The DeepSeek Doctrine - by X.PIN and CT Zhao - X.PIN
SubscribeSign in
The DeepSeek Doctrine<br>What Liang Wenfeng’s Four-Hour Investor Meeting Reveals About AGI, Open Source, and Restraint.
X.PIN and CT Zhao<br>Jul 23, 2026
Share
Late on July 22, we came across an article titled “A Four-Hour Investor Meeting With Liang Wenfeng,” published by elsewhere , a WeChat publication.<br>The piece contains excerpts from a recent conversation between DeepSeek founder Liang Wenfeng and investors during the company’s latest fundraising round. Liang addresses almost every important question surrounding DeepSeek: why build foundation models, why embrace open source, and why continue pushing costs lower.<br>But the transcript reads less like a fundraising discussion and more like a philosophical inquiry. At its center is one of the deepest questions in AI: what happens to human labor when intelligence becomes cheap?<br>We translated the article. It offers one of the clearest glimpses yet into the thinking behind China’s open-source AI movement.<br>It may be the closest thing the AI industry has to scripture.
Last month, elsewhere reported on DeepSeek’s fundraising. The most discussed part of the story was the now almost mythical four-hour investor meeting.<br>Since then, remarks attributed to Liang Wenfeng have circulated widely. We also gathered fragments of the conversation from multiple sources.<br>Throughout the meeting, Liang said “no” repeatedly. No, he is not a genius. No, DeepSeek will not chase unreasonable profits or pursue user growth for its own sake. No, it will not close-source its models. And no, it will not build 3D generators, video generators, world models, or the next super app.<br>For Liang, restraint is a strategy: a way to improve DeepSeek’s chances of eventually reaching AGI.<br>Across the limited material we were able to review, the same words appeared again and again: models, cost, AGI, time, and open source.<br>Liang spoke cautiously for most of the meeting. His language was plain and understated. But when the conversation reached the subjects he cared about most, a sharper confidence appeared.<br>“As long as I can keep the team stable, I will be able to build AGI,” he said. “It is that simple.”<br>Below are 52 remarks we collected from the meeting. Some wording may differ slightly from the original, but we have preserved the meaning as accurately as possible.<br>DeepSeek Has One Main Goal
1. Now isn’t the time to make as much money as possible from products. Products are one step on the road to AGI, but we don’t need to spend too much time building consumer or enterprise products. If you control the more advanced technology, the products below it become much easier to build. To us, products are a byproduct of the journey toward AGI.<br>2. A lot of things aren’t part of our main plan. That includes 3D generation and video generation. The same goes for world models. We don’t think they have much to do with how intelligent a model can ultimately become.<br>3. Multimodality matters a lot for products and consumer users. But it’s still only one component. It isn’t our main goal, and it isn’t intelligence itself.<br>4. There are ways to reduce hallucinations in large models, but it’s a long-term problem. Internally, we see hallucinations mainly as a product issue. We’ll work on it, but it isn’t our main focus right now.<br>5. At this stage, coding agents are still the top priority. Given the situation in China, the best approach is probably to focus on a general-purpose agent. Specialized agents for finance, healthcare, and other industries can come later.<br>6. If the AI era creates many trillion-dollar companies, it would be enough for DeepSeek to become one of them.<br>Continuous Learning, Self-Improving AI, and Embodied Intelligence
7. AI doesn’t lack taste or intuition right now. What it lacks is the ability to keep learning.<br>8. Humans keep learning over time. But with AI, you have to provide all the relevant context every time you ask it to do something. That’s almost impossible. This is why AI can’t truly replace employees yet. The next generation of models needs to learn continuously. Otherwise, it isn’t really a new generation.<br>9. We want our next model to help us develop future models. Put simply, our first goal isn’t to make the model useful for everyone else. It’s to make it useful for us. We think that’s the fastest path to AGI.<br>10. No one has found a good solution yet, because “learning” isn’t just one thing. It involves many different processes.<br>11. DeepSeek’s long-term goal is AGI. If the road to AGI is a staircase, last year’s step was chain of thought, or CoT. This year’s step is agents. After agents, the next problem is continuous learning.<br>12. Once we achieve continuous learning, we may enter a gradual singularity. Models could eventually do everything humans can do, including building more advanced AI models. In other words, AI could speed up AI research. Embodied intelligence comes after that.<br>13. Intelligence may eventually...