DeepSeek Liang Wenheng’s 4-hour internal exchange highlights: 1. Products are merely byproducts on the path to AGI. At this stage, we don’t pursue maximum profitability—leveraging technological superiority to build low-level applications is a dimensional advantage; there’s no need to expend excessive effort on C-end or B-end markets. 2. 3D generation, video generation, and even world models are not on our core path—they have little to do with the upper limits of intelligence. 3. Multimodality is important for C-end products, but it’s merely a component, not the core or intelligence itself. 4. Large model hallucinations are a complex, long-term challenge with inevitable solutions. Internally, we treat them as product issues to be resolved, but they are absolutely not our current priority. 5. The top priority right now is the Coding Agent. The most rational approach in China is to fully focus on building a general-purpose Agent; vertical domains like finance and healthcare are lower priority. 6. In the future AI era, there will inevitably be trillion-dollar giants; DeepSeek only needs to secure one seat. The path is clear: continuous learning → AI self-iteration → embodied intelligence. 7. AI currently lacks nothing in taste or intuition—it only lacks the ability for “continuous learning.” 8. Humans can learn continuously, but AI today relies on full context and cannot truly replace human workers. The next-generation model must cross the threshold of “continuous learning.” 9. The first goal of the next-generation model: we ourselves find it useful for assisting our own development. This is the fastest route to AGI. 10. The “learning” mechanism is extremely complex—no one globally has yet found a perfect solution. 11. DeepSeek’s ultimate vision is AGI. Our climbing strategy: last year was CoT (Chain-of-Thought), this year is Agent, next is “continuous learning.” 12. Mastering continuous learning will trigger an incremental singularity: AI will be able to take over all human work, even autonomously developing cutting-edge AI models. Only after this step can embodied intelligence be achieved. 13. The ultimate form of intelligence will most likely be embodied. What humans truly crave is “human labor,” not “computers.” A full shift to commercialization remains far off. 14. We reject profit-maximizing pricing—we only seek reasonable profits. 15. A good model should be widely used. We once proactively cut model prices to one-fourth of their original level—the entire company celebrated. 16. Cost leadership is an inevitable result of architectural optimization. Under scarce compute resources, extreme computational efficiency means we can train larger models. We win through efficiency; big companies win through resource accumulation. 17. Outsiders see our model as extremely difficult to replicate; internally, it’s remarkably easy. Lowering prices makes competitors suffer, but for us, selling APIs requires no heavy operations or sales teams—users vote with their feet. 18. Commercialization is ongoing, but it’s never our ultimate goal. The day we fully pivot to commercialization is still extremely distant. 19. If something has massive commercial value, there’s no need to rush for positioning. DeepSeek is a product of our times—not blind imitation or trend-chasing. 20. Restraint is strategy: we gain more by letting go. Open source is giving: internally, it unites our team and fosters achievement; externally, it benefits society and supports peers. 21. AI is a grand narrative that will occupy 10% of global GDP. To succeed commercially, open source is essential. Attempting to monopolize profits will lead to historical obsolescence. 22. Our open source has no tricks—we release models identical to those we deploy internally; we never withhold anything. 23. We’re not afraid of open source being copied or competed with. Big companies are bloated; small ones lack capital; very few competitors have both the willingness and capability. 24. As long as we stick to “reasonable profits,” open source won’t harm our business model. If you want 100x returns? Open source isn’t for you. 25. We don’t create enemies. We don’t oppose any internet companies—big or small—and gladly help everyone (including Alibaba, Zhipu, Moonshot, etc.) improve. 26. We must rewrite the U.S.-China AI narrative: achieve a technology gap reduction of 6 months—or even 3—with just a fraction of the compute resources. 27. The gap between U.S. and Chinese AI lies only in resources, not technology. We believe in Scaling Law—the current scale limits are solely due to insufficient resources. 28. Talent is zero-differentiated! The U.S. and China share the same talent pool. China does not lack talent—shortages are merely temporary and阶段性 (phase-specific). 29. Anthropic overtaking OpenAI is only temporary. In the future, OpenAI and Google will likely alternate as leaders. 30. China’s large model space is overcrowded and resources are overly fragmented. The endgame will inevitably consolidate—two big players and two small ones will suffice. 31. I assert: large model companies will never capture the majority of profits in the AI industry. 32. The ultimate competition among large models hinges on three things: first, cost (how low can you go at equal quality); second, timing (months or even days make all the difference); third, user experience. 33. We have zero interest in becoming the next super app—being the next ByteDance or Tencent holds no appeal for us. 34. We don’t chase immediate gains—we’re laser-focused on the bigger prize: AGI. 35. Last year we chased C-end; this year we chase B-end—we don’t care about either. Internally, we obsess over the AGI roadmap. Ironically: the less we chase fame and profit, the more we tend to receive. 36. Last year’s sudden viral success was completely unplanned—an accident outside our script. 37. The only non-negotiable底线 (bottom line): team stability. This was our greatest risk—but our latest funding round has fully resolved it. 38. We avoid creating enemies, build goodwill everywhere, and proactively empower others—all business actions aim to create the most stable and pure R&D environment for our team. 39. Management is both disciplined and unconventional: 50% of time spent top-down on “serious work”; 50% bottom-up on “wild exploration”—no requirements, no restrictions, just pure frontier experimentation. 40. We rarely work overtime—research requires mental relaxation; we’re extremely focused, and even when products have flaws, we don’t rush to fix them—this too is restraint. 41. Our organization evolves dynamically with company growth. Structural forms may emerge later, but we’ll never adopt rigid traditional hierarchies—the vision-driven core remains unchanged forever. 42. Our original intent was never profit or IPOs—the first core team wouldn’t have joined if they only wanted wealth. We’re doing this out of deep goodwill toward the world. 43. We don’t obsess over KPIs—we’re driven purely by vision. This is our double-edged sword—and DeepSeek’s defining characteristic. 44. Vision isn’t written on paper—it’s embedded in how we act. Even with thousands of different individuals, our direction remains unified. 45. Once admired Jack Welch; now most of his theories feel outdated—but one truth remains eternal: a company’s core is its vision. How you act matters far more than what you say. 46. AGI is the only ultimate payoff. If we have capacity, we do other things; if not, we cut them decisively. Restraint has been written into our vision from the start.47. The AI cake is too large. The more restrained and less greedy we are, the higher the probability of accomplishing this great endeavor. 48. Practicing restraint aligns with my intuition. Beyond our vision, we don’t possess any extraordinary advantages over others. 49. Two years ago, we had no money, no cards, no fame. My favorite narrative was: “A group of the most ordinary people achieved something extraordinary,” not “geniuses changing the world.” 50. Open-sourcing and low pricing are both forms of restraint. In the short term, we earn less; in the long term, they dramatically increase our chances of achieving AGI. Strategic restraint is our ultimate weapon. 51. Open-sourcing makes our employees proud, benefits our peers, and empowers the public. This restraint is infinitely amplifying our probability of achieving AGI. 52. If your vision is merely “take more,” you’ve already lost at the starting line. This is the fundamental logic of how the world works.