1. When we first started this company, our original intention wasn't to think about how much money we'd make in the end, or to go to the capital market, or to list, or whatever. The initial dozens of people never thought like that; if they had, they wouldn't have come.

2. We did this with great goodwill towards the world. We felt it was useful to humanity; it's something beyond money. Our original intention, our vision, and the vision we've maintained until now were not done in a way that maximizes commercial interests.

3. Managing a large company doesn't rely on your rules and regulations; it relies on vision. A vision isn't a slogan hanging on the wall; a vision is how you act, not what you say. It's about how you actually operate.

4. We are unorganized; we are driven by vision, organized by a single vision. We don't do things in a "I want to achieve certain KPIs, no assessment" way. There is only vision.

5. This vision isn't even written down. Nothing has ever been written out. This vision exists in our methods of doing things and our attitude towards the world.

6. We don't have many other advantages. We don't have special skills. We aren't richer than others, nor do we have better personnel than other companies. Actually, we don't. When we founded this company two years ago, we didn't have much money, many cards (GPUs), much fame, or much influence. We were just a group of very ordinary people.

7. The more restrained you are, the easier it might be to succeed, or at least so far it has been proven and explained. Otherwise, there's no way to explain why we succeeded: we didn't have any weapons, our starting point was very low, resources were scarce, and our people were essentially a random group of ordinary folks.

8. AI is too big, and the stakes are too high. We are very restrained. As long as we can succeed, the final benefits will be huge. Even a small share would be very valuable. So right now, there's no need to consider which part of these benefits to take or how to take them. I think we really don't need to consider this because the potential benefit is large enough.

9. Last Spring Festival, user numbers suddenly surged, but we didn't pursue retaining these users, monetizing them, or grabbing commercial benefits from them. We didn't fight for users or make money; instead, we worked hard to serve them well.

10. We never had the idea of becoming the next Super App, competing with anyone, or becoming the next ByteDance or Tencent. We absolutely didn't have such thoughts. I believe the opportunities for AGI later on should be very large, and the opportunities for AGI will always be very large.

11. Restraint is a strategy. It means sometimes you can give up some things to gain more of others. Not open-sourcing is the same; it can be seen as our pressure or our concession.

12. I understand this restraint, in the long run, increases our probability of achieving AGI. When considering a matter, I have no doubt that AGI will have huge commercial value. On this basis, my priority isn't how to add a bit more share or take more share; my priority is how to increase the probability of us succeeding.

13. We have always been very restrained, unwilling to become rivals with any major or minor internet company. I hope to empower them, or assist everyone in doing this thing, hoping to help everyone do this.

14. I feel that by maintaining this attitude previously, we haven't lost anything because of it. We haven't received less because we open-sourced, or because of our goodwill or helping others. Instead, we might have gained extra points. This seems counter-intuitive, but it is indeed so.

15. We aim for AGI, but we have always been doing commercialization, which is why we have C-end users and B-end revenue. Historically, this strategy has been successful.

16. If you can describe a problem clearly and provide complete context and instructions, it exceeds humans. But there's a definition and a premise here: you must give it complete context and complete instructions.

17. AI cannot replace your employees. But if AI has continuous learning capabilities, and learns at your company for two months like your employees, it could replace everyone in the world. So we are still one step away: continuous learning.

18. We can understand AI development as a series of steps. The step taken last year was Chain of Thought (CoT). Because we found that through CoT, intelligence can reach a higher level.

19. This year's step is Agents. Because we found that using Agents, even more tasks can be done, its capability range is larger, and its intelligence ceiling is higher. Agents use CoT, and CoT also uses previous steps. The previous step is language models, so no step is wasted.

20. After Agents, we think the problem to solve should be continuous learning. That is, how to let models learn continuously, rather than requiring strong training. It should be able to learn continuously for a relatively long time, like humans.

21. After continuous learning, we might reach a singularity. This singularity is when the model can learn continuously, it can already do everything humans can do. It can develop its own versions, research itself, and develop its next version, creating more advanced AI models.

22. This singularity isn't really a singularity; it's also a gradual process. This process might be a relatively long gradient, not a突变 (sudden change). But habitually, we all think it might be a singularity.

23. This is our speculation. We think the timeline should be: first solve learning-to-learn, then reach that intelligent singularity, the self-iterating singularity, and then embodied intelligence. After embodied intelligence, it enters the real world, can do housework for you, and care for the elderly.

24. If we first solve continuous learning, then solve that self-iterating singularity, then solve embodied intelligence, the path is quite easy. Because later on, you can use previous technologies to help develop subsequent technologies.

25. We only focus on the main line of AGI. The AI field is broad, and there are many things we feel are not on this main line, such as 3D and video generation. I think they might not have much relation to the main line of intelligence, so we won't do them.

26. When video generation first came out, it was very popular, as if it was mandatory. If you didn't do it, you weren't an AI company. So I was very confused. Actually, if you think carefully, it has little to do with the roadmap of intelligence.

27. Commercially, it's a good business. But it has nothing to do with intelligence. We won't do it just because it's a good business; we only do it if it's part of the intelligence roadmap.

28. From our judgment, World Models and Intelligence are not the most important things at this stage. The most important are AI training and how to solve continuous learning after AI training. This is our company's judgment, although every company's judgment is different.

29. We currently believe a narrative that AI can accelerate AI research. That is, it's not linear, because you can use AI to accelerate your own research, so later on it might be non-linear.

30. I think embodied AI definitely needs to enter, ultimately embodied. Because for a normal person, their need isn't a computer, right? Normal people eat, drink, play, wear, live, and travel; they don't need computers. They need... so they still need embodied intelligence to solve specific human labor needs.

31. What do we hope AGI can do? It can help me iterate the next version of the model, just like iterating the next version of the model. If there is embodiment, what we hope it does is also to iterate the next version of embodiment, to make the next version of robots.

32. The core capability of the next-generation model must have continuous learning capabilities; otherwise, it can't be called a next-generation model. Before that, what we can do is reduce costs, improve effects, and speed up. But for a major breakthrough, it should possess continuous learning.

33. The current limitations of Agents' capabilities are because they cannot learn continuously; they cannot effectively learn continuously. If we could finish continuous learning first, AI's capabilities would be very strong, greatly improving our own research efficiency.

34. Once continuous learning is achieved, general intelligence might be very easy, and using it would be easy. So I say this is a result we hope to see; we save effort, making it relaxed. Otherwise, going for general intelligence manually now is tiring and bitter, data-intensive, labor-intensive, and not cost-effective.

35. My experience before taught me that the vision of AGI is powerful. This talent advantage isn't that my people are smarter than others, but how I organize these talents, how to motivate them, and how to cooperate.

36. Gathering smart people together doesn't mean they will naturally cooperate, naturally run passionately towards a goal, and complete it. So you need a vision.

37. Our greatest core interest is maintaining team stability. This is our greatest core interest, and can even be considered the only core interest. As long as I can maintain team stability, I will definitely succeed, definitely achieve AGI. It's that simple.

38. Money is definitely not a problem, resources are not a problem, other factors are easy to obtain. For us, there is only one core interest, one thing we cannot yield on: we must maintain team stability.

39. This is also a huge challenge we face, or rather, I think it's the biggest risk. Of course, this risk has been largely alleviated by our recent financing round. Because everyone got relatively many options, and the amounts are still quite large.

40. In terms of team stability, as long as the most important employees and the oldest employees remain stable, others are unlikely to leave. Others, even if they have fewer options or lower income, won't leave. Because they aren't solely chasing money; everyone hopes to do this in an environment where AGI can be achieved.

41. Everything else is a matter of time. At most, it might delay us by half a year or a year, but it won't mean failure. Definitely不缺 money, definitely不缺 resources. Actually, these are not lacking.

42. Our gap with the US is mainly in resources. The gap in people is not very large. There's almost no gap in people because it's the same batch of people, possibly Chinese. When Chinese went abroad, some stayed domestically, some stayed abroad, some went abroad. It wasn't that smart people went abroad; no, it wasn't.

43. Talent is not the bottleneck; resources are the biggest bottleneck. Resources first affect talent cultivation. Because computing power is scarce, we have fewer opportunities for experiments, so our overall talent lags behind the US. The talent gap is essentially due to the computing power gap.

44. The shortage of AI talent is also阶段性 (phased), and we have already seen it significantly alleviated. Because there really aren't few AI people; every company can quickly cultivate people. Cultivating people is fast.

45. There are too many companies doing models domestically now, still too many. Maybe just three in the US, but China has too many doing base models. Eventually, it definitely won't require that many people to do base models; it will definitely converge.

46. Our company's management actually has two lines: one top-down, one bottom-up. Bottom-up means everyone decides what they want to do, does it themselves, no one manages them, no KPIs.

47. Generally, we hope employees have half their time unassigned, free to do what they want. This is a research scope, letting them explore themselves, exploring what they feel is important, without prior requirements.

48. We generally don't work overtime much. Overtime has two reasons. First, doing research requires a relatively relaxed environment. If you push too hard, you can't do research. Since it requires your own interest, and you need to think about these problems平时 (normally), it must be in a relatively relaxed environment to possibly explore.

49. Second, we are very focused. Being very focused means we have few things to do. Then I don't have so many things to do, so I don't need overtime. This is consistent with the restraint mentioned earlier.

50. Our company is fundamentally built on consensus. I don't decide everything alone; instead, I seek consensus. My authority within the company and my influence within the company are built on consensus.

51. This decision-making mechanism is actually a consensus-seeking mechanism. It's not that I can push something through; it must be consensus for me to push it forward, and then I will push it.

52. As personnel increases, we will make this adjustment. We should have to make this adjustment immediately, because I am already making this adjustment. If we don't make this adjustment, many things can't be pushed forward. Indeed, many departments should have organizational structures.

53. How many cards do we need? Now definitely more is better. Within our承受能力 (affordability/capacity), definitely more cards are better, which is unquestionable. So our current strategy is, within reasonable prices, buy as many cards as possible.

54. Actually, spending so much money is very difficult. Buying that many cards is hard, and prices are also high. You can't say spend very high prices to buy; you must ensure the price is reasonable. If we can spend twenty billion this year, then our procurement department's performance is super good.

55. The biggest gap between us and the US is in resources. Computing power resources: on one hand, cards are hard to buy domestically; on the other hand, our capital investment is less than the US. We have much less in capital investment. Based on talent salaries, the proportion here is very low. Look at their salaries opening at hundred million USD types, but calculating it, talent salaries still account for a small portion; the bulk is computing power.

56. All the differences we see, including talent differences, model capability differences, application differences, can be considered due to differences in computing power resources.

57. Our gap with the US might be lagging the US by 12 months, maybe 12 to 18 months, or 6 to 12 months. Simply put, we lag the US by two years, yet accomplish this with only one-twentieth of the US's computing power.

58. This narrative is lagging one to two years, but using only one-twentieth of its computing power. In the future, we want to rewrite this narrative: we use one-fraction of its computing power, but shorten the time further, to 6 months, 3 months. I think this is a goal.

59. Scaling, we believe in Scaling. Definitely, the larger the scale, the better the effect, unlocking more functions. What stops our Scaling is actually computing power, not that we don't want to Scale, but that we don't have enough computing power to do this Scaling.

60. We train such large models not because I think such large models are enough, but because I happen to have this much resource. I calculate based on my resources, what size model I can accept and train. That's how it's calculated; it's not that this model is enough.

61. Silicon Valley says Scaling has hit a ceiling, but that's for Silicon Valley; for Chinese people, we are far from that. We haven't even reached that degree of Scaling. This Scaling includes Data Scaling, Model Size Scaling, and Training Costs.

62. NVIDIA CUDA's moat is being rapidly eroded. On one hand, now there is AI, and after having AI, building this ecosystem is much easier than before, because AI can write code.

63. The market for compute cards is already larger than gaming cards, so there's no reason these two still need to be coupled. The current trend is that in the future, they will no longer be coupled. Then dedicated chips, whether Huawei or NVIDIA itself, will be dedicated chips in the future, not the previous things.

64. Domestic AI chip substitution currently has a historic opportunity. We believe that within the next year, we can see one thing verified: domestic chips' ecosystems are completely fine. Previously it was thought there were problems, thought unusable, hard to use, but in the future, I think within a year, we can reverse this perception, or use facts to reverse these.

65. Domestic AI chips' hardware and ecosystems are fine; the only problem is insufficient capacity. Adapting to domestic cards has no obstacles; NVIDIA can't stop it. If in a normal commercial environment, I can buy NVIDIA cards, then domestic substitution is relatively difficult; but under the circumstance that NVIDIA cards can't be bought, everyone is forced to do domestic chips.

66. During V3 training, it still used NVIDIA cards, but no longer used NVIDIA's ecosystem. V3 used NVIDIA cards, but didn't use NVIDIA's ecosystem; instead, we first wrote a high-level compiler called TileLang, and then completed all other matters based on TileLang's ecosystem, already almost not relying on NVIDIA's ecosystem.

67. I am relatively optimistic about domestic computing power. I think in this regard, NVIDIA is digging its own grave. Huawei's 950 super node can completely substitute NVIDIA's GB200, GB300 in performance and price.

68. Four Huawei cards equal one NVIDIA card.

69. Our gap with the US in chips, I think the ecosystem gap will no longer exist in the future, but in chips it is four times plus two years.

70. We are mainly cooperating with Huawei now. Huawei adapts themselves, but we will participate in this ecosystem ourselves, deeply participating in Huawei's inside. Huawei's problem is still insufficient capacity.

71. I don't really believe that after five years, we will still be stuck on capacity issues. Now definitely stuck on capacity issues, this year, next year, the year after, I think might still be stuck on capacity issues, but after five years, I think maybe not, I'm still relatively optimistic.

72. The gap in final effects among various models should be comprehensive. Comparing model effects, definitely must compare at the same cost; this is meaningful. Because comparing two cars, you also compare cars of the same price.

73. Anthropic surpassing OpenAI now, is this long-term? I think this isn't long-term; this is definitely phased. OpenAI and Google, in the future, will likely alternate rising.

74. When playing a leading role in global AI division of labor, Chinese companies are very likely to play a role as the largest producer. Common sense says, our production capacity is largest, including chips, chips might have our largest production capacity, our electricity is the most.

75. Chinese people will make this product the cheapest, and then in terms of effect, after all, foreign goods, many goods produced in China and the US don't have too big a difference now. Future AI might be like this too, but Chinese-produced AI might be cheaper. This cheapness might be systematically low, similar to how services provided by China in other industries are cheaper.

76. The final gap should be three aspects: one aspect cost, one aspect time, one aspect user experience. Besides this, there might be no gap.

77. Cost is definitely a difference. I think cost might be ranked first in distinction. Then second is time, when you can achieve it. Early by a few months, late by a few months, it's different.

78. OpenAI initially thought it really could monopolize this world, but actually it will encounter many, many challengers. It will encounter challenges, so it won't be so relaxed. The US will encounter challenges, so in the future it might still encounter Chinese challenges, because Chinese people are willing to take less, thus providing this service to you.

79. Those who take more will be defeated by those who take less. Even you don't really need to take more; if the vision is to take more, you will be defeated by those whose vision is to take less. Actually, nobody has gotten money yet; it's just a vision. Your vision is to take more, you lose first, you will face greater difficulties.

80. For us, we don't want to take the most profit, or calculate pricing for maximum return, but only earn a reasonable return. This is an explanation. I believe this matter; I'm not looking for reasons for this, because there's no need to look for reasons.

81. I think in many experience aspects, possibly we can do better than the US. In products, product capabilities won't necessarily be worse than the US. Costs should also be lower than the US, so China will still have competitiveness.

82. Cost is easy to understand, because they don't need to do it, so they don't develop this capability. They definitely don't value this matter as much as we do. We can treat it as a very important matter, but for them, this is unimportant.

83. Large models, maybe not saying two big companies, two small companies, might already be enough. The gap has only two things: one is time, one is cost. So it's not that any one has huge profits; I think there won't be huge profits. People who control costs well earn a bit more, people who control costs poorly earn a bit less, that's all.

84. Half of our company probably thinks OpenAI is better usually. Actually, Anthropic has a first-mover advantage, but this first-mover advantage should disappear soon; it's not an advantage it can hold long-term. These three are all very capable; among these three, its efficiency is the highest, the cost it spends, the money it burns should be the least.

85. Multimodal layout, we have always been doing it. For products, it's very important; for C-end user products, it's very important. But for the upper limit of intelligence, it's a component, not the main line itself.

86. We should launch related models; that is, our V4, V4's subsequent versions will support native multimodality. But for multimodality, for intelligence, it's a component; we don't treat it as intelligence itself.

87. Can only say Language Model Scaling, I currently haven't seen an upper limit. Our current intellectual level, or the US's intellectual level, haven't seen an upper limit.

88. Many people inside us have such thoughts: first must be useful to ourselves, first is for us to use. Then this is the fastest method to achieve AGI. When we ourselves find it useful, that means possibly others find it useful too, but first must guarantee we ourselves find it useful.

89. The first goal of the models we make isn't that everyone uses them well, but that we ourselves use them well. First is useful to ourselves. After being useful to ourselves, when I develop the next version of the model, it will be faster.

90. We call this "drawing lottery." The threshold is very low, anyone can draw, but who can draw what, this I might not know if it depends on talent or what. So here doesn't require us to allocate resources. Just that, the place where we differ from other companies is that we spend time discussing this problem, thinking about this problem, then treating it as an important matter.

91. Our API pricing is a reasonable profit, roughly ten months to recover costs buying a batch of equipment back from the market, I think this is a reasonable profit.

92. If maximizing profit, should set prices higher. Because in this price range, user demand is inelastic, that is, if I double the price again, or raise the price another fold, token consumption volume difference isn't big.

93. One of our models, initially we worried demand would be too much, so set prices relatively high, team members weren't very happy. Later I lowered the price again, down to a quarter, everyone was very happy.

94. To B business's upper limit should still be demand. Under the background of this generation of AGI, AI technology, To B demand should be limited. It will grow rapidly, but isn't an infinitely large matter; ultimately constrained by demand, not computing power.

95. I currently think it should be achievable, want both. Suppose I can have several hundred million USD B-end revenue this year, plus we have C-end users, then this itself already has a certain commercial foundation. With next year we have B-end revenue, if this demand can increase further, the company is not far from net profit, might already be net profit.

96. Worst case selling API, might still support a listed company. That is, if technology has no new progress later, our technology freezes here, then finally we fully sell API, do these services well, I think it's enough.

97. Looking at our current situation, I think the most reasonable approach should be fully doing general Agents, other Agents' priorities should be lower, including finance, doctor these Agents. Must do Coding first, because Coding Agents can do many things, there are also many vertical Agents. At this stage we think most important should still be Coding Agents.

98. I think low cost is first a result. Our models indeed keep moving in a lower cost direction in model architecture, this relates to our vision. We still have many algorithmic methods, costs can go down further.

99. Another reason costs go down is, the lower the cost, the bigger model I can train, the bigger model I can undertake. On same computing power, under limited computing power circumstances, if my calculation efficiency is higher, I can undertake bigger models.

100. I think we will open-source, and our strongest models probably will also open-source. Because I don't see what benefits closed-source has, don't see inevitable benefits. ByteDance's models are closed-source, what benefits do they have? I don't see what benefits.

101. Even if models open-source, telling everyone everything, this threshold is also very high. Others using it, this threshold is also very high. They using it, is very hard; secondly, they using it, also must have very low costs, also very very hard, not that easy.

102. Open-source won't affect revenue. Open-source, I think has no impact on our business model.

103. I don't worry about others deploying our models, then competing with us, not worried at all. We also hope they can deploy. We try our best to help open-source community, assist everyone to deploy our models.

104. When dealing with outsiders, our attitude is: we only do AGI's main line. When dealing with outsiders, we are very willing to assist, help anyone, even our competitors, including Alibaba, Zhipu, Moonshot, to do better. Because we don't lose anything, we originally are open-source.

105. Are the open-source models we give, same as the models we deploy ourselves? Same. We won't say open-source a worse model, then when we deploy ourselves use a better model, won't be, same.

106. Data should almost equal half the model. Earlier there was a data labeling problem. We in data annotation, this relates to our capital investment. With our capital investment structure, can't support so much high-quality data annotation cost, because cost is very high.

107. US data annotation cost and China data annotation cost have no difference. China going to label data doesn't have cost advantage, especially labeling high-end data won't have cost advantage, making it hard for us to invest to label data like the US. This road in China is very hard, because labeling data is really too expensive, whether we outsource or label ourselves, both are very uncomfortable.

108. Now basically walking on two legs. Not saying we completely can't label, but because labeling data has some low cost, some high cost. We first label low cost.

109. You can also think, now our company half people are labeling data. Half core researchers, most important people, half are labeling data. We concentrate on labeling data. Solving AI this problem, at this stage relies on labeling data.

110. High-quality data annotation bottleneck, I think is time, that is needing time. Because for OpenAI, for Anthropic, they are earlier, then capital more, cards more.

111. Large model hallucination problem relatively affects user experience. Hallucination problem also has a method to solve, but this is a long proposition. Hallucination problem can be considered a solvable problem through better Post-training, a solvable, improvable problem.

112. First, we have no object to imitate. Every step is us starting from actual situation, seeking truth from facts, making decisions according to actual situation, finding how we should do. So it's a product of an era, or a reaction of reality, it's not a result of imitation.

113. We explicitly want commercialization. We ultimately still need to survive, we after all are a company, government won't give us a penny.

114. We essentially are still a company, just saying we consider earning which money, when to earn money, how much money, relying on what to earn money, we have取舍 (trade-offs). Many companies do great, because they have a pursuit beyond profit. That pursuit finally not only didn't affect its commercialization,反而 (instead) made its commercialization better.

115. For partners, actually this financing is carefully selected. First I think, interests are relatively consistent, that is most consistent with our interests, most hostile-free to us, or say most hoping we can succeed. Not everyone hopes we can succeed, because we still harm many other people's interests.

116. AI now lacks not taste and intuition, lacks continuous learning capability. AI's taste and intuition no problem. You let it write an article, its taste and intuition, I think no problem.

117. I hope only do one piece. I think AI this matter is big, doesn't need me... I only do one piece. If focusing, and I think here business interest already large enough, if AI era produces many trillion-level companies, I think we are one of them.

118. We hope to support more people, but we don't have that much energy. We have this willingness, and won't have interest conflicts, but whether we do is another matter. But at least here side has no interest conflicts, we hope win-win cooperation.