Wenfeng Liang: Four-Hour Investor Meeting Transcript | elsewhere

@elsewhere<br>Last month, *elsewhere* reported on DeepSeek's fundraising story. The most discussed part was undoubtedly the legendary four-hour investor meeting.<br>In the month since, Wenfeng Liang's various quotes have circulated widely, and we've gathered some of their contents from multiple sources.<br>Throughout, Liang said "no" many times: not a genius, not chasing unreasonable profits, not pursuing user numbers, not closed-source, not doing 3D / video generation / world models, not building the next super app. Restraint, he said, is a strategy — traded for a higher probability of achieving AGI.<br>From the limited material we've seen, the high-frequency words include: model, cost, AGI, time, open source.<br>Most of the time, Liang's tone was measured, his language plain and unadorned. Only on a few matters he cared deeply about did a certain sharpness emerge: "As long as I can maintain team stability, I will definitely achieve AGI. It's that simple."<br>Below are the 52 quotes we've collected. Some phrasing may differ slightly from the original while preserving the meaning.<br>DeepSeek Has Only One Main Thread<br>1. Now is not the time for product-driven profit maximization. The path to AGI requires passing through product as a stepping stone, but we don't need to pour too much energy into C-end or B-end products. When you're positioned at a higher technical level and work on relatively lower-level technology, it's dimensional reduction. Products are byproducts on the road to AGI.<br>2. Many things aren't on our main thread — 3D, video generation; world models too, which don't have much to do with the upper limits of intelligence.<br>3. Multimodality matters a lot for products and for C-end users. But it's just a component, not the main thread or intelligence itself.<br>4. Of course there are ways to solve LLM hallucination, but it's a long-term proposition. Internally, we classify hallucination as a product problem — we'll address it, but it's not the priority.<br>5. At this stage, Coding Agent is what matters most. Looking at the domestic situation, the most sensible approach is to go all-in on general-purpose Agents; finance, healthcare, and other vertical Agents are lower priority.<br>6. If the AI era produces many trillion-dollar companies, it would be great if DeepSeek is one of them.<br>Continuous Learning First, Then AI Self-Iteration, Ending at Embodied Intelligence<br>7. What AI lacks right now isn't taste or intuition — it's the ability to learn continuously.<br>8. Humans can learn continuously, but for the same task, you have to give AI all the context. That's nearly impossible, so AI can't replace employees. Therefore, the next-generation model must have continuous learning capability to qualify as next-generation.<br>9. We hope the next-generation model can help with our own development. Put simply, the first goal of our models isn't that everyone finds them easy to use — it's that we find them easy to use. That's the fastest path to AGI.<br>10. No one in the world has found a good method yet, because "learning" is composed of many things.<br>11. DeepSeek's long-term vision is AGI. If you picture the path as climbing stairs, last year's step was CoT (Chain-of-Thought), this year's step is Agent. After Agent, the problem to solve is continuous learning.<br>12. After achieving continuous learning, we may reach a gradual singularity: models can do everything humans can, including developing more advanced AI models themselves — AI accelerating AI research. Only after this step comes embodied intelligence.<br>13. The endpoint of intelligence may all be embodied. Because for a normal person, what they need isn't a computer — it's human labor.<br>Still Far From a Full Pivot to Commercialization<br>14. We only take reasonable profit, not profit-maximizing pricing.<br>15. For one of our models, we initially worried about too much demand and set the price high, then cut it to one-quarter — many people in the company group chat cheered. Because this is exactly why we worked so hard to make the model good: so everyone can use it fully.<br>16. Low cost is an outcome — our models have been architecturally moving toward lower costs all along. We also want costs to be affordable, especially against a backdrop of compute scarcity.<br>Another reason: the lower the cost, the larger the model you can afford. When compute is limited, higher computational efficiency lets you train bigger models. Big companies can solve problems by adding resources; we prioritize cost efficiency.<br>17. From the outside, it may look like we chose a very hard model. But actually we're doing it very easily. A price cut is definitely not good news for our competitors — they're certainly not cheering. I don't find selling APIs that attractive. I just need a few people to maintain the API, no customer service needed, no sales, users come on their own.<br>18. We've been commercializing all along, just not with commercialization as the goal. The...