142. 雨森的创投观察第2集：Harness、下一个字节、2026大机会和Stanley Druckenmiller

Channel: Zhang Xiaojun Podcast Source: https://www.youtube.com/watch?v=XEhf371Aeso

Transcript

[00:00] Hello, everyone. I'm Xiaojun. Today is the second episode of our series, "Dai Yusen's Creative Observation". In the first episode of Dai Yu Sen's "Innovative Observation", which is episode 124 of our program, Dai Yu Sen, a partner of Zengge Fund Management, said the key word of the year is "The Year of R". He is more cautious than ever, thinking that the year of R will be a year of reality and return to the original state, and he emptied all the stocks in the second-tier market before the New Year. Now, in the past half year, many people say that some of his opinions are a bit flattered. Yu Sen has gone through a brief inner struggle and decided to continue to record our series. He will continue to share his investment thoughts. Then the next is my interview with Yu Sen. I look forward to our joint progress in 2026 with AI. I was affected by my blog. I don't dare to add to it as aggressively as my friends. This year, I have a lot of friends who have made a lot of money. For me, I must have added some space, but it's not that aggressive. Wait a minute. Are you talking about the impact of your blog or the impact of your previous point of view? Does this word have an impact on you or does your point of view itself have an impact? I think when you express your point of view more strongly, you will be restricted by what you say. And if this sentence is expressed publicly, there will be more restrictions. This is also the reason why I was wondering if I should record the second episode of BOK. But then, Mr. Wa from GEEK convinced me. He said, if you think your ideas are being ignored, you don't want to say it anymore. In fact, you are really bound by your ideas. But in fact, you should be a person who continues to learn and evolve. Hello, I'm Dai Yusen a partner of Zengge Fund's management partner. - - - When you are constantly being hit in the face, it means that the industry is changing very fast. If you make early investment, the industry will change very fast, which means it has a lot of opportunities. There are also many people who are hit in the face by AI. Actually, when we recorded it, it was around November or December. At that time, the most popular company was OpenAI. Because the DAO of OpenAI was about $800 million, and it was about $1 billion. At that time, I remember that OpenAI announced that it would cooperate with a company, and the share price would rise by tens of points. It announced who would buy who's profit, such as Oracle. The share price immediately jumped by tens of points. Then in December last year, GMI Live 3 was released. At that time, everyone said that Google had become very powerful. Everyone said that Google had TPU, Google had the most computing power, Google also had a lot of talent, and Google had the original Dobermuttai model. So at that time, Google replaced OpenAI with the model network. Then in January this year, we found that Cloud is very capable of coding. Of course, the following open cloud, right? Then Cloud Code, everyone went to learn this agent coding. So in about March this year, Zopic's income was very fast. It's also more than OpenAI in the second-hand market. So in March, the trend was that Zopic was 100 million. But now in May, we found that Codex is also very good. The number of new users has exceeded the number of Cloud Codes. Then, we found that the Opus 4.7 was a bit down. Of course, it was caused by a series of system problems. But we found that GPT-5.5 was also very good. So now we feel that OpenAI has a kind of rollback or a trend of "you are chasing me". So I feel that in the field of AI, even if which model is better, this problem is in our past six to eight months. Ask when it's different, there may be different answers. So I think face-sacking is a normal thing. And if someone is afraid of being hit in the face, they will never say it, they will never judge it, they will never be hit in the face. So I think it's also a step-by-step review and thinking. It will be helpful for myself and our communication. I went back and looked at it. Last time you said very accurately that it will be brought back in the second half of this year. Are you still making this judgment now? Because at least not in the first half of this year. He's even crazier. Yes, I think this is also what we call "strong opinion, weakly held" in the second-tier market. You need to have a strong point of view, but you don't want to be kidnapped by your own point of view. To be honest, after I recorded that podcast, I also thought that people are easily kidnapped by their own words. For example, when you look at the void, you can easily look at the void and do the void. I'm not talking about the void, but you might be empty, right? Or you want to buy. But why is it called weakly held? It's because if you have a strong opinion, you'd better know how your opinion came about. I have this kind of opinion because of one, two, three different reasons. So when these reasons change, I think a really smart person should actually adjust his or her opinion. So I think it's harder to change what you say. But this is something I'm going to try to overcome. So we did see that I had a more important judgment at the time. I think OpenAI, everyone expects its income to be at a risk of falling out. Because at that time, everyone thought that OpenAI's core income was divided into three pieces. One is the income from subscription, one is the income from advertising, one is the income from e-commerce, and the other is the income from corporate services. I think the first two pieces were really seen after six months. In fact, it's not as fast as everyone thought. Because of the cost of its Pro users, including its advertising e-commerce, the progress is actually lower than expected. But indeed, the income from ad-topic coding in the corporate service has increased significantly. It was certainly not seen at the end of last year's November and early December. This is indeed a improvement from quantity change to quality change after the release of Cloud 4.5 and 4.6, which has led to many people who could not complete their tasks with ad-topic coding before. Now they can do it very well. So when this happens, naturally, your judgment of the future trajectory will also change. The world is basically a BS world. The probability of what happened before has an impact on the later time judgment. So now I'm going to push the timeline of this overall return back. At the same time, I will definitely make some adjustments in my position and in my own investment. This is the charm of AI. Because it often makes a big change. We feel very curious about new things every day. You said the second-tier market was empty at the end of last year. How has it changed today? Of course, we don't make any investment suggestions, just as some sharing. First of all, my idol in the second-tier market is Stanley Drunkmeat. I think this is a big change. It has a profound impact. At the same time, we also saw in the stock market that Meta, Google, and more and more companies are all eating crabs. Everyone is using Agile Decoding together to try to innovate with many tokens. Many of my friends also have a strong sense of it. Some people burn tens of thousands of dollars a month. So this will definitely continue to drive token consumption. So I also added some things that everyone can think of, whether it's storage, light, CPU, and these so-called "stab-head" parts. Because now the whole big part of the turtle is in the multi-tool hardware, especially the bottleneck part of multi-tool hardware. But in this respect, I was really influenced by my bloggers. I don't dare to add as much as my friends. This year, I have a lot of friends who make a lot of money. For me, I must have added some more layers, but it's not that extreme. Wait a minute, are you talking about the impact you received from the blog or the impact you received from the previous point of view? Does this word say the impact it has on you or does your point of view itself have an impact on you? I think the more strongly you express a point of view, the more you will be restricted by what you say. And if these words are expressed publicly, there will definitely be more restrictions. This is also the reason why I was wondering if I should record the second episode of Broker. But then, Mr. Wa from GEEK convinced me. He said, if you think your ideas are being ignored, you don't want to say it anymore. In fact, you are really bound by your ideas. But in fact, you should be a person who continues to learn and evolve. In fact, it is also a non-standard learning method, right? You will continue to express your thoughts and get feedback from them. This is actually very much like the model of enhanced learning. You need to get a feedback signal, a high-quality feedback signal, and sometimes a negative feedback signal. I think it's very important. This is also where I think the market is very conscious and where the bloggers are very interested. Because in daily life, there may not be so many people who criticize me directly, but the market will not say good things to you. If the market says you are wrong, you are wrong. Of course, there are many listeners who say, "Look, is this being slapped?" So I think this is a good practice for thinking and thinking. - One is the subscription fee for ordinary users representing OpenAI. I think it's hard to raise the price. It's hard to raise the price when the user pays 20 US dollars a month. It's hard to raise the price of hundreds of US dollars a month for ordinary users. Secondly, there were already 50 million paid users. I think the paid users have already selected the people who are willing to pay for the daily chatbot of AI. It's hard to raise the price. The DAO of OpenAI, the number of paid users, and the number of up-puts of each paid user, I think the number of paid users will be slower. I think this is what we've seen in the past six months. Its growth has been limited to a certain level. The second is that OpenAI was working hard to attract advertisers and e-commerce companies. They recruited a lot of people and launched a lot of new product models. I think it's not easy to put advertisers in a very innovative user product and make an e-commerce. Even a company as powerful as ZJ He spent a long time on the Internet and network exploration of TikTok and TikTok. Facebook and Google also spent six or eight years after the establishment of the company to become a super-sounding commercial. So I think it was too optimistic to insert ads in chatGPT. In fact, it's not that simple. I think this is also verified. So I think these two, based on the chat LGBT as the leader of AI, as the most concerned transformation channel, I think this judgment is not a big problem. But obviously, for Android or coding, it was completely unexpected that such a big change would happen in such a short time. This is the part that is wrong. But I can also open up and talk about why I didn't see this coming at that time. Because I think if it is intelligence, it is actually a sudden change of course. This is very different from the Internet. The Internet, for example, in the last century, had e-commerce, advertising, door-to-door, and real-time communication. Many business models have already emerged. But the Internet needs to connect more people, have faster network speed, and generate higher edge efficiency. Its business model has appeared for a long time. But it's a continuous process of improvement. But for AI, it's basically governance. We can think of a human being as saying that his or her governance must break through certain barriers to be valuable. Although AI's governance has been improved from cat governance to, say, IQ 80, IQ 90, it's a great achievement. But if you want to hire an employee, you must want his or her IQ to be at least 100. um to really complete high-value coding applications. There is actually no clear expectation behind it. Otherwise, Cloud 4.5 would not be called 4.5, but 5. Because they themselves can also see this version number. In fact, there is not much expectation. Then we saw that there was no change in the way of looking at it in the retrospective. Of course, there is no new model structure, and there is no new way of building. But it is in the process of 4 to 4.5, then 4.5 was further enlarged by 4.6, which brought the fundamental change of the coding experience. So, of course, I think from the back, it's common to think that data is still a big part of it, right? High-quality user coding data makes the model better at programming. At the same time, there are many parts of agentic I/O that improve the model's ability to do this. So the original agentic loop that couldn't run can run. At this point, AI can start to complete longer-term, higher-value programming work. I don't know when this will happen, but it's obvious when you look at the node. When do you think you started to truly understand Anthropik? First of all, we can't say that we understand, right? Because SELB is a company that is relatively closed to China. So many of our information is actually second-hand, maybe not first-hand. So it's easier to understand with the example of Shunyu, right? So I can't understand it like this. But this year in March, after I did further exchange with Gui Gu, I really think that everyone often mentions the ability and form of organization. Or to be more specific, the AI that we are exploring, for example, before the emergence of Child LGBT, we didn't know what AI could do. It's a super-personal, super-personal, super-personal, super-personal, super-personal, super-personal, super-personal, super-personal, super-personal, super-personal, super-personal, super-personal, super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super-personal super Because I remember when OpenAI first started, there were three different directions: to make robots, to make world models, to play games, to play I/O. At that time, the language model was only done by a small team. But this way of supporting free exploration is very valuable in the exploration period. But now AI has entered a process of mainline and intermediate refinement. I understand that Anthopic chose to do coding. It's not that he came up and said, "I want to do coding." It's because he gradually found that when his training data has more and more coding, his coding performance is getting better and better. And he actually chose to do enterprise, do productivity, and do the same route as OpenAI does large-scale products. But when he has a more clear direction of coding, I think Anthopic is actually a company's organizational ability from top to bottom in this process. Everyone they go in for an interview has a value-added interview. This is actually a kind of common direction of recognition. So I think the point of the point is also very important. He is not so strong. There is a personal heroism. We have seen a lot of celebrity researchers in OpenAI. They may each have their own direction to do. But if you are a star, I am also a star. In fact, each may have their own ideas. The resource allocation seems to be the same as Sam. He used to be the YC president. He's kind of like an angel investor. In Weibo, there are actually a lot of people who will constantly encourage various small projects. So we see that OpenAI often has a lot of projects made. Sarah. Sarah. But from the perspective of AmpliPay, it's not just about hiring people, the company's direction is consistent. And it seems that Dario will have a thinking memo with the company every two weeks. It's a very similar organization. It's like some Chinese companies with strong combat power. And then it actually fell very quickly. We saw that in March, Cloud Code should have been in the top ten, and it fell to dozens of different functions. Sometimes when I go into the main cloud, it upgrades again. It's actually an organization that controls sand and bury. Many products in OpenAI are a bit like controlling sand and not burying. It feels like no one is protecting it afterwards. I think this is actually a development-type organization that is suitable for the exploration period. This high-speed racing machine is actually suitable for the obvious bet at the beginning, and then from top to bottom, the values are more consistent, and the direction is also more accurate. So I think this is indeed an advantage of the organization in this regard. But indeed, I think you and Guangbei also talked about this point, that is, OpenAI still has a lot of innovative seeds that are growing, right? Suppose that in the next few years it will not change, then the advantage of this kind of specialized will be very large. But if the next model is born in a year or two, then this time it may return to a research stage. This is actually the second R in my R last year, which is Research. Because when I saw that there were many new new labs born, everyone was actually betting on one thing, that is, the next generation may need this kind of new lab again, which is more scattered, more free, and at the same time has a lot of funding to explore the organization to be born. But indeed, under the current generation, a big change in the topic of coding was directly realized. I think this is a change that many people have not seen. I also saw another observation on Weibo. Many people thought that coding was a vertical field before last year. Everyone is linking coding and medical finance. Although I remember that Guangming and I have actually talked about a lot of things, coding is a more general ability. But even if I talk to some research lab senior researchers, if he doesn't do coding himself, he often thinks that coding is a vertical direction. But now I found that coding is not a vertical field. Coding is a horizontal field. So it can actually strengthen the work of the office, strengthen the medical field, and strengthen the speed of research. So this actually made coding very different. But at the same time, it's also very obvious when you do it. But before that, it wasn't that obvious. Then, For Adobe, we were worried about whether it was a company that sold APIs. Especially before Cloud Code became so popular. Cloud Code was also so popular. Yes, it was. I remember last year, he was worried that if you sold APIs, and Codex sold APIs cheaper, would you not be good enough? At that time, I was worried that better models would come out or cheaper models would come out. There was no threshold for selling APIs in the market. But in fact, you will find that Cloud Code is very important. If there is no Cloud Code, and Ensobic only sells its API, I think it will be a very different company. Cloud Code brings a large number of users to use data. Then, after all aspects of understanding, this kind of data flow really allows it to train better coding models to achieve this kind of data flying wheel improvement. And in fact, you see, Codex has also grown very strong recently, right? So you will find that Cloud Code, Codex, are actually all harness, right? That doesn't mean that you have a powerful model, you can and sell it to the API, you have a strong advantage. Users still need to use a product. Users don't have to use the API. So in this, it has both a good model and a long-term horizon mission, such as Harness, which is Cloud Code. I think this is also the right thing for the company. But Harness is made by a model company. Yes. Is it possible that the future Harness will be made by a model company? I don't think so. Because Cloud Code can actually be used for other models. So from this point of view, it's not as strong as the model. And then obviously, we saw OpenCloud in the beginning of the year. OpenCloud is a one-man organization. It has once become one of the very popular harnesses, right? You will find that at least it is not that only model companies can make good harnesses. Of course, if you are a model company, you know your own model development direction, you know its parameters and all kinds of details. So you can say that your harness will do well. But I think OpenCloud is another example. It has a lot of innovation. It's not something that only the model companies can think of. In fact, everyone can think of it. So I think that among them, there are only a few good harnesses that are made by the model companies. OpenCloud or Manas are actually all harnesses. It's just that they used to be shell-shaped. Last year, everyone said that this is shell-shaped, it's not important. Now everyone says, "Oh, it's called harness." It's very important. I think everyone's judgment on this matter, their judgment on its importance, is actually changing constantly. Okay, we'll talk about this later. We were talking about Anthropic and OpenAI Codex. Do you think Anthropic's advantage is stable? Do you think Anthropic is overestimated today? Is OpenAI underestimated? This is interesting. I think you should look at it from a short-term, medium-term, long-term perspective. For example, if it's very short-term, it may be discounted in three months. Suppose they go public today, everyone thinks it's at least $20,000 a dollar. Now Anthopic's new stock price is $9 trillion. OpenAI is about $9 trillion, right? But they're actually expected to go public in the second half of the year or early next year. In the current market, I estimate that the market will double, and it may go up to $30,000 billion. From this perspective, it may be underestimated. Because the ARR expected at the end of the year is $1 billion. Now it is 10 times that. $10 billion. It can't be said to be particularly expensive. If you look at it from a very short-term market sentiment, maybe Anzobic and OpenAI are actually underestimated. But if you look at it from a year or two, I still think they may have been overestimated. Because the problem of return is not really solved. It's just turned into another form. We can talk about this later. Simply put, you may find that their valuation will drop significantly in two years. But if you look at it from the perspective of 10 years, I think the winner in it, or maybe both are winners, maybe both will exist for a long time, it may be a company worth 50,000 or 100,000 billion. So I think it depends on how to think about this problem in the short term, medium term, and long term. Another view is that now, in fact, the first few model companies have not opened a real big gap. So you will find who will be the most powerful model at the end. Because each model is a relatively large improvement. The post-production has the advantage of post-production. But if you use this to over-push, there may be some problems. How do you see the competition between Cloud Code and Codex? At present, many companies will have brand locks. For example, I use Cloud Code very well. Then do I have the motivation to try Codex? Maybe not. So this brand will have a very important advantage. The brand's new positioning is first released. And user connectivity. User connectivity is based on habits. For example, I have a lot of skills in my cloud code. I have a lot of my configuration in Cloud AMD. At this time, if I use Codex, I may have to redo it. So you see Codex is actually a bit of a price war now. Their cost is basically 50% cheaper than Cloudcore's same level. It's now in a long time horizon. From a reasoning perspective, some of my friends who use both of them also react that Codex is actually pretty good. So if someone comes up now and uses Codex, he might also think it's pretty good. So I think this is another channel fight. Because now when the two are doing a lot of tasks, there is no such fundamental difference. So at this time, brand channels and prices become very important. But if you ask me this question two months ago, then at that time, Codex was going to fall behind Cloud Code a lot. That was the appearance of GPT-5.5, which actually changed a lot of people's understanding of Codex. So this is a good horse to match, right? Of course, now their model and Harness often have binding relationships. Binding through its coding plan. You have to see whether it is because it provides a very discounted model or Harness itself does a good job. But overall, I think now, in mid-May, everyone's gap is not as big as two months ago. You also mentioned earlier that the return issue has not been really resolved. It's like this. We invested in these capex, and it eventually turned into a return, right? This is the essence of the return issue. The increase in income of Anthropic has made many people think that the return issue has been solved. But my view is that the income of Anthropic is not the final return. It is actually the investment of its customers. We bought such a token. We are not playing around. We want to get the result at the end. So this is to further push the return issue to the customers of Anthropic. Let's think about it. The token of Anthropic is actually the investment of its customers. The tokens that its customers bought and burned are actually the software that they created. This is its output. This output is not enough for software only. This software has to sell money or be able to sell books. It can bring about a profit improvement. This is the result. We can think of it as a three-step chain called input, output, and result. Now everyone is desperately investing because everyone thinks that this investment will eventually have results. I think the result will be an increase in profits. What about the increase in profits? Or the increase in income? Or the decrease in cost? Or the decrease in cost? Or the effect? Or two together? We have to think about it. First, if we think that the company of the cycle is established, we have to see its end users can finally have new profits as a result. So if you look at it from the perspective of real-life, the income of increasing prices should actually be born through new products, new services, and expanding new business areas. Then we can think of a problem that mobile Internet has been in a stagnation in the past few years. Everyone thinks that this product seems to have no new product form. Is this because there are no programmers to do the functions? Or because of the lack of good product managers to discover new needs and make new products? I think a lot of times it's not about the lack of programmers to write codes, but we don't know what to do. Then we think about another thought experiment, which is for a slightly larger company, suddenly there are ten times more engineers, will their income increase? I think many times it won't. In fact, we can often see that many people say that after using agent-led coding, my efficiency has increased by ten times. But with so many ten times, we look at the company's income, does the company have a lot of new products, sell to many new users, bring a lot of new income? We haven't observed a big improvement in the final result. Will it be because the time is not yet right? Yes. It's your theory of boiling water. Yes, I think this is the appearance of new products and the discovery of new functions. It's a gradual process. It's not like your programmer suddenly doubled, and your new functions came out a lot, right? So it's a gradual release, a gradual discovery of new products and new needs. But the tokens you burned are burned now. So I think it will have a time error. When we built the GPU Group and the Decentralized Center, we thought we would earn the money back in six years or even longer. But today we are out of tokens. Can we earn it back in two years? I think the patience will be slower than we thought. And now we are in China and China, we all see that the original shares are not so well-off. I think there is an important reason for not being well-off. The original programmers are expensive. I'm a very expensive programmer, so I still let it do my core business. But now you suddenly have a lot of cheap programming capabilities, but you don't know what new things to do. So we now find that the companies that have rules and regulations are actually rolling up against each other and entering each other's territory. For example, Lafbo was originally a website, and now Lafbo can do PPT. Then Gamma was originally a PPT, and now Gamma can do websites. Because you find that when you can do web coding, when you can use edge-in-edge coding, But you don't know what to do with the new one. The first reaction is to do it for others. So now, it's getting faster and faster to copy an already-existent ability. But if you want to do something that everyone didn't expect, it's still very difficult. Because it's not a programming problem. It's a innovation problem. It's a product problem. So I think the effect in the discounted efficiency value, or this effect is to increase income. Increasing income is a gradual process. It's not that you can increase your income immediately by increasing a lot of programming capabilities. So now everyone finds that the discount is the core. Right? First, if there is a large-scale financial resource, the income of these programmers is also the income of other companies. This actually has a huge impact on the entire economy. Of course, in Gui Gu, I feel that everyone is very worried about being laid off. Because everyone is anxious. This anxiety is real. Why? Because everyone sees that the main part of their work has been made by AI. But I think in this year, you will find that A programmer may have 80% or 90% of the work in the programming, but there is also 1% or 20% of the work. Many are communicated, coordinated, or even managed by the trust between the human language. Although these parts may only take part of his work, he can't be replaced directly by agent. So my opinion is that it is easy to replace his programming power, but many times, for example, a person in an organization, his relationship, his people trust each other, communicate, It can't be replaced So you can't just get rid of this person In fact, you will find that in the stock market Of course, some of them are already more popular This part may be black or not It may be able to be removed Basically, the big companies in the stock market have 15% of the people who are able to be removed at any time It's easy to get rid of But I don't think it's that easy to get rid of it This is a bit like when you drive yourself When you were young, you could already solve 80% or 90% of the high-speed road driving But can we just remove people at this time? Actually, no, because there are still some corner cases. Similarly, I think in a job, even in software engineers, there are still some corner cases. This makes this person not so easy to be fired. The third is that the fire is after all a job you can only do once. You can only fire one person at a time, right? So he can save it and invest it, but he can't continue to increase the cost of spending more. So I feel that the return problem is not really eliminated. It is in the form of the income of Azopi. Because as long as Azopi has income, its top-level hardware companies, memory companies, and blockchains can have a lot of orders. But if the customers of Azopi don't make money in the end, will they buy so many Azopi tokens? or coding tokens. I think this is very worth observing and asking questions. But now, people who have used AI coding are far more than those who found their own token that has no value. It's a bit like a tide. There are countless people who want to eat crabs. There are actually many people who think crabs are not delicious and they leave. But now, there are far more people who have come in than those who have left. So, Ensobic, including OpenAI, its AI has grown very fast. We also saw the Chinese coding model. They actually also sell very fast. As long as there is a profit, it can be sold. But I think you have to stop the leak in the end. People who come to eat crabs and use AI coding are the ones who can make money in the end. And the cycle of making money can't be too long. I can't make money in two years with a token today. I think most companies can't stand it. In fact, we may also look at it ourselves. Because we also make many AI coding attempts. For example, in March, In February, I think we all raised a lot of crayfish, right? We also burned a lot of tokens. But it's very likely that you will find that the number of tokens you burned in March is much higher than now. Because many AI tokens may not bring about the value of the essence. I think this is actually what we have observed. Many companies have also encountered this. But now I really think that small companies have a lot of opportunities for individuals. Because the original individual innovation and the innovation of small companies are often stuck in the programmers. So they have ideas that no one else has. So at this time, coding has made a big change. But I think many big companies should not have no one to do it, but they don't know what to do. And the organizational efficiency is very low. But I think what you just said is about programmers. AI coding is a substitute for programmers, or an increase in the number of programmers. But since we say coding is a lie, it should have an impact on more industries. Yes, I think programmers should take it as an analogy. The second step is the office worker. For example, you can write more reports. You originally wanted to write a report, but now you have the AI to write it for you. Can you get off the job? I think there's something interesting about this. For example, when we do investment institutions, everyone has a lot more reports to read. But when you invest in a decision, you still have to be responsible for the decision. For example, you used to look at your loan from one company, now you can look at ten companies. But can you take on the responsibility of the investor? Can you make ten times more money? I don't think so. In this, I actually thought of a framework called that an organization can carry a pot is limited. As long as your AI can't help you finish a job, it can write more than 100 reports. But people still have to be responsible for the result of the report by its driving behavior. For example, if I invest in this company, I still have to be responsible for my investment decision. AI can't be responsible for it. I can't fire it, I can't turn it off. When a person can take on the responsibility, it's a certain time. Because this determines his salary or determines the loss he can bear. AI can't give you a single discount. So when a host has relatively limited responsibility, you just write better reports and better models. I think this may not allow a person to spend more than ten times as much money in a short time. Unless we can get a report now, we don't need to read it. It must be correct. Then we can use it to guide. But at present, we also see that there must be a certain number of hallucinations or a lot of work. It still needs people to check. At this time, it will come back to a calmness of people. AI wrote a lot of reports, but people still need to check. So the time to check may be less Overall, the improvement of production does not mean the final result So back to the framework of the three steps, investment, output, and result I think the investment is going on a lot more than the output But the output did not bring any results I think it needs time to verify Because AI is still changing rapidly, let's look at the problem of return. Will it be a little early? I think it will be earlier. But I think why we should look at it now, and at the end of last year, because this number is very large. If this number is hundreds of billions of dollars, then in fact, everyone may think it is worth it. But now the profit of AI on hardware this year is 7 trillion dollars. You can also see that recently, the workers of Samsung Airways have been receiving millions of dollars in bonuses every year. Because Samsung Airways' profits have been nearly unimaginable. And then the traditional CSP, which are a few big cloud companies, their profits are already lower than the profits of pure companies. The hardware section has already made a lot of investment. At the same time, as expected, Antelope's AR is worth $1 billion by the end of the year. It means that she sells $1 billion of tokens every month. How much money did the people who spent 10 billion to buy tokens earn? I think this question needs to be answered faster Otherwise, the number is too big to ignore So I think that the current hyperscalers like Amazon have already been raising money to build their data center After it is raised with capital leverage, the demand for return The time requirement for answering this question is actually getting shorter Let's talk about the whole of 2026. Because when we talked at the end of last year, I remember you said it was the Year of Return. But at that time, there were many people saying that 2026 would be an outbreak of applications. And then some people predicted that this year would be a big battle of the super app. But over the years, it's completely different from what people thought. So what do you think of the whole trend this year? Can you summarize and explain? In fact, we have always been very clear about the focus of the attention of the AI revolution. We have always been optimistic about the huge improvement in AI productivity. We have always believed that only agents are the main forms of AI productivity improvement. Because agents don't need human attention to liberate human attention. So I also took a look at it. Because I have recorded some different podcasts from 2023 to now. Then I looked at the titles I mentioned before For example, in 2023, our title was called "The Big Model is just the beginning, and the future of mankind is in cooperation with multidimensional technologies." Then in 2024, because many people questioned the commercial value of the big model, I said that the improvement of the model ability can unlock more AI values. Especially when we talk about programming capabilities, machine learning capabilities, thinking capabilities, these three capabilities can unlock the original idea of ​​AGENT in 2025. So based on this, we have also invested in Manage, GenSpark, and Agen, which is the first Harness company. And we also invested in Kimi, which is the most concerned about agentic coding technology model company in China. They also made a huge transformation in 2005, and now it has become what we think is the best open source coding model in the world. And then in 2025, I think the ability of models will be improved. After that, we will no longer need so much attention. Attention is all you need. Because my summary is that all the work that humans did in the past, such as writing codes, doing a PPT, doing anything, all need human attention. So as long as AI can't release people's attention, it can't bring about fundamental improvement. But now we see that before many people go to bed, agents set up their tasks. 需要是醒来之后发现A- 把他试个做了 所以这个其实是真正的 因为模型能力的提升 带来了真正A-的到来 所以我们能实现 不需要注意力 从而能够实现一个人干很多事儿 So at this time, I also recorded a video last year, which said that the appearance of agents means that we will all be the boss of AI. I think a lot of people now have a lot of interesting things. It seems that everyone is in this year. It's a stage where agents are in the lead, right? So everyone thinks there's something they can't do. Let them do this and that later. Later, I have a friend who said a sentence. I think it makes sense. He said that many people are enjoying the pleasure of being the boss. I said, "I've been a boss for many years, I know how to perform well, so I'm not always the boss." So if we put it back, we are not surprised. We have always been optimistic about the improvement of AI productivity. We have always been optimistic about the form of AI productivity as an agent. We have been investing in such companies. But before, everyone said that this is a shell, this has no barriers. Now everyone says that Harness is still very important. Harness Engineer. So if I go back to the R6.0 AI world, I think the improvement of model capabilities brings the true landing of agent. The form of agent landing, it actually lands on a product such as an application or Harness. Speaking of which, because we have talked about Harness many times, can you tell us about the programming process of this model that you understand? Of course, this word is also a show, right? I think if we used a product before, whether it is an AI product or not, it's a bit like driving myself. Now I think we are building a bit like we look at the F1 race room, the driver is driving, it's just a model of the driver. It's a good one, but we need to get a bunch of people to take care of it, change its fetus, and let it control it in the tunnel. So it's a bit like the word "harness", right? Because it comes from the word "mage". It's a kind of thing that you can make something very powerful, and you can let it operate within your set rules. I think there are several parts to this. The most basic part is the model, right? This is what we've been discussing the most. Is there a good model? But the model needs to have context first. Because there are many things that the model doesn't know. You need to give it this context. This context may be a real context, may be a group of information. This is very important. Then outside the context, you have to establish a tool for it. What kind of agenda can it run in the loop of agent queue? From this, they can do more things in a good harness. We actually saw why OpenCRAWL came out at that time. Everyone was very surprised. It's because OpenCRAWL gives the model a lot of full lines. It unlocked a lot of scenes that we didn't expect. For example, operating on Mac files. And then it has a heartbeat heartbeat MD. It checks the model every 30 minutes to see if it has done anything. Many of these mechanisms are very simple to speak of. But it actually makes the model be able to do better on long-term tasks. So this is the harness. There is a layer of other things that many people are doing now, which is the Thunderbox, which is its runtime. In this, we found that models need more complex applications. I remember that when the Maynard came out last year, they made the Thunderbox. They were working with E2B at that time. The model can use software online in the Thunderbox. Of course, now you will find that I want to have a long-lasting Thunderbox. I want to match an AI with a more long-lasting computer. So now we see that DigitalOcean has also been doing this training. They also have a lot of growth. At the same time, we see that AI has a computer for a long time. They have been running more tasks in this computer. This has become very important. So this is actually all structured outside the model city. Now it's just that the Harness is doing it on the model site. But obviously there are OpenCRA, Hermes, Manus, James Park, etc. So Context needs users to pass on this product to him. This model will not fall from the sky by itself. And there are many companies in the round-time industry doing this. So I think we have exposed more and more layers outside the model. Some of them are made by the model factory, some are made by third-party entrepreneurs. And the more value is in the few shells exposed outside the model. So you think the "but" has become more valuable? I think the "but" has become more valuable. Do you still agree with the concept of "models are products"? Yang Zhiling is very firm in his opinion. I think the core of what he said is that the core ability of a product is provided by a model. But obviously, what we are dealing with now is not a model or an API of a model. Now, if you have an API, you can of course connect it. But what we are dealing with is always a product. Even if the first AI wave of the wave is the "shut up" of LGBT people, But ChediGPT is also a product. It actually puts the InstructGPT model as a post-training chatbot. If it's not a harness, it only allows users to interact with the model API themselves, there will definitely be no such big change. For example, you can say that ChediGPT is a model-based product. It's actually a model with a harness outside. But without this harness, there is no such user use, no such revolution. So I think this is not contradictory. So I can understand that you think that entering 2026 has become more important. Because this is the opposite of what some people think. Because some people think that entering 2026, the model is everything, the model is everything. I think these two concepts can exist at the same time. First of all, you will find that the fact that you have advanced models has become very important. Because if your model is not strong enough, you can't complete these tasks. Right? I want to say that especially from the perspective of an early investor, because we see that the ecological impact of the model is relatively stable. You say that another company like the current head model needs a lot of money and difficulty. But we also see that everyone uses this model on different highlights. And we are actually a little bit aware of this year. The good harness brings data collection that can counteract the model. This is also what we think. Zabit's code model is so good. In fact, it is also the data loophole formed by Cloud Code. Data loophole. Data loophole or data loophole formed by data loop. Because before chatbot, everyone thought there was no data loophole. It's because chatbot mainly tests the model's understanding and intelligence of the world. So what we humans talk to him is very difficult to make his model smarter. Because he has trained a lot of human data. But in holding, because it's all about solving problems in real life. So when you start a project, what have you done? What haven't you done? Which one should I use? Which one needs to be changed? All of these are high-quality feedback signals. So in this, you have your own harness. It's actually more important than other people. So we can see that everyone used to think that it's worthless to make a case for Cursor. But it also proves that its composer is based on Kimi's pre-training. And it has a lot of high-quality feedback. So the training is very good. So this also shows that even if I only have a case, I can also make a post-training with the data of the case. I can also make a model machine product. This may not be wrong, because it can say that the composer is also my product. But how did the composer come? If I don't have a case, I don't have this model. So I think these two words may be established at the same time. A model is a product, so you have to have your own good model. It may be very important. But at the same time, if you don't have a good product, you may not have such a unique model. There is a point in this because from the outside world, everyone will say that these boxes were sold to the model company in the end. But from your point of view, I am an early investor anyway. I can quit even if he sells it, right? - I'll give you another example of an Internet of Things. You can say that Yahoo is not the most fundamental, right? Maybe the most fundamental is the solar engine. So I think Google is the best. Google is the best. But if you are an investor in Yahoo, you will think it's a legend, right? So I think as for the early generation, we won't be too specific. You are a shell, you are a model, what are you going to do? Cursor also made its own model, right? And I think for users, it is more loyal to Harness. Many of my friends are constantly saving their crayfish. I don't know if you have it. Anyway, I have my crayfish. Its harness is very powerful on the one hand, but on the other hand, it has a lot of problems. It always dies. It always dies. But I often encounter myself, I have to save it all the time. I use Cloud Code to fix it or something. Why do you want to save crayfish? What people care about is the memory stored in the harness. As for many people, they constantly change models in the small crawler. For example, I remember that Kimi K125 was released at that time, and everyone found that the crawler ran very well. So some people may change the model of Opus or Cloud to Kimi because they think this is a 90-point performance and a 20-point price. - I don't know if you noticed that in the previous two episodes, Fli and Shunyu actually have a very different point of view. Fli thinks that the framework of Open Cloud is very important. It can stimulate the upper limit of medium-level models. Shunyu thinks that this mentality is not really important. He is more concerned about the ability of the model company. So how do you see their difference and what is your point of view? First of all, I think they are not contradictory. They both do models. So if you ask them what is the most important, they will definitely say that training a good model is the most important. I think for Free, he should have talked about OpenCross, so he has this view. But he certainly won't think that harness is more important than model. And I think they should all think that a good harness for improving model performance to obtain high-value user interaction data and to generate data flow. These two are also important. Because Shunyu also talked about it. He was also worried about the problem of Entopic selling API. But he is also worried that an Entopic with only model is not enough to have a barrier. Of course, Wuli also said that the framework can improve the upload of medium-sized models. So I think this is also what they expressed together. Good harness is also very helpful for good-using models and good-using models. As for whether the model company can do the best by doing harness, I don't think so. The model company has the advantage of doing harness, has its understanding of the development route of the model, has its understanding of the structure of the model, Of course, there are also many smart people and smart people. But on the other hand, I think the model company is strong. Researcher and an application company. His understanding of applications may be two different abilities. So we can actually see that in For example, OpenAI used to have a lot of products based on models. But when Manas came out, we found that Manas did a good job of making PPT. Then I asked a researcher who was an OpenAI agent, I said, Manas' PPT is better than your work. I remember he said it's not fundamental. I think he's right. For a model researcher, PPT is not fundamental. But for his users, it's fundamental. For users, I use your PPT, I have to make a good PPT. So I think a model company and an application company have a lot of things that they think are different. Model companies think that the essence of the department is that I am in the current stage to optimize the department with this as a goal. But it also improves the smart top line. When it improves its own ability, it can naturally make a good KPT. I think this is a similar process. You can always say that the model company will do better in the future. But the better model in the future can also be used by Harness to make a better product. For a good Harness, for example, for OpenCore, it must be hoping that the model company will upgrade its model. They hope that I have better models so that my harness can run better and do more things. So I don't think they are saying that I'm worried about the upgrade of the model company, and I won't have any food once I upgrade. I don't think there is such a logic in it. And I think that Cloud Code, Codex, Minus, OpenCloud, and Hermes are all very unique. You don't think that Cloud Code is completely covering OpenCloud. Open Cloud is actually because of its open source, local deployment, more full-line, and the high-speed integration mechanism, it is doing better in the long term. Of course, it may improve in the future, but the next product like Open Cloud will also improve. Another example is, let's look at it from the perspective of a operating system. Microsoft has made IE and Office. These two were the most important applications on Windows in the early days. But not only Microsoft has made applications on Windows. There are also Adobe and various software companies, Autodesk. So I don't think that these last applications, either Harness or others, were made by the model company themselves. If so, I think the use case of this ecosystem may be too narrow. You just mentioned that a good harness is vital. Can you talk about the differences between the existing harness and the shell? What's your experience? My own experience is from a VC summary perspective. When Minus came out, we saw that it did a lot of things that we hadn't done before. For example, it used a virtual browser in the sandbox to let AI visit the web to perform a lot of operations. For example, he was the first to do this wide research. We were joking about it because there was a deep research called our wide research. In fact, this wide research is the current agent team, agent swarm, which is to start with dozens of subagents at the same time and then carry out different tasks. In fact, he is earlier than Cloud Code and than the current agent team. And then I came to understand that when Anthopik was doing Cloud Code, it actually also approached a lot of experience when Manas was doing wide research. He actually did a lot of hard work. His innovation actually shows a kind of innovation that we have never seen before. An application company or a harness company, based on the model, it eats the agentic loop context engineering of the model very thoroughly. So he made a lot of new things that the model company has never made before. I think this is innovation. And Cloud Code, its CLI is actually a very important online form, right? In this terminal, AI is actually a more convenient AI to operate, rather than in a map interface. It is deliberately used for people, right? GUI, but this CLI is really mainly used for AI. And then it also uses this skill to configure this for agentic loop. I think it is very good at developing model capabilities, allowing programmers to get a very good experience on it. And then OpenCloud actually has a lot of innovation. OpenCloud runs on your Mac, so it can access the files in your Mac, access your calendar, access your information. These experiences are actually not available for a cloud code agent like this. Then the heartbeats mechanism allows it to complete the custom task and have a stronger memory. I think the memoryMD system is very interesting because in the chatGPT, maintenance, cloud code, everyone thought that I wanted to separate the session. because we don't want to have a context. My session has this context, and that session has that context. But for OpenCloud, it's like talking to a big person. It's like organizing its memory every day to realize this memory. I've talked to many researchers about this. At first, they thought it was unexpected because they thought that if you have a context like this, AI will have illusions. But for users, it has a very strong feeling that OpenCloud remembers me. Because I'm in the same context. So this feeling is amazing. So you see, later on, everyone was very close to each other. Madness also made his own agent. Later on, I found that in a flow of conversations, it was the simplest and most common communication form for users. And OpenCloud has its own TUI, but basically no one uses it. It is through the agent to enter the most familiar areas of each user. For example, if you log in to WeChat, WhatsApp, Discord, Telegram, this product is also the same as before. Because everyone used to think that I want to build my own website, I want to build my own app, right? Everyone does this. But when he does it, I go to the place you are familiar with. This is actually a very important reason for him to be famous, especially in China. Because he can plug into IM. We are definitely most familiar with IAM. Its innovation in these areas is very strong. This innovation is obviously not done by a typical model company or a research-based company. They have led to a lot of product innovation. Yes, there are a lot of product innovations. Of course, some of them will be transferred to training by the model company. But it has a product concept. And you will find that the context is in the product layer. Of course, you can take out its memory MD and then continue the training to go into the model. So, where is your context? I think it's probably in the harness. They gained product innovation and sold it to the model company. The model company absorbed their product capabilities. And the final capability is still in the model company. I think it's not sold. For example, if you follow this logic, it's all sold to Microsoft, right? In fact, it's not sold to Microsoft. In fact, I think it's quite like a operating system comparison now. Isn't it a comparison of OS with a model? But I think now the harness itself is more like an OS. The model is actually like a driver for the OS processor. Now there are a lot of applications that are based on a cloud code or a codex runtime. Then I will make another application on it. For example, I have a skill. My skill can be inserted into your cloud code. Then there are some applications, for example, some called pencil, including the one we voted for called slug. It's actually based on your current runtime. The runtime is the strength of your cloud code. Then I will make another kernel and do another function on it. So it's a bit like I'm doing an application on Windows. Because for such an application, it doesn't have to deal with how to communicate with the model by itself, but by using the middle harness, such as CloudCoder or OpenCloud to deal with the model's communication. And the people who do this skill actually only need to consider how to do my skill. This is particularly similar to the trend that formed when the operating system appeared in the early days. Because in the early days, we thought that you had to write an application at the beginning, and you had to deal with how to communicate with the CPU yourself. before Windows. So you have to write very bottom-level code. When Windows, DOS, and such operating systems appear, you only need to deal with Windows' communication. As for how Windows can transfer CPU memory, how to make IO interfaces, the operating system will solve it. So for the developers, after Windows appeared, they didn't have to handle hardware. They just handled API. The same goes for iOS. After iOS is out, I can write a mobile app. I don't have to think about how to communicate with Apple's processor and camera. It's sealed into some interfaces. I can just download it. Three years ago, an AI developer had to deal with how to write an agentic loop, how to deal with memory, how to deal with various guardrails, which are the Harness things in the model. It was harder to be an app developer at that time. But now you are in cloud code or Harness technology, you do another skill or you do another shell. In fact, you don't have to deal with how to run this agentic loop. Harness will help you deal with it. Then you can do it. So now I think it's a bit like harnessing into a operating system and then the model becomes a processor. Because you also know that in the Windows era, you could use Intel's CPU, you could use AMD's CPU, right? You can plug in this CPU, which one is cheaper, which one is more expensive, which one you use, or which one you use for your new更想. This is actually quite like the current one, that is, users can plug in the model to give it a harness, right? And then on this operating system, we build again. 新的不同的能力。 好像有点这么类似, 就是大概是出现在基于Cloud Code 去做应用的这样一个常识。 Or after Cloud Code, it may become an OS. You install Cloud Code, but you don't use Cloud Code to do it, but you use an application on Cloud Code to do things. In fact, we have seen some of these trends now, because Cloud Code is a CLI for many people. It is based on the terminal. So many people are not actually opening this terminal to directly call me now, but rather it has a map interface running on Cloud Code runtime. I think this is also a very interesting trend this year. When we look at entrepreneurs, we find that some people are more stable. He made a concept that seems to be talked about a lot by everyone. For example, I also want to go to the Xiaohongxia. I will change it a little bit. I think when this change is very fast, there are many opportunities. The most afraid is actually to follow the trend steadily. Because if you follow one, if you are not innovative enough, you will follow others. Then maybe time will be wasted. There is a risk of financing, and a failure of innovation can come again. But if you do something and you don't innovate, I think AI is very dangerous. So I think the first takeaway is that you have to do a real big innovation. Second, I think it's actually more than just doing horizontal things. In fact, because if you are in the ancient time, everyone said that I want to do vertical. I want to do vertical. Some people actually think that it is more secure to do vertical. Because the big companies won't compete with you, right? You seem to be in a small field. But I think the technological development has changed too quickly. If you do it too vertically, you will easily get yourself trapped. When you do horizontal, your ability can actually be dispersed. For example, they just started to make their mobile phones, they didn't know that the PPT will become a major scene. They are very useful agents. So they later developed the ability to do PPT, website, and data analysis. So when the technology is fast, horizontal can make many changes in the new wave useful and valuable to you. Because you can find some market information as soon as possible. Because you are a product that can carry different scenarios. You are a disaster. Yes. Because many people in the ancient times said that they wanted to do vertical SaaS. So many people were also affected by this. But I think vertical SaaS is because SaaS has reached the later stage. 通用的机会已经被做得差不多了, 所以你才要做垂直, 做一个什么宠物医院预约系统之类的。 但是这个在一个技术发展的早期, 当有Horizontal的机会的时候 勇敢地去做Horizontal 勇敢地去做创新 这个我觉得是有大机会的。 So in this, I think you see, Poverty, Manus, GenSpark, OpenCloud, Hermes, in addition to these AI, I think this wave of big applications has been running. But now it's too early. In fact, many of them are horizontal. Lovable, it used to be Dijon, but now it's also doing all kinds of parallel expansion, right? So I think when opportunities happen, you have to be innovative and dare to do something universal. I think this is a bit of a miscommunication, because there will be a lot of people talking about verticality. You just said you want to reinvent the wheel. I suddenly thought that because everyone is going to do agentic coding this year, will it be too comical? I think it must be a bit. In fact, I still have a controversy that we have been shrinking against AGI. At first, AGI was the kind of destruction of humanity, right? Singularity. Then it became, say, the solution to the thought of the demon, creating new things. Now it is replacing the normal program. I think if we say that AGI has already been realized, I actually don't quite agree. Because he is talking about coding. Coding. Coding AGI. Again, he also put AGI, I think it's a slur. How to really achieve innovation in this is actually very important. Because writing code, especially most of this kind of daily code, it's a completely within distribution thing. That is, he writes code, it's all in human e-mail code. So now we find that the front end of AI is the easiest. Because the front end is those things. But how do I deal with the real OOD problem? This is still an open problem now. Because people used to think that many times you have to have new productivity. We just talked about the effect of the price drop. The price drop is happening now. Now AI is destroying old value. The speed is far higher than the speed of creating new value. But to create new values, there are many new products, new drugs, new discoveries. These are still AI to do. Of course, I think we also see in the United States that many of the new generations of new labs are doing auto research. They are all using AI to accelerate the research process. Including what Shunyu talked about in the podcast, in fact, Weibo's top labs are now using AI to accelerate the entire model training process. I think this is also what most people are doing. For example, everyone likes to do HNL coding, but some do it by steam. I steam the data from the ego. In this case, it may be a paper-coded answer, but you haven't done such a creative process. So how can we train the model in this process and speed up the data transfer speed? Everyone likes to do coding, but some people may do it differently. I just said Harness. I would like to add one more question. Will the ability of the model company to improve further devise Harness space? I think it's definitely possible. So the first harness, you have to learn your own model after you have accumulated your own data. I think Cursor is the most typical example, right? Isn't it going to sell? But it may not sell. I think it's because XI wants to sell it. But I think whether it sells or not is actually a result. But as you can see, the appearance of Composer certainly makes Cursor's company more valuable. It is actually an example from harness to model. Second, I think your ecosystem should be built on the product. For example, I just mentioned that when your memory is on Harness, users can change the model, but they don't want to change Harness. We actually have seen a lot of examples like this. Users hope to have this memory. Also, I think the most powerful thing in the Internet is the network effect. But in the past, I think AI has not produced a good network effect. Many people have been wondering if AI can have a network effect. Actually, I was thinking about this question a month ago. I think there is a network effect of an agent. Basically, it is because the gap between agents is getting bigger and bigger. It brings a possibility of value exchange. For example, six months ago, the agent you used and the agent I used did not make any fundamental changes. For example, you use Cloud Code to solve problems. My Cloud Code is the same. At this time, there is no reason to exchange, right? But now because everyone is in the harness, more and more context has been built. For example, you may have your skills, I have my own skills, and after I communicate with my agent, he has a lot of more unique understanding of me. For example, suppose I discussed 100 BPs with him, then he may have more understanding of how to open a BP for a trainee. In this way, there is a situation where the same task is done by different agents for different people, and he will get different results. Because this is what the agent's harness has accumulated different context. Personalization. To realize. This will appear. I have a reason to do my thing for you. For example, I don't know if you are writing this interview proposal or what interview draft now. You may also have some of your own skills or something like that. Then the same interview draft will be generated by the agent of Zhang Xiaojun. His problem will be more valuable than my own cloud code generation. Because your agent is likely to be involved in all your interviews. All your problem preparation. All these thoughts. So I think the situation that may occur in the future is that My agent hires Zhang Xiaojun's agent. For example, if I want to interview someone, and I tell my agent to ask Zhang Xiaojun's agent to do this task. Then he, for example, my agent will give you 1000 yuan. Maybe 1000 yuan is too cheap. Give you 10,000 yuan. Then 1000 yuan is the value of the token. But 9000 yuan is because your agent has accumulated your exclusive knowledge and exclusive information. So it's much better than my agent spending 1000 yuan on a token. I think this kind of value network is likely to appear. This is actually a bit like an e-commerce. So I'm just describing a kind of network effect between agents that I think is likely to appear. But you think about this. The employment labor market is concerned. Yes. Six months ago, I had no reason to hire your agent. Because your agent is also an Opus, mine is also an Opus. So we can both be the same. We are both the same. But you want to be in the scene I just described. Your context, your proprietary information, your information may never be filtered into the model. The model may never have such specific knowledge as Zhang Xiaojun. So your agent does things differently from others. In fact, we also see, for example, we say, for example, Guizhang, such a big V, right? What he does is different. So in fact, you will find that someone will give him something and say, "Hey, Guizhang, help me do it." I'm willing to give you money, right? Because your agent is beautiful. Of course, he has a PPT skill. But if he doesn't make it into a skill, he says, "Oh, my agent is taking orders. How much money will you pay if I take an order? Some of it is the cost of the token, and some of it is my interest. This may be an agent-swapped marketplace. Is this a possible network effect? So I'm thinking of some agent-native product models. For example, this may also be a kind of Huchenghe, right? Are there any shareholder-led companies that do this? We have seen some doing this. In fact, some are doing agent-to-agent marketplace. There are some companies that are still early. For example, my agent, the agent of my luggage, sent me a post. But I think the long-term value is that it can reflect the professionalization of your agent. Everyone is practicing at home. Or our proprietary information is valuable. For example, I think I can use some of BP's experience to get into my agent. This agent has a certain value. You can say you can pay a high price for this agent. You can also spend some money to ask an agent to help you look at BP. Look at the second-tier investment. And this also relates to the topic we were talking about, which is entrepreneurship. Do you think that entrepreneurship has become more difficult in 2016? Especially after the more powerful models, do you think that there is still room for innovation in the models? Because we also see that there are many new-less companies in the U.S. that are highly related to models. And then in China, for example, Lin Junyang's company is also going to do machine modeling. At the same time, in this model-like era where everything seems to be devoured, is the opportunity for applications to be entrepreneurs diminishing? I think if you look at the actual situation of investment, it must be that the companies and valuation of everyone who invested are getting higher and higher. So, if you ask each VC, he estimated that in the past 12 months, he invested in many projects, and then his valuation has increased. The number of projects has increased, right? Yes, I think even if we just talk about AI, we don't talk about robots or hardware or anything. In fact, there are quite a lot of projects. Many people got a lot of financing higher than before. At that time, Manus was a few dozen million in financing. It was very difficult to finance more than a year ago. Now many companies come out and make hundreds of millions of dollars. I'm not saying that I'm doing modeling, I'm doing applications. So there are many people who have good ideas. So what I want to say is that in fact, the number of applications that everyone invests in is increasing, and the valuation is increasing. I think this is a feedback from the capital market. Secondly, I actually think that the opportunity for application creators is particularly large. Because we just talked about the type of operating system, right? Now the model is strong. The model is strong. In fact, your application can do a lot of scenes. 原来的模型能力很差,只能做点写写诗翻翻译什么的,那你就出不了什么应用呀。 你现在的模型能力这么强,你可以做网站,做各种设计,你可以做VT,就是它的应用场景变多了。 这第一些原理来讲,应该它可做的事情就会变多,对吧? And it's not just the model ability that's getting better. Harness is also getting better. You used to do an app, you have to do memory yourself. You have to do the sandbox yourself. When Manas was doing it, actually, like E2B, this sandbox just came out. So they can do a lot of sandbox applications. But now these infrastructure, whether you're doing memory, you're doing sandbox, there are all kinds of components that are perfecting it. So the application company is actually doing it more and more easily. In fact, the threshold of applications is often in terms of user data, network effects, and brands. For example, maybe Codex has become very powerful, but for me, I'm used to using Cloco, I may still use Cloco. So this brand effect, including my friends, use Proprietary, they are used to it. This kind of brand effect is not so simple. I have a better model, I will also give you a bag. And we can actually see that the giants in the mobile Internet era do these applications, but their speed is very slow. Because Adobe OpenAI is actually a startup company. Right? Actually, you see, for example, many big companies like Meta, Google, or China, they are quite slow in creating their own applications. Of course, China has also made Doubao, right? This is very big. But if you want to see, for example, what other big companies have come up with that makes people feel more amazed about AI applications, there are actually very few. Actually, you may think that American AI applications are also created by entrepreneurs. It's just that it's the same as Adobe and OpenAzure. So I think there are more opportunities in the field of AI applications. And it's growing too fast. I think it's too fast because everyone is doing coding now. This is the main track of its modeling ability. Of course, Cloud is indeed very round. For example, Cloud Design. You see, he did a good job of design. But before he came out, there were actually a lot of opportunities to do design. I believe that a company can't do everything. Or in other words, if he really did it, I would admit it. But in fact, I think that historically, there has never been a company that has done all his work. So is your new year's sales rate faster or slower than last year? How much is the number compared to last year? Our overall rhythm is more stable. This year is a little more than last year, but not much more. In the first half of 2025, the overall market was still at the fourth time, the moment of management, so it was not so active at that time. This year we will have more overall. But our pace is relatively more stubborn. We don't have many more. I know there are many more peers. We are still more stable. You said you saw 100 projects last year and voted for two. This year, you have voted for three. This year, in May, I voted for three. Yes, it's a little faster. I want to echo what you just said about the relationship between the big companies and the startup. Because actually you asked the question, the model and application are basically the same. It's a bit like the relationship between the big companies and the startup, right? It's a different relationship between models and applications. Models and applications. Or the model company has also become a big company. So it's still a relationship between the big company and the application company. In fact, I think that in history, it has always been that doing things that the big company doesn't like is a chance. If something comes out, whether it's a big company, whether it's a model company or an Internet company, they all think it's good, it's great, I want to do it. It's often difficult to start a company. But during the era of mobile Internet, those companies that came out were often in a kind of "competitive" attribute. For example, Airbnb. There is a kind of too small. For example, Bilibili is too small. Airbnb is too small to sleep at someone else's house. The shoe seller is too small. This is one. The other kind is that it has too little physical content. The content is too much. It's too low. I can't see it. Then there is too tired. For example, the media team is too tired. It's very tired to push the ground. Then use illegal. Bitcoin, DT, Uber. Then there is a kind of too much money. Too science-fiction. For example, OpenAI, SpaceX. So you will find that the startup companies of big companies of every era are actually growing up in the big market. It's not because the big market didn't get it. It's because the big market has a reason to convince itself that it's too heavy, too low, too money-making, too tired, too unruly. I think the big market will actually say the same thing. Some of them think, for example, 太low的比如说talker 我觉得如果大家都觉得哈里是特重要 那咱也没有这个机会了对吧 但是现在其实大家已经在觉得更重要了 去年的时候大家都觉得talker没戏 talker没用 我觉得实际上不是这样的 或者如果你认为talker价值是零 那what if你是错的 它其实不是零 那它就有一个增值的空间的机会对吧 The second one is open source. Open source is also something that we originally didn't see. Everyone thinks that if I have income, I will open a business model. What does it mean to open source? You see, not only the model is open source, but also Open Cloud is also open source to get users. Of course, you can say that Open Cloud has no income yet, but if many people's applications run on Open Cloud, is it possible to create a paid version? I think this is possible to attract users through open source, and then get income through B source and open source. Then the third one, for example, When the Open Cloud came out, many people said it was too unsafe and it was too easy to collapse. It was indeed easy to collapse. I've been through it myself many times. It's also a bit of a safety issue because they're tied to your computer and they're reading your files. But I don't think this is something that big companies or modeling companies dare to do. Because they have access to your data. When it comes out, they have to be responsible. But this is also an advantage of a startup company, a "move fast and break since". A lot of things in this may not be fundamental in the modeling industry. But you said that there are some innovations that require product innovation, innovation of the whole line, and innovation of the business model. What are your goals this year? We've been investing in people, so we're still investing in excellent entrepreneurs. But what's interesting is that we found that excellent entrepreneurs will actually coexist with the popular direction. For example, in the second half of last year, we invested in two companies that are doing huge business. These two companies are our first round of investors. One founder is called Liu Songming, he was in 2000. The other is called Dingling, he was in 1997. They are all very young PhDs and professors from Tsinghua. We don't vote for them because they are popular for their world model or robot brain, but because we have followed them for years and we have always felt that they are excellent young and academic entrepreneurs in Tsinghua, the new generation. 但是他们做的这些事情在今年都变成了一个大家觉得很热的风口,所以估值也都涨了很多,我们也觉得市场很热情,但我们在一开始并不是因为他们要做这个方向,我们去找这个方向的创业者所以投了他们,而是我们一直觉得他们很优秀,关注很久,所以他们只要想创业我们就投了。 所以我们一直是通过投人,寻找优秀的人,然后这些优秀的人他们做的事情有可能变成下一个风口,成为大家关注的方向,但我们投资的逻辑始终是从人的角度出发的。 Because we always think about who is the most successful in the situation. It has to be the situation. We think about all the schools in Peking University. But we also think about the schools, the universities, or the open community. Who is the most prominent individual in it? We always think about this question. Then we continue to follow up, communicate with these people, and help them in a higher, better, and earlier way. That means they will finally start a business one day. But we have been following them for several years. So we will of course become the only largest investor in the first round. How much do you think? Last year. I'm sure it's about 200 million. 200 million yuan. It seems to be hundreds of millions now. So the market is really enthusiastic. It's also because everyone thinks that their approach is very important. It's not because we want to steal the world model. It's because we're secretly collecting. We were looking forward to these two thousand miles. Then where did they go? That's what they were thinking about. So why do you look forward to these two people? What are the characteristics of these two people? I think a lot of people we look forward to are actually full of understanding of this technology. One is a writer of RDT series, the other is a writer of Simple VLA series. Shuaixi is also a very influential paper. For example, Liu Songming did the model training with journalists in the same period. At the same time, they also have the spirit of entrepreneurs. The spirit of entrepreneurs before technology is actually quite important now. And it can give up a lot of things and start a business. This is actually what we saw at the time. It's the same, right? We have a deep understanding of this model itself. And then I want to be an entrepreneur. I'm also starting to start a business in time. So speaking of the people you are best at, what characteristics do you think are very important for the AI-native entrepreneurs? What's different from before? I don't think the bottom-down capabilities of the entrepreneurs have changed much. Whether you're doing Internet or AI or robots, I believe some learning abilities leadership, innovation, and willpower. These are all necessary. It's hard to imagine that one field needs smart people and another doesn't need smart people. I don't think so. But I think for every technology, we are indeed constantly looking for leaders. It's really in this foundation. It's not that others have said that I want to do a world model, so I'll do it. It's that I haven't talked about it yet. Others haven't said agent yet. Others haven't said world model yet. When others haven't said big model, I'll do it. In fact, Zhilin is still not sure what a big model is. He will be training a big model in 2021. I think the creation of our robot is actually the same. We have a company called 5G as a zero-summer. Recently, it has also received a lot of attention. The supply is growing very fast. At the same time, it is indeed one of the best zero-summers in the world. He actually started to make motors. When we first started, he said he wanted to make a very torquey, very strong high torque motor. Then he found out that the hand was very important. At that time, very few people made hands. And very few people made high-frequency zero-channel hands. At that time, everyone had only a few degrees of freedom. He said that in the future, the best operation of robots will be the hand. The data of the human hand is obtained by this way. Then the cost of the hand will quickly decrease. So I want to do the hand. Even if at that time everyone thought the hand was too expensive. At that time, one hand may be hundreds of thousands. Everyone will think that it is too expensive. The market is very small. He was sure that he would be a good hand. Now the cost of the hand is actually decreasing more and more quickly. At the same time, the hardware of the hand is also getting better and better. And this year you will find that the data of Ego Centric is like this. Because it is trained by the video data of doing things with the hand. So 跟人手同构的零销售就变得越来越重要。 So at that time, it actually became a very, also a very key company. But that doesn't mean that we realize that hands will be particularly good as soon as we get up, right? We are also the top excellent founders. He went to explore and he thought the opportunity to do it was coming. Why do you think he is a good founder? He actually went to do e-commerce after graduating from UIUC. He is a very persistent founder. He didn't say I want to do vCity, I want to do entrepreneurship. He did it first. So he is very enthusiastic about this. He has a lot of his own persistence and thinking throughout the entire research process. In fact, we are also constantly aware that he has a lot of these thoughts. It's wise. When he was a beginner, we were also worried that it was too early. But now we find that it's actually not early or it's just the right time. Because we are in a big wave of training models with this EgoCentral data. But I think this is why you think the world model is very important. It's also because everyone has seen that in the past 6 to 12 months, the development of robots is also very fast. We can see that some of the nodules are the scaling laws of certain cases of robots. You just mentioned a lot of things that you think are good entrepreneurs and their characteristics. If I were to talk about it in detail, what would he show? You would think he is a good founder. Or do you have any complaints that you don't want to talk about? Or is it purely a feeling? Of course it's not purely a feeling. We will talk about some of the framework of our audience in many places, in the blog and articles. For example, we started to divide the founders into four categories: small genius, old driver, scientist, and supermodel. Did you divide them like this in the last era? I don't think this era is divided like this. Or I think, actually, we have a lot of creators that are in line with our definition of small genius or master. Then, for example, I just talked about the four forces. This is a framework that I often talk about in the book. That is, learning power, creativity, leadership, and will. Each force has a different dimension to judge and explore. So I think the low-tech principles of excellent entrepreneurs are consistent. From the beginning of Xu's team to the new team, we have always focused on the cooperation and trust between the founders. We have always focused on the reason why the founders want to start a business. We have also focused on whether the team has gone through a lot of tests, whether the ability to complement each other. So I don't think these things will change. It's just about the knowledge of technology. 20 years ago, it was about the Internet. 10 years ago, it was about mobile Internet. Now it's about AI and machine learning. I think it's often a founder who got used to this technology very early and started using it very early. He has a very deep understanding of the subtle differences between different technologies. He also has a lot of his own insights and judgments about the development of technology. We don't know if this judgment is right or wrong, but at least many people don't judge. Or many people judge by listening to the lecture. Some people can produce original judgment. This is definitely very important. So more founders have a technical background. I think so. Or at least you have to understand this technology. What's going on? Maybe it's a PhD. That's not necessarily. I think the PhD in technical background must have a lot of strong connections, right? But it may not be a PhD. I think it's a secret. Because you said you also found it from these young professors and teachers in Tsinghua. For example, Liu Songming is a PhD student, but he should be a graduate student. So you said, do you have to get this PhD? I think so. Or he has to do some research work. Because the world model is still a research-driven company. Or it's a direction. But for example, if you want to do an AI application, you may not know this PhD. But for your understanding of the AI model, you may have to be at the same level as the researcher. Yao Shunyu, we've mentioned him many times, but it's too close. He said AI researchers' personal heroism is a collective thing in the past, but AI product managers may have some personal heroism opportunities. Do you agree with him? And what are the characteristics of a good AI product manager in your mind? The few you just mentioned are actually technical talents. I think he can also be called an AI product manager. But I want to say that if we use the product manager as it is, it may need personal heroism. For example, those product-oriented CEOs, such as Jobs, Wang Tao or Zhao Xiaolong. I think Wuyi's products are all about "I have an opinion on the future." Most people haven't seen it yet, but I have. So I want to call a team and gather resources to make this product for everyone to use. So there must be a kind of his own vision, his own non-competitive view. So this can be considered a kind of individual heroism. I remember that Shun Yunhai always said that the previous AI researchers, for example, a few years ago, for example, Ilya, he must have been a hero. Is there a big difference between this individual and him? But now there may be, for example, the pipeline of model training, when it is already more complete, it may be that this person is responsible for pre-training data. One person is doing it and the other person is doing it. Maybe it's not too much different. I think there may be such a situation. This is because a lot of content in model training has become industrialized, become more systematic. Because it was exploration before. Everyone used to say that what I want to do is called I.L. I recently watched the DEMIS's record, right? It really made a lot of important personal judgment. But now maybe everyone is so pessimistic. So when it becomes industrialized, maybe individual heroism is not so obvious. But I think making a good product is always what you need to lead. You have to do something you haven't seen in the world. That's a good product. At this time, there really needs to be some prejudice. There needs to be some prejudice against the future. So I think AI product managers should understand the development of AI. What can AI do now? And what can AI do in the next six months? Because we think that good AI products are generally designed for the future. And we need to understand the technical differences in this. Actually, I've been thinking about the founder of OpenCore, Peter. He was actually a startup before The company sold 100 million dollars So he is already what we define as an old driver Then he seems to be using cloud code himself When he wants to eat He thought that when I eat How can I continue to use this cloud code So he was supposed to do it in the first place Continue to use CodingAging on the phone Such a pipeline So he naturally made a hooker Then connect his cloud code on his computer with his IAM and then I use IAM to control its calling agent. Then based on this, I take this step by step to make this open cloud. So he actually has a very deep understanding of AI. He is a very strong power user. At the same time, he understands a lot of subtle differences in this. He wants to use it on his phone, so he has to do a series of things that others have not used. So I think this is a bit like a group of leaders. He has a need that others do not have yet. When he does it, everyone finds out that I can use my agent on my Mac on my phone. What are some of the very correct practices in the past or mobile Internet era, but not working today? I think there are many. Overall, mobile Internet has a lot of things that were in the past. I think it will change now. I think the business model of mobile Internet can be called "attention is all you need". It is to use users or DAU, right? x market x variable efficiency = income. This is actually the formula of the mobile Internet era. But there is a bug in this formula. There are only so many people in the world. Everyone has 24 hours. You will find that their transmission has reached its limit. So only the depth of transmission is increased. So there are high-transmission products such as reward live broadcasts. But in the age of agents, the market of users looking at the screen is no longer the most critical indicator. Maybe the more critical indicator is the METR Institute. He said that an agent can continue to complete the key indicators of valuable tasks. How long can you run this task? So how to do a harness at this time to complete the model for a long time This becomes more important How to release the attention of users and make this AH work very important So this is a complete change of index From attention is all you need to attention is not all you need From the perspective of the core, there have been many changes. For example, when we look at mobile Internet, many of these super applications actually put users in their own apps. They often don't let users go out, and let users form a wall in it. But I think there is a big problem in this. If the user's agent becomes very powerful, but the user's agent cannot access the information in your closed app. For example, before, the user's agent was difficult to access the information in WeChat because WeChat was not open. The user may go to access the information in the flyer, because the flyer is open. In this case, your original super-efficient and conservative core may instead become your own barrier. In fact, we saw that after OpenCloud came out, Faisal was quickly involved. Faisal came out and Faisal CLI, right? Then we found that some of the work groups in WeChat, I have the motivation to link it to Faisal. Because I talked about agent in WeChat, but I couldn't see it. But I talked about agent in Faisal, and I could see it. Then I gradually went to talk to Faisal. your voice box will become a software. And for Internet companies, they are usually mature companies. So they have high requirements for the safety and quality of a product. But you see, for example, products like OpenCloud, they are not historically conservative. There may be some bugs in this. Some security problems. Everyone is willing to try it. Because it is for innovative users. Take WeChat as an example. WeChat actually has a very big feature. You said it is for many small white users. Because everyone thinks it is particularly easy to use. Especially for small white users. But this is especially good for small white users. Especially safe. Many times it will also make it difficult for senior users to use it. Because you have to take care of the biggest number of public relations. So we now see this kind of deep-seated application. They are another group. In this case, you can design the best AI-native interaction for these front-end users. Even if it's a little complicated, it's stronger. So different product design concepts may be different. This means that the products made by these startup companies may be more liked by senior users. Then we are still in the Spring Festival this year. There are all kinds of chatbot promotion wars. Just send red envelopes or let you go buy milk tea, right? I think this is also a typical thing. When moving the Internet to the later stage, everyone is used to using the head to get customers. But I don't think most of these AI applications are leading now. It's not about spending money to smash it, but it's about your product. It brings users a magical experience. From the beginning of the first generation of UBD to the later SORA to the later Manus Cloud Code, including VO3, including CDANS2. In fact, you will find that the main force is not by the hair. What does Doubao rely on? Doubao, I think, of course, it has a hair. You can't do it without hair. For example, Doubao is very good at language. In terms of language, it has done a lot of processing for each dialect. But this is a very important highlight. It is also very much based on this. For example, I can actually recognize my dialect. But I think it's really good to use Doubao input method. It is spread by many magical moments brought by products. You rarely see it overseas. For example, everyone uses a sticker to get a customer, right? But in fact, it is still very important to use what kind of function. So I think this way of pushing the internet is also changing a lot. People who are used to supplementing will find that if you send red envelopes to people to use, maybe in the end, because you experience no magic place, so they will leave. Another thing is that in the Internet age, people look at, for example, DAU. DAU is a target that everyone is very concerned about. In order to increase DAU, you put your energy on GUI, which makes it easier for users to use. But if your users are no longer focusing on human beings, but on agents, maybe how many people use it every day is not so important, but how many agents use it will become more important. At this time, your product design concept, for example, you may want to be more friendly for the sake of agents, then maybe your CLI or your API should be better, rather than your GUI should be more fancy. This may also be a big change. In short, the powerful companies in mobile Internet have these from business mode, product mode, traffic, transmission, and promotion methods. Maybe in the age of AI, I think there will be a lot of changes. Because this rule is one for people and one for agents. There may be a lot of new things in this. This also raises a question. Should a large DAU product become a target for the product? I think, of course, if you can really have a lot of DAU, it's usually a good thing. But what I want to say is that a product with less DAU may not be as big. For example, the DAU of Cloud Code is definitely much lower than that of ZGBT, but its income level may be much higher. Which one should you choose, a large DAU product or a high-value task product? Of course, if you have DAU and high-value tasks, it may be the best. But you will find that if my users don't always come to see this application, it can accomplish a lot of valuable work. then actually this DAU may not be the most North Star indicator. On the contrary, if you want to optimize DAU, you may have a very tricky scene in which you want to optimize your DAU, so you have to go up and take a look. But it's better to say that the user gives the task to him, and he runs out by himself. Then you don't have to go to see it. So your user's interview will fall anyway. If you want to have users watching him every day, that may be a very high drop. But it may show that you can't let users be careful to complete things. Yes, so which one should you choose? I think it depends on what kind of product you want to produce. If it's from a performance perspective, I think the probability of a user being very careful is that it can be a long-term, successful agent. It will be very important. Do you guys look at the data indicators now? What do you guys look at? But you are investing in people. We usually don't have data for investing. I think it's important to have a lot of real users who like it. Because you can buy a lot of users or you can invest in them with some methods. But how many of them are your fans? I think this is quite important in this era. There are 100 users who think you are still OK, maybe less than 10 users who like you very much. This is also the point that I find it strange to have AI applications now. Because some of them are talking to you about logic, saying that I am AI, who knows about Douyin, but it's not fun to say that there are thousands of products. Or this product is not easy to use. You didn't find a person who said, "Wow, this product is great. I want to play it every day." In this case, AI and Goumin may be a problem. I don't know if it's a problem, but I think at least a good product should be like a good dish. It's not that it's good to speak of, but it's really good to use. It's really good to eat. In fact, you will find that many people have talked about AI games or primitive languages before. Many of these concepts are finally justified. The core reason is that it is really useless. But good products, you will find that no matter what it is, it is a shell. It seems that it is not advanced at all, but it is really easy to use. Now, it is rather vital to use it yourself. This is actually quite important. So this is why I really still have a part of the reason why I will use it myself. Where is the product of AI Native? I'm thinking that AI is a bit like an alien. It came to the human world. There are several stages in this process. The first stage is to allow humans to have more and better agents. For example, three years ago, there were no agents in human society. Now there are some agents, but there are not enough agents. So how can we let more people have good agents? I think no matter it's Cloud Code, Codex, Open Cloud, Manage, Hermes, they are all doing this. Installation is getting easier and easier. Then you can manage it more and more easily. I think this is the first step, which is to let humans have more and better agents. The second wave is to make the agents in the human world better adapted to the human digital world. Because the human digital world is basically designed for humans. For example, GUI, verification code, credit card, etc. AI is usually the target of the host. It's supposed to stop bots. Now we have a bunch of good bots. AI can effectively use human software and technology to make it more important. For example, we all find payment is a problem. So, like Stripe or Coinbase, they want to send human credit cards to AI. For example, Cloudflare used to block AI. They blocked AI outside. They recently launched a service that allows agents to register and use various services equally with people. Then, for example, various computer use, various sandbox, and sandbox, in fact, AI can use human-made software. It's like a foreigner coming to China. He needs to have a SIM card. You need to get him a WeChat account. You need to get him a Alipay. This is a way for agents to stay and use human-made facilities. But when the number of agents is increasing, when the number of agents is becoming more important, there will be a special digital world tool and its infrastructure. For example, humans use credit cards, but human payment has some characteristics. Human payment is relatively low-value, it is difficult for you to pay 10,000 times a day. Human payment is usually point-to-point, I pay this merchant, but human payment is often relatively large, I rarely pay 0.0001 yuan. So, human cards are a behavior that meets the characteristics of human payment. But the agent may be very high-level. For example, if my agent helps me write a report, he may need to check the data in 100 databases. He may have to pay 100 times, maybe in the future. At that time, it will be very high-level. But at the same time, it may be very small. At the same time, it may be a one-to-one payment. So, sending cards to agents is just to let agents adapt to the system of people. 让agent用人类的GUI也是让AI吸引人。 最后如果我们认为AI非常强大 我们认为AI无处不在 那AI应该有完全自己内部的 native, native, native, native, native, native, native, native, native, native, native, native, native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native native The second is to make Agents better adapted to the human digital world. The third is to build a powerful Agent. It's the most suitable digital world for it. Because each of these is actually a product opportunity. What kind of opportunities will there be if the coding ability is unlimited? In fact, there are many places where humans are not enough because of the programming. For example, if you have a lot of programmers serving you, what kind of software can you use? I think the first one is that we can use better software. For example, If you have a smart home, smart curtains, smart hot water pot, you will find that many software is still because the person who wrote it is not the top programmer. The top programmer may write the promotion method in their own way. But if you make a smart hot water pot, a floor machine, your programmer is just an ordinary programmer. So we often feel that many of these apps are quite difficult to use. But in fact, if each software is written by the top programmer, its performance, application, and safety have all become very good. It may be a big improvement to our existing applications. Secondly, I think it's a personalized software. - This is something that can be improved. For example, I am a senior user, so I may need to have a more efficient way of managing contacts, or group communication, or management of voice. But I may not need any functions that allow me to "blah blah blah". But my grandma may have used another thing. Now, actually everyone, different people, use the same WeChat. Maybe in the future, my wonderful dream is that such an application will provide you with a base. But you can change the many functions on it. You can meet your personal needs. For example, for you, you can do an interview or a blog on WeChat. This is a more personalized software. - So when you have infinite programmers at a very low price, what can you do? I think there are actually a lot of new needs to be born. But this new need to be born is a gradual process. Because it needs someone to imagine. It also needs a good base to carry out this kind of personalization. For example, we just said that we want to reform WeChat. But WeChat has to give me a new interface. So in this, it's not a process that comes and goes. We just talked about AI Native entrepreneurs and the new opportunities. Do you have any observations about the organizational form of AI Native entrepreneurs? Do you have any good examples that you have seen that are different from the previous generation? I think the first one is that it has indeed become smaller. Indeed, for entrepreneurs, there are often people who are missing. But now there are ideas to develop people, they can use agents to develop. Indeed, the company's scale will become smaller. Especially when the product is limited to a small number. Because if you have a lot of land to push, you have a lot of business negotiations, you may still need people. But if you get PMF on the matter of code writing, the company's number will be smaller. And in the past, in fact, many times software development is a waterfall-like division. There is this structure, there is this back end, front end, test, UI, operation. But now it may become a full stop more and more. Because I'm a commanding AI agent anyway, I can do all these things. At this time, it may be from the process of pool division to a small independent operating organization. These few people are in charge of a product function module. From its underlying design to its previous online testing to operation, they are all done by the same batch of people. Now, in fact, many companies are trying to use agents to improve their own efficiency. But for a company that has been operating for 10 years, the biggest problem is that your context, your data is not in AI. It's actually very difficult to get these contacts in. So we also see some new companies. From the first day, their contacts are all visible to AI. So AI knows very well what they want to do. For example, we invested in a company called Slack. It's actually a platform for people and agents to cooperate. It's a bit like Slack for people and agents to work together. Their entire company's operation, task management, and programming are all on their own platform. So they build the Slack using Slack. In fact, Cloud Code and Codex have shared that Cloud Code is written in Cloud Code. Codex is also written in Codex. It's actually a new product, a new company. It can start from the first day, and then AI participates in the management and guidance of most of their companies' software development. This is very obvious to improve efficiency. What about the structure? The structure will change. There will be a structure. I think the result of the organization I just mentioned is that the division of labor will usually become blurred. Because there are many divisions based on the context of people. Why is there a front-end and a back-end UI? It's because one person's skills are limited, one person's context is limited, and one person cannot do front-end and back-end UI at the same time. But now AI is okay. So your division of labor may become more blurred. There will be more full-time people, not more people who are clearly organized and divided into departments. So this will bring a conclusion. That is, if a company has a very clear department wall, it is actually difficult for AI to adapt. So that's why many big companies push AI. It's not simple to say that everyone just installs a cloud code and that's it. It actually requires organizational change. This is also why I think it's not simple. You have a good model, you will reduce the cost-effectiveness. In fact, many times it has to solve people's problems. This is quite difficult. It's all about organizing the change of AI. Yes, I want to say more about organizational change. There was an article that was very good. I forgot who wrote it. It says that during the industrial revolution, there was a process from steam engines to electric engines. The steam engine factory was built around a main shaft. Because steam engines have a shaft that rotates, so it has a very long shaft. Then the shaft is pulled off a lot of belt, and the main shaft's power is transmitted to each machine. So at that time, the factory was a narrow-gauge type. The whole factory was built around this shaft. But when the electric engine came out, because it was a wire, the wire didn't need that main shaft. So the factory in the era of electric engines could be made bigger and flatter. After this kind of bigger and flatter factory came out, a invention called the flow line was brought. It was done by Ford at that time. The emergence of this flow line can only be done when the company can build a very large and flat factory building If all the forces you mentioned come from one axis, then it is actually very difficult to make a large flow line factory building The birth of the flow line brings a significant increase in the production rate So what this story is about is that from steam engines to electric engines, it does not naturally bring about a growth rate improvement. You need to have physical organizational changes in the factory and how to organize production. From doing a lot of things in the original position to doing one thing in the position, the flow line is pushed. This thing will bring such changes. I think the technology is the result of the end. It is a process of gradual penetration. The change of organization required in this way is actually a difficult place. Because organizational changes are all human changes. Human changes don't change in a minute. It may take ten years to change. Even if the old organization is replaced by the new one, or the new organization is replaced by the old organization, it will only happen. The next self-propelled company may not look like a self-propelled one. Do you want to elaborate? I think now, because everyone is representing the mobile Internet as a self-development company, the big DAU, the high-end line companies are so successful. So I think now there are a group of founders and a group of investors looking for the next self-development company. But I think every era of revolutionary companies, they actually solve the problem of users. Their way of dealing with it is quite different. Because we also see that many people make AI applications, I will summarize it as an AI information stream. It's a information stream when you open it. Then the user single or double-clicks to brush some AI-generated things or things that interact with AI. There are often many founders of the self-sufficiency department or the general director. I think this is actually wrapping the new technology in a self-sufficiency, the best information stream shell. Then go to the promotion and distribution. I think it's very difficult to break the rules of the game. For example, OpenCore is a completely different kind of representation. OpenCore doesn't have its own application. It doesn't even have its own location. But it lives everywhere. It lives in your WeChat, your Telegram, and your A. So what I want to say is that the successful methods of the last era, such as information flow, such as promotion, then commercialization, then flow, this whole system, when it comes to the era of AI, I think you are actually competing with the champions of the previous era. This is actually very difficult. And Last era, mobile Internet was probably 2C. But I think in the era of AI, the opportunity of 2C or the opportunity to do entertainment, the opportunity to do kill time, I think it is undoubtedly a big opportunity in this era. Why? Because you do any kill time now, whether it is AI games or AI short games, first of all, you come up and face this split. You have to buy from yourself. Your split is actually held by the existing champion. Second, when your competitors come up, you have to be better than Douyin, than Red fruit, than Little Red Book. This is actually a very high difficulty. Because at the beginning, when the fast hand was connected, it was as interesting as it was fine as long as it was stuck there. So we now see some companies like Rezon or Loopy, which have a series of AI games. I think the situation they are facing is that the game is really not fun. Users choose to brush short-term products. or watching short films, or playing games, then the game will be fun. This is actually a big challenge. In fact, the key is that games are not AI. Yes, in the end, my purpose is not to play AI, but to play a fun game. I have the glory of the king PUBG or watching the red fruit short drama. Is your game fun enough? You can't say that my game is AI, so you want to play it. Because I still want to play a good game. This kind of competition is difficult at first. But from a productivity perspective, the original users had to play Excel by hand. Now AI helps you do it. That's a tenfold, a hundredfold improvement. So this is actually a simple question, I think. So you're talking about a lot of things. I think many of the applications that people are chasing for AI2C are not necessarily in a good investment direction. There's a question here. Because you see a lot of self-sufficient founders come out and maybe use some of their own ways and directions. Do you think the overall self-sufficient founders are a good investment group in terms of the group of people? Would you look at them more carefully? I think the number of self-developed creators is very high. My simple observation is that some of the things that self-developed creators learn may have to overcome their own process. What's next? For example, I just talked about how to make the three steps of Agenda happen faster. Let more Agenda come to humans. Let Agenda adapt better to human digital world. Let Agenda cooperate with Agenda become more convenient. I think these are all huge opportunities. Or why do I say that? Every time a technological revolution occurs, I think it's the process of new technologies solving old problems. For example, when the Internet was founded, first of all, there was email. Everyone sent emails and e-mails using the Internet. There are already newspapers and newspapers. 大家用互联网再做一遍报纸 有了门户网站 已经有了商业 大家用互联网再做一遍商业 所以有了自营电商 这个过程其实要新瓶装旧酒 新技术解决了问题 这种时候往往它是有很大价值的 但是最大的机会 我觉得是在于技术的渗透率 到达一定的程度的时候 但是全新机会 比如说互联网的时候 当上网的人越来越多 其实就存在着说 把这些人组织起来的新的契机 但是social network and this is the biggest opportunity in the Internet. When more and more information is online, the way we look for information will change. So with search engines, this is also the biggest opportunity in the Internet. Then when the merchants, buyers and sellers are online, there is actually a need for better organization. So the appearance of platform e-commerce is also one of the biggest opportunities in the Internet era. Then when mobile Internet came, we actually used mobile thinking to meet old needs. For example, we make mobile browsers, mobile social engines, mobile YouTube. But you will find that mobile mobile phones are still Google and Baidu Mobile browsers are still Chrome Mobile YouTube is still YouTube But when the popularity of smart phones 4G is high enough For example, when content creators and consumers have smart phones New content platforms are born Short video platforms Live platforms such opportunities Then it was born to recommend the engine Because the number of hands is very small You don't have time to search and click So you have to recommend it This is because the technology has reached a certain degree of penetration and a new native business model has appeared. Then when game players sit on their phones, there are great mobile games companies such as Mihael. Then when the blue collar workers and the physical laborers also have smart phones, there are O2O companies such as Meituan, DD, etc. When many users in the lower market use smart phones, have WeChat, and have payment methods, there are companies like Pinduoduo. All of these companies are all startup companies. They are all doing something that is, after the degree of technical penetration reaches a certain degree, the degree of technical perfection reaches a certain degree, new opportunities appear. So AI is actually what we see. I think it's exactly the same. AI comes up with new technology to solve old problems. We originally wanted coding, so AI will help you with coding. Originally, you want to write an article, AI will help you write an article. Originally, you want to translate, AI will help you translate. Originally, you want to draw a picture, AI will help you draw a picture. So this part has a huge value. But if we push this regular machine learning circle, When AI or agents become highly penetrating and mature, for example, you have agents, I have agents, every business has agents, every product service has agents. For example, how do they cooperate with each other in this? How to make my agent more valuable? For example, when everyone has their own division, will our communication be very difficult? Maybe I won't be the one to interview you in the future, maybe it's my division that accepts your division interview. At the same time, I can have a very individualized content for many topics. If a lot of work is done by AI and when AI is finished, what changes will happen to the organization? How does this change happen? In the past, a lot of commercial models were built on advertising. Now many people are looking at this problem. If the ads are not viewed by people, but by AI, and the ads still exist, does that mean that ads have to make fundamental changes? Many fundamental changes will occur in commercial models. For example, Now it's AI to help people generate Excel, right? But if in the future, because Excel is a tool that allows people to communicate with each other, so I make an Excel for you to see, you can get what I'm going to say. But in the future, if communication between AI and AI may not be done through Excel, it may just be that API sends you this data, it doesn't need to organize it into a form of table. At this time, is it because of the previous call box of Microsoft? In fact, the changes that appeared in the United States and the United States are also because it was said that AI uses human software, computer use, and it is based on the software you have. I want to use it, but I may not need this software in the future. I will directly communicate with your underlying database, and your interface will become useless. So I think a lot of this is a new opportunity that appears after the ability of AI to a certain extent. There are many things that can be done in this and the probability is that it is a startup company. Today is the first decision you made. I think today is the first decision. There must be a lot of things to do. AI does what humans should do. But I think there are some things that have started to appear. When you have AI, I have AI. For example, the slug I voted for is actually a collaboration between the three of us. But the three of us may have five agents. Then the eight of us are agents in a channel to cooperate. You can give my agent a name, right? For example, you take a look at this file and give me a response. In fact, it is very similar to our daily work, right? But these are very original attempts Including what I started to say Agent is an agent-generated marketplace You have an agent, I have an agent Then can our agents do some valuable transactions individually? This is actually a very interesting attempt I think it's just very early now Including the mount book that was bought by Meta He said that everyone's agent went to post Of course, that still feels a bit cosplay It's actually cosplay Because what is posted is designated by its owner But if your agent can run a long-term horizon task for a month, he can go to the middle of the day and do something. It seems to be very normal. It's like going to a bar where foreigners often go. Gui Gu, the most popular topic of VC investment this year is New Labs. What do you think the topic of China is? New Labs is really popular because it seems that DD has made more than 60 New Labs in recent statistics. I think this is also a reflection of their observations. The first is that they think that the existing investment model companies are really big. So if you want to innovate, you need a more free and funded special organization. But I would say that Kui Gu is not only investing in New Lab. Because the main investor who invests in New Lab is a type of VC. If there is another type of VC, such as Benchmark, he will think that this company is too expensive. He doesn't want to invest in a company that is research-driven and has no clear direction. He would prefer to invest in a product-based company. In fact, you see YC's Demo Day, 70% is still various vertical SaaS. In fact, I think this is a bit like habit motivation. Because Kui Gu likes vertical SaaS. So many people use AI to do this kind of SaaS. It's just that the original human-dual code is now added to AI and some agents. The application level is more than just various vertical SaaS. Then there are many new labs in the field of partial research. AI research, AI for science, robot, etc. China, I think, will be much more popular in terms of hardware, right? This year we saw that the robots have been very popular for the past 12 months. Of course, now from the original human-shaped hardware to the robot brain, which is the world model. Anyway, the world model, of course, has many different definitions for this word, but there are many world models. For example, doing AI hardware or portable or said to be a bug or something, there are many. AI applications are also a lot. On the one hand, I think it's related to productivity. On the other hand, I can summarize AIQ time. I think there are a lot of words to be said. Of course, there are also many hard technologies. Quantum computing, text and language transformation are also very popular recently. Overall, I think we are relatively hard. On the one hand, because hard technology is still a bit of a problem in the entire capital market, including the results from the listing. On the other hand, many people still want to find the next word. This is also my feeling. He's obsessed with himself. Why didn't he look for SpaceX? I think I saw two people yesterday who were hand-picking rockets after the war. Actually, there are quite a lot of rockets. There should be quite a few commercial launch companies. Of course, China's commercial launch will be a more cooperative area that needs more cooperation from the government. Of course, there are 10 rocket companies. Among the top players should also have 5 or 6. Didn't you study in Guigu for a while? Do you have anything to share with us? Some of the directions of the future. I think the two biggest feelings are um Many people in Guigui want to go to the bank to steal old shares. Because if you want to destroy my work, I might as well become your shareholder. I want to get on the bus, right? Maybe Koreans want to go to Sanxinghai to do the construction work. This is the love of the Bipa. This may actually be a big impact on Guigui because it saw the potential career and the replacement. And I think the world model is indeed a very hot concept in the past six to ten years in China and the United States. I think in China, in fact, many times you will find that I have a lot of hardware, but this hardware has to be useful. So it has to be able to do the manipulation, to be able to help people really do things, otherwise I can't buy it at home and put it there. This is a practical perspective. From the perspective of the model, it is indeed a model of training through large-scale egocentric, which may represent the past with Ego Scale, which is to see how the hands are doing things. Video training model, and this kind of large-scale and insidious model like the Journalist, which is a model of training with data operated by a parent. There are several different routes on the robot, and there is a stability of the robot doing some tasks, and there are some principles on the flexibility of the tasks that have not been seen before. So I think in Gui Gu, everyone wants to put the language model of this mode to be parallel to the model of the robot. By collecting a large amount of data, and following a certain scaling law, we can train the robot to have zero-shot ability. For example, if I see a few-shot model, I will learn it. And improve its stability. So the speech I recently did today is about the process of the robot as a pre-training, fine-tuning, and I/O three steps to match the original model. If you are starting with a model, you will want to build up this match. So the world model is actually a very hot, very big word. But indeed, the world model is now also in a concept dispute. What is the world model? We don't know. Maybe everyone thinks that no matter what way it is used, its goal is to be like a language model. It is like predicting the next token. The world model is to predict the state of the next world. But what route do you use to go here? What kind of data do you collect? Or what kind of data do you scale? There are many differences in this. But I think this is also something that everyone is doing with a lot of effort. So the second is coding. The second is the world model. And the third is auto research. Because you will find that what you want to do is still a bit recursive. AI can be self-adapted and self-improvement. The process of AI improvement. Whether it's in scientific research or AI research, there are companies doing this kind of thing. The recursive of Tianwen Dong this year is also public. They seem to be doing self-improvement AI. This is also a very important direction for the future. Of course, I think China also has some companies that are considering this direction of scientific quality. About these three are the areas that I think Gui Gu has become very popular in the past six months. What's the difference with China? I want to talk about these three things in a more basic way. In the case of Gui Gu, it is rare to have a horizontal, or a level of common application. I will find that it seems that there is no forbidden product in the Mana 4.0. Of course, it may be Cloud Code and Codex. But in fact, you will find that there are many people using Probability. The ER of Probability has also reached $400 billion. And then Devin actually has more than $400 million. And then OpenCore, these apps are also used by a lot of people. So I feel that Gui Gu is actually very focused on models and verticals. In fact, I think there are still a lot of opportunities for horizontal. Do you personally invest in apps in China? I invested in Kimmy for the model. And then the rest of the companies are all app companies. Is the world model a model or an app? The world model is a model. You invested in two models. Yes. Three models. From the perspective of training models, yes. OK. Yes, you can understand it that way. Is the world model a machine model? It's a robot technology model. A robot-based technology model. You laugh because the stubbornness will be higher? I think the word "world model" is too vague. Because it has to be summed up with a hype word. But maybe because it's convenient to communicate, right? Maybe everyone says this is a world model. But every world model company, many times, the approach he talks about will be very different. For example, Xie Sai Ling would think that these virtual models are not virtual models. But for example, is VLA a virtual model? You can also think that it is a virtual model. But Sai Ling might think that there are so many languages in it, this is not good. Sai Ling also thinks that video generation is not a world model. But now some people don't think so. So I think there is still an open problem in this. There is no chain. Will there be a lot of domestic hardware opportunities? What AI-native hardware products do you think are good? You mean the kind of portable recording equipment? Not the kind of robot equipment, right? Yeah, the consumer electronics. I actually have a discussion. I think many consumer electronics companies will repeat the same process of new consumption. If you remember, the new consumption was very popular in 2020. At that time, all the categories were made at once, from skin care products to cream to toothpaste to everything, but later on, many people found that it was not a traditional category to make some small innovations, and then find a live broadcast on the Internet, find a product to bring goods, it became a completely different new category. I think there are many AI hardware now, in fact, it has not solved some fundamental problems. It may add a chatbot to the hardware, or some AI sensing capabilities, communication capabilities. But I'm a little worried about whether or not we can create a new demand category. And in general, it's harder to make hardware than software. Because hardware has a supply chain, it needs to be opened, designed, and delivered in a slower way. It needs more capital and use. It's more troublesome to sell. So hardware is actually more difficult than software. In fact, many people are discussing it now. For example, is there a special machine for A-Series? Some people also sell lobster machines, right? Special lobster machines. But in fact, I think why did everyone buy a Mac mini or a mobile phone? In fact, existing computers and mobile phones are quite balanced in terms of size, performance, power consumption, and heat dissipation. So I would rather think that in a relatively long time, the good body of agent is actually the mobile phone and cloud. 因为现在有很多 可能几十个各种可穿戴 然后录音 然后比如说给你更多的context 或者这些应用 我觉得在这里面 所谓做可穿戴的 其实都会面临问题 就是说你要把一个东西穿戴起来 真的是一个很大的一个 就是你要很坚持的习惯 尤其是你别在衣服上什么的时候 你今天换件衣服 你要把它又要别上去 这个里面 它得跟你解决很大的 - So I have some cautious views on this area, especially the impression of the wearable belt. No head? We can say that there is basically no head. OK, like Lucky. Anyway, we don't have heads. Not to mention that there are some companies, but I don't think we really see Of course, we are from the perspective of the head of the company. I think it's not easy to do from the perspective of the market. But on the contrary, for example, I have this whoop in Aura. They are actually a bit like AI. But it's not actually an AI hardware. It's a health hardware. I think it's really useful. If it wasn't for the AI I use, I would have made a lot of money. How much did you invest in the robot? We invested in the main company of robots, such as 5G's Lin Xiaoshou and Fangzhou Robot. Two world-class companies. We also invested in Feixi before, and it was divided into the middle. But we didn't invest in the humanoid robot. But I think the humanoid robot is still in a very early stage. It's mainly scientific research. And then what is the story he told? I think it's useful by manipulation. That is to say, it can be realized by manipulation. So I think it is indeed a model. No matter who makes it or how to do it. But the ultimate goal he has is that robots can control these things. He needs it very much. But I think human robots are definitely in a very exciting stage of capital. So his valuation is relatively high for large sales. I don't know if this can continue to be so high. But I think at least we can see whether human resources are useful or not. I think it's still in the process of development. I think human resources may be useful eventually. But it needs to have good hands. So we stole hands. It may need a good arm or a good mechanical structure. It also needs better models. So we also stole models. We must have missed a lot of opportunities. But I think we are also observing. What was the most popular project last year? I think it's the follow-up project. Because everyone is more formal. and there are many information flows that become very convenient. So it's really fast, for example, VC pays attention to the field or the field he likes, he becomes very consensus. So we will see, for example, many do various things, of course, that can be passed on to recording, right? Because everyone will say, because agent needs a lot of context, so I will record the context of the recording. and then there are many people who do various fields of medicine. I think we want to vote for the first-tier. We hope that he has a different view. He does something different. Even if this thing sounds very strange, it's not that this thing is already a track. For example, it's called the record track. For example, the plug is actually very powerful because it is a product model. But later I found that many people also do various plugs. It grows in various shapes and sticks to the back of the phone. I think that way it becomes a follow-up. This is a bit of a follow-up. So This is one of the reasons why we passed the test. We saw that it was really because of the fire. There was a VC who talked about it and a broker who talked about it. So it was just a better product. So can brokers make a judgment in investment? I think it's definitely a very good channel for information transmission. So many people listen to Xiaojun's blog. I think everyone is coming to the conclusion of the front page faster and faster. Will you finish listening? I actually don't listen to blogs. You are all turning... I'm reading. You are turning into text with an app. Yes. Oh, you think it's more efficient? For some very good ones, for example, the one you talked to Sally about, I just listened to it. You listened to it for seven hours. Why didn't I believe it? I listened to it over and over again. Because I like to read more than listening. I don't drive, I don't cook. I don't have any common scenes where my hands can't reach out. Now I have my own AI app. So the bloggers I'm interested in will check every day if there's a new transcript. Then they will automatically pull it down and put it in my Notion. So I go to Notion every day to see if my AI has made me a new blogger's text. This text is so long. Yes, it's very long. Can you finish it? It must be shorter than what I heard. I can finish it. For example, your two-hour blog, I definitely don't need to read it for two hours. You are indeed a person who can read 100 books a year. This should be like this. For example, your seven-hour broadcast, I should not need seven hours to read it. What have you changed in the past six months? AI has developed so fast. What is the change in the investment? First of all, I did a lot of small tools with agent-based coding. I used to be a person who didn't know how to write code. Then I used Cloud Code at the end of last year. I found that it could do 80% of the tasks. But it is not as good as the online posting. I use it every day. But I found that the completion rate is completely different in the early years of this year. Of course, a lot of things I do now are still about getting all kinds of information. For example, my blog, my newsletter, my meeting notes, and all kinds of content I write. I have several passwords to organize them together every day. Then I organize a daily newspaper for myself. For example, the person I meet every day. Maybe there is a person I meet three times in three hours. Maybe the notes on his content will be organized under this person's bar. and build a long-term tracking system for a person, a company, and a stock. I think this has really improved my acquisition and collection of information. So I think AI is indeed a great improvement in the investment industry. Because the investment industry is indeed looking at the information and analyzing it. Then I have been thinking about the problem lately. More and more naturally, we put a lot of thinking outside the bag to add LGBT or AI. For example, if there was a problem before, maybe everyone would go to understand it and think about it. Now many people are asking AI, right? What's going on? I think this is a lot of benefits for you, but there may be a lot of problems. Because if you get the answer directly and you don't think about it, your quality of answer and how you think about it later may be relatively... Your brain hasn't changed, your whole body hasn't updated, so you may not really understand this thing. I later thought that if a person has been riding a wheelchair, his leg muscles may be weakened and shrink. If a person uses AI to bring this kind of so-called thinking for a long time, in place of his own thinking, then this thinking ability may also be weakened. So you see, we used to export physical labor to machines, so we are now estimated to be a thousand years younger than humans. Many people may be weaker physically because we sit for a long time or in the car, right? Then we export knowledge and memory to the Internet. So now a lot of things, but I can't remember things, but I'm going to search. Now we are thinking about exporting to AI. Andrew Kibasi said one thing, he said you can export thinking, but you can't export understanding. But I think thinking may not be completely exported. I thought about it later. You see, we have a gym to exercise. I'm going to exercise deliberately. Maybe when you think a lot about VivaLog AI, you may need a gym for your thoughts. You deliberately think. For example, in order to prepare our e-book, I also have a point of view on many things. But this point of view has no arguments. It's actually a process of forcing me to think. When I think, I don't let AI summarize it. I can talk more clearly about things. Thinking requires deliberate practice. And innovation is also something that can be practiced. Because there is a very obvious characteristic now that young people have a lot of ideas. But as a middle-aged person who is almost 40 years old, your ideas may be much less than when I started my business at the age of 20. I feel like I'm now using the subscriptions of all major coding agencies. There are a lot of AI waiting to work for me every day. What do you want to do? In fact, I think this is a problem that many people encounter. I'm thinking that now there are a lot of AI programmers waiting for you to drop the money. But But what to do? Including myself, I have this problem. Maybe it's not a problem, but a problem. This also needs to be practiced deliberately. For example, maybe do some simple web apps first. Right? Sync my meeting, sync my schedule, sync my subscription, and start. I think this is not a matter of one-off. It's hard for you to write a complicated application as soon as you get up. You don't know what to write. But maybe you can start from a simple one. Step by step. It's actually the same as fitness training. Start from simple, then stretch, then do push-ups, and then gradually become complicated. So I feel that in the era of AI, learning to use AI is actually a deliberate practice. For many people, they don't know what to do with AI. But I think this matter can be gradually improved. I have a friend who doesn't want to reveal his identity. Maybe it's Mr. Lin. He has to burn a thousand dollars of tokens every month. He is very hard to solve various problems with AI. What did he even do? He made a scanner. Scanning which people are out of date every day and no one is registered. Then he would see what the AI registered back today and what the domain name was. Then he would see what the domain name was made for. I said, "You're a bit like Omar Khasee, right?" "I went to the fish market today and saw some fish, so I made some dishes." He was a bit like forcing himself to do it deliberately. For example, he bought a cat-related thing last time. He made a real-time monitoring system for his cat at home. This thing is very fun anyway. Because I admire him very much. Because I think he is deliberately practicing his own AI innovation. What do you think he does? For example, my family, including my wife and my parents, all have smart keys, such as OOP, AURA, and I made a dashboard to put all the data APIs together to compare and see. And recently, I, Mr. Lin, and our other friends are using the SLOCK we voted for to create a tool that encourages friends to compare their health data. And their original goal was to make me lose weight. They said, "I want to make a dashboard that shows everyone's sleep well. So you have to sleep more." I said, "Okay, I'll develop it together." So we also used agentic to collaborate. But all of these things don't seem to have any direct value. But I think in this process, for example, we know how difficult it is to collaborate with agents. In fact, I think it's a very interesting experience. How do you think society should digest the changes and impacts of this technological change, including unemployment? I think unemployment is really inevitable. Because the technology is expanding faster now. It will be faster than many people learning new technologies and adapting to it. So I think from a personal perspective, learning AI, thinking about the ability to use AI, and the safety of your profession are actually very important. Because it's harder to test who can handle AI better. But at present, it is true that AI destroys value faster than creating value. When work is replaced quickly, the company will become more efficient. But for me personally, one is learning AI, the other is in the field where AI is more difficult to replace. For example, things that are talked about between people are actually difficult to replace with AI. Because it requires trust between people or people to take responsibility. In this case, I guess it's hard to say that because there will be unemployment, AI will not develop. Because no one can afford the price of not developing. It may indeed be better with some kind of UBI or tax form to distribute to a certain degree of interest. Recently, someone in South Korea has been proposing that Samsung and Hades make too much money. In fact, this is also a move for social stability. But I think what we see now is the first-hand or direct impact of AI. AI can write programs, so programmers may lose their jobs. But in the long run, I am very optimistic because we look at it from the point of view of new work. It takes time, but it will occur in a lot of places you don't want to go. I gave an example before. The industrial revolution started with the invention of steam engines in order to draw water from the mine. Because the coal mine was getting deeper and deeper, the deeper it was, the easier it was to get water, and the water was easy to collapse, so it was very dangerous. At that time, steam engines could only do straight lines, so they could only draw water in this scene. Later, Watt actually invented the cooling system and the propeller, and he turned this network-referral movement into a round-trip movement. So steam engines can do more things, especially to drive the steam engine. At that time, the invention of the Stirlitz steam engine was the first step from the invention of steam engines to the development of the steam industry. After the progress of the fashion industry, everyone has changed their clothes. This leads to a demand for color. Because you didn't need to distinguish colors when you made a few clothes But how to distinguish colors when you have dozens of clothes The dye at the beginning was natural dye For example, various colored rouge grass, this kind of worm, etc. Very expensive, very little production For example, purple was the most expensive color Because purple is relatively rare in nature At that time, the chemical synthesis dye process began to develop Now the chemical dye factory like Bassefou and Bayer In fact, it was a dye factory at that time It's a human-made dye factory Of course, it can't be a direct comparison. But historically, the development of the textile industry has been greatly affected by the growth of the dye industry. The dye industry also brought about the birth of the chemical industry. After the birth of the chemical industry, later the chemical industry produced gasoline, plastic, and fertilizer. This series of very important chemical industry products. It also brought about the improvement of the production of agricultural products in humans. Then the plastic brought about many new applications. So with the invention of a steam engine, many of these things spread. At first, we didn't expect that the real estate industry would bring about an improvement in the housing industry or in the chemical industry. This requires time and a lot of companies to find the right time. When the ability to supply is greatly improved, it will bring about a new opportunity for production. So there will be a lot of new work in this. So I think in the long term, there is no doubt that new work will be produced. But it will take a little time to spread. In the process, maybe someone wants to learn to use AI. Another aspect is that I want to explore what new opportunities AI brings. But the current problem is that the change is too fast. Indeed, the industrial revolution is a few decades of time. But AI is now one year. Everyone feels that writing a program has really been replaced a lot. So maybe the degree will be bigger. Whether it's investment opportunities or thinking about society, thinking about people and their professional relationships, there will be a big change. The biggest change may be how humans think about the relationship with AI. I don't know if you use Cloud, will you have a strong sense of trust and dependence on it? After it can help you do a lot of things well, you will naturally trust it. Then you will have a cooperative relationship with it, because you have done a lot of things with it. At this time, I can't help it. It's more. Yes, you will be more in a very hard way. Then this kind of dependence will actually make a lot of changes in your life. We actually didn't prepare this at all. But in fact, many people, you see, I observed it with crayfish at that time. Many people actually did not do any important work at that time. Sometimes it's like a training program. I found that he remembered a lot. Then he can help you do a lot of things. You have your memories and can help you do a lot of things. At the same time, OpenCloud also has a feature. It's that the file is written in a more human way. - I think a lot of people have been asking a question lately. We don't give any suggestions because there are many things we have experienced in the past. Many people say that my child should learn some major in the future. What do you think? Because my child just went to kindergarten. I think this is actually a very difficult question. I think it's really important to train his agency. In my blog two years ago, I said that people will have the taste of agency. This is what many people say. You have the taste of working in AI. But I realized that it's hard to test people. AI is so smart. It has seen so much good data. Its test is definitely far superior to that of ordinary humans. So I realized that TEST may not be a special place for humans. AI-made PPT is very good. I saw it. So I don't think TEST is reliable. Maybe for now, it's still an agency. No matter how well it does, what you do is what you tell it. So I just said that innovation should be deliberately practiced. Do what you do deliberately. Many people, including myself, if you ask me what I want to do, Actually, you may not know. You may not know what you are going to do. But if you think about it, right? You may find that there are many things that can be better. I think this is still a good question to ask. If someone answers all your questions, but you know you have to ask questions. You need to have problems that you want to solve, problems that you want to change. But now the next step is also Proactive Aging. Proactive Aging means that I need to have stronger AI. I'll help you. Yes, at this time, it may be more challenging. So I thought later that if the current AI is based on the current model structure, it can solve the data that humans already have, that is, within distribution. But humans are interesting. There are a few humans who can create things out of the distribution. They can create things that humans have never seen before. You will find that AI can't tell a original joke now. It can only refreeze a joke that humans have already told. But it can't tell an original joke. So this is interesting. Why can some people create OOD content? For example, Jaloha can do a group theory at the age of 20. Why? I think there are many studies on the human brain, including Yang Le Kun, or there are many Saturn. For example, the intelligence of the rat, we humans have not learned yet. The intelligence of the cat, how to explore the world to form a process of self-learning and evolution, this may not find a good answer. So from this perspective, if you do things, your ability is all in a human distribution, a center of political distribution, it is actually easy to be replaced. But if a person can produce content that is out of distribution, it can create content that has never been created before, then this may not be easily replaced. But this may be curiosity, creativity, and how to reflect these in education. I think these are all very big topics. But I think that in human history, people have been constantly accumulating the real important abilities of mankind. For example, after the industrial revolution, human physical strength became less important, and their brain power became very important. So at that time, people started to study to change their fate. At that time, people started to pay attention to education. Before that, if you didn't have education, you had physical strength. But after the Internet was invented, I think it separated human memory and ability. Knowledge and ability. Because knowledge can be found. knowledge has been replaced by the ability to do things, such as writing codes and doing things. But after AI came, I think it separated the ability from the execution and judgment. First, the execution power did something, and this became more and more AI-based. But what do I do with judgment? and TASTE. But maybe TASTE will be replaced by AI if it is done correctly. Because AI can go to the lab to get a lot of rewards. So what to do? This action will become at least at present, AI can't move by itself. I think this is what I think. But I have no answer. I think it seems to make a person more agency, including myself. So I say start with more AI, do more strange things. Will you continue to record when you are often hit in the face after recording? I told Mr. Wa that I would not record it next time. Because it would limit my ability to make a flag. Mr. Wa immediately said that he was convinced by me. Because if you don't record it after you feel that you are hit, you are really beaten. So I think to keep summarizing my thoughts is a process to let myself think about some unthinkable problems and maybe think about them more clearly. You can practice. You can practice. Because when you express it, you will find that some places you know very clearly that you understand or you think you understand. But some places are confused. This is very clear when you say something. So I think this is also a process of my own organization. I think it's very valuable. And I still think that as an early investor, being hit on the face is often happy. Enjoy being slapped. So will we continue to do this? Yes, we can. Next time, we may be hit again. I don't know what other guests will think. Anyway, I feel this is normal. Because early investment is really a habit of being hit. I think they have also accepted it. 好了,今天的节目就是这样。 这里是商业访谈录,是一档由语言及世界工作室出品的深度访谈节目。 你可以到公众号关注我们的工作室获取更多的信息。 我们的公众号是语言及世界,languageisworld。 我们希望和你一起从这里探索新的世界。