4小時、118個回答，梁文鋒內部交流回應一切

Content Compiled by: Gu Lingyu

Original Editors: Xu Qingyang, Su Yang

Source: Tencent Technology

DeepSeek recently completed its first external financing round since its inception. The total fundraising amount exceeded 50 billion RMB (approximately 7.4 billion USD), with a pre-investment valuation of about 367.5 billion RMB (approximately 54.3 billion USD). Among the investors, DeepSeek founder Liang Wenfeng personally contributed 20 billion RMB, Tencent contributed 10 billion RMB, CATL contributed 5 billion RMB, while NetEase, JD.com, and IDG Capital each invested 3 billion RMB. The National Artificial Intelligence Industry Investment Fund contributed 1 billion RMB.

Prior to this, Liang Wenfeng had proposed a principle of "no financing, no IPO, no commercialization." This large-scale fundraising marks DeepSeek's official entry into the capital market and has also sparked widespread industry attention regarding its commercialization path and technological vision.

In a recent investor meeting, Liang Wenfeng elaborated in detail on DeepSeek's organizational culture, open-source logic, technology roadmap, and views on the industry's competitive landscape.

Below is a transcript of Liang Wenfeng's speech during the nearly 4-hour meeting obtained by Tencent Technology, organized by topic into 118 points. The text retains the original meaning as much as possible with only slight edits.

01 Vision and Restraint

1. When we first started this company, our original intention wasn't about how much money we would eventually make, going to the capital market, or going public. The initial few dozen people never thought about this. If they had, they wouldn't have joined.

2. We started this with immense goodwill towards the world. We believe this is useful for humanity; it's something beyond money. Our original intention, our vision, and the vision we maintain to this day are not based on maximizing commercial interests.

3. Managing a large company doesn't rely on rules and regulations but on vision. Vision is not a slogan on the wall; it's how you act, not how you speak, and how you actually operate.

4. We have no formal organization; it's vision-driven, organized by a vision. We don't operate by "achieving certain KPIs" without assessment; we only have a vision.

5. This vision isn't even written down. Nothing was ever documented. The vision lies in how we do things and our attitude towards the world.

6. We don't have many other advantages. We have no special skills. We aren't wealthier than others, nor do we have better talent than other companies. Two years ago, when we founded this company, we didn't have much money, many GPUs, fame, or appeal. We were just a group of very ordinary people.

7. The more restrained you are, the easier it might be to succeed, or at least this has been proven true so far. Otherwise, there's no way to explain how we succeeded: we had no weapons, a very low starting point, few resources, and our people were just a random group of ordinary individuals.

8. AI is too big; the stakes are too high. We are very restrained. As long as we succeed, the benefits will be enormous. Even a small share will be huge, so there's no need now to think about which part of the benefits to take or how to take them. The benefits are already large enough.

9. Last Spring Festival, we suddenly gained many users, but we didn't seek to retain them, monetize them, or seize commercial interests from them. We didn't compete for users or make money, but we worked hard to serve them well.

10. We have no intention of becoming the next super app, competing with others, or becoming the next ByteDance or Tencent. We believe the future opportunity for AGI is immense, and it will always be immense.

11. Restraint is a strategy. Sometimes you can give up something to gain something else. Open-sourcing is the same; it can be seen as our pressure or our concession.

12. I understand that such restraint can, in the long run, increase the probability of us achieving AGI. When considering something, I have no doubt AGI will have enormous commercial value. On this basis, my priority isn't how to increase my share but how to increase the probability of success.

13. We have always been very restrained, unwilling to become rivals with any internet giant or startup. I hope to empower them or assist everyone in doing this.

14. I believe that this attitude hasn't caused us to gain less. Open-sourcing, our goodwill, or helping others hasn't resulted in us gaining less. On the contrary, it might have added value. This seems counterintuitive, but it's true.

15. We aim for AGI but have been commercializing, hence our C-end users and B-end revenue. Historically, this strategy has been successful.

02 AGI Roadmap

16. If you can describe a problem clearly, providing complete context and instructions, it already surpasses humans. But this definition has a premise: you give it complete context and instructions.

17. AI cannot replace your employees. But if AI has continuous learning ability, it could learn like an employee for two months and replace everyone. So we are still one step away from continuous learning.

18. AI development can be understood as a staircase. Last year's step was chain-of-thought (CoT). We found that CoT can achieve higher intelligence levels.

19. This year's step is Agent. We found that with Agent, even more tasks can be done, expanding its capability range and raising its intelligence ceiling. Agent uses CoT, and CoT uses the previous steps, like language models, so no step is wasted.

20. After Agent, the next issue to solve should be continuous learning—how to make the model learn continuously without strong training, like a human learning over a long period.

21. After continuous learning, we might reach a singularity. Once the model can learn continuously, it can do everything humans can, develop its own versions, research, and develop more advanced AI models.

22. This singularity isn't a singularity; it's also a gradual process. It might be a long gradient, not a mutation. But we habitually think of it as a singularity.

23. This is our speculation on the timeline: first solve learning to learn, then the intelligence singularity of self-iteration, then embodied intelligence. After embodied intelligence, it enters the real world, doing chores and providing elderly care.

24. If we solve continuous learning first, then the self-iterating singularity, then embodied intelligence, the path becomes easier. Later technologies can help develop subsequent ones.

25. We only focus on the main track of AGI. AI is vast, and many things, like 3D and video generation, are not on this main track. We won't pursue them.

26. Video generation was very popular at first, as if it were a must-do, or you wouldn't be considered an AI company. This is strange because, upon reflection, it has little to do with the intelligence roadmap.

27. Commercially, it's a good business, but it has nothing to do with intelligence. We won't do it just because it's a good business; we only do things on the intelligence roadmap.

28. From our judgment, world models and intelligence are not the most important things at this stage. The most important are AI training and solving continuous learning after training. This is our company's judgment, though different companies may differ.

29. We believe in the narrative that AI can accelerate AI research. It's not linear; using AI to accelerate research makes it potentially non-linear later.

30. I think embodied intelligence is necessary eventually. For a normal person, needs aren't just computers. People need food, clothing, housing, and transportation—all requiring human-like solutions—so embodied intelligence is needed to address specific human labor demands.

31. What do we want AGI to do? Help iterate the next model version. With embodiment, we want it to iterate the next version of embodiment and the next robot.

32. The core capability of the next-generation model must be continuous learning to be called next-generation. Until then, we can only improve cost, effectiveness, and speed. A major breakthrough requires continuous learning.

33. Current Agent capabilities are limited because they can't learn continuously. If we solve continuous learning first, AI's capabilities will be very strong, greatly improving our research efficiency.

34. If continuous learning is achieved first, general intelligence becomes easy. This is the outcome we hope for—easier and more efficient. Otherwise, manually building general intelligence is labor-intensive, data-intensive, and low in cost-effectiveness.

03 Team and Talent

35. My past experiences taught me that the AGI vision is very powerful. Talent advantage isn't about having smarter people but about organizing, motivating, and collaborating with them.

36. Gathering smart people doesn't automatically lead to collaboration and passionate pursuit of a goal. You need a vision.

37. Our core interest is maintaining team stability. This is our most important, perhaps only, core interest. If we maintain team stability, we will succeed in achieving AGI.

38. Money, resources, and other factors are easy to obtain. For us, only one core interest is non-negotiable: maintaining team stability.

39. This is also a major challenge, or perhaps the biggest risk. However, this risk has been significantly mitigated by our recent financing, as everyone received substantial options.

40. For team stability, as long as the most important and longest-serving employees are stable, others are unlikely to leave. Even with fewer options or lower pay, they stay. Everyone comes not just for money but to work in an environment that can achieve AGI.

41. Everything else is a matter of time, causing at most a half-year or year delay, but not failure. We definitely have enough money and resources.

42. The gap between us and the US mainly lies in resources, not people. There's almost no gap in talent; it's the same group of Chinese people. Some stay in China, some go abroad—it's not that smarter people go abroad.

43. Talent is not the bottleneck; resources are. Resources affect talent development; fewer GPUs mean fewer experiments, so our talent lags behind the US. The talent gap is essentially due to the computing power gap.

44. The shortage of AI talent is temporary and has already been significantly alleviated. AI talent isn't scarce; companies can quickly train people.

45. Currently, too many companies in China are building models. In the US, there are about three; in China, far too many are building base models. Eventually, the number will converge.

46. Our company management has two lines: top-down and bottom-up. Bottom-up means everyone decides what to do independently without supervision or KPIs.

47. We generally expect employees to have half their time unscheduled, free to do whatever they want. This is a research environment for them to explore what they think is important without prerequisites.

48. We rarely work overtime. There are two reasons: First, research requires a relaxed environment. Pressuring people hinders research. Since you need personal interest to think about problems, a relaxed environment is necessary for exploration.

49. Second, we are very focused. Being focused means we have few things to do, so overtime isn't needed. This aligns with our restraint.

50. Our company is built on consensus. I don't decide everything alone; I seek consensus. My authority and influence within the company are based on consensus.

51. This decision-making mechanism seeks consensus. I can only push something forward if there is consensus; only then will I push it.

52. As our team grows, we will make adjustments. We need to do this soon, as I'm already making adjustments. Without this, many things can't proceed. Many departments need organizational structures.

04 Computing Power and Resources

53. How many GPUs do we need? More is always better, within our capacity. Undoubtedly, the more GPUs, the better. Our current strategy is to buy as many GPUs as possible at a reasonable price.

54. Actually, spending all that money is very difficult. We can't buy enough GPUs; it's hard, and prices are high. We can't pay exorbitant prices and must ensure fairness. If we can spend 20 billion RMB this year, our procurement department would have done an excellent job.

55. The biggest gap between us and the US is resources. On one hand, we can't buy enough GPUs domestically; on the other, our capital investment is less. Capital investment is much lower, and talent salaries are a small part; the bulk is computing power.

56. All differences we see—in talent, model capability, applications—can be considered due to differences in computing resources.

57. The gap between us and the US is about 12 months, maybe 12 to 18 months, or 6 to 12 months. Simply put, we lag by two years, having achieved our results with one-twentieth of the US's computing power.

58. The narrative is that we lag by one to two years but use one-twentieth of the computing power. In the future, we aim to rewrite this: use a fraction of their computing power but shorten the gap to 6 months or 3 months. That's a goal.

59. We believe in Scaling. Larger scale leads to better results and unlocks more features. The only thing preventing Scaling is computing power; it's not that we don't want to Scale, but we lack the resources.

60. Training such a large model isn't because we think it's sufficient, but because we have exactly these resources. We calculate the model size based on our resources, not because this size is enough.

61. When Silicon Valley says Scaling has hit a limit, it's for them. For us in China, we are far from that limit. Scaling includes data scaling, model size scaling, and training costs.

05 Domestic Chips and Ecosystem

62. Nvidia CUDA's moat is rapidly eroding. With AI, building an ecosystem is much easier since AI can write code.

63. The computing card market is now larger than the gaming card market, so there's no reason they need to be coupled. The trend is decoupling. Dedicated chips, whether from Huawei or Nvidia, will be specialized, unlike previous ones.

64. There's a historic opportunity for domestic AI chip replacement. We believe that within a year, one thing will be verified: the domestic chip ecosystem is perfectly fine. Previously thought problematic, it will be proven otherwise.

65. The hardware and ecosystem of domestic AI chips are fine; the only issue is insufficient production capacity. Adapting domestic chips has no barriers that Nvidia can block. In a normal business environment, Nvidia's chips would be preferred, but since they are unavailable, everyone is forced to adopt domestic chips.

66. During V3 training, we used Nvidia GPUs but not its ecosystem. We used a high-level compiler called TileLang, building our ecosystem based on it, almost independent of Nvidia's ecosystem.

67. I'm optimistic about domestic computing power. Nvidia is digging its own grave here. Huawei's super nodes, like the 950, can fully replace Nvidia's GB