4시간, 118개의 답변, 량원펑(梁文锋) 내부 교류에서 모든 것에 응답하다

Content organized by: Gu Lingyu

Original editors: Xu Qingyang, Su Yang

Original source: Tencent Technology

DeepSeek recently completed its first external financing since its establishment. The total fundraising amount for this round exceeds 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 investor lineup, DeepSeek founder Liang Wenfeng personally invested 20 billion RMB, Tencent invested 10 billion RMB, CATL invested 5 billion RMB, while NetEase, JD.com, and IDG Capital each invested 3 billion RMB. The National Artificial Intelligence Industry Investment Fund invested 1 billion RMB.

Prior to this, Liang Wenfeng had stated the principle of "no financing, no IPO, no commercialization." This large-scale financing 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 conference, Liang Wenfeng elaborated in detail on DeepSeek's organizational culture, open-source logic, technology roadmap, and views on the industry's competitive landscape.

The following is a compiled transcript of Liang Wenfeng's remarks from the nearly 4-hour conference, obtained by Tencent Technology. It is categorized by topic, totaling 118 entries. The text strives to preserve the original meaning as much as possible, with only slight edits.

01 Vision and Restraint

1. When we first started this company, our initial intention was not about how much money we would eventually make, or going to the capital market to list. The first few dozen people never thought about it that way. If they had, they wouldn't have come.

2. We started this with great goodwill towards the world. We believe this is useful for humanity; it's something beyond money. Our initial motivation, our vision, and the vision we maintain to this day are not built around maximizing commercial interests.

3. Managing a large company isn't about rules and regulations; it's about vision. A vision isn't a slogan on the wall. Vision is how you act, not what you say. It's how you actually operate.

4. We have no formal organization; we are vision-driven, organized around a shared vision. We don't operate by "achieving certain KPIs" or without assessment; we only have the vision.

5. This vision isn't even documented. It was never written down. The vision exists in the way we do things and our attitude towards the world.

6. We don't have many other advantages. We are not particularly skilled, not richer than others, nor do we have better personnel than other companies. In fact, we don't. When we founded the company two years ago, 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. 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, very few resources, and our people were just a random group of ordinary individuals.

8. AI is too big, and the potential rewards are too immense. We are very restrained. As long as we can succeed, the final rewards will be enormous. Even if you take just a small piece, it's huge. So there's no need to think about which part of the reward to take or how to take it right now. It's simply not a consideration because the rewards are large enough.

9. Last Spring Festival, we suddenly had a surge in users. But we didn't pursue retaining these users, monetizing them, or competing for commercial interests by cashing in on users. We didn't compete for users or try to make money, but we worked very hard to serve the users well.

10. We have no intention of becoming the next super app, competing with someone, or becoming the next ByteDance or Tencent. We have absolutely no such thoughts. I believe the opportunity for AGI in the future is vast, and it will always be vast.

11. Restraint is a strategy. Sometimes you can give up something in exchange for something else. Not open-sourcing is similar. It can be seen as our pressure, or as our concession of profits.

12. I understand this restraint as something that, in the long run, can increase our probability of achieving AGI. When considering things, I have no doubt that AGI will have tremendous commercial value. Based on that, I don't prioritize how to increase my share or grab a larger portion. My priority is how to increase the probability of my success.

13. We have always been very restrained and unwilling to become an adversary to any internet giant or startup. I hope we can empower them, assist them in doing this, and help them achieve their goals.

14. I feel that by adhering to this attitude, we haven't lost anything. Despite open-sourcing, our goodwill, or providing assistance to others, we haven't obtained less. On the contrary, it might have been a plus. This seems counterintuitive, but it's indeed the case.

15. Our goal is AGI, but we have always been commercializing, which is why we have C-end users and B-end revenue. Historically, this strategy has been successful.

02 AGI Roadmap

16. If you can describe a problem very clearly, providing full context and instructions, it already surpasses humans. But there's a definition, a prerequisite: you provide full context and complete instructions.

17. AI cannot replace your employees. But if AI has the ability for continuous learning, learning at a company for two months like an employee, then it could replace everyone. So we are still one step away from the next stage: continuous learning.

18. AI development can be understood as a ladder. Last year's step was the Chain of Thought (CoT). We discovered that through CoT, we could achieve a higher level of intelligence.

19. This year's step is Agent. We found that using the Agent method, even complex tasks can be done. Its scope of ability expands, and its intelligence ceiling is higher. Agent uses CoT, and CoT uses the previous step, which is the language model. So no step is wasted.

20. After Agent, the next problem we should solve is continuous learning – how to enable the model to learn continuously, not by giving it a very strong training session, but allowing it to engage in long-term continuous learning like a human.

21. After continuous learning, we might arrive at a singularity. This singularity is when the model, capable of continuous learning, can do everything humans can do. It can then develop its own version, conduct its own research, and develop its next, more advanced AI model.

22. This singularity is not a singular event; it's also a gradual process. This process might be a relatively long transition, not a sudden mutation. But habitually, we tend to think of it as a singularity.

23. This is our speculation; this is the timeline we envision: first, solve learning to learn, then reach the intelligence singularity capable of self-iteration, and then embodied intelligence. After achieving embodied intelligence, it enters the real world, capable of doing housework and providing elderly care.

24. If we solve continuous learning first, then the self-iterating singularity, and then embodied intelligence, the path becomes much smoother. Because later on, you can use the earlier technology to help develop the later technology.

25. We only focus on the main line of AGI. The AI field is vast. Many things, like 3D or video generation, we believe are not on this main line and are not significantly related to the path of intelligence. We won't pursue them.

26. Video generation was very popular when it first emerged, as if it were a must-do, like you wouldn't be an AI company without it. I find this strange. If you think carefully, it has little relationship with the roadmap for intelligence.

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

28. In our judgment, world models and general intelligence are not the most critical things at this stage. The most important things are AI training and, after training, how to solve continuous learning. This is our company's judgment; of course, each company has its own.

29. We currently believe in the narrative that AI can accelerate AI research. It's not linear because you can use AI to accelerate your own research, making it potentially non-linear later on.

30. I think we must eventually enter embodied intelligence. For a normal person, their needs are not just computers, right? They need food, clothing, housing, transportation. They need embodied intelligence to address specific human labor needs.

31. What do we want AGI to do? We want it to help iterate the next version of the model, just like how we iterate the model now. If we have embodied intelligence, we want it to iterate the next version of the embodiment itself, to build the next generation of robots.

32. The core capability of the next-generation model must be continuous learning for it to be called next-gen. Until then, what we can do is reduce costs, improve performance, and increase speed. But for a major breakthrough, it must possess continuous learning.

33. Current Agent capabilities are limited because they cannot effectively learn continuously. If we can solve continuous learning first, AI's capabilities will be incredibly strong, greatly enhancing our own research efficiency.

34. Once continuous learning is developed, general intelligence might become much easier. Use the model with continuous learning to do it easily. So I say this is the outcome we hope to see, making things easier and more relaxed for us. Otherwise, manually developing general intelligence now is laborious, painful, data-intensive, and human-intensive, with poor cost-effectiveness.

03 Team and Talent

35. My previous experiences taught me that the vision of AGI is very powerful. The talent advantage isn't about my people being smarter than others; it's about how I organize, motivate, and facilitate collaboration among these talents.

36. Gathering smart people doesn't automatically mean they will collaborate and passionately pursue a common goal. You need a vision.

37. Our core interest is maintaining team stability. This is our paramount core interest, arguably our only core interest. As long as I can maintain team stability, I will definitely succeed and achieve AGI. It's that simple.

38. Money is certainly not an issue; resources are not an issue; other factors are easy to obtain. For us, there's only one core interest we cannot compromise on: we must maintain team stability.

39. This is also a very significant challenge we face, perhaps the biggest risk. Of course, this risk has been substantially mitigated with our recent financing round. Everyone received quite a significant amount of options.

40. Regarding team stability, as long as the most important, longest-tenured employees are stable, others are unlikely to leave. Even if others have fewer options or lower income, they won't leave. Because everyone isn't solely driven by money; they all want to work in an environment where they can help build AGI.

41. Everything else is a matter of time. At most, it might set us back by half a year or a year, but it won't prevent us from succeeding. We definitely won't lack money or resources. These are not lacking.

42. The gap between us and the US is mainly in resources. The talent gap is not significant. There's almost no gap in talent because it's the same group of people, likely Chinese. When Chinese go abroad, some stay abroad, some return. It's not that smart people go abroad; that's not the case.

43. Talent is not the bottleneck; resources are the biggest bottleneck. Resources first impact talent cultivation. With less computing power, we have fewer opportunities for experiments, making our talent pool generally lag behind the US. The talent gap is essentially a computing power gap.

44. The shortage of AI talent is also temporary, and we've already seen it being significantly alleviated. There is no real shortage of AI people. Every company will quickly train their people; training is fast.

45. There are too many companies building foundational models in China right now. The US probably has three. China has too many doing basic models. In the end, not so many people will be needed for foundational models; the field will inevitably converge.

46. Our company's management actually follows two lines: one from top to bottom, and one from bottom to top. From bottom to top, everyone decides what they want to do, does it themselves, with no supervision and no KPIs.

47. Generally, we expect employees to spend half their time unassigned, free to do whatever they want. This is a research scope allowing them to explore on their own, based on what they think is important, without any prerequisites.

48. We don't usually work much overtime. There are two reasons. First, research requires a relaxed environment. If you push too hard, you can't do research. Since you need to have a personal interest and think about these problems regularly, a relaxed environment is necessary for exploration.

49. Second, we are very focused. Being very focused means we have very few things to do. So there isn't much to do, and overtime isn't needed. This is consistent with the earlier point about restraint.

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

51. This decision-making mechanism is essentially one of seeking consensus. It's not that I can push through anything. Only when there is consensus can I push it forward, and only then will I do it.

52. As the team grows, we will make adjustments. We should probably make adjustments immediately because I'm already doing so. Without these adjustments, many things can't progress. Many departments indeed should have an organizational structure.

04 Computing Power and Resources

53. How many GPUs do we need? The more, the better, undoubtedly, within our affordable range. So our current strategy is to buy as many GPUs as possible at a reasonable price.

54. Actually, it's very difficult to spend all that money. You can't buy that many GPUs. They are hard to acquire and expensive. You can't just pay a sky-high price; you must ensure the price is reasonable. If we can spend 20 billion RMB this year, that would be an outstanding performance by our procurement department.

55. The biggest gap between us and the US is in resources. On one hand, GPUs are hard to buy in China; on the other, our capital investment is less than in the US. We invest significantly less capital. The proportion of talent salaries here is very low. Look at the salaries they offer, like 100 million USD packages. But in total, talent salaries still account for a small portion; the bulk is computing power.

56. All the differences we see, including differences in talent, model capabilities, and applications, can be attributed to differences in computing power resources.

57. We are probably 12 months behind the US, maybe 12 to 18 months, or 6 to 12 months. Simply put, we are about two years behind the US, but we achieved this using only one-twentieth of the computing power.

58. The narrative is being one to two years behind but using one-twentieth of the computing power. In the future, we want to rewrite this narrative: using a fraction of the computing power but closing the time gap significantly, to 6 months or 3 months. I think that's a goal