Liang Wenfeng 4-Hour Investor Meeting Transcript

Last month, "elsewhere" reported on DeepSeek's Funding Story. Among them, the most discussed was the legendary 4-hour investor meeting.

In the month since, various quotes from Liang Wenfeng have circulated in the industry, and we have collected some of them from multiple sources.

During this time, Liang Wenfeng said "no" very many times: not a genius, don't earn unreasonable profits, don't pursue user volume, don't close source, don't do 3D / video generation / world models, don't make the next super app. He said restraint is a strategy used to exchange for the probability of achieving AGI.

In the limited material we understand, the following high-frequency words exist: models, costs, AGI, time, open source, etc.

Most of the time, Liang Wenfeng's tone leaves room, using simple and plain words. Only when talking about the few issues he cares about does he show some sharpness: "As long as we can maintain the stability of the team, I will definitely achieve AGI. It's that simple."

Below are 52 quotes we collected. Some sentences may differ slightly from the original while preserving the original meaning.

DeepSeek Has Only One Main Line

1. Now is not the time for maximizing returns on products. On the road to AGI, we must pass through the step of products, but we don't need too much mind and energy to make consumer-side or business-side products. When standing at a high level of technology to make relatively lower-level technology, it is a strike from a higher dimension. Products are a byproduct on the road to AGI.

2. Many things are not on our main line, such as 3D, video generation; also like world models, which don't have much relationship with the upper limit of intelligence.

3. Multimodal is very important for products, very important for consumer-side users. But it is just a component, not the main line and intelligence itself.

4. Of course there are methods to solve the hallucination of large models, but this is a long-term proposition. Internally, hallucination is categorized by us as a product problem, will be solved, but is not the focus.

5. The most important thing at this stage is still Coding Agent. Looking at the domestic situation, the most reasonable approach should be to go all out on general Agents, while finance, medical, and other Agents have lower priority.

6. If the AI era will produce many trillion-level companies, it would be great if DeepSeek is one of them (if).

First Achieve Continuous Learning, Then AI Self-Iteration, The End Goal is Embodied Intelligence

7. AI now doesn't lack taste and intuition, what it lacks is the ability for continuous learning.

8. Humans can learn continuously, but for the same thing, you have to give AI all the context. This is almost impossible, so AI cannot replace employees. Therefore, the next generation of models must have the ability for continuous learning to be called the next generation of models.

9. We hope the next generation of models can help our own development. Simply put, the first goal of the models we make is not for everyone to use well, but for us to use well. This is the fastest way to achieve AGI.

10. Currently, the whole world has not yet found a good method, because "learning" is composed of many things.

11. DeepSeek's long-term vision is AGI. If comparing the implementation route to climbing stairs, last year's step was CoT (Chain-of-Thought), this year's step is Agent. The problem to solve after Agent is continuous learning.

12. After achieving continuous learning, it may come to a gradual singularity: models can complete all things humans can do, including developing more cutting-edge AI models themselves. That is AI can accelerate AI research. Only after this step is done is it embodied intelligence.

13. The endpoint of intelligence may all be embodiment. Because for a normal person, their demand is not computers, but manpower.

Still Very Far From Completely Turning to Commercialization

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

15. One of our models was initially worried about too much demand, price set relatively high, later dropped to one quarter, many people in the company group cheered. Because this is the purpose of us making this model so carefully: to allow everyone to use it fully.

16. Low cost is a result, our models have been moving towards lower costs in architecture. We also hope costs are affordable, especially against the background of scarce computing power.

There is another reason: the lower the cost, the more able to bear larger models. When computing power is limited, high calculation efficiency can train larger models. Big companies can solve by adding resources, we prioritize cost efficiency.

17. Outside may look like, we chose a very difficult model. But actually we do it very easily. Price cuts for our competitors, definitely not a good thing, they definitely won't cheer. I don't think selling API has that much attraction. I just need to get a few people to maintain this API, even no customer service, no sales needed, users will come on their own.

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

19. I don't even need to consider occupying a position in it then. As long as there is such a big business opportunity, there must be a way. DeepSeek is a product of an era, a reflection of reality, not a result of imitation.

Open Source is the Sweet Point for a Company of Our Scale

20. Restraint is a strategy: give up some, exchange for more other things. Open source is yielding profit: internally, employees have a sense of achievement, company has cohesion; also good for society, peers or ordinary people will be happy.

I have no doubt AGI will have very large commercial value. On this basis, what I prioritize is not taking more share, but increasing the probability of success.

21. To make AI work commercially, open source is beneficial. This sounds a bit counter-intuitive. Historically a software company, market one year maybe just billions of USD, open source and it's gone. But AI thing is big enough, eventually it may take up ten percent of human society GDP. If we want to monopolize this benefit, then definitely will be abandoned by history. This is an objective law, this is a historical view.

22. The open source models we give are the same as the models we deploy ourselves. We won't open source a worse model, deploy better models ourselves.

23. I don't worry about others deploying our models to compete with us. Not every company has this willingness and ability to reach this goal. Startups if too small, no power to do this; big companies again hard to organize. This belongs to a sweet point for a company of our scale.

24. Open source has no impact on our business model, premise is "only earn certain profit". If you want to earn 100x profit, open source definitely will affect you.

25. We are unwilling to become opponents with any internet big factory or small factory. Under this premise, we are very willing to assist, help anyone, even including Alibaba, Zhipu AI, Moonshot AI do better.

The Gap Between China and US is Not in Talent

26. In the future we want to rewrite the China-US AI narrative: using a fraction of computing power, shrink the gap shorter, shrink to 6 months, 3 months.

27. The gap between China and US AI is mainly in resources. We believe in Scaling, definitely the larger the scale, the better the effect. Training such large models, not because I think such large models are enough, but only have this much resources.

28. Talent has almost no gap—just the same batch of people. Domestic talent is not lacking. Talent shortage is temporary, historically never long-term lack of a certain type of people situation.

In Model Factory Competition, Cost Ranks First

29. Anthropic now exceeds OpenAI, not long-term, is temporary. OpenAI and Google future high probability will alternate rise.

30. Domestic doing model companies too many, every family does same thing, resources relatively scattered. Eventually will converge, but needs process. If every family only takes reasonable profit, don't need so many people do large models: maybe two big companies, two small companies enough.

31. I absolutely don't think large model companies can take (AI industry's) most profits.

32. Large model competition final gap will reflect in three aspects: costs, time, user experience.

Costs rank first—same quality service, you can at what cost provide. Second is time, early few months late few months not same. User experience will have some stickiness and barriers, but not essential.

No Intention to Become the Next Super App

33. We don't want to make the next Super App. Make the next ByteDance? Next Tencent? Completely no such thoughts.

34. Not fighting for this thing, is because behind there are watermelons, front may be all sesame. Of course maybe this sesame is relatively big, but I think all don't count big.

35. Last year everyone grabbed Chatbot, grabbed C-side traffic, this year is grabbing To B revenue. But we don't think it important, company internal truly care about is AGI roadmap, next step technology how breakthrough. Very strange, most want to get things don't get, not that much care about instead easy to get.

36. Last year Spring Festival hot, not in our script.

Maintaining Team Stability is Core

37. Only one is cannot concede: must maintain team stability. This is also we face a very big risk. Of course this risk with this one financing, got relatively big relief.

38. We do many things all for maintain team stability. We unwilling with any internet big factory or small factory become opponents, we hope give them empowerment, help them. We don't want make enemies. This way our own environment will also better.

39. Some people feel our organization is from top down, some people feel is from bottom up, I feel both correct. From top to bottom call "do right things", generally hope "right things" don't occupy employee all time half. He still has half time is from bottom up, is not arranged, want research what just research what, can go oneself explore, according to he feel what important, just go explore what, no front requirements.

40. We generally also not too much overtime. First reason is do research needs a relatively relaxed environment. Second reason is we very focused. Products many all not perfect, but we also no go fix it, also count as a kind of restraint culture.

41. Organization is dynamic, not is unchanging. Future with company become big, may have some adjustments. But won't completely become traditional hierarchy, but may have some necessary structures. But vision driven won't change.

With Goodwill Towards the World

42. We do this company, original intention no want earn how much money, want to capital market go list. Most beginning few tens people completely no so thought. If he so think, he just won't come. We are holding a towards this world very big goodwill to do this thing, feel this for humans useful.

43. "Want to achieve XX KPI" not is our way, we are a rely vision driven organization. This has benefits, also has drawbacks. Future will think methods maximize strengths avoid weaknesses, but this is our feature.

44. Vision even also not is written, but is in the methods of doing things, in the attitude towards the world. May our company every person to this vision understanding also not same, but in big direction is consistent.

45. About twenty years ago, I in management most worship is Jack Welch (GE former CEO). Now look, he said most things may already all not correct, but he one point said correct: a company most important is vision.

Vision not is hang on wall slogan, not is how say, but is how do.

Restraint Enables Us to Achieve AGI More Likely

46. AGI is the most profitable. That other things, if have energy, we will do, no energy, we just no do. Restraint is part of our vision.

47. AI this thing too big, benefits also too big. Just can achieve, casually divide one point benefits just very big. And you more restraint, just more possible can achieve this thing.

48. I feel this is accord intuition—at least is accord my intuition. Except vision, we no have very many other advantages.

49. Two years ago establish this company time, no have many money, no have many cards, no have what popularity, again no have what appeal, we just are a group very ordinary people. I like narrative is "a group ordinary people made not ordinary things", not is "a group geniuses made not ordinary things".

50. Open source also belongs to restraint part. On pricing, we also definitely not is with company income highest or profit highest as starting point. From short term speak, price high one point, income more one point; but from long term speak, really not good say. To me speak, restraint is a strategy.

51. Open source and low price, let employees feel very have achievement sense, organization therefore has cohesion, to society has benefits, other peers and ordinary people all will very happy. So this kind restraint, from vision come speak, can increase we achieve AGI probability.

52. If your vision is take more, you just first lost, you may facing will be bigger difficulties. This world is just like this.