation equipment can be made by others. As long as the price is reasonable, why do you make it yourself?
So I hope I don't have to make chips, I hope I can buy chips at a reasonable price, so I don't have to make chips. We want to make just one piece. AI is a big thing and doesn't require me to do everything. If we focus, and I think the business benefits here are big enough-if the AI era will produce many trillion-dollar companies, I think we are one of them. We have been making a small piece of it, but one of them has no problems. I am not confident that I will do anything else. This is our attitude.
For example, at least in the two businesses of To B and To C, what you can see at present: if you really want to do a To C closed-loop, we will do it well; if you really want to do a To B closed-loop, you may not be as good as us. Good job. The more you want…
To B can say more. The upper limit of To B's business should still be demand. In the context of the current generation of AI technology, To B's demand should be limited: it will grow rapidly, but it is not an infinite thing, and will ultimately be subject to demand, not computing power. Under current technical conditions, how much income will ultimately depend on demand. Because you think about it this way: I can recoup the cost in ten months, and if there is a demand, I will definitely keep buying cards; I haven't bought more now because there is not so much demand. The demand should grow-as technological breakthroughs continue, the demand will grow.
14. Multimodal, Scaling and Low Cost
Multimodal layout, we have been doing it. It is very important for products and C-end user products. But for the upper limit of intelligence, it is a component, it is not the main thread itself. As components, we will definitely do multimodal and are doing it. We should discuss the relevant model: V4 and subsequent versions of V4 will support native multimodal. But our understanding of multimodal is that for intelligence, it is a component, and we don't regard it as intelligence itself-it has the same status as search. Search is also a component, and multimodal is also a component.
Scaling, we believe in Scaling: The larger the scale, the better the effect, and more functions can be unlocked. What prevents us from Scaling is actually computing power-it's not that we don't want to scale, it's that we don't have so much computing power to do this Scaling.
We have not touched where this upper limit is. We train such a big model not because I think such a big model is enough, but because I happen to have so many resources. I calculated based on my resources: how big is the model I can accept and train? This is how I calculated, but this model is not enough.
At present, the benefits of the model are still very obvious. We haven't had a chance to touch this Scaling wall yet, it's still far away from us. When Silicon Valley said that Scaling came to an end, that was for Silicon Valley; for China, we are still far away from that, and we have not scaled to that extent at all. This Scaling includes Scaling of data, Scaling of model size, and training costs. We are still far from exploring the upper limit of Scaling-we don't have that much computing power-but we will spare no effort to advance this upper limit of Scaling. Including after financing, we have more computing power and may train a larger model to buy it in the shorter time, the better; if we can spend all the money within half a year, that will be the best, but in fact we can't.
The core capabilities of the next generation model, I think it must have the ability to continue learning before it can be called the next generation model. Before that, all we can do is cost, and then we can do it better and faster. But to make a big breakthrough, it should have continuous learning.
Why do we seem to care so much about the computational efficiency of the model? I did find that not everyone cared much about model efficiency. For a commercial company, there is no incentive to pursue the efficiency of the model-several startups, you haven't heard them say that low cost is what they pursue, because that is not in their interests: low costs, what can you still make? You won't get any money.
Rather, it is part of our vision. Our classmates care about this cost because our classmates are ordinary people. He knows that using this will cost money. He can have empathy: others will have to spend money to use it, so if it is cheaper, others will be able to accept it. It will be higher. So many of our classmates hope that we can reduce costs. But if you think about it from a business perspective, you don't actually think about it that way-from a business, from a cost perspective, this is not the first priority.
For startups or large companies, the cost of this service is simply not high. But I hope it is relatively light, and I hope it is affordable-especially in the context of a shortage of computing power in China-that is affordable and can be used on domestic cards.
Low cost is a result first and foremost. Our model has indeed been architecturally moving in a lower-cost direction, which has something to do with our vision. We still have many methods in algorithm, and the cost can still go down. There is another reason why costs go down: the lower the cost, the larger I can train and afford larger models. With the same computing power and limited computing power, if my computing efficiency is higher, I can take on a larger model. For large companies, this is not necessarily the case-resources can be added and can be solved by adding resources-but we will give priority to cost efficiency.
15. Those who take less defeat those who take more
The range of data is relatively wide, and the data should be almost half of the model.
Why do I feel that if AI can account for 20% of GDP, if I want to account for 5% of this, it will never work? Because I will definitely be defeated by another person-if the other person says,"I only need one percent," then he will definitely be defeated. If my goal is to account for 5% of the GDP of all mankind, in theory this account is still valid-look at OpenAI, it seems to be able to calculate this account, and in theory it is no problem. But he has a problem: he will be defeated by another person willing to account for only one percent. Because another person said, I am doing so well, but I only need to get 1% of global GDP, and then I will defeat him. At this time, if another person comes out and says, I only need 0.1%, then the person in front will be defeated again.
From a macro perspective, no matter where these percentages are taken from, there is no difference between them.Those who take more will be defeated by those who take less.You don't even have to really take too much-if you take too much vision, you will be defeated by people who take too little vision. In fact, no one got the money, it was just a vision; if you took more, you would lose first, and you would face greater difficulties. This is the world.
OpenAI thought from the beginning that it could really monopolize the world, but in fact it would encounter many, many challengers. He will encounter challenges, and he will not be so relaxed. For the challenges that the United States has encountered, it may encounter challenges from China in the future-because China people are willing to take less, they can provide you with this service. In China, there will be people who will be willing to take less. But in the end, there will be a balance here: because after you take too little, the company's business logic will not hold true and you will not be able to survive. So if you take too little, you can't live; if you take too much, you will be defeated by those who take too little.
So for us, we do not want to make the most profits and maximize the profits, but only make a reasonable income. This is an explanation. I believe in this matter, I am not looking for reasons for this matter-there is no need to find reasons. I did it anyway, and I must have a reason. This reason may not be very common, but I don't think the company is used to it.
16. Organization method: two lines, no overtime
The management of our company is actually two lines: one line is from top to bottom, and the other is from bottom to top.
Bottom-up: Everyone does what they want to do, no one cares about them, no KPI.
From top to bottom: It's formal-if we want to do something collectively, we need the cooperation of the whole company. For example, if we want to launch V4, then we have to divide the labor, and everyone has to do part of it. That goes from top to bottom, and this one we call "formal".
Generally, we hope not to take up half of employees 'time--No more than half officially.He still has half of his time, which is not scheduled, and he can do whatever he wants. This is a scope of research that allows him to explore on his own. There are no pre-requisite requirements for what he explores according to what he thinks is important. As long as the company can support him and computing power can support him in doing it; or he does not need computing power and needs to do very little, then he does not need to coordinate at all. This is how we organize now.
Some people think we are from top down, some people think we are from bottom up, and I think both are right. One of my criteria is: it is best not to exceed half of the formal level.
We don't usually work too much overtime. There are two reasons for less overtime. First, doing research requires a relatively relaxed environment. If you press hard, you can't do research-that is, you have to have this interest and think about these issues in ordinary times, so you have to be able to explore in a relatively relaxed environment. This is due to the need of research culture. Second, we are very focused. Being very focused means that we have very little things to do, so I don't have so much to do and don't need to work overtime.
This is in line with the previous restraint: because I restrain, many times I stop doing it. The less things I have to do, the less work everyone gets. You see that many of our products are imperfect, and we have not made up for them. This is also one of our culture.
questioning session
Question: Can you share more on the timeline of when continuous learning leads to breakthroughs? Including achieving continuous learning, what architectural, algorithmic innovations, and other key elements are needed?
Answer: There are few people and research is needed. Now the whole world is studying this issue. For investors, what we see most now is Agents; but for those of us who study, what we see now is more about learning and how to solve the problem of learning.
In fact, learning may not be a technology, it is a problem. There may be many technologies on how to solve this problem-it is not one thing, it can be many things. Or, AGI is made up of many things, requires models, and then requires many other things. It is also an engineering and algorithm issue. This question is quite professional, but there are many methods and a lot of research.
Question: Thank you, Mr. Liang. First of all, I would like to respond and express my gratitude. What you said at the beginning moved me very much and inspired us a lot: this team, with the greatest goodwill, hopes to promote the development of this society and human intelligence and make a little contribution. This is very similar to the corporate culture of the company we serve, which is "cultivating oneself and cultivating others". I deeply understand why you led your team to do open source. I vividly feel that if I go to a bird paradise ecology like a banyan tree, it will benefit all things without competing for it, and it will be accepted by everyone. All things will coexist symbiotically and ultimately be everywhere. We will also express our recognition, support and respect for this mission and vision with this investment.
the first question, about ecological co-construction: Now that open source is available, how many partners, talents, and teams are there in this industry that can better replicate our open source models and achievements? If you want to further develop this ecosystem in the future, in which areas do you feel you need more high-quality talents to connect to our model and reproduce it? Or is GPU computing power relatively scarce now? Will everyone be a model matrix approach in the future-we will make the basic model better and better, and partners and teams from all walks of life will make vertical industry models and application models in the model matrix? How is the current development of this area? In two or three years, what kind of ecology do you think it will grow into?
the second questionRegarding AI hardware: AI companies around the world may invest hundreds of billions of dollars per unit. China currently seems to have some shortcomings in hardware computing power. How long do you think it will take to solve this problem, support the development of our AI, and prevent computing power and hardware shortcomings from slowing down AGI? Is this something that China will achieve in the future and will definitely achieve? It is just a matter of time and capital investment? At the same time, it may be twofold: on the one hand, the advancement of models will improve intelligence, and the consumption of hardware and computing power by a single task or some intelligent units will gradually decrease, and so much computing power will no longer be needed because the progress of models will make it smart calculations; on the other hand, technological advances in hardware will make computing power more energy-efficient. Could this be a path with both sides facing each other? In addition, whether it is 960 or H200 to invest hundreds of billions of dollars in computing power, you just mentioned that it is amortized in three years. Do you think its actual life cycle is four to five years? Will the Wanka cluster built now actually have relatively less advanced computing power in three years? Will there be a situation where there is not enough now and it will become redundant in three years? **
Answer:Thank you. The first issue is ecological.
We now feel that the problem faced by every company is not enough talents. But I think this talent shortage will be phased. In the early stages of the development of every industry, we were not enough talents. When we first started building websites, there were very few people building websites and there was a shortage of talents; later, when the Internet wanted to be a server, there was also a shortage of talents. However, this talent shortage will be solved very quickly, in just two or three years, because a large number of people will be trained.
The shortage of AI talents is also phased, and we have already seen it be greatly alleviated. There is really no shortage of AI people. Every company will cultivate people soon, and cultivating people is very fast. Therefore, as a whole, the AI industry, whether it is an ecosystem, model company or anything, there is no shortage of talents. The shortage of talents is definitely a short-term phenomenon. There has never been a long-term shortage of a certain type of people in history. I still remember saying more than ten years ago that there was a shortage of pilots and that the training period for pilots was very long, but it was quickly solved.
So no one needs to worry about the shortage of talents. Also, there are too many companies making models in China, or too many. There may be only three in the United States, but China has too many models. In the end, it will definitely not require so many people to do the basics, and it will definitely converge. Resources are also relatively scattered and wasteful to some extent-equivalent to every company doing the same thing, but in the United States, only three companies do it, and resources are only concentrated in these three companies; China's resources are very scattered, and each company receives less. I think this will definitely converge, but it requires a process and will definitely converge in the end.
You don't need so many. Now people may think that the profit rate of doing this thing is very high, so they must do it themselves; but when he finds out that this thing may not be so profitable, he may not do it. I don't believe there are such high profits because it does not conform to objective laws. This means that we are at a stage: if there is a very high profit margin, this must not comply with objective laws, and we should be a reasonable profit. This is the current situation of the industry, and I think it will definitely restrain itself.
To make the part of the model bigger, don't say that a certain company monopolizes it or "I want to take all the profits". This will definitely not work. If each company only earns a reasonable profit, then there is actually no need for so many people to make a big model. In the end, there were three or four competing companies in China, and the competition was sufficient, and the price was definitely enough to fight a price war. A large model may be enough for two large companies or two small companies.
As for ecology, I don't have too many ideas. We hope to support more people, but we don't have that much energy. We have the will and there is no conflict of interest, but whether we do it is another matter. But at least there is no conflict of interest here, and we hope for win-win cooperation. First of all, I definitely don't think that the big model company can take most of the profits. This is impossible-there are so many big model companies, and the gap is not so big now. The gap is only two things: one is time and the other is cost. So no one will have huge profits. People who control costs well earn more, and people who control costs poorly earn less. That's all.
ask: In terms of data, I believe that model companies already have channels to obtain the data that is publicly available. It is just a matter of time and cost. When we actually came to AGI in the future, everyone imagined that the ideal state was that the model could self-iterate, self-learn, and train itself. Do you think simulation data can be used in this area, or does real data have the highest quality? If real data is still needed, will it limit AI's intelligence to the level of human past-because it relies on real data that humans really once had? Or can we break this upper limit and make model capabilities surpass all realities in the past by simulating data, simulating data, creating data, etc.?
a: I think it can be surpassed. There are two things. For example, in Go, AlphaGo played a game that humans have never seen before-he must have surpassed humans to a certain extent. But it may also have upper limits and limitations, but we can't see this limitation now. Therefore, we generally believe that it can be transcended based on the knowledge that humans already have and that we can already speak out.
As for whether the follow-up depends on real data or simulated data, and whether it will work-I think there will be many methods.
So the gap between us and the United States may be 12 months, 12 to 18 months, or 6 to 12 months behind the United States. Simply put, it means being two years behind the United States, and then using only one-twentieth of the United States 'computing power to do this. This narrative is:One to two years behind, but only one-twentieth of its computing power. In the future, we will rewrite this narrative: use a fraction of its computing power, but shorten this time to 6 months or 3 months. I think this is a goal. and that we can even surpass them in some ways. However, when there is still an order of magnitude gap in overall computing power, it is unrealistic to comprehensively surpass it; but in certain key areas and trade-offs, some transcendence is possible.
Question: Two quick technical questions. First, when we talked about the technical route just now, we mentioned that the core issue we need to solve at this stage is continuous learning, which is also a hot topic in foreign research at present, called Recursive Improvement. At present, what is the biggest technical difficulty? From your perspective, when will it be resolved? Second, you just mentioned that we should first solve continuous learning and then develop intelligence. What are the technical roots behind this? Does this mean that after solving the continuous learning problem, DeepSeek will also develop general intelligence in the future?
Answer: The difficulty with the technical problem is that we have not found a way to work very well yet. The whole world has not found a good solution yet. Everyone is exploring it, and it is still in the exploring stage-I don't know who can find the next way to solve this problem. We have a lot of ideas, a lot of ideas that seem to have a future now, but none of them have yet worked. This is the first one.
Second, we now value internally a narrative that trains our next version of the model, which we hope will help us with our own development. It can improve DeepSeek efficiency-our model first and foremost improves DeepSeek's own efficiency, allowing us to develop the next version of the model, it can provide more help. Or to put it simply: the first goal of our model is not to use it well for everyone, but to use it well for ourselves. First it's useful to ourselves, and then it's faster when I develop the next version of the model.
The idea of many of us within us is this: First of all, it must be useful to ourselves, first of all, it must be used for ourselves-this is the fastest way to implement AGI. When we are easy to use ourselves, it means that it may be easy to use others, but we must first make sure that we are easy to use ourselves. This narrative is a bit strange, but it is true that many people think this: not that users use it easily, but that they hope to help ourselves more, so that we can implement AGI much faster. The logic of this narrative is: to help us achieve AGI-but first to help us achieve it, and we need this help to achieve AGI. It is now very certain that we do need artificial intelligence to help us achieve AGI. Although it doesn't work independently and just works with people, it's already very useful.
(About continuous learning first, then universal intelligence) Because solving the problem of continuous learning can greatly accelerate our research and development progress. If I solve continuous learning first, then the problem of general intelligence will be a problem. I have the assistance of AI. If AI can continue to learn, its ability should be very strong. Today's Agent capabilities are limited because it cannot continue to learn and cannot effectively continue to learn. If we can finish continuous learning first, AI's ability is very strong, and it can greatly improve the efficiency of our own research. By continuing to learn and do it first, general intelligence may be easy-it will be easy to do it with it. So this is the result we prefer to see: if we save more effort, we will relax. Otherwise, now you have to do general intelligence manually. It is a relatively tiring and bitter task, a data-intensive and labor-intensive task, and the cost performance is not high. I at least think it has to be able to continue to learn.
(Regarding whether domestic hardware can catch up) It may take several years for domestic hardware to catch up. In China, we must first solve the ecological problem-because ecology is a matter of confidence-and then solve the problem of production capacity. I think it should be solved step by step. I don't quite believe that we are still stuck with production capacity five years later. Now it is definitely stuck with the problem of production capacity. This year, next year, and the year after that, it may still be stuck with the problem of production capacity. However, after five years, I think it may not be the case. I am still relatively optimistic.
(Regarding the future organizational structure and personnel size) Our previous organizational structure was very decentralized because there was no organizational structure. But after people expand further, some changes will definitely need to be made. It can only be said that many changes will need to be made here, but it is difficult to express them all at once. In the end, we should still be divided into different departments: some departments need to establish a more rigorous hierarchical structure; other departments may still maintain a relatively loose and relatively flat structure. As the number of people increases, we will make this adjustment-it should be said that we will have to make this adjustment immediately, and I am already making it. Without this adjustment, many things cannot be moved forward. There are indeed many departments that should have organizational structures.
(Regarding the release rhythm) The version of V4 we are currently releasing online is still relatively crude, and many capabilities require time. For me, a relatively comfortable publishing pace is about one edition every two to three months. The last release may be at the end of April, and the next release may be at the end of June, which is roughly like this. If there are no surprises, each edition should still be better than the last.
At this scale of 50B activation, I feel that in the end we will not be much different from the current wave of open source-there may not be much difference in reasoning speed and effect. But there should be a big gap with its large size, undisclosed model. That gap should not be possible with our activation parameters. It must be a larger size model, such as 150B. So 150B, based on our current training progress, if we are optimistic, we can start training at the end of this year, at least on Thursday next year... the gap is still quite large. This is the gap with OCE.
Question: You often say that the process of AGI implementation is an incremental process, rather than a mutation. So I can understand, is this a process without critical points?
Answer: It has no critical point, but it is nonlinear. We now believe that one narrative is that AI can accelerate AI research. It's not linear because you can use AI to accelerate your own research, so it may be nonlinear in the end.
Question: Understand. So the previous conclusion is that it is enough for the language model to continue to scale, and this state can be achieved?
Answer: I can only say that I don't see any upper limit for Scaling the language model. Our current intelligence level, or the intelligence level of the United States, does not see an upper limit.
Question: Understand. You said that the model activated by the American 800B can be trained better than us, but it can't be used either, so we can only train it and it's difficult to give it to everyone because it's really a bit expensive. One point I was very curious about before: It has been nearly 100,000 years since humans have the ability to speak, and it may have evolved for 3.7 billion years. But when it comes to training AI, the order can be reversed, but it may still end up entering-and it may not be called the world model-the physical model or the embodied part, right? After this cap?
Answer: Yes, I think I still have to enter the body, and ultimately it will be the body. So for our company, naturally, the end point may be physical. Because for a normal person, his need is not a computer, right? A normal person, he eats, drinks, plays, clothes, housing and transportation, and does not need a computer. What he needs is to address specific human needs. If the goal is to solve manpower needs, I don't think it can be circumvented.
Question: Understand. Staged, if you reach a state where you can self-evolve and self-evolve better, close to that point of rise, I'm curious, what do you want the first AI to land in this state to be?
A: It might be different now. If there is no embodiment, what is our definition of AGI, or what do we want AGI to do? It helps me iterate the next version of the model. Then if we have an avatar, what we hope it will do is the same: let it iterate on the next version of the avatar, and it will make the next version of the robot.
Question: I'm still very curious. In a previous interview with DeepSeek, it was mentioned that taste and intuition are very important in some important direction choices and research choices, rather than simple engineering optimization. If AI can evolve on its own later, will this taste, taste, and intuition still matter? Or rather, what would be important?
Answer: AI now does not lack taste and intuition. What it lacks is the ability to continue learning. There is no problem with AI's taste and intuition-if you ask it to write an article, I don't think there is any problem with its taste and intuition.
Question: Continuous learning, continuous learning. You also mentioned that many researchers also mentioned that it is an unresolved research problem. The coding agent-especially catching up with MILES and Scaling to Office level and MIS level-is a relatively certain goal. Regarding an unresolved research goal and a relatively certain Scaling goal, how do you think research resources should be allocated, especially the talent resources of researchers, and how can we achieve the best balance and effect?
Answer: Model Office is a relatively definite goal; model MIS, I don't think it's a definite goal. It can only be said to be a goal. Model Office, I think it should be relatively certain.
This (continuous learning) does not take up resources. To do this kind of free research, it doesn't consume cards, it only requires very few cards, it requires ideas; it also doesn't consume human resources, because there is no need for someone to be there all the time. It's not a project, it requires a lot of people to be there thinking about it. So there is no need to allocate any resources to it because it does not need resources. You need resources to train the model, make the model, issue the model, and do efficiency experiments on the model. The one I just mentioned does not consume much resources on people or cards.
We call this "prize winning": the threshold is very low, and anyone can touch it, but who can find out what, maybe I don't know whether it depends on talent or what.
So we don't need to allocate resources here. It's just that what makes us different from other companies is that we spend time discussing this issue, thinking about it, and treating it as an important thing. In the company, it's an important issue that we spend time thinking about, but it doesn't require a lot of resources to do.
The hallucination problem of large models affects the user experience. There is also a way to solve the illusion problem, but this is a long-term proposition. The illusion problem can be regarded as a problem that can be solved through better Post-training. It is a problem that can be solved and improved, but everyone has not put a lot of effort into it. Or, for me, hallucinations are a problem, but we conclude it may be a product problem. We will solve it, but it is not the key issue.
There is also the problem of marking data. Our data annotation is related to our capital investment: with our capital investment structure, we cannot afford the cost of so many high-quality data annotation because the cost is very high. The cost of labeling data in the United States is no different from that of China. China's standard data has no cost advantage, especially for high-end data, which makes it difficult for us to invest in de-labeling data like the United States. This road is very difficult in China because the data is too expensive-whether it is for us to bid externally or ourselves, it is very difficult.
So now we basically walk on two legs. It's not that we can't bid at all, but that some data are low-cost and some are high-cost, so we bid for low-cost ones first. So you can also think that half of the people in our company are marking data now: half of the core researchers and the most important people are marking data, and we focus on marking data.To solve the AI problem, at this stage, we rely on standard data. The core is data.
Question: The first question. You just said that China's model is definitely more efficient than the American model, and there are other aspects that may be stronger than the United States in the future. In what aspects do you think we will be better than the United States in intelligence or other aspects?
A: I think in many aspects of experience, it is possible that we can do better than the United States. Not to mention my own experience-my product experience is quite good-I don't think our user experience will necessarily be worse than the United States. In terms of products and product capabilities, they are not necessarily worse than those of the United States, and costs should also be lower than those of the United States, so China may still be competitive.
In other aspects, if there are structural advantages, I don't think there is any. But in terms of cost and product, I think it does have certain structural advantages. Cost is easy to understand: they don't have to do it, so they don't develop this ability. They definitely don't pay much attention to this thing as we do-we can treat it as a very important thing, but to them, it's not important. The same is true for products. Many companies 'product capabilities are still acceptable. So these two aspects may have structural advantages.
Question: Second question. You just talked about post-training. The relatively high cost of our investment is, like Anthropic and OpenAI both invest huge amounts of money. After this financing, will we increase investment in post-training?
Answer: The gap is mainly in the labeling of high-quality data, mainly in AI research. Investment will definitely be increased. However, the bottleneck for labeling high-quality data is not capital investment, I think it is time, or it takes time. For OpenAI, for foreign countries, and for Anthropic, they are all earlier, and then they have more capital and more cards. We can think that we only started doing it in the last six months, so in terms of time, I think it takes more time. This has little to do with capital investment-even if there is no more capital investment, the original capital will be enough for it to expand at the fastest speed. But there is an upper limit to this speed. The bottleneck is not stuck in the money or the card. It is indeed in a rapid expansion process. Therefore, we feel that within a year, the issue of high-quality data will be relatively good, and it should be predictable domestically. I don't think it may be that difficult, but it does take time.
Question: Third question. We have seen Anthropic and others use their own models to make their own products, launching many vertical financial, legal... and even moving in the direction of medical care in the future. In terms of this vertical application, will we also consider it at some stage?
Answer: I haven't thought clearly about what the domestic business model will look like in the future, or what the smoothest path will be. We have not reached that stage yet. The domestic situation is not necessarily the same as the foreign situation. I think it is still difficult to judge what the domestic situation will be like at the time. Judging from the current situation, I think the most reasonable approach should be to make a general Agent, and the priority of other agents should be lower, including agents such as finance and doctors. Coding must be done first, because Coding Agents can do many things, and there are many vertical Agents. At this stage, we think the most important thing should still be the Coding Agent.
Question: In fact, we also do DeepSeek related things. I admire you for always doing DeepSeek in a very pure scientific research way. But now this industry has indeed reached the capital market and the road of capitalization. You are also a very responsible person, whether to your friends or investors. Between the direction of pure scientific research, pure AGI and the capital market-you will definitely have to go to the capital market in the future, and there will definitely be public shareholders-how do you think the balance will be in the future? What do you think about this?
Answer: Now I think it should be possible, so I want everything. If I can have hundreds of millions of dollars in B-side revenue this year, and we have users on C-side, then this already has a certain commercial foundation. Next year, we will have income from the B side. If this demand can increase further, the company will not be far from net profit. It may already be net profit, and it may not be a stage of pure burning money. So I feel that the movements that can be done and operated later should be relatively large.
Or, worst case scenario: Selling APIs may be able to support a listed company. If there are no new advances in technology and our technology will be frozen here, then in the end, we will try our best to sell APIs and do these services well. I think it will be enough.
So I still feel confident. It's really not that difficult. Because it is indeed in a place with high leverage and in a very fast-growing field, it may not be that difficult. I can only say thatWe hope to have bigger dreams, but we also have results that we can guarantee.
Question: I think the biggest difference between DeepSeek and other companies is that our organization is different from others. However, the organizational form is not only related to the goals, but also considers the organization's own boundaries and efficiency. From a macro perspective, will this organizational form have a good learning target? Historically, Bell Labs may be an ideal form? Or do we feel that we have no ideals ourselves and rely more on us to gradually explore them? Maybe in a new era, we can only rely on ourselves to explore-because our organizational form is definitely different from several companies in the United States, and the three companies themselves are different, but they will at least explore from the perspective of a commercial company. I mainly want to ask this question about the organization.
Answer: First of all, we have no object to imitate. At every step, we proceed from actual situations, seek truth from facts, make decisions based on actual situations, and find out what we should do. So it is a product of an era, a reaction to reality, not a result of imitation.
In this case, the optimal solution may be like this; this is the path we choose ourselves. We must have thought about and chosen every step, and the result of our choice is this way. It is not because we see who chooses this way, but because we choose this way because we analyze the pros and cons. It may be the same in the future, where we are not imitating anyone.
I think we are still different from Bell Labs: it clearly doesn't need to be commercial, but we clearly want to be commercial. We still have to survive in the end-we are a company after all, and the government will not give me a penny. So we can have very big missions, but in the final analysis, it is a company that needs to think about how to live. So the B-terminal is definitely important to us, because we may have to rely on it to survive in the future. Just saying it doesn't matter now because it's a cost line now.
So I think we are still different from Bell Labs, we are still essentially the same company. There have also been many companies in history that have pursuits other than profit, but it cannot be said that they have pursuits other than profit. They are not a company.Many companies do great things because they have a pursuit other than profit; in the end, that pursuit not only does not affect its commercialization, but can make it commercialized better.We are still essentially a company, just saying that we are thinking about what money to make, when to make it, how much money to make it, and what to make it-just that we have trade-offs.
Question: Two small questions. First, you mentioned that MILOS may not be a definite goal, but you will definitely move in this direction. You also mentioned that the next generation activation parameters may be around 150 to 250B-do you think 150 to 250B is benchmarking against O4.7, or may it be benchmarking against something else? Second, we now use some compilation languages other than CUDA on the entire inference. It turns out that based on the Nvidia ecosystem, PTX may be used more; now that we use more TileLang and others, will it significantly reduce our efficiency in infusion or reduce efficiency in the short term? What do you think of the efficiency loss caused by this change in compiled languages, or is it actually a state that can improve efficiency in the long run?
Answer: It is to improve efficiency, it is to significantly improve efficiency.
Question: There is no negative impact?
Answer: Yes, it is to greatly improve efficiency. So this is an opportunity. It is equivalent to the past that you were inseparable from the ecology of CUDA. Now we can abandon its ecology and use a simpler method, just use TileLang. It is a high-level language, and writing programs is fast. The amount of code to write is very small, so I can rewrite it again.
Question: Understand. So these two are actually relatively large opportunities brought by AI as you mentioned, not a shortcoming that needs to be made up for in the short term.
Answer: Yes, it is a big opportunity for technological development. We also have a project that is using AI to write TileLang.
Question: Is that faster?
Answer: Now all TileLang is written by humans, but it is much faster than the original ones who wrote CUDA.
Question: Understand. So even if it is the execution efficiency of the underlying hardware, it will not affect?
Answer: I think a loss of 1% to 2%, which is acceptable.
Question: I understand, I understand. Thank you, very clear.
Do others have any questions?... No problem, let's call it a day. Thank you very much for your time.