{"id":55033,"date":"2026-07-23T11:13:35","date_gmt":"2026-07-23T03:13:35","guid":{"rendered":"https:\/\/www.1ai.net\/?p=55033"},"modified":"2026-07-23T11:13:44","modified_gmt":"2026-07-23T03:13:44","slug":"%e6%a2%81%e6%96%87%e9%94%8b-4-%e5%b0%8f%e6%97%b6%e6%8a%95%e8%b5%84%e4%ba%ba%e4%bc%9a%e8%ae%ae%e5%86%85%e5%ae%b9%e6%9b%9d%e5%85%89%ef%bc%9adeepseek-%e4%b8%8d%e8%bf%bd%e6%b1%82%e6%88%90%e4%b8%ba","status":"publish","type":"post","link":"https:\/\/www.1ai.net\/en\/55033.html","title":{"rendered":"Liang Wen Sing 4 hour investor conference was exposed: DeepSeek does not seek to be the next byte or communicator. Restraint, open source and low cost are the core AGI strategy"},"content":{"rendered":"

July 23rd DeepSeek<\/a> FounderLeung Man Fung<\/a>This was preceded by a four-hour investor conference that was circulated within the industry. Elsewhere recently released a compilation of the conference. At this meeting, Liang Wenfeng revolved around the direction of DeepSeekAGI<\/a>THERE WAS AN EXCHANGE OF TOPICS SUCH AS ROUTES (GENERAL ARTIFICIAL INTELLIGENCE), OPEN SOURCE STRATEGIES, COMMERCIALIZATION PLANNING AND COMPETITION IN THE AI INDUSTRY\u3002<\/p>\n

\"4d7f58d1j00tly9l00k7d000rs00fp-1\"<\/p>\n

The following is a summary of the statements made by Liang Wenfeng during the nearly four hours of the exchange, organized by theme, totalling 118 articles, the text of which remains as intended as far as possible, with only minor editing\u3002<\/p>\n

01 Vision and restraint<\/p>\n

We started out as a company without thinking about how much I'm going to make, how I'm going to go to the capital market, how I'm going to go on the market. Dozens of people at first had never thought so, and if he did, he would not come\u3002<\/p>\n

2. We are doing this with a great deal of goodwill towards the world, and we feel it is useful for humanity, and it is something other than money. The original purpose of our departure, our vision and the vision we have maintained to the present time are not done in a way that maximizes the benefits of business\u3002<\/p>\n

3. The management of a large company depends not on your regulations, but on your vision. Vision is not a sign on the wall, vision is how you do it, how you say it, or how you actually run it\u3002<\/p>\n

4. WE ARE UNORGANIZED, DRIVEN BY A VISION, ORGANIZED BY ONE VISION. WE DIDN'T DO IT IN A WAY THAT'S \"WHAT I'M GOING TO ACHIEVE, THERE'S NO ASSESSMENT\" BUT A VISION\u3002<\/p>\n

5. This vision is not even written, it is not written, and nothing has ever been written. That vision is in our way of doing things and in our approach to the world\u3002<\/p>\n

We do not have many other advantages, we do not have any skills, we do not have money, we do not say that our personnel are better than others, and we do not. When we founded this company two years ago, we didn't have much money, we didn't have many cards, we didn't have much visibility, we had no appeal, we were a very common group\u3002<\/p>\n

7. The more you exercise restraint, the easier it may be to do, or, at least, to prove it so far, and to explain it so far. Otherwise, there is no way to explain why we can do it: we have no weapons, very low starting points and very few resources, and our people are simply a random group of ordinary people\u3002<\/p>\n

8. THE MATTER OF AI IS TOO BIG AND TOO GOOD. WE HAVE EXERCISED GREAT RESTRAINT, AND IF WE CAN DO SO, THE FINAL INTEREST WILL BE GREAT. IT'S A BIG PROFIT, SO I DON'T HAVE TO THINK ABOUT WHAT PART OF IT, HOW TO TAKE IT, BECAUSE IT'S BIG ENOUGH\u3002<\/p>\n

We suddenly had a large number of users last spring, but we did not go after me to keep them, or use them to cash them, or say I was going to rob them and cash them on. We didn't rob the users, we didn't make money, but we tried to find a way to serve the users\u3002<\/p>\n

10. We would not have thought that I was going to be the next SuperApp, and then I was going to compete with who, and I was going to be the next byte, and the next to be the next to be the arraignment. I think the AGI behind is supposed to be very big, and the AGI behind is always very big\u3002<\/p>\n

Restraint is a strategy. Just sometimes you can give up something for something else. It's the same thing if we don't open it up. It's our pressure or our concession\u3002<\/p>\n

THIS RESTRAINT, I UNDERSTAND, INCREASES OUR CHANCES OF BECOMING AGI IN THE LONG RUN. WHEN I WAS THINKING ABOUT ONE THING, I HAD NO DOUBT THAT AGI WOULD BE OF GREAT COMMERCIAL VALUE. ON THAT BASIS, THEN, MY PRIORITY IS NOT HOW MUCH I INCREASE MY SHARE, HOW I INCREASE MY SHARE, AND HOW I INCREASE THE PROBABILITY THAT I CAN DO IT\u3002<\/p>\n

We have been very restrained and unwilling to compete with any major or small Internet factory. I hope that I can give him the power, or that I can help you to do it, and that I can help you to do it\u3002<\/p>\n

14. I feel that we have taken this attitude before, and that we have not, in fact, had nothing less to gain from it, not because I was open to it, not because we were good or because I was helping others. There may be more points. This looks against instinct, but it is\u3002<\/p>\n

15. WE ARE TARGETING AGI, BUT WE HAVE BEEN COMMERCIALIZING, SO WE HAVE C-END USERS AND B-END INCOME. FROM HISTORICAL EXPERIENCE, THIS STRATEGY HAS BEEN SUCCESSFUL\u3002<\/p>\n

02 AGI ROAD MAP<\/p>\n

16. If you were able to describe a problem in a clear way, it would be more than human beings to give it a complete context and direction. But here's a definition, and there's a premise: you give it a complete context, give it a complete order\u3002<\/p>\n

17. AI CANNOT REPLACE YOUR STAFF. BUT IF AI IS CAPABLE OF CONTINUOUS LEARNING, IT'S, LIKE YOUR STAFF, TWO MONTHS AT THE COMPANY, THEN WE CAN REPLACE THE WORLD'S PEOPLE, SO WE'RE STILL SHORT ON THE NEXT STEP\u3002<\/p>\n

18. THE DEVELOPMENT OF AI CAN BE UNDERSTOOD AS A LADDER. LAST YEAR'S LADDER WAS THE THOUGHT CHAIN. BECAUSE WE'VE FOUND THAT, THROUGH A CHAIN OF THOUGHT, WE CAN GET A HIGHER LEVEL OF INTELLIGENCE\u3002<\/p>\n

This year's ladder is Agent, because we found out that even more can be done in Agent's way, it will be more capable and its upper limit of intelligence. Agent had to use Cot, and then Cot had to use the front ladder, which was the language model, so it didn't walk in vain\u3002<\/p>\n

20. After Agent, we felt that the problem to be addressed was continuous learning, that is, how to make the model sustainable, rather than saying that you were going to give it a strong training, and that it should be able to do a longer, continuous learning like a human being\u3002<\/p>\n

21. After continuous learning, perhaps we will come to a strange point. The strange thing is, when the model can continue to learn, it can do everything that humans can. It is able to develop its own version, to re-examine it, to develop its next version and to develop better artificial intelligence models\u3002<\/p>\n

22. This singularity is not a singularity but a gradual process. This process may also be a long gradient, not a mutation. But customarily, we all thought it might be a strange thing\u3002<\/p>\n

23. This is our assumption that the timetable should be: to solve the problem of learning, and then to the singularity of intelligence, the singularity of self-repeatation, and then to be smart. When it's smart, it walks into the real world and gives you housework and old age\u3002<\/p>\n

24. It would be easy to say that we should address the problem of continuous learning before addressing the singularity of a self-oversight and then being smart. Because when you get to the back, you can help develop the back\u3002<\/p>\n

WE ONLY DO AGI'S MAIN LINE. IT'S A WIDE FIELD, AND THERE'S A LOT OF THINGS THAT WE DON'T THINK IT'S ON THIS MAIN LINE, LIKE 3D, VIDEO GENERATION, AND I DON'T THINK IT'S MUCH OF A SMART LEAD, AND WE'RE NOT GONNA DO IT\u3002<\/p>\n

26. VIDEO PRODUCTION IS HOT AT THE BEGINNING, AS IF IT HAD TO BE DONE. IF YOU DON'T, YOU'RE NOT AN AI COMPANY. SO I WAS WONDERING, ACTUALLY, IF YOU THINK ABOUT IT, IT HAS NOTHING TO DO WITH SMART ROAD MAPS\u3002<\/p>\n

27. Commercially, it is a good business and commercially a good business. But it has nothing to do with intelligence. We will not do it because it is a good business, we will do it only because it is something on the smart road map\u3002<\/p>\n

28. FROM OUR JUDGEMENT, WORLD MODELS AND INTELLIGENCE ARE NOT YET THE MOST IMPORTANT AT THIS STAGE. THE MOST IMPORTANT THING IS FOR AI TO BE TRAINED AND HOW TO DEAL WITH CONTINUOUS LEARNING AFTER AI IS TRAINED. THIS IS OUR COMPANY'S JUDGEMENT, AND OF COURSE EACH COMPANY'S JUDGEMENT IS DIFFERENT\u3002<\/p>\n

29. WE NOW BELIEVE MORE IN THE NARRATIVE THAT AI CAN ACCELERATE AI ' S RESEARCH. SO IT'S NOT LINEAR, BECAUSE YOU CAN USE AI TO SPEED UP YOUR OWN RESEARCH, SO IT'S PROBABLY NONLINEAR BACK THERE\u3002<\/p>\n

30. I feel certain that I will enter, and eventually. Because for a normal person his needs are not computers, right? Because of normal people, he eats, eats, eats and eats, he doesn't need a computer. What he needs is that he needs to be smart enough to address specific human needs\u3002<\/p>\n

WHAT DO WE EXPECT AGI TO DO? IT HELPS ME WITH THE NEXT VERSION OF THE MODEL. IT HELPS ME WITH THE NEXT VERSION OF THE MODEL. AND IF YOU HAVE A BODY, WE WANT IT TO DO THE SAME THING, AND LET IT DO THE NEXT ONE, AND IT DOES THE NEXT ONE\u3002<\/p>\n

32. The core competencies of the next generation model require continuous learning to be called the next generation model. Until then, all we can do is to reduce costs and then do better and faster. But for a major breakthrough, it should be continuous learning\u3002<\/p>\n

33. Agent ' s current capacity is limited by its inability to continue learning and its inability to do so effectively. If continuous learning can be done first, then AI's ability is very strong, and it can greatly enhance the efficiency of our own research\u3002<\/p>\n

34. Continuing learning begins with universal intelligence, which can be easily used. So I say it's the result we want to see, we're more energy-efficient, we're easy. Otherwise, now you're going to manual general intelligence, which is a more tiring, bitter, data-intensive, labour-intensive thing, and it's not very expensive\u3002<\/p>\n

03 Team and talent<\/p>\n

35. THE EXPERIENCE BEFORE US HAS REVEALED TO ME THAT THE AGI VISION IS STRONG. THIS TALENT IS NOT ABOUT MY PEOPLE BEING SMARTER THAN HIM, BUT ABOUT HOW I ORGANIZE THEM, HOW I INSPIRE HIM, HOW I COOPERATE\u3002<\/p>\n

36. It is not that you bring smart people together that he can naturally work together, that he can naturally travel very passionately towards a goal, that you need a vision\u3002<\/p>\n

37. OUR GREATEST CORE INTEREST IS TO MAINTAIN TEAM STABILITY. THIS IS OUR GREATEST AND PERHAPS ONLY CORE INTEREST. AS LONG AS I CAN KEEP THE TEAM STABLE, I CAN DO IT. I CAN DO IT. I CAN DO IT\u3002<\/p>\n

38. Money was certainly not a problem, resources were not a problem, and other elements were readily available. For us, there is only one core interest, and only one that cannot give in: we must maintain team stability\u3002<\/p>\n

39. This is also a very big challenge, or, I think, the greatest risk. Of course, this risk has been lifted relatively significantly with this recent financing. Because you have more options and more money\u3002<\/p>\n

40. IN TERMS OF TEAM STABILITY, AS LONG AS SOME OF THE MOST IMPORTANT, THE OLDEST, ARE STABLE, OTHERS ARE LESS LIKELY TO LEAVE. OTHERS WOULD NOT LEAVE, EVEN IF THEY HAD LESS OPTIONS AND LESS INCOME. BECAUSE HE DIDN'T COME ALL THE WAY HERE, AND EVERYONE WANTED TO DO IT IN AN ENVIRONMENT WHERE HE COULD BE AN AGI\u3002<\/p>\n

41. The rest is a matter of time, and the rest leads us at most to six months and one year, but not to words. There must be no shortage of money, there must be no shortage of resources, but they are\u3002<\/p>\n

The gap between us and the United States lies mainly in resources, and then not so much in people. There is almost no difference, because it's the same people, probably Chinese. When the Chinese went out, there were some who stayed in the country, some who stayed abroad, and some who went abroad, he did not say that he was smart enough to go abroad, and he did not\u3002<\/p>\n

43. Talent is not a bottleneck; resources are the biggest bottleneck. Resources affect talent development in the first place, because we have less ability to calculate and fewer opportunities to experiment, so we have a difference in talent overall compared to the United States. The talent gap is also intrinsically due to the arithmetical gap\u3002<\/p>\n

44. THE SHORTAGE OF AI TALENT IS ALSO OF A PHASED NATURE AND, AS WE HAVE SEEN, HAS BEEN SIGNIFICANTLY REDUCED. BECAUSE THE AI ARE REALLY GOOD, EVERY COMPANY WILL SOON BE ABLE TO RAISE PEOPLE AND PREPARE PEOPLE\u3002<\/p>\n

There are a little too many domestic companies that are now modeling. There's probably three in the United States. There's too much in China. I'm sure you won't need so many people to do it\u3002<\/p>\n

46. THE MANAGEMENT OF OUR COMPANY IS IN FACT TWO LINES: ONE FROM TOP TO BOTTOM AND ONE FROM BOTTOM TO TOP. BOTTOM UP, IT'S EVERYBODY WHO WANTS TO DO WHAT THEY WANT TO DO, WHO DOESN'T CARE, WHO DOESN'T HAVE KPI\u3002<\/p>\n

In general, we hope that the staff will have half of the time left unscheduled and will do whatever he wishes. This is the scope of the study, which allows him to explore on his own, what he considers important, and what he explores, without any pre-existing requirements\u3002<\/p>\n

48. We do not generally work overtime. Overtime is for two reasons. The first is that research requires a more relaxed environment. If you push hard, you can't do research. Because since it's your interest to think about it, you have to think about it yourself, so it's in a more relaxed environment that you can explore\u3002<\/p>\n

49. Second, we are very focused. Our focus means that we have little to do. I don't need to work overtime. This is in keeping with the restraint in the front\u3002<\/p>\n

Our company as a whole is based on consensus, and I am not saying that I decide everything by myself, but that I seek consensus. My authority and influence within the company are based on consensus\u3002<\/p>\n

51. This decision-making mechanism is in fact a mechanism for seeking consensus, not to say that I can advance a thing. It must be a consensus that I can push and then I will push\u3002<\/p>\n

52. As the number of staff increases, we will make this adjustment. We should be doing this right away, because I'm already doing this. Without this adjustment, many things cannot go forward. There are indeed many sectors that should be structured\u3002<\/p>\n

04 Calculations and resources<\/p>\n

How many cards do we need? The more the better. There can be no doubt that, within the limits of what we can afford, the more the card is the better. So our strategy now is to buy as many cards as we can at a reasonable price\u3002<\/p>\n

54. In practice, it is very difficult to spend so much money, to buy so many cards, to buy so many cards, to buy them at very high prices, and to ensure that they are reasonable. If 20 billion dollars could be spent this year, the performance of our procurement department would be excellent\u3002<\/p>\n

The biggest gap between us and the United States is over resources. Calculator resources cannot be bought at home, on the one hand, and our capital investment is less than in the United States, on the other. At the capital investment level, we are much lower, and talent-based wages are very low. You see, they pay a billion dollars, but by the way, they still pay a small percentage of the people, and they still count\u3002<\/p>\n

All the differences we have seen, including the difference in talent, the difference in modelling capabilities and the difference in application, can be considered to be due to the difference in computing resources\u3002<\/p>\n

57. Our gap with the United States may be 12 months behind, 12 to 18 months behind, or 6 to 12 months behind. In short, it was two years behind the United States, and then it was done with only one-tenth of the United States\u3002<\/p>\n

58. The narrative is one to two years behind, but only one in twenty. So, in the future, we're going to rephrase this narrative by using a fraction of its calculus, but we're going to shrink it to six months and three months, and I think it's a goal\u3002<\/p>\n

59. Scaling, we are the letter Scaling, must be the bigger, the better, the more able to unlock more functions. It's not that we don't want Scaling, it's that we don't have the ability to do it\u3002<\/p>\n

60. Not because I think this size is enough, but because I happen to have so much resources. I'm counting on my resources, how big a model I can accept, how much I can train, and it's not enough for this model\u3002<\/p>\n

When Silicon Valley was talking about Scaling, it was for Silicon Valley; for Chinese, we were far from that, and we had no Scaling to that extent. This Scaling includes data Scaling, model size Scaling, and then training costs\u3002<\/p>\n

05 Nationally produced chips and ecology<\/p>\n

62. THE CUDA MOAT IN WEIDAY HAS BEEN QUICKLY DISMANTLED. ON THE ONE HAND, THERE'S AN AI NOW, AND THEN THERE'S AN AI, AND I'M GOING TO BUILD THIS ECOLOGY A LOT EASIER THAN BEFORE, BECAUSE AI CAN WRITE CODE\u3002<\/p>\n

63. The market for calculating cards is now larger than the game card, and there is no reason why the two must be combined. The trend is that there will be no more coupling. So special chips, whether Wigand or Yin Weidar himself, will be special chips, not those before\u3002<\/p>\n

64. THERE IS NOW A HISTORIC OPPORTUNITY FOR THE REPLACEMENT OF THE NATIONAL AI CHIP. WE BELIEVE THAT WITHIN THE NEXT YEAR WE CAN SEE ONE THING PROVEN: THE ECOLOGY OF THE NATIONAL CHIP IS COMPLETELY CLEAN. IT WAS THOUGHT TO BE PROBLEMATIC, NOT USEFUL, BUT IN THE COMING YEAR I FEEL THAT WE CAN REVERSE THAT PERCEPTION OR REVERSE IT WITH THE FACTS\u3002<\/p>\n

65. THE HARDWARE AND ECOLOGY OF THE NATIONAL AI CHIP ARE NOT PROBLEMATIC, BUT THE ONLY PROBLEM IS THE LACK OF CAPACITY. THERE IS NO OBSTACLE TO THE SUITABILITY OF THE NATIONAL PRODUCTION CARD, AND IT CANNOT BE STOPPED BY THE BRITISH. IN A NORMAL BUSINESS ENVIRONMENT, I CAN BUY A CARD IN INVERDA, SO IT'S HARD TO REPLACE IT; BUT IN A SITUATION WHERE YOU CAN'T BUY A CARD IN WEIRDA, EVERYONE HAS TO MAKE A NATIONAL CHIP\u3002<\/p>\n

66. The V3 training was based on the British card, but no longer on the British ecology. V3 uses the British card, but instead of the British ecology, we write an advanced compiler named TileLang, and then do everything else on the basis of TileLang's ecology, it's almost no longer dependent on the English ecology\u3002<\/p>\n

67. I AM MORE OPTIMISTIC ABOUT THE STRENGTH OF NATIONAL PRODUCTION. I THINK, AT THIS POINT, YIN WEIDA IS DIGGING HIS OWN GRAVE. WIGAND'S, 950'S, CAN BE COMPLETELY EQUAL TO GB200 AND GB300 IN PERFORMANCE AND PRICE\u3002<\/p>\n

68. Four valor cards are for a card at the top of the card\u3002<\/p>\n

The difference between us and the United States on the chip, and I think there will be no ecological difference, but four times plus two years on the chip\u3002<\/p>\n

70. We are working primarily with China. China fit for them, but we'll be part of the ecology, and we'll be part of it. The problem is still the lack of capacity\u3002<\/p>\n

71. I am not convinced that five years from now we will be stuck on the issue of capacity. It's definitely about capacity, and this year, next year, and the year after, I think it might still be about capacity, but five years later I think it may not be, and I'm optimistic\u3002<\/p>\n

06 Competition patterns and industry judgements<\/p>\n

72. The gap between the end-effects of models should be an integrated one. Comparing the effects of the model must be done at the same cost, and that's what makes sense. Because you compare two cars with the same price\u3002<\/p>\n

Anthropic now exceeds OpenAI, isn't it long? I don't think it's long-term. It's definitely staged. OpenAI and Google, there's still an increase in the future\u3002<\/p>\n

74. WHEN THE DIVISION OF LABOUR DOMINATES GLOBAL AI, IT IS LIKELY THAT CHINESE COMPANIES WILL HAVE THE MOST PRODUCTIVE ROLE. NORMALLY, WE HAVE THE GREATEST CAPACITY, INCLUDING CHIPS, CHIPS THAT MAY HAVE THE GREATEST CAPACITY AND POWER\u3002<\/p>\n

75. THE CHINESE WOULD MAKE THE PRODUCT THE CHEAPEST AND THEN, IN TERMS OF ITS EFFECTS, AFTER ALL, MANY OF THE GOODS PRODUCED ABROAD WERE NOT VERY DIFFERENT FROM THOSE PRODUCED IN THE UNITED STATES. THAT MAY BE TRUE FOR AI IN THE FUTURE, BUT CHINESE-MADE AI MAY BE CHEAPER. THIS MAY BE A SYSTEMIC LOW, AND IT MAY BE THE SAME AS THE CHEAPER SERVICES PROVIDED BY CHINA IN OTHER INDUSTRIES\u3002<\/p>\n

76. The final gap should be three: cost, time and user experience. Beyond that, there may be no difference\u3002<\/p>\n

77. Costs are certainly a difference, and I think that they may be the first. And the second is time, when you can do it. A few months early, a few months later, it's different\u3002<\/p>\n

78. OpenAI had felt at the outset that he could really monopolize the world, but in fact he would have many and many challenges. He'll have a challenge and he won't be so easy. The United States will face challenges, and then he may face challenges from China in the future, because the Chinese are willing to offer you this service with less\u3002<\/p>\n

Those who take much more are defeated by those who take less. You don't even really have to take much, and if you take much more, you'll be defeated by the few. It's just a vision. You take much, you lose, you face even greater difficulties\u3002<\/p>\n

80. For us, we are not the most profitable price, or the price of maximizing gains, but only one reasonable gain. That's an explanation. I believe in it. I'm not trying to justify it, because I don't have to\u3002<\/p>\n

81. I think it is possible that in many experiences we can do better than the United States. Product capacity is not necessarily worse than in the United States. The costs should also be lower than in the United States, so China would still be competitive\u3002<\/p>\n

82. The cost is well understood because they do not have to do it, so they do not develop it. They certainly don't value this. We can think of it as a very important thing, but it's not important to them\u3002<\/p>\n

Large models may be enough without two large and two small companies. The gap is only two things: time and cost. So I don't think there's a big profit. People with good cost control earn a little more, and those with bad cost control earn a little less, that is all\u3002<\/p>\n

07 Model R & D and Technology<\/p>\n

84, maybe half of our company, half of our people usually think OpenAI is better. In fact, Anthropic has a pre-emptive advantage, but this pre-advanced advantage should soon be lost, not an advantage that it can hold in the long run. All three are very good, and the three are the most efficient, and it costs the least, it spends the least, it burns the least\u3002<\/p>\n

85. MULTIMODEL LAYOUT, WE'VE BEEN DOING IT. IT IS IMPORTANT FOR THE PRODUCT; IT IS IMPORTANT FOR THE C END-USER PRODUCT. BUT FOR THE UPPER LIMIT OF INTELLIGENCE, IT'S A COMPONENT, IT'S NOT THE MAIN LINE ITSELF\u3002<\/p>\n

86. WE SHOULD HAVE THE FOLLOWING VERSIONS OF OUR V4 AND V4 THAT SUPPORT THE ORIGINAL MULTI-MODEL. BUT WE'RE MULTIMODULAR, AND FOR INTELLIGENCE, IT'S A COMPONENT, AND WE DON'T THINK OF IT AS INTELLIGENCE ITSELF\u3002<\/p>\n

87. Scaling, which can only speak the language model, is not on the upper limit. None of our current intellectual levels, or the resulting intellectual levels in the United States, see a ceiling\u3002<\/p>\n

88. THE IDEA OF MANY WITHIN US IS THAT IT IS FIRST TO BE USEFUL TO OURSELVES AND FIRST TO OURSELVES. AND THEN THIS IS THE FASTEST WAY TO GET TO AGI. WHEN WE'RE GOOD FOR OURSELVES, IT MEANS THAT OTHERS MAY BE GOOD FOR US, BUT FIRST WE HAVE TO MAKE SURE WE'RE GOOD FOR OURSELVES\u3002<\/p>\n

89. The first goal of our model is not that we use it well, but that we use it well. First, it will be useful to us. When it's useful to us, I'll be faster when I'm developing the next version of the model\u3002<\/p>\n

We call it the Touch Award. The threshold is low, anyone can touch it, but who knows what to do, and I don't know whether to look at talent or what. So there's no need for us to allocate resources. It's just that, unlike other companies, we're going to spend time talking about it, thinking about it and thinking about it as something important\u3002<\/p>\n

08 Commercialization and pricing<\/p>\n

91. OUR API PRICING IS A REASONABLE PROFIT, AND I THINK IT IS A REASONABLE PROFIT THAT WE HAVE TO BUY BACK A SHIPMENT OF EQUIPMENT ON THE MARKET AND RECOVER COSTS FOR 10 MONTHS\u3002<\/p>\n

if profits are maximized, prices should be set higher. because in this price range, the demand of the user is inelastic, i.e. half the price, or double the price, i.e. token's consumption is not very different\u3002<\/p>\n

One of our models, which at the outset we feared that there was too much need, started with a higher price set, and the team was not happy. Then I dropped the price again, down to a quarter, and everybody was happy\u3002<\/p>\n

94. To B operations should be capped or required, and in the context of the current generation of AGI, AI technology, To B needs should be limited. It will grow rapidly, but it will not be an infinity, and it will end up being conditioned by demand, not by calculation\u3002<\/p>\n

95. I NOW FEEL THAT IT SHOULD BE POSSIBLE TO DO SO. IF I COULD HAVE A FEW BILLION DOLLARS OF B-SIDE INCOME THIS YEAR, PLUS THAT WE HAVE USERS AT C-END, THEN THERE IS ALREADY A CERTAIN BUSINESS BASE. IF THIS DEMAND CAN BE INCREASED NEXT YEAR, THE COMPANY WILL BE CLOSE TO NET PROFITS, WHICH MAY ALREADY BE NET PROFITS\u3002<\/p>\n

96. SELLING API IN THE WORST-CASE SCENARIO MAY ALL SUPPORT A LISTED COMPANY. IF THERE ARE NO NEW ADVANCES IN TECHNOLOGY, OUR TECHNOLOGY IS FROZEN HERE, AND WE END UP SELLING THE API, AND I THINK WE CAN DO IT\u3002<\/p>\n

97. If we look at the current situation, I think that the most rational approach should be to do everything in common, and other Agent priorities should be lower, including finance and doctors. Coding first, because Coding Agent can do a lot, and there's a lot of vertical Agent. At this stage, we think it's still Coding Agen\u3002<\/p>\n

I think that low cost is first and foremost an outcome. Our models have indeed been moving in a lower-cost direction on the model architecture, which is relevant to our vision. We have a lot of algorithms, and the costs can go down\u3002<\/p>\n

99. Another reason for moving down is that the lower the cost, the more I can train a larger model, the more I can take on a larger model. In the same calculus, in the case of limited computing, if I am more efficient, I can take on a larger model\u3002<\/p>\n

09 Open Source Policy<\/p>\n

100. I think we're open, and then our strongest model may be open. Because I don't see the benefits of being closed, and I don't see the benefits. Byte, it's closed-sourced. What does it do? I don't see any good\u3002<\/p>\n

101. You tell people everything, even if it's a model open source. The threshold is very high for others to use. It's hard for him to use it; and secondly, it's not easy for him to use it at very low cost and hard\u3002<\/p>\n

Open sources do not affect income. Open source, I don't think it has anything to do with our business model\u3002<\/p>\n

I'm not worried about other people deploying our models and then competing with us at all. We also hope that they can be deployed. We have helped open-source communities as far as possible to enable us to deploy our models\u3002<\/p>\n

104. IN OUR DEALINGS WITH THE OUTSIDE WORLD, OUR ATTITUDE IS THAT WE ARE ONLY AGI ' S MAIN LINE. IN OUR DEALINGS WITH THE OUTSIDE WORLD, WE ARE WILLING TO ASSIST AND HELP ANYONE, EVEN OUR COMPETITORS, INCLUDING ALI, THE SPECTER, THE DARK SIDE OF THE MOON, TO DO BETTER. BECAUSE WE DIDN'T LOSE ANYTHING. WE WERE OPEN SOURCES\u3002<\/p>\n

The open source model that we gave us is the same as the model that we deployed ourselves? We don't say open source is a near model, and then we deploy ourselves with a better model\u3002<\/p>\n

10 Data and post-training<\/p>\n

The data should equal almost half of the model. There is also a question of target data. In terms of data labelling, this is about our capital investment. With the structure of our capital investment, we cannot support the cost of so many high-quality data, because it is costly\u3002<\/p>\n

The cost of labelling data in the United States is no different from the cost of labelling data in China. China does not have a cost advantage, especially on top-end data, which makes it difficult for us to invest in such data as the US. It's hard in China, because it's too expensive, and it's hard for us, for us, and for us\u3002<\/p>\n

108. The walk is now essentially on two legs. It is not that we are completely unattainable, but because some of the target data are low and some are high. We start with low-cost\u3002<\/p>\n

109. YOU CAN ALSO ASSUME THAT HALF OF OUR COMPANY IS NOW TESTING DATA. HALF OF THE CORE RESEARCHERS, THE MOST IMPORTANT, ARE IN THE TARGET DATA. WE CONCENTRATE ON THE TARGET DATA. THE SOLUTION TO AI'S PROBLEM IS, AT THIS STAGE, THE TARGET DATA\u3002<\/p>\n

110. The bottlenecks in high-quality data are, I think, time, that is, time. Because for OpenAI, abroad, and for Anthropic, they are earlier, and then they have more capital and more credit cards\u3002<\/p>\n

111. The hallucinogenic problems of large models affect the experience of users more. There is a solution to the problem of hallucinations, but it is a long-term proposition. The problem of illusion can be considered to be a problem that can be solved through better Post-training and that can be solved and improved\u3002<\/p>\n

Organization and corporate positioning<\/p>\n

112. First, we do not have a copycat. Each step is for us to be realistic, to be realistic, to make decisions and to find out what we should do. So it's an age, or a reflection of reality, and it's not an imitation\u3002<\/p>\n

113. We clearly want to commercialize. Ultimately, we have to live, we are a company and the Government will not give me a penny\u3002<\/p>\n

114. We are in essence a company. We simply say we are considering what to make, when to make, how to make and what to make. Many companies do great things because they have a profit other than profit. Instead of affecting its commercialization, that pursuit will make it more commercial\u3002<\/p>\n

115. For partners, our financing is in fact carefully chosen. First of all, I think that there is a greater convergence of interests, the best of our interests, the least hostile to us, or the best hope that we can succeed. Not all of us want us to succeed, because we are at the expense of many others\u3002<\/p>\n

116. AI DOES NOT NOW HAVE A LACK OF TASTE OR INTUITION, BUT A LACK OF CONTINUOUS LEARNING. THERE'S NO PROBLEM WITH THE TASTE AND INTUITION OF AI. YOU LET IT WRITE, IT'S TASTE AND INTUITION, AND I DON'T THINK IT'S A PROBLEM\u3002<\/p>\n

117. WE WOULD LIKE TO MAKE ONLY ONE PIECE. I THINK THIS IS A BIG THING FOR AI, AND I DON'T NEED IT... I'M JUST DOING ONE. IF THE FOCUS, AND I THINK THE BUSINESS INTERESTS HERE ARE BIG ENOUGH, IF IT WERE THE AI ERA, THERE WOULD BE HUNDREDS OF BILLIONS OF COMPANIES, AND I THINK WE'RE ONE OF THEM\u3002<\/p>\n

118. We hope to be able to empower more people, but we do not have that much energy. We have that will and there will be no conflict of interest, but it is another matter whether we do it or not. But, at least here, there is no conflict of interest, and we want cooperation to win\u3002<\/p>","protected":false},"excerpt":{"rendered":"

On July 23rd, the news that recently, the content of a four-hour investor ' s conference about DeepSeek founder Liang Wenfeng was circulated in the industry. Elsewhere recently released a compilation of the conference. During the meeting, Liang Wenbing exchanged views on the direction of DeepSeek, the AGI (General Artificial Intelligence) route, open source strategy, commercialization planning and competition in the AI industry. The following is a summary of the statements made by Liang Wenfeng during the nearly four hours of the exchange, organized by theme, totalling 118 articles, the text of which remains as intended as far as possible, with only minor editing. Vision and restraint<\/p>","protected":false},"author":1,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[148,146],"tags":[151,3606,5762],"collection":[],"class_list":["post-55033","post","type-post","status-publish","format-standard","hentry","category-headline","category-news","tag-agi","tag-deepseek","tag-5762"],"acf":[],"_links":{"self":[{"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/posts\/55033","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/comments?post=55033"}],"version-history":[{"count":0,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/posts\/55033\/revisions"}],"wp:attachment":[{"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/media?parent=55033"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/categories?post=55033"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/tags?post=55033"},{"taxonomy":"collection","embeddable":true,"href":"https:\/\/www.1ai.net\/en\/wp-json\/wp\/v2\/collection?post=55033"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}