4 giờ, 118 câu trả lời, Lương Văn Phong trao đổi nội bộ, giải đáp mọi thắc mắc

Author: Gu Lingyu

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), with a pre-investment valuation of around 367.5 billion RMB (approximately $54.3 billion). Among the investors, DeepSeek founder Liang Wenfeng personally contributed 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.

Previously, Liang Wenfeng had established a principle of "no financing, no IPO, no commercialization." This large-scale fundraising marks DeepSeek's official entry into the capital market, sparking widespread industry attention on its commercialization path and technological vision.

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

The following is a compiled transcript of Liang's remarks during the nearly four-hour meeting, obtained by Tencent Technology, organized by theme into 118 points. The text retains the original meaning as much as possible, with only minor edits.

01 Vision and Restraint

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

2. We are doing this with a great deal of 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, is not built on maximizing commercial interests.

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

4. We are essentially organization-less, driven purely by vision. We operate not by "achieving certain KPIs" or "no assessments"; there is only the vision.

5. This vision isn't even written down. Nothing has ever been written down. The vision exists in our methods of doing things, in our attitude towards the world.

6. We don't have many other advantages. We have no special skills. We aren't wealthier or have better people than other companies. We truly don't. Two years ago, when we founded the company, we didn't have much money, many GPUs, any reputation, or any influence. 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 proven true so far. It's logical. Otherwise, there's no explanation for how we managed to succeed: we had no weapons, a very low starting point, very few resources, and our team was essentially just a random group of ordinary people.

8. The AI field is too vast, the potential benefits are too enormous. We are very restrained. As long as we can succeed, the benefits will be enormous regardless. Even taking a small piece yields huge returns, so there's no need to think about which part of the benefits to take or how to take them. This is completely unnecessary because the benefits are large enough.

9. Last Spring Festival, we suddenly gained many users, but we didn't pursue retaining them, monetizing them, or fighting for those commercial interests. We didn't compete for users, we didn't try to make money from them, but we worked hard to serve them well.

10. We don't have the ambition to become the next super app, compete with anyone, or become the next ByteDance or Tencent. We have absolutely no such thoughts. I believe the future AGI opportunity is immense, it will always be immense.

11. Restraint is a strategy. Sometimes, you can give up something to gain more in other areas. Not open-sourcing is similar. It can be seen as our pressure, or as our act of concession.

12. I understand that this restraint, in the long run, increases our probability of achieving AGI. When considering matters, I have no doubt that AGI will have tremendous commercial value. Based on that, my priority isn't how to increase my share or take a larger piece; it's how to increase the probability of our success.

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

14. I feel that by maintaining this attitude, we haven't lost anything. We haven't gained less because of open-sourcing, our goodwill, or our help to others. If anything, it might have been a plus. This seems counterintuitive, but it's true.

15. We aim for AGI, but we have always been commercializing. That's 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 complete context and instructions, AI has already surpassed humans. But this definition comes with a premise: you provide complete context and complete instructions.

17. AI cannot replace your employees. However, if AI acquires continuous learning ability, learning at a company for two months like an employee, it could replace everyone. So, we are one step away from that: continuous learning.

18. AI development can be understood as a staircase. Last year's step was Chain of Thought. We discovered that CoT could elevate intelligence to a higher level.

19. This year's step is Agent. We found that with Agents, even more tasks can be accomplished, expanding the capability range and raising the intelligence ceiling. Agents utilize CoT, and CoT uses the previous step – language models. Therefore, no step is wasted.

20. After Agents, the next challenge we need to solve is continuous learning: how to enable models to learn continuously without needing a strong training signal each time. They should be able to engage in long-term continuous learning like humans.

21. After continuous learning, we might arrive at a singularity. This singularity is that when a model can learn continuously and do everything humans can, it can then develop its own versions, conduct research, and create the next, more advanced AI model.

22. This singularity isn't actually a singularity; it's a gradual process. It's likely a long, gradual change, not a sudden mutation. But habitually, we tend to call it a singularity.

23. Our speculation for the timeline is: first solve learning to learn, then reach the intelligence singularity for self-iteration, and then embodied intelligence. After embodied intelligence, AI enters the real world, capable of household chores and elderly care.

24. If we solve continuous learning first, then the self-iterating singularity, and then embodied intelligence, the path becomes easier. Later technologies can be developed using earlier ones.

25. We only focus on the main line of AGI. AI is vast, and many things seem off this main line. For example, 3D and video generation, which I believe are not strongly related to the main intelligence line, so we won't pursue them.

26. Video generation became very popular initially, as if it were a must-do, and not doing it meant you weren't an AI company. I find this strange. If you think carefully, 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 something just because it's a good business. We only do things if they are on the intelligence roadmap.

28. In our judgment, world models and intelligence are not the most important things at this stage. The most important things are AI training and solving continuous learning post-training. That's our company's judgment; every company has different judgments.

29. We currently believe in a narrative where AI can accelerate AI research. It's non-linear because you can use AI to accelerate your own research, potentially leading to non-linear progress later.

30. I think embodied AI is ultimately necessary. For a normal person, their needs aren't just a computer, right? They need food, drink, entertainment, housing, and transportation – not a computer. They need embodied AI to solve real human labor needs.

31. What do we hope AGI can do? Iterate the next version of the model, just like it can iterate the current one. With embodiment, we hope it can iterate the next version of the robot itself.

32. The core capability of the next-generation model must be continuous learning for it to be called next-gen. Until then, we can only focus on reducing cost, improving effectiveness, and increasing speed. But a major breakthrough requires continuous learning.

33. Current Agents are limited because they can't learn continuously effectively. If we can achieve continuous learning first, AI's capabilities would be extremely powerful, greatly enhancing our own research efficiency.

34. Once continuous learning is achieved, general intelligence might become easy, making everything else easier. This is the outcome we hope for; it's less labor-intensive for us. Otherwise, manual work towards general intelligence is strenuous, data-intensive, labor-intensive, and not cost-effective.

03 Team and Talent

35. My past experience taught me that the vision of AGI is powerful. The talent advantage isn't that our people are smarter, but how we organize, motivate, and collaborate with them.

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

37. Our greatest core interest is maintaining team stability. This is our paramount core interest; arguably, it's the only one. As long as we maintain team stability, we will definitely succeed, definitely achieve AGI. It's that simple.

38. Money is certainly not a problem, resources are not a problem. Other elements are easily obtainable. For us, there is only one core interest, one non-negotiable: we must maintain team stability.

39. This is also a significant challenge we face, perhaps the biggest risk. Of course, this risk has been considerably mitigated by our recent financing round. Everyone received relatively substantial options, with significant amounts involved.

40. Regarding team stability, as long as the most important, longest-serving employees are stable, others are unlikely to leave. Even if others have fewer options or lower pay, they won't leave. They aren't all purely motivated by money. Everyone wants to be in an environment that can achieve AGI.

41. Everything else is a matter of time. At worst, it might delay us by half a year or a year, but it won't prevent success. We definitely won't lack money or resources; these are not concerns.

42. The gap between us and the US is mainly in resources, not significantly in talent. There's almost no gap in talent because they are the same people, likely Chinese. When Chinese people go abroad, some stay, some return. It's not that the smart ones go abroad. No.

43. Talent is not the bottleneck; resources are the biggest bottleneck. Resources first affect talent development. With less computing power, we have fewer opportunities for experiments, so our talent lags behind the US overall. This talent gap is essentially a computing power gap.

44. The shortage of AI talent is also temporary, and we've already seen significant easing. AI people are truly not scarce. Companies can train people quickly; training is fast.

45. There are too many model-making companies in China right now. Way too many. The US might have three. There are too many companies in China building foundation models. Eventually, fewer people will need to build them; the field will converge.

46. Our company management has two lines: top-down and bottom-up. Bottom-up means everyone does what they want, with no one managing them and no KPIs.

47. Generally, we expect employees to have half their time unscheduled, free to do whatever they want. This is a research domain for them to explore, based on what they think is important, without prerequisites.

48. We generally don't work overtime much. There are two reasons. First, research requires a relaxed environment. Pressuring people makes research impossible. Since employees need intrinsic interest to think about these problems, a relaxed environment is necessary for exploration.

49. Second, we are very focused. Being very focused means we have fewer things to do. So, there isn't much work, and overtime isn't needed. This aligns with the principle of restraint.

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

51. This decision-making mechanism is about seeking consensus. I can only push something forward if there is consensus, and only then will I push it.

52. As the team grows, we will make adjustments. We need to do this soon because I'm already making adjustments. Without adjustments, many things can't move forward. Many departments indeed need an organizational structure.

04 Computing Power and Resources

53. How many GPUs do we need? Right now, the more, the better. Within our capacity, more GPUs are definitely better, no doubt. So, our strategy is to buy as many GPUs as we can at a reasonable price.

54. Actually, spending all this money is very difficult. We can't buy enough GPUs; they are hard to acquire, and prices are high. We can't just pay exorbitant prices; we must ensure the price is reasonable. If we can spend 20 billion this year, it would mean our procurement department performed exceptionally well.

55. The biggest gap between us and the US is in resources. Computing resources are limited domestically, and our capital investment is less. Our capital investment is much lower, and talent salaries are a small fraction of total spending. Look at their salaries, like $100 million, but talent costs are still a small part; the bulk is computing power.

56. All the differences we see – talent, model capability, applications – can be attributed to differences in computing resources.

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

58. The narrative is lagging behind by one to two years while using one-twentieth of the computing power. The future goal is to rewrite this narrative: using a fraction of their computing power but narrowing the time gap to six months or three months.

59. We believe in scaling. The larger the scale, the better the results and the more features are unlocked. What prevents us from scaling is computing power. It's not that we don't want to scale; we lack the computing power to do so.

60. Training such a large model isn't because we think it's sufficient, but because it's the largest model we can afford with our resources. The model size is determined by our available resources, not by the model being deemed sufficient.

61. When Silicon Valley talks about the end of scaling, it's for Silicon Valley. For the Chinese, we are far from that point. We haven't scaled to that extent at all. This scaling includes data scaling, model size scaling, and training costs.

05 Domestic Chips and Ecosystem

62. N