DeepSeek has recently completed its first external financing round since the company’s founding. The round raised more than RMB 50 billion (approximately US$7.4 billion), valuing the company at roughly RMB 367.5 billion (approximately US$54.3 billion) before the investment. Among the investors, DeepSeek founder Liang Wenfeng personally contributed RMB 20 billion; Tencent invested RMB 10 billion; CATL invested RMB 5 billion; NetEase, JD.com, and IDG Capital each invested RMB 3 billion; and the National Artificial Intelligence Industry Investment Fund invested RMB 1 billion.

Prior to this, Liang Wenfeng had repeatedly maintained the principles of “no fundraising, no IPO, and no commercialization.” This major financing round marks DeepSeek’s formal entry into the capital market and has sparked widespread discussion about its commercialization strategy and long-term technological vision.

At a recent investor meeting, Liang Wenfeng elaborated on DeepSeek’s organizational culture, open-source philosophy, technology roadmap, and views on the competitive landscape of the AI industry.

The following is Tencent Technology’s edited transcript of Liang Wenfeng’s remarks during the nearly four-hour meeting. The content has been organized by topic into 118 entries. The wording has been preserved as faithfully as possible, with only light editing for clarity.

1. Vision and Restraint

1. When we founded this company, we never started with the idea of how much money we would eventually make, whether we would enter the capital markets, go public, or anything like that. The first few dozen people who joined us never thought that way. If they had, they wouldn’t have come.

2. We are doing this with profound goodwill toward the world. We believe this is something beneficial to humanity—something that transcends money. Our original intention, our vision, and the vision we have maintained ever since have never been driven by the goal of maximizing commercial gain.

3. What governs a large company is not rules and regulations but vision. Vision is not a slogan hanging on the wall. Vision is reflected in how you act, not what you say. It is embodied in the way you actually operate.

4. We don’t really rely on organizational structures—we are driven by vision. Vision is what holds the organization together. We don’t operate by saying, “We have to hit this KPI,” and we don’t have performance assessments. We have only a shared vision.

5. This vision isn’t even written down. We’ve never formally documented it. It is embedded in the way we do things and in our attitude toward the world.

6. We don’t have many special advantages. We don’t possess extraordinary abilities. We aren’t wealthier than anyone else, nor do we have better people than other companies. Two years ago, when we founded this company, we had little money, few GPUs, no brand recognition, and no particular ability to attract talent. We were simply a group of very ordinary people.

7. The more restrained you are, the more likely you are to succeed—or at least that has been true so far. Otherwise, there’s no way to explain why we succeeded. We had no special weapons, started from a very low base, had extremely limited resources, and our people were simply an ordinary group brought together by chance.

8. AI is simply too big an opportunity. The potential benefits are enormous. That’s why we’ve been extremely restrained. As long as we succeed, the eventual rewards will already be immense. Even taking a tiny share of those rewards would be enormous, so at this stage there’s no need to think about which portion of the value we’re going to capture or how we’re going to capture it. The opportunity itself is already large enough.

9. Around last year’s Spring Festival, our user base suddenly surged. But we never tried to maximize user retention, monetize those users, or seize commercial opportunities from that traffic. We weren’t trying to compete for users or make money from them. Instead, we focused on one thing: doing everything we could to serve them well.

10. We’ve never had the ambition of building the next super app or competing with anyone to become the next ByteDance or the next Tencent. We’ve never thought that way. I believe the opportunities presented by AGI will be enormous, and those opportunities will remain enormous for a very long time.

11. Restraint is a strategy. Sometimes you give up certain things in order to gain something much greater. Open source is the same. You can see it either as a burden we take on or as value that we’re choosing to give away.

12. My understanding is that this kind of restraint increases our long-term probability of achieving AGI. I have no doubt that AGI will have tremendous commercial value. With that as the premise, my first priority is not how to capture a larger share of the market or a greater share of the profits. My first priority is how to maximize our chances of actually succeeding.

13. We’ve always been very restrained. We don’t want to become the enemy of any internet company, large or small. I hope we can empower others. I hope we can help everyone accomplish this and contribute to the industry’s progress.

14. Looking back, I don’t think this approach has caused us to lose anything. We haven’t lost out because we’re open source, because we’ve acted in good faith, or because we’ve helped others. On the contrary, it may even have worked in our favor. It seems counterintuitive, but that’s genuinely been our experience.

15. Our ultimate goal is AGI, but we’ve also been pursuing commercialization all along. That’s why we have consumer users and enterprise revenue. Judging from experience so far, this strategy has been successful.

2. AGI Roadmap

16. If you can describe a problem clearly enough—providing complete context and complete instructions—then today’s AI already outperforms humans. But there’s an important premise here: you must provide complete context and complete instructions.

17. AI cannot replace your employees today. But if AI acquires the ability to learn continuously—if it can spend two months learning at your company just as a new employee would—then it could replace virtually anyone. So the next major step we’re missing is continuous learning.

18. The development of AI can be understood as climbing a staircase. Last year’s step was Chain of Thought (CoT). We discovered that reasoning through chain-of-thought enables AI to reach a significantly higher level of intelligence.

19. This year’s step is the Agent. We’ve found that an Agent-based approach allows AI to handle a much broader range of tasks and raises its upper limit of intelligence. Agents rely on Chain of Thought, and Chain of Thought relies on the previous step—the language model. None of these stages is wasted; each builds upon the one before it.

20. After Agents, the next problem we need to solve is continuous learning—how to enable a model to keep learning over time, rather than depending solely on a single, large-scale training run. It should be able to learn continuously over long periods, much like a human being.

21. Once continuous learning is achieved, we may arrive at what people call the singularity. At that point, a model capable of continuous learning would be able to perform everything a human can do. It could develop its own next version, conduct its own research, and create even more advanced AI models on its own.

22. That “singularity” isn’t really a single moment. It’s more likely to be a gradual process—a long period of continuous change rather than an abrupt leap. We simply call it a singularity out of convention.

23. Our view is that the sequence should be: first solve learning to learn; then reach the self-iterating intelligence singularity; only after that should we move into embodied intelligence. Once AI becomes embodied, it can enter the physical world—it can do household chores for you or help care for the elderly.

24. If we first solve continuous learning, then self-iteration, and only then embodied intelligence, the rest of the journey becomes much easier. That’s because each stage can be used to accelerate development of the next one.

25. We focus only on the main path toward AGI. AI is an extremely broad field, and there are many areas that we don’t believe belong to that main path—for example, 3D generation or video generation. In our view, those aren’t closely related to the core problem of intelligence, so we won’t pursue them.

26. When video generation first emerged, it became incredibly popular. It almost seemed as though every AI company had to do it, and that if you weren’t building video generation, you weren’t a real AI company. I found that strange. If you think about it carefully, it really has very little to do with the roadmap toward intelligence.

27. Commercially, it’s a good business. There’s no denying that. But it isn’t fundamentally about intelligence. We won’t pursue something simply because it’s a good business. We’ll only do it if it’s part of the roadmap toward intelligence.

28. In our judgment, neither world models nor embodied intelligence are the most important priorities at this stage. What matters most right now is AI training, and after training, solving the problem of continuous learning. That’s our assessment as a company. Of course, every company has its own view.

29. We increasingly believe in the narrative that AI can accelerate AI research. In other words, progress won’t be linear. AI can help speed up its own development, so eventually the pace of advancement may become nonlinear.

30. Ultimately, embodied intelligence is indispensable. Think about an ordinary person’s needs. People don’t actually need computers. What they need is to eat, travel, enjoy life, and take care of daily necessities. Those needs ultimately require embodied intelligence capable of acting in the physical world.

31. What do we hope AGI will do? First, we hope it can help us develop the next generation of AI models. And once embodied intelligence arrives, we hope it can also build the next generation of embodied systems—that it can design the next generation of robots.

32. The defining capability of the next generation of models must be continuous learning. Without that, it isn’t truly the next generation. Before reaching that point, all we can really do is reduce costs, improve performance, and increase speed. But genuine breakthroughs require continuous learning.

33. Today’s Agents are limited because they cannot learn continuously. If we first solve continuous learning, AI’s capabilities will become extraordinarily powerful. More importantly, it will dramatically improve the efficiency of our own research.

34. Once continuous learning is achieved, general intelligence should become much easier to realize because AI itself can help accomplish it. That’s the outcome we’d most like to see—it makes the whole process much less labor-intensive. Otherwise, building AGI manually is an exhausting, data-intensive, and labor-intensive endeavor with relatively poor cost-effectiveness.

3. Team and Talent

35. What our past experience has taught me is that the vision of AGI is incredibly powerful. A talent advantage doesn’t mean my people are smarter than yours. It comes from how you organize talented people, how you inspire them, and how you enable them to collaborate.

36. Simply putting smart people together doesn’t mean they’ll naturally work well together or become passionately committed to a common goal. That’s why you need a shared vision.

37. Our greatest core interest is maintaining the stability of the team. That’s our most important interest—arguably our only real core interest. As long as we can keep the team stable, I’m confident we’ll succeed. We’ll achieve AGI. It’s as simple as that.

38. Money isn’t the issue. Resources aren’t the issue. Everything else is relatively easy to obtain. For us, there is only one non-negotiable priority: maintaining the stability of the team.

39. That is also our greatest challenge—or, put another way, our biggest risk. Fortunately, our recent financing has substantially reduced that risk because everyone has received meaningful stock options with considerable value.

40. As far as team stability is concerned, as long as the key employees and the earliest members of the company stay, most others won’t leave either. Even if some people receive fewer stock options or lower compensation, they still won’t leave, because they didn’t come here just for the money. They came because they want to work in an environment where AGI can actually be built.

41. Everything else is simply a matter of time. At worst, it might delay us by six months or a year, but it won’t prevent us from succeeding. We’re certainly not short of money, and we’re certainly not short of resources. Those are not the constraints.

42. Our primary gap with the United States lies in resources, not in people. The difference in talent is actually quite small. In many cases, they’re the same people—just Chinese researchers who ended up in different places. Some stayed in China, some remained overseas. It wasn’t that the smartest people all went abroad. That’s simply not true.

43. Talent isn’t the bottleneck. Resources are. Resource constraints affect talent development because limited computing power means fewer opportunities to conduct experiments. That’s why, overall, our talent base still lags behind the United States. Fundamentally, the talent gap is itself a consequence of the compute gap.

44. The shortage of AI talent is only temporary, and we’ve already seen significant improvement. In reality, there’s no shortage of people. Companies can train AI researchers very quickly.

45. Right now there are simply too many companies in China building foundation models. There are probably only three major players in the United States, whereas China has far too many. Eventually the market will converge. There simply won’t be a need for so many companies developing foundation models.

46. Our company is managed through two parallel systems. One runs from the top down, while the other runs from the bottom up. The bottom-up system means everyone is free to pursue what they want to work on. Nobody tells them what to do, and there are no KPIs.

47. Ideally, we want our employees to have about half of their time completely unscheduled. They should be free to explore whatever they think is important. Research requires that kind of freedom—there are no predefined assignments.

48. We generally don’t work overtime. There are two reasons. First, research requires a relaxed environment. If people are under constant pressure, genuine research becomes impossible. Since researchers need to think deeply about problems on their own, they need an environment that allows for exploration rather than relentless pressure.

49. Second, we’re extremely focused. Because we’re so focused, there simply aren’t that many things we’re trying to do. As a result, there’s no need for excessive overtime. This is entirely consistent with the philosophy of restraint I mentioned earlier.

50. As a company, we’re fundamentally built on consensus. I don’t make every decision by myself. My authority and influence within the company come from building consensus, not from issuing orders.

51. Our decision-making process is essentially a process of building consensus. It’s not that I can simply push something through because I want it done. Only when there is consensus can an initiative move forward, and only then will I push it.

52. As the company grows, we will make adjustments to the organization. In fact, we’re already doing so. Without these adjustments, many things simply can’t move forward. Some departments genuinely need a formal organizational structure.

4. Computing Power and Resources

53. How many GPUs do we need? As many as possible. Within what we can afford, the more GPUs we can acquire, the better. There’s no question about that. Our current strategy is straightforward: at a reasonable price, we’ll buy as many GPUs as we can.

54. In reality, it’s surprisingly difficult to spend that much money because there simply aren’t enough GPUs available. They’re hard to buy, and they’re expensive. We can’t just pay any price—we still have to make sure the price is reasonable. If we manage to spend RMB 20 billion on procurement this year, then our purchasing department will have done an exceptional job.

55. Our biggest gap with the United States is in resources. On the one hand, GPUs are difficult to obtain domestically. On the other hand, our capital investment is much smaller than that of U.S. companies. Salaries account for only a small share of AI investment. People talk about companies offering US$100 million compensation packages, but even then, payroll is only a small fraction of the total. The overwhelming cost is computing power.

56. Every difference we observe—whether in talent, model performance, or applications—can ultimately be traced back to differences in compute resources.

57. Compared with the United States, we’re probably 12 months behind—or perhaps 12 to 18 months, maybe even 6 to 12 months. Put simply, we’re roughly two years behind, but we’re accomplishing the same work with only one-twentieth of their computing power.

58. That’s the narrative today: we’re one to two years behind, yet we’re doing it with just one-twentieth of the compute. In the future, we want to rewrite that narrative. We still want to use only a fraction of their computing power, but reduce the time gap to six months, or even three months. That’s one of our goals.

59. We believe in scaling. The larger the scale, the better the results, and the more capabilities can be unlocked. The only thing preventing us from scaling is computing power. It’s not that we don’t want to scale—we simply don’t have enough compute to do it.

60. We train models at the scale we do not because we think that size is sufficient, but because that’s what our available resources allow. The size of our models is determined by what we can realistically afford to train—not because we’ve concluded that’s the optimal size.

61. When people in Silicon Valley say scaling has reached its limits, they’re speaking from Silicon Valley’s perspective. For China, we’re still a long way from that point. We haven’t yet reached that level of scaling—whether in terms of data, model size, or training investment.

5. Domestic Chips and the Ecosystem

62. NVIDIA’s CUDA moat is rapidly eroding. One reason is the emergence of AI itself. AI now makes it much easier to build alternative software ecosystems because AI can write code.

63. The market for AI accelerators has already surpassed the gaming GPU market. There’s no reason these two product lines should remain tied together. The trend is that they will become completely separate. Whether it’s Huawei or NVIDIA, future chips will increasingly be purpose-built AI chips rather than general-purpose GPUs as we’ve known them.

64. We believe domestic AI chips are facing a historic opportunity. Within the next year, I think we’ll see one thing conclusively demonstrated: the ecosystem surrounding domestic chips is fully viable. Many people currently believe these chips are difficult to use or that the ecosystem is immature. Within a year, I believe the facts will overturn that perception.

65. There are no fundamental problems with either the hardware or the software ecosystem of domestic AI chips. The only real constraint is manufacturing capacity. As for software compatibility, there are no major obstacles, and NVIDIA can’t stop that trend. In a normal commercial environment, where NVIDIA chips are readily available, domestic substitution would be much more difficult. But when NVIDIA chips can’t be obtained, everyone is forced to adopt domestic alternatives.

66. During the training of V3, we still used NVIDIA hardware, but we no longer relied on NVIDIA’s software ecosystem. We wrote our own high-level compiler called TileLang, and built the rest of our software stack around it. As a result, we now depend very little on NVIDIA’s ecosystem, even though we’re still using NVIDIA hardware.

67. I’m optimistic about China’s domestic computing infrastructure. In this respect, I think NVIDIA is digging its own grave. Huawei’s supernode architecture—its 950 Supernode—can fully match NVIDIA’s GB200 and GB300 in both performance and cost.

68. Four Huawei accelerator cards are equivalent to one NVIDIA card.

69. As far as chips are concerned, I don’t think there will be any long-term ecosystem gap between China and the United States. The remaining gap is simply one of hardware performance—roughly a factor of four—and about two years of technological lead time.

70. Our primary collaboration is with Huawei. Huawei handles much of the adaptation work themselves, but we’re also deeply involved in building that ecosystem. Huawei’s main challenge is still insufficient production capacity.

71. I don’t believe production capacity will still be our bottleneck five years from now. Certainly, it’s a bottleneck today—and probably will remain one this year, next year, and perhaps the year after. But five years from now, I’m not so sure. I’m actually quite optimistic.

6. Competition and Industry Outlook

72. Ultimately, the differences in model performance will come down to an overall combination of factors. Any comparison between models only makes sense if they’re evaluated at the same cost. It’s like comparing two cars—you compare vehicles in the same price range.

73. Has Anthropic now surpassed OpenAI? Is that a long-term trend? I don’t think so. It’s only temporary. In the long run, OpenAI and Google will most likely continue overtaking one another in cycles.

74. In the global division of labor in AI, Chinese companies are likely to play the role of the largest producers. That’s simply because China has the greatest manufacturing capacity. The same is likely to be true for AI chips—we may ultimately have the largest production capacity. We also have the largest electricity supply.

75. Chinese companies will build the lowest-cost products while steadily improving performance. Today, many products made in China are already comparable in quality to those made in the United States. AI may evolve in the same way. Chinese AI is likely to be cheaper, and that lower cost will probably be structural, much like China’s cost advantage across many other industries.

76. In the end, competition will come down to three things: cost, timing, and user experience. Beyond those three factors, I don’t think there will be major differences.

77. Cost is certainly the biggest differentiator. I’d rank it first. The second is timing—when you’re able to achieve something. Being a few months earlier or later can make a meaningful difference.

78. At first, OpenAI believed it could monopolize the world. In reality, however, it has encountered—and will continue to encounter—many challengers. Once competition appears, things become much harder. The United States will face competition, and eventually it may also face competition from China, because Chinese companies are willing to accept lower profits in order to provide the same services.

79. Those who seek to take more will ultimately be defeated by those willing to take less. In fact, you don’t even have to be making excessive profits. If your vision is centered on taking more, you’ll lose to people whose vision is centered on taking less. At this stage, no one has actually made much money yet—it’s a matter of mindset. If your guiding principle is maximizing profit, you’ve already put yourself at a disadvantage.

80. For us, the goal isn’t to maximize profits or optimize pricing for the highest possible return. We only seek a reasonable profit. That’s genuinely what I believe; I’m not inventing a justification after the fact. There’s simply no need to.

81. In many aspects of the user experience, I think we may actually outperform the United States. Our product capabilities are not necessarily inferior. Our costs should also be lower. That’s why I believe China will remain highly competitive.

82. The cost advantage is easy to understand. They don’t need to focus on cost optimization, so they never develop that capability. For us, it’s a top priority. For them, it isn’t.

83. In the end, the market for foundation models probably only needs a handful of companies—not dozens. The real differentiators are only two things: time and cost. I don’t think any single company will enjoy enormous monopoly profits. Companies with better cost control will earn somewhat more; those with poorer cost control will earn somewhat less. That’s all.

7. Model Development and Technology

84. Within our company, about half the people believe OpenAI still has the stronger models. Anthropic currently benefits from a first-mover advantage, but I don’t think that advantage will last very long. It isn’t something they can hold onto indefinitely. All three companies—OpenAI, Anthropic, and Google—are outstanding. Among them, the most important difference is efficiency: who spends less money and burns less capital to achieve comparable results.

85. We’ve always been working on multimodality. It’s very important from a product perspective, especially for consumer-facing products. But in terms of raising the ceiling of intelligence, it’s just one component—it isn’t the main path.

86. We’ll certainly release multimodal models. V4 and its subsequent versions will support native multimodality. But from the standpoint of intelligence, we see multimodality as a module rather than intelligence itself.

87. As far as language-model scaling is concerned, I haven’t yet seen any upper limit. At our current level of intelligence—or even at the level reached in the United States—I don’t think we’ve encountered the ceiling yet.

88. Many people within our company share the same philosophy: the first priority is that our models must be useful to ourselves. That’s the fastest path toward AGI. If they’re genuinely useful for us, they’ll probably be useful for others as well—but first they have to work for us.

89. The primary goal of the models we build isn’t to make something everyone else finds easy to use. The first goal is to make them useful for ourselves. Once they’re useful internally, we’ll be able to develop the next generation of models much faster.

90. Internally, we call this “drawing the lottery.” The barrier to entry is very low—anyone can try. But whether someone discovers something valuable depends on factors I can’t fully explain; perhaps it’s talent, perhaps something else. That’s why we don’t allocate resources to it in advance. What distinguishes us from many other companies is that we’re willing to devote time to discussing these questions, thinking deeply about them, and treating them as genuinely important.

8. Commercialization and Pricing

91. Our API pricing is designed to generate a reasonable profit. Roughly speaking, if we buy a batch of equipment on the market, recovering the investment in about ten months represents a reasonable level of profitability.

92. If our objective were to maximize profits, we would price the API much higher. At the current price range, demand is relatively inelastic. Even if we increased prices by 50 percent—or even doubled them—token consumption probably wouldn’t change very much.

93. When we first released one of our models, we were worried demand would be overwhelming, so we initially set the price relatively high. The team wasn’t very happy about that. Later, I reduced the price to one-quarter of the original level, and everyone became much happier.

94. The ceiling for enterprise AI (To B) is ultimately determined by demand. Given the current generation of AGI and AI technologies, enterprise demand is still limited. It will grow rapidly, but it isn’t unlimited. In the end, demand—not computing power—will be the limiting factor.

95. My current view is that we should pursue every opportunity that is realistically achievable. If this year we can generate several hundred million dollars in enterprise revenue while continuing to grow our consumer user base, then we’ve already established a solid commercial foundation. If enterprise demand expands further next year, profitability won’t be far away—we may already be profitable.

96. Even in the worst-case scenario, simply selling APIs could support a publicly listed company. If technological progress were to stop tomorrow—if our technology froze where it is today—we could still focus entirely on API services, deliver them well, and build a sustainable business.

97. Given where we are today, I think the most sensible strategy is to devote ourselves fully to building a general-purpose Agent. Other vertical Agents—whether for finance, medicine, or other industries—should have lower priority. Coding should come first, because Coding Agents unlock many broader capabilities. At this stage, the Coding Agent is our top priority.

98. Low cost is first and foremost an outcome. We’ve consistently designed our model architecture to reduce costs, and that’s closely tied to our vision. Beyond architecture, we also have many algorithmic approaches that can drive costs down even further.

99. Lower costs have another important benefit. The lower our costs, the larger the models we can afford to train. With the same amount of computing power, higher computational efficiency enables us to train larger models.

9. Open-Source Strategy

100. We will continue to open-source our models, and I believe that even our most advanced models will eventually be open-sourced. I simply don’t see any compelling advantage to keeping them closed. Take ByteDance’s models, for example—they’re closed source. What clear advantage has that brought them? I don’t see one.

101. Even if you open-source a model and disclose everything, the barrier to entry remains extremely high. It’s difficult enough to make the model work well. It’s even harder to make it work efficiently and at low cost. None of this is easy.

102. Open source doesn’t reduce our revenue. I don’t believe open sourcing has any negative impact whatsoever on our business model.

103. I’m not worried at all that others will deploy our models and compete with us. In fact, we hope they do. We’ll do everything we can to help the open-source community deploy our models successfully.

104. Whenever we interact with the outside world, our position is clear: we focus exclusively on the main path toward AGI. We’re willing to help anyone—including competitors such as Alibaba, Zhipu, or Moonshot AI—build better products. We lose nothing by doing so, because we’re open source to begin with.

105. Is the model we release as open source the same one we deploy ourselves? Yes—it is exactly the same. We don’t open-source an inferior model while keeping a better one for internal deployment. They’re identical.

10. Data and Post-Training

106. Data accounts for almost half of a model’s success. Before that, there’s also the challenge of data annotation. Our investment structure simply can’t support annotation at the scale of the highest-quality datasets, because the cost is extremely high.

107. The cost of data annotation in the United States isn’t significantly different from that in China. China doesn’t enjoy a cost advantage, especially for high-end data annotation. That makes it difficult for us to invest in annotation on the same scale as American companies. In China, this is an extremely expensive path, whether we outsource the work or do it ourselves.

108. At present, we’re pursuing a two-pronged approach. It’s not that we can’t annotate data—we simply prioritize the lower-cost annotation tasks first, because some kinds of annotation are much more expensive than others.

109. You could say that about half the company is engaged in data annotation. Roughly half of our core researchers—the people most critical to our work—are focused on it. At this stage, solving AI depends heavily on high-quality data annotation.

110. The biggest bottleneck in high-quality data annotation is simply time. OpenAI, Anthropic, and other leading companies started earlier, have more capital, and possess more computing resources.

111. Hallucinations remain one of the biggest factors affecting user experience in large language models. They can be mitigated, but it’s a long-term challenge. Better post-training methods can improve the problem considerably, even if they don’t eliminate it entirely.

11. Organization and Company Positioning

112. We have no company to imitate. Every decision we’ve made has been grounded in reality. We assess the actual situation, seek truth from facts, and determine the right course accordingly. DeepSeek is a product of its time and circumstances, not the result of copying someone else’s model.

113. We’ve always intended to commercialize our technology. At the end of the day, we’re still a company, and we have to survive. The government isn’t going to give us money.

114. Fundamentally, we are still a business. The difference is that we’re selective about what money we earn, when we earn it, how much we earn, and how we earn it. Many great companies have pursued goals beyond profit, and those goals ultimately strengthened—rather than weakened—their commercial success.

115. We selected our financing partners very carefully. Above all, we wanted investors whose interests were closely aligned with ours—people who genuinely wanted us to succeed and had the least incentive to work against us. Not everyone wants us to succeed, because our success inevitably threatens the interests of others.

116. Today’s AI doesn’t lack taste or intuition. What it lacks is the ability to learn continuously. If you ask AI to write an article today, I think its taste and intuition are already quite good. Continuous learning is the missing piece.

117. We only want to focus on one part of AI. The field is vast; it doesn’t require us to do everything. If we remain focused—and if the commercial opportunity within our chosen area is already enormous—that’s enough. I believe the AI era will create many trillion-dollar companies, and we intend to be one of them.

118. We hope to support many more people, although our resources are limited. We have the willingness to do so, and there is no conflict of interest in helping others. Whether we have enough capacity to do it is another matter. But at the very least, we believe in cooperation and mutual success rather than zero-sum competition.

Editor: Zhiyu Wang