4 hours, 118 questions: Liang Wenfeng’s internal exchange addresses everything

Content curated by: Gu Lingyu

Original editing: Xu Qingyang, Su Yang

Original source: Tencent Technology

DeepSeek recently completed its first external financing round since its inception. The total funding amount exceeded 50 billion RMB (approximately $7.4 billion), with a pre-investment valuation of about 367.5 billion RMB (approximately $54.3 billion). In the investor lineup, 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, and the National AI Industry Investment Fund contributed 1 billion RMB.

Previously, Liang Wenfeng had proposed the principle of "no financing, no IPO, no commercialization." This large-scale financing marks DeepSeek's official entry into the capital market, also sparking widespread industry attention on 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 curated transcript of Liang Wenfeng's speech from the nearly 4-hour conference, obtained by Tencent Technology, organized by topic into a total of 118 items. The text retains the original intent as much as possible, with only minor edits.

01 Vision and Restraint

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

2. We approached this with great goodwill towards the world, believing it would be useful for humanity, something beyond money. Our starting intention, our vision, and the vision we maintain to this day, is not built around maximizing commercial interests.

3. Managing a large company relies not on rules and regulations, but on vision. Vision isn't just a slogan on the wall; it's about what you do, not what you say – how you actually operate.

4. We are essentially organization-less, driven and organized purely by a vision. We don't operate based on achieving specific KPIs or having assessments; we only have a vision.

5. This vision isn't even formalized or written down anywhere. It exists in our methods of doing things and our attitude towards the world.

6. We don't have many other advantages. We are not particularly talented, we weren't richer than others, nor did we have better personnel than other companies. Two years ago when we founded this company, we didn't have much money, many GPUs, much 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 so far and is logically explainable. 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, the stakes are too high. We are very restrained. As long as we can succeed, the eventual rewards will be enormous. Even a small share is huge, so there's no need to worry about which part of the benefits to take or how to take them. We don't need to think about it at all because the potential rewards are big enough.

9. Last Spring Festival, we suddenly had many users, but we didn't pursue retaining them, monetizing them, or competing for commercial interests by cashing in on them. We didn't aggressively chase users or make money, but we worked hard to serve our users well.

10. We don't have the idea of becoming the next super app, competing with someone, or becoming the next ByteDance or Tencent. We have no such thoughts at all. I believe the opportunities for AGI are enormous, and will always be.

11. Restraint is a strategy. Sometimes you can give up something to gain something else. The decision not to open source is similar; it can be seen as our pressure or as our concession of profit.

12. I understand this restraint, in the long run, can increase our probability of achieving AGI. When considering things, I have no doubt AGI will have immense commercial value. On that basis, my priority isn't how to increase my share or get a bigger piece, but how to increase my probability of success.

13. We have always been very restrained, unwilling to become adversaries with any internet giant or startup. I hope we can empower them, assist them in doing this, and help everyone with this endeavor.

14. I feel that by maintaining this attitude, we haven't missed out on anything. Because of our open-source approach, our goodwill, or our assistance to others, we haven't lost anything. On the contrary, it might have been a plus. This seems counterintuitive, but it's true.

15. We aim for AGI, but we have always been commercializing, which is why we have C-end users and B-end revenue. Historical experience shows this strategy is successful.

02 AGI Roadmap

16. If you can describe a problem very clearly, providing complete context and instructions, it already surpasses humans. But there's a premise: you must give it complete context and complete instructions.

17. AI cannot replace your employees. However, if AI had continuous learning ability, coming to your company to learn for two months like an employee, then it could replace anyone. So, the next step is still continuous learning.

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

19. This year's step is the Agent. We found that using the Agent paradigm, even more complex tasks can be handled, expanding its capability range and raising its intelligence ceiling. Agents utilize CoT, and CoT itself relies on the previous step – the language model. So no step is wasted.

20. After Agents, the next problem we need to solve is continuous learning: how to enable the model to learn continuously, rather than requiring a one-time heavy training. It should be able to engage in prolonged, continuous learning like a human.

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

22. This singularity isn't really a single point; it's also a gradual process. It might be a relatively long, gradual change, not a sudden mutation. But conventionally, we tend to call it a singularity.

23. This is our speculation on the timeline: first, solve learning to learn, then reach the intelligence singularity (capable of self-iteration), and then embodied intelligence. After embodied intelligence, it enters the physical world, capable of doing housework or providing elderly care.

24. If we solve continuous learning first, then the self-iterative singularity, and then embodied intelligence, the path becomes smoother. Later stages can use preceding technologies to help develop subsequent ones.

25. We only focus on the AGI mainline. The AI field is vast, and many things are not on this mainline, such as 3D generation or video generation. I believe these might not be closely related to the main intelligence track, so we won't pursue them.

26. Video generation was very popular initially, seemingly a must-do for any AI company. I found 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 it if it's on the intelligence roadmap.

28. In our judgment, world models and intelligence aren't the most important things at this stage. The most important are AI training and how to solve continuous learning after training. This is our company's judgment; of course, different companies have different views.

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 research, potentially making it non-linear later on.

30. I think embodied intelligence must eventually be pursued. For a normal person, their needs aren't just a computer. They need embodied AI to solve specific labor needs.

31. What do we hope AGI can do? Help us iterate the next model version. With embodiment, we hope it can do the same: iterate the next version of embodiment and build the next robot.

32. The core capability of the next-generation model must be continuous learning. Until then, we can only optimize cost, improve effectiveness, and increase speed. But a major breakthrough will require continuous learning.

33. Current Agent capabilities are limited because they cannot learn continuously, or effectively. If we can solve continuous learning first, AI capability would be very strong, greatly enhancing our own research efficiency.

34. Once continuous learning is achieved, general intelligence might become much easier and less labor-intensive. Otherwise, manually creating general intelligence is a data-intensive, labor-intensive, and low-efficiency task.

03 Team and Talent

35. My past experience taught me that the AGI vision is very powerful. The talent advantage isn't that my people are smarter, but how I organize, motivate, and coordinate them.

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

37. Our core interest is maintaining team stability. This is our single most important interest, arguably our only core interest. As long as I can keep the team stable, I will succeed in achieving AGI.

38. Money is definitely not an issue, resources aren't an issue; all other elements are easily obtained. For us, there is only one core interest we cannot compromise on: maintaining team stability.

39. This is also a massive challenge, or perhaps our biggest risk. Of course, this risk has been significantly mitigated by our recent financing round, as everyone received relatively substantial options.

40. For team stability, as long as the most important, longest-tenured employees are stable, others are less likely to leave. Even with fewer options or lower income, they won't leave because it isn't purely about money. Everyone wants to work in an environment capable of achieving AGI.

41. Everything else is a matter of time, potentially delaying us by half a year or a year, but it won't lead to failure. We certainly don't lack money or resources.

42. The gap between us and the US is mainly in resources. The talent gap is not significant. There's almost no talent gap because it's essentially the same pool of people – Chinese. Some Chinese stay in China, some go abroad. It's not that the smarter ones all go abroad.

43. Talent is not the bottleneck; resources are. Resources first affect talent cultivation. With less computing power, we have fewer opportunities for experiments, creating a talent gap that fundamentally stems from the computing resource gap.

44. The shortage of AI talent is temporary, and we've already seen it significantly alleviated. AI talent is not really scarce; companies can quickly train people. Training is fast.

45. There are too many model companies in China right now. The US probably has three. China has too many working on base models. Eventually, fewer people will be needed for base models; it will converge.

46. Our company management has two lines: top-down and bottom-up. Bottom-up means everyone decides what they want to do, manages themselves, with no KPIs.

47. Generally, we expect employees to have half their time unscheduled, free to do whatever they want. This is a research scope for them to explore based on what they think is important, with no pre-set requirements.

48. We usually don't require much overtime. Overtime has two reasons. First, research requires a relaxed environment. Pressuring people makes research impossible. They need intrinsic interest and the freedom to ponder problems in a relaxed setting for exploration.

49. Second, we are very focused. Being focused means we have fewer things to do, so there's less work overall. This is consistent with the principle of restraint mentioned earlier.

50. Our company operates on consensus. I don't make all decisions 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 forward an initiative if it has widespread consensus, and only then do I push it.

52. As our team grows, we will adjust. We need to do this adjustment soon, as I'm already working on it. Without this adjustment, many things won't move forward. Indeed, some departments should have an organizational structure.

04 Computing Power and Resources

53. How many GPUs do we need? The more, the better, undoubtedly, within our affordability. Our strategy is to buy as many GPUs as possible at reasonable prices.

54. Actually, spending all this money is very difficult because it's hard to buy enough GPUs. Prices are high, and we can't pay exorbitant prices; we must ensure fair value. If our procurement department can spend 20 billion RMB this year, their performance would be outstanding.

55. The biggest gap between us and the US is in resources. On one hand, we can't buy enough GPUs domestically. On the other hand, our capital investment is lower than in the US. Personnel costs, even with high US salaries (e.g., $100 million), are a small fraction; the bulk is computing power.

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

57. The gap between us and the US is likely 12 months, perhaps 12 to 18 months, or 6 to 12 months. Simply put, we trail by about two years, but we achieved this while using only one-twentieth of their computing power.

58. This narrative is: trail by 1-2 years but use one-twentieth the computing power. Our future goal is to rewrite this narrative: use a fraction of their computing power while narrowing the time gap to 6 months or 3 months.

59. We believe in scaling. Larger scale yields better results and unlocks more capabilities. What prevents us from scaling is simply computing power; it's not that we don't want to scale.

60. Training a model of this size isn't because we think it's sufficient; it's because we have this many resources. We calculate the maximum model size we can afford based on our resources.

61. When Silicon Valley says scaling is hitting its limits, that's for Silicon Valley. For China, we are far from that point. This scaling includes data, model size, and training costs.

05 Domestic Chips and Ecosystem

62. NVIDIA CUDA's moat is quickly eroding. With AI now available, building an ecosystem is much easier than before because AI can write code.

63. The computing card market is now larger than the gaming card market, so there's no reason for the two to remain coupled. The future trend is decoupling.