DeepSeek Liang Wenfeng's 4-hour internal communication tips: 1. The product is just a byproduct leading to AGI. At present, not seeking to maximize profits and using high-level technology for low-level applications is a dimensionality reduction blow, without the need to spend too much energy on the C-end and B-end. 2. 3D、 Video generation and even world models are not on our main line, and they have little to do with the upper limit of intelligence. 3. Multimodality is important for C-end products, but it is only a component, not the main line and intelligence itself. 4. The illusion of a large model is a long proposition that must have a solution. Internally, it will be addressed as a product issue, but it is not the current core focus. At present, the top priority is the Coding Agent. The most reasonable strategy in China is to fully conquer general agents, with priority given to vertical fields such as finance and healthcare. In the future era of AI, there will inevitably be trillion level giants, and DeepSeek only needs to occupy a place. Clear path: continuous learning ->AI self iteration ->embodied intelligence. Currently, AI does not lack taste and intuition, the only thing it lacks is the ability to "continuously learn". 8. Humans can continue to learn, but AI currently relies on comprehensive context and cannot truly replace employees. The next generation of models must overcome the hurdle of 'continuous learning'. 9. The first goal of the next generation model is to use it that we find useful to assist our own development. This is the fastest shortcut to AGI. The mechanism of "learning" is extremely complex, and currently there is no perfect solution found globally. 11. DeepSeek's ultimate vision is AGI. Stair climbing strategy: Last year it was CoT (Siwei Chain), this year it is Agent, and the next step is "continuous learning". 12. Overcoming continuous learning will usher in a gradual singularity: AI can take over all human work, and even independently develop cutting-edge AI models. Completing this step is the embodiment of intelligence. 13. The ultimate goal of intelligence is likely to be embodiment. What humans truly crave is' manpower 'rather than' computers'. The complete shift towards commercialization is still far away. 14. Refuse to maximize profit pricing, we only make reasonable profits. A good model should be fully utilized. We took the initiative to lower the model price by a quarter, and the whole company cheered for it. Cost leadership is an inevitable result of architecture optimization. Under the shortage of computing power, ultimate computational efficiency means the ability to train larger models. We win by efficiency, while big companies rely on resource accumulation. 17. Our model may seem extremely difficult to the outside world, but internally it is actually extremely easy. Lowering prices makes our competitors uncomfortable, but for us, selling APIs does not require heavy operations and sales, and users will vote with their feet. Commercialization has always been pursued, but it is never the ultimate goal. The day when we will completely turn to commercialization is still extremely far away. As long as there is significant commercial value, there is no need to grab a spot in advance. DeepSeek is a product that conforms to the reality of the times, not blindly imitating trends. 20. Restraint is strategy: giving up in exchange for more. Open source means giving up benefits: it brings together people's hearts and achievements, and benefits society and peers externally. 21. AI is a grand narrative that accounts for 10% of human GDP. To succeed in business, open source is essential. Attempting to monopolize profits will inevitably be abandoned by history. 22. Open source has no tricks. Our open-source model is completely consistent with our own deployment, and we will never hide anything. 23. Fearless of open source being plagiarized and competing. Large factories have bloated organizations, small factories lack financial resources, and there are few willing and capable competitors. As long as we adhere to the principle of "only making reasonable profits", open source will not harm the business model. Want to make a hundredfold profit? Open source is indeed not suitable for you. 25. Do not make enemies. We are not willing to oppose any large, small, and medium-sized Internet manufacturers, but are willing to help everyone (including Ali, Zhipu, and the Dark Side of the Moon) do a better job. 26. It is necessary to rewrite the AI narrative between China and the United States: with a fraction of computing power, shorten the technological gap to 6 months, or even 3 months. The gap between China and the United States in AI lies only in resources, not in technology. Adhering to the Scaling Law, the current scale limitation is only due to the limited resources available to do so. 28. Zero talent gap! China and the United States are the same group of people. There is no shortage of talent domestically, and the shortage is only a temporary phenomenon. Anthropic's ability to surpass OpenAI is only temporary. In the future, OpenAI and Google are likely to alternate and lead the way. 30. The domestic large model track is too crowded, and resources are extremely scattered. The final outcome will inevitably converge, two large factories and two small factories are enough. I assert that big model companies will never take away the vast majority of profits in the AI industry. 32. The ultimate competition of big models only looks at three points: first, cost (how low the cost is under the same quality), second, time (the difference between morning and evening months), and third, user experience. I never want to create the next super app. Becoming the next ByteDance or Tencent? We have no interest at all. 34. Don't argue with the sesame seeds in front of us, because we are fixated on the watermelon behind us (AGI). Last year we competed for the C-end, this year we compete for the To B end, and we don't care. Internally, only stubbornly adhere to the AGI roadmap. Interestingly, the less one seeks fame and fortune, the easier it is to obtain them. The sudden explosion of popularity during last year's Spring Festival was completely unexpected beyond the script. The only bottom line that cannot be compromised: team stability. This is the biggest risk, fortunately the latest round of financing has completely resolved it. 38. Not making enemies, building good relationships, and actively empowering, all business actions are aimed at creating the most stable and pure research and development environment for the team. 39. Management is both positive and negative. 50% of the time is spent "doing business" from top to bottom, and 50% of the time is spent "tinkering" from bottom to top - without requirements or restrictions, exploring the forefront to the fullest. 40. Very little overtime. Research requires a sense of relaxation; We are extremely focused and do not blindly remedy any defects in our products, which is also a form of restraint. Organizations dynamically evolve as the company grows. In the future, there will be necessary structures, but traditional rigid hierarchies will never be established, and the vision driven core will never change. 42. The original intention was not to make money and go public. If the first core team only wants to make money, they won't come at all. We are doing this with great goodwill towards the world. 43. Don't focus on KPIs, rely solely on vision driven. This is our double-edged sword, and it is also a characteristic engraved in DeepSeek's bones. 44. The vision does not fall on paper, but is integrated into the attitude towards work. Even if there are thousands of people and faces, the overall direction will always be the same. Once admired Jack Welch, most of his theories seem outdated now, but there is one thing that is always correct: the most core of a company is its vision. How to do it is far more important than how to say it. AGI is the only maximum benefit. If you have spare energy, do something else; if you don't have energy, cut it down decisively. Restraint has long been written into our vision. The AI cake is too big. The more restrained and not greedy one is, the higher the probability of achieving this great thing in the end. 48. Adhering to restraint is in line with my intuition. Apart from our vision, we do not have any extraordinary advantages over others. Two years ago, I had no money, no cards, and no fame. My favorite narrative is: 'A group of the most ordinary people did the most extraordinary things,' rather than geniuses changing the world. 50. Both open source and low price are restraints. In the short term, it may earn less money, but in the long term, it can greatly improve the success rate towards AGI. Strategic restraint is our ultimate weapon. 51. Open source makes employees proud, benefits peers, and benefits the public. This restraint is infinitely increasing our probability of achieving AGI. If your vision is only to 'take more', you have already lost at the starting line. This is the underlying logic of the world's operation.