DeepSeek Liang Wenfeng 4-hour internal exchange highlights:

1. The product is only a byproduct on the way to AGI. At this stage, we don’t seek to maximize revenue. Standing at a technical high point and building low-level applications is a dimensionality reduction—there’s no need to spend too much effort on the consumer (C-end) and business (B-end).
2. 3D, video generation, and even world models are not in our main line. They have little relationship with the upper limit of intelligence.
3. Multimodality is very important for C-end products, but it’s only a component, not the main line—nor intelligence itself.
4. Model hallucinations are a long-standing problem with a long solution; it must have an answer. Internally, we treat it as a product issue to be solved, but absolutely not the current core focus.
5. At this stage, the top priority is the Coding Agent. The most reasonable approach in China is to fully focus on building general-purpose agents; vertical domains like finance and healthcare come later in priority.
6. In the future AI era, there will definitely be trillion-level giants. DeepSeek only needs to secure a seat. The path is clear: continuous learning -> AI self-iteration -> embodied intelligence.
7. AI currently lacks neither taste nor intuition; the only missing ability is “continuous learning.”
8. Humans can continuously learn, but AI currently relies on full context and cannot truly replace employees. The next-generation model must clear the hurdle of “continuous learning.”
9. The first goal of the next-generation model: make it something we ourselves find useful, using it to assist our own development. This is the fastest route to AGI.
10. The “learning” mechanism is extremely complex, and globally, nobody has found a perfect solution yet.
11. DeepSeek’s ultimate vision is AGI. The stair-climbing strategy: last year was CoT (chain-of-thought), this year is Agent, and the next step is “continuous learning.”
12. Cracking continuous learning will bring a progressive singularity: AI can take over all human work, and even independently develop cutting-edge AI models. Once you get through this step, then comes embodied intelligence.
13. The endpoint of intelligence will most likely be embodied. What humans truly crave is “labor,” not “computers.” A complete shift to commercialization is still far away.
14. We reject profit-maximizing pricing; we only make reasonable profit.
15. Good models should be fully utilized. We once proactively cut the model price to one quarter, and the whole company cheered for it.
16. Cost leadership is an inevitable outcome of architectural optimization. With compute being scarce, extreme computational efficiency means you can train bigger models. We win through efficiency; big companies win by stacking resources.
17. The outside world finds our model extremely hard to copy. Internally, it’s actually extremely easy. Cutting prices makes competitors uncomfortable, but for us, selling APIs doesn’t require heavy operations and sales—users will vote with their feet.
18. Commercialization has always been in progress, but it is never the ultimate goal. The day we completely shift to commercialization is still extremely far away.
19. As long as there is massive commercial value, there’s no need to grab a position early. DeepSeek is a product of adapting to the real world of the times, not blind imitation.
20. Restraint is strategy: trade what you give up for more. Open-source as a way to share benefits: internally unite people and build achievements; externally benefit society and help peers.
21. AI is a grand narrative that takes up 10% of human GDP. To succeed commercially, open-source is indispensable. If you try to monopolize profits, you will inevitably be discarded by history.
22. There are no tricks in open-source. The models we open-source are exactly the same as the ones we deploy ourselves—we never keep anything hidden.
23. We are unafraid of open-source being copied and competing. Big companies are bloated; small companies lack financial strength; opponents with both willingness and capability are few.
24. As long as we stick to “we only make reasonable profit,” open-sourcing won’t harm the business model. Want to make 100 times the windfall profit? Then open-source really isn’t for you.
25. Don’t make enemies. We don’t want to go against any internet companies of any size; instead, we’re happy to help everyone (including Alibaba, Zhipu, Moonshot, etc.) do better.
26. We must rewrite the AI narrative between China and the US: using only a fraction of the compute, shorten the technical gap to 6 months, or even within 3 months.
27. The AI gap between China and the US is only in resources, not in technology. Believing in Scaling Law, the current scale limitation is simply because the available resources are only enough to go this big.
28. Zero gap in talent! China and the US share the same pool of people. China doesn’t lack talent; shortage is only a temporary, stage-based phenomenon.
29. Anthropic overtaking OpenAI is only temporary. In the future, it is likely to be a duo-race where OpenAI and Google alternately lead.
30. China’s large model track is too crowded, and resources are extremely scattered. The endgame will inevitably converge: two big companies and two small companies are enough.
31. I assert: large model companies absolutely cannot take away most of the profits in the AI industry.
32. The ultimate competition among large model players comes down to three things: first is cost (how low it is under the same quality), second is time (being earlier by a few months makes all the difference), and third is user experience.
33. We never want to be the next super app. Becoming the next ByteDance or Tencent? We’re not interested at all.
34. We don’t fight for the sesame seeds in front of us, because we’re dead set on the watermelons behind them (AGI).
35. Last year we fought for the C-end; this year we fight for To B—we don’t care. Internally, we only obsess over the AGI roadmap. The interesting part is: the less we seek fame and profit, the easier it often is to get them.
36. The sudden breakout last Spring Festival was entirely an accident outside the script.
37. The only non-negotiable bottom line is team stability. This is the biggest risk; luckily, the latest round of funding has completely resolved it.
38. Don’t make enemies, build good relationships broadly, and proactively enable others—every commercial move is to create the most stable and pure R&D environment for the team.
39. Management is both right and wrong. 50% of the time, work top-down on “proper matters”; 50% of the time, work bottom-up on “messing around”—no requirements, no restrictions, explore freely the frontier.
40. Very little overtime. Research needs a sense of looseness; we’re extremely focused. Even if the product has imperfections, we don’t blindly patch them up—this is also a kind of restraint.
41. The organization dynamically evolves as the company grows. In the future, there may be a need for structure, but we will never do traditional rigid hierarchies. The vision-driven core will never change.
42. The original intent was absolutely not to make money and go public. If the first core team members only wanted to get rich, they wouldn’t have come. We are doing this with immense goodwill toward the world.
43. Not fixated on KPI, driven purely by vision. This is our double-edged sword—and also DeepSeek’s signature etched into its bones.
44. The vision doesn’t stay on paper; it is embedded in our attitude toward doing things. Even with a thousand people having a thousand faces, the overall direction is always the same.
45. I once admired Jack Welch. Now it seems most of his theories are outdated, but there is one sentence that is always correct: the company’s most core thing is vision. How to do it matters far more than what to say.
46. AGI is the only biggest payoff. If we have spare capacity, we do some other things; if we don’t have the energy, we cut them decisively. Restraint has long been written into our vision.
47. The AI “cake” is too big. The more restrained and non-greedy we are, the higher the probability we ultimately do this big thing.
48. Sticking to restraint aligns with my intuition. Besides vision, we don’t have any “unearthly” advantages over others.
49. Two years ago, we had no money, no cards, and no name. My favorite narrative was: “a group of the most ordinary people did the most extraordinary thing,” rather than “geniuses change the world.”
50. Open-source and low pricing are both forms of restraint. In the short term, it means we make less money; in the long term, it can greatly increase the success rate of reaching AGI. Strategic restraint is our ultimate weapon.
51. Open-source makes employees proud, helps peers, and benefits the public. This kind of restraint is infinitely amplifying our probability of achieving AGI.
52. If your vision is only “take away more,” you already lost at the starting line. That is the underlying logic of how the world runs.