Last month, "elsewhere" reported on DeepSeek's financing story. The most discussed topic among many was the legendary 4-hour investor meeting.

During the past month, various quotes from Liang Wenfeng have circulated widely, and we have collected some of them.

During this time, Liang Wenfeng repeatedly emphasized "no": no genius, not earning unreasonable profits, not pursuing user volume, not closing the source, not engaging in 3D/video generation/world modeling, not creating the next super app. He stated that restraint is a strategy used to increase the chances of achieving AGI.

From the limited materials we have learned about, the following high-frequency terms emerged: model, cost, AGI, time, open source, etc.

Most of the time, Liang Wenfeng maintained an open tone with simple and plain language. Only when discussing a few issues that he cared about did he reveal a bit of sharpness: “As long as I can maintain the stability of the team, I will definitely be able to achieve AGI. It’s that simple.”

Here are 52 collected statements. Some phrases may slightly deviate from the original meaning while retaining its essence.

DeepSeek has only one main line

1. Now is not the time to maximize revenue from products. The path to AGI must pass through the product step, but we do not need to invest too much thought and energy into creating C-end and B-end products. Approaching relatively lower-tier technologies from a high-tech perspective is a dimensionality reduction attack. Products are a by-product on the path to AGI.

2. Many things are not on our main line, such as 3D, video generation; similarly, world models have little relation to the upper limits of intelligence.

3. Multimodality is very important for products and for C-end users. But it is just a component, not the main line and intelligence itself.

4. The hallucinations of large models can obviously be addressed, but this is a long proposition. Internally, we classify hallucinations as a product problem that will be resolved, but it is not the focus.

5. At this stage, the most important thing is still the Coding Agent. Given the domestic situation, the most rational approach should be to focus on creating a universal Agent, while prioritizing the development of other Agents like finance and healthcare lower.

6. If the AI era produces many trillion-dollar companies, it would be great if DeepSeek is one of them.

First achieve continuous learning, then AI self-iteration, the endpoint is embodied intelligence

7. AI currently lacks taste and intuition; what it lacks is the ability to learn continuously.

8. Humans can learn continuously, but in the same matter, AI requires all the context. This is nearly impossible, so AI cannot replace employees. Therefore, the next-generation model must have continuous learning capabilities to be called a next-generation model.

9. We hope the next-generation model can assist us in our own development. To put it simply, the primary goal of the models we create is not for everyone to use effectively but for us to use effectively. This is the fastest way to achieve AGI.

10. Currently, there are no good methods found worldwide because "learning" consists of many components.

11. DeepSeek's long-term vision is AGI. If we compare the implementation route to climbing stairs, last year’s step was CoT (Chain-of-Thought), and this year’s step is Agent. The next challenge after Agent will be achieving continuous learning.

12. After achieving continuous learning, we may reach an incremental singularity: the model can accomplish all tasks that humans can do, including developing cutting-edge artificial intelligence models. Hence, AI can accelerate AI research. Only after this step is completed can we talk about embodied intelligence.

13. The endpoint of intelligence may be embodied. Because for a normal person, their need isn’t a computer, but human capacity.

We are still far from completely shifting to commercialization

14. We only earn a reasonable profit, not maximizing profit through pricing.

15. At first, we were concerned that the demand for one of our models was too high, and the price was set relatively high, but later dropped to a quarter, much to the joy of many in the company group. This is the purpose of diligently creating this model: to allow everyone to use it fully.

16. Low cost is a result; our models have always been architected towards lower costs. We also hope costs are manageable, especially against the backdrop of computing power shortages.

Another reason: the lower the cost, the larger the models we can support. When computing power is limited, high computational efficiency enables training of larger models. Large companies can rely on adding resources; we prioritize cost efficiency.

17. It may appear to outsiders that we have chosen a very difficult model. But actually, we find it quite easy. Lowering prices is definitely not good for our competitors; they are certainly not cheering. I don’t think selling APIs is that attractive. As long as I can get a few people to maintain this API, even without customer service or sales, users will come on their own.

18. We have always been commercializing, just not with commercialization as the primary goal. The time point for DeepSeek to completely shift to commercialization should be quite far away.

19. I don’t even need to consider occupying a position at that time. As long as there is a significant business opportunity, there will be a way. DeepSeek is a product of an era, a response to reality, not a result of imitation.

Open source is the sweet spot for a company of our scale

20. Restraint is a strategy: giving up some things in exchange for more of others. Open source means benefit-sharing: internally, employees gain a sense of achievement, and the company receives cohesion; it also benefits society, making peers or ordinary people happy.

I have no doubt that AGI will have significant commercial value. Based on this premise, my priority is not to capture more market share but to increase the probability of success.

21. To succeed in AI commercially, open source is advantageous. This sounds somewhat counterintuitive. Historically, for a software company, the market may only be worth tens of billions of dollars annually; once it’s open-sourced, that disappears. But AI is large enough that it could ultimately take up ten percent of human society’s GDP. If we want to monopolize this benefit, we will definitely be abandoned by history. This is an objective law, it’s a perspective of history.

22. The open-source models we provide are the same as those we deploy ourselves. We will not open-source a lesser model while deploying a better one.

23. I am not worried about others deploying our models to compete with us. Not every company has the willingness and capability to achieve this goal. Startups that are too small lack the power to do this; large companies find it challenging to organize. This is a sweet spot belonging to companies of our scale.

24. Open source does not impact our business model as long as we "only earn a reasonable profit". If you aim for a hundred-fold profit, open sourcing will indeed affect you.

25. We do not wish to become competitors with any of the major or minor internet companies. Under this premise, we are very willing to assist and help anyone, including Alibaba, Zhizhu, and the Dark Side of the Moon, to do better.

The gap between China and the US is not in talent

26. In the future, we want to rewrite the narrative of AI between China and the US: using a fraction of the computing power to shorten the gap to six months, three months.

27. The gap in AI between China and the US primarily lies in resources. We believe in scaling; the larger the scale, the better the results. Training such large models is not because I think that they are big enough, but because there are only so many resources.

28. There is almost no gap in talent—it’s the same group of people. There is no shortage of talent domestically. Talent shortages are temporary; historically, there has never been a long-term lack of any specific type of person.

Competition in model-making, with cost as the top priority

29. Anthropic is currently surpassing OpenAI, but this is not long-term; it is temporary. OpenAI and Google will likely alternate rising in the future.

30. There are too many companies in the domestic model-making sector, all doing the same thing, leading to resource dispersion. Ultimately, this will converge, but it will take time. If each company only aims for a reasonable profit, there will not be a need for so many people to work on large models: possibly two large companies and two small companies would be sufficient.

31. I absolutely do not believe that large model companies can take most of the profits in the AI industry.

32. The final gap in large model competition will be reflected in three aspects: cost, time, and user experience.

Cost is the top priority—what cost can you offer for the same quality service? The second is time; being a few months earlier or later makes a difference. User experience may have some stickiness and barriers, but it’s not fundamental.

Not interested in becoming the next super app

33. We do not want to become the next super app. To become the next ByteDance? The next Tencent? We have absolutely no such thoughts.

34. Not competing for this is because there are still watermelons behind; what’s in front might be just sesame seeds. Of course, this sesame seed might be relatively large, but I don’t think it’s that big.

35. Last year, everyone rushed for chatbots and C-end traffic; this year, they are competing for B2B revenue. However, we do not consider this important; what the company internally truly cares about is the roadmap to AGI and how to achieve breakthroughs in the next steps of technology. It’s strange that what we most wish for is unattainable, while what we don’t care much about is easier to obtain.

36. Last year's Spring Festival phenomenon was not in our script.

Maintaining team stability is core

37. There is one thing we cannot compromise on: we must maintain team stability. This is a very significant risk we face. Of course, this risk has been largely alleviated with this round of financing.

38. We do many things to maintain team stability. We do not want to become competitors with any major or minor internet company; we hope to empower and assist them. We do not want to create enemies. This way, our own environment will also be better.

39. Some people feel that our organization is top-down, while others feel it is bottom-up; I think both views are correct. Top-down is called “doing the right thing,” usually hoping that “the right thing” does not take up half of the employees’ time. The other half is bottom-up, unstructured time, where they can research whatever they want, following their perceived importance without prior constraints.

40. Generally, we do not have much overtime. The first reason is that research requires a relatively relaxed environment. The second reason is that we are very focused. Many products are not complete, but we have not sought to remedy that, which also reflects a culture of restraint.

41. Organizations are dynamic; they are not static. In the future, as the company grows, there may be some adjustments. But it will not completely turn into a traditional hierarchical system; some necessary structures may arise. But the vision-driven aspect will not change.

With goodwill towards the world

42. Our intention in starting this company was never to make a lot of money or to go public on the capital market. The original few dozen people never thought this way. If they had such thoughts, they wouldn’t have come. We are doing this with a great goodwill towards the world, believing it is useful for humanity.

43. “Achieving XX KPI” is not our way; we are a vision-driven organization. This has its advantages and disadvantages. In the future, we will find ways to leverage our strengths and avoid our weaknesses, but this is our characteristic.

44. The vision is not even formally written; it exists in the ways we do things and our attitude towards the world. Perhaps everyone in the company has a different understanding of this vision, but the general direction is consistent.

45. About twenty years ago, the person I admired most in management was Jack Welch (former CEO of GE). Now, looking back, most of what he said may no longer be correct, but he did say one thing correctly: the most important thing for a company is its vision.

The vision is not a slogan hanging on the wall, it’s not about what to say, but about what to do.

Restraint enables us to achieve AGI

46. AGI is the most profitable. As for other matters, if we have the energy, we will do them; if we don’t have the energy, we won’t. Restraint is part of our vision.

47. The AI matter is too big, and the stakes are too high. As long as it can be achieved, sharing a bit of profit is already significant. The more you restrain yourself, the more likely you will succeed in this endeavor.

48. I think this is intuitive—at least in line with my intuition. Apart from the vision, we do not have many other advantages.

49. When this company was established two years ago, we had little money, few resources, no fame, and no appeal; we were just a group of very ordinary people. The narrative I like is "a group of ordinary people achieving extraordinary things," rather than "a group of geniuses achieving extraordinary things."

50. Open source is also part of restraint. In terms of pricing, we definitely do not start from the highest revenue or profit point for the company. In the short term, a slightly higher price means more revenue; but in the long term, it’s hard to say. For me, restraint is a strategy.

51. Open source and low prices give employees a sense of achievement, which enhances organizational cohesion and benefits society, making peers and ordinary people happy. Therefore, this kind of restraint, from a long-term perspective, can increase our probability of achieving AGI.

52. If your vision is just to gain more, you will lose first; you may face greater difficulties. That’s how the world works.