NetEase Technology launched a series of dialogue column Attitude AGI. In the past three years, AI changes have been surging, the global scientific and technological order is being reconstructed, and the road to AGI may be approaching quietly. With AGI as the topic, this column will talk to 100 AI experts, entrepreneurs and investors to try to break the clouds and witness the arrival of the AGI era. Phase II Dialogue: Founder and CEO of Baichuan Intelligent Wang Xiaochuan 。
Produced | Netease Technology's Attitude AGI column
Author | Cui Yuxian
Edit | Ding Guangsheng
On the way to AGI, a super model+super application dual wheel drive is required. On May 22, 2024, Baichuan Intelligent, which has been established for more than a year, released the latest generation base Large model Baichuan 4, And launched the first AI assistant "Baixiaoying" after its establishment.
Wang Xiaochuan, the founder and CEO of Baichuan Intelligent, said that the model technology changes very quickly. Compared with Baihuan3, the ability of Baihuan4 is improved. The science of Baihuan4 ranks first in China at present in SuperCLUE. The liberal arts has obvious advantages and is in the best position in the world.
Baixiao should be the first application launched by Baichuan Intelligent, positioning AI assistant. Compared with other assistants, Baixiao should not only make progress in the field of original voice interaction or picture recognition and voice broadcast, Wang Xiaochuan believes that Baichuan Intelligent has put forward a more important concept: understand search and ask questions. However, Wang Xiaochuan also said frankly that the app released this time is still a long way from the super app.
As for the hot big model price war, Wang Xiaochuan said that this is the range of big factories and is only limited to the actions of cloud manufacturers, so start-ups will not follow.
"In my opinion, first of all, we are too optimistic about the future of this era, and would rather not lose any opportunity, and would rather be involved at zero price. This side reflects that we have enough vision for the AI era. In addition, price reduction is only an action of cloud service providers, and for start-ups, it has gone out of the way."
Wang Xiaochuan said that Baichuan Intelligent would not be involved in the price war because To B is not the company's main business model.
The following is a part of the dialogue between Wang Xiaochuan, founder and CEO of Baichuan Intelligent, held by NetEase Technology and other media (sorted out):
1. Super app: It's too early to consider revenue to solve user pain points
Question: In fact, there are quite a lot of applications on the market now, especially those issued by many large manufacturers last year. Will it be a bit late for you to send them now?
Wang Xiaochuan: I think the opposite is true. It's early. When we launched the application, we still need more time to polish in the model application today.
A model comes out of the market. If today is a million DAU, it is far from a super application. I think the various applications that you have published before are just models. We think that we have enough experience, such as input methods, search engines, and browsers. We have a deep understanding of how an application can become a widely used product, but now no one has.
I think that the "100 Primary Schools should" released this time is far from the image we want, but in this industry, we also need an opportunity to post it, so that the industry has a basic understanding and the team can turn around. This is my answer to the first question. It is not a question of whether we are early or not, but that the whole industry has not yet reached a mature state.
Question: What is the expectation of super applications?
Wang Xiaochuan: The first sentence just came in that the future is called super model plus super application. Now it is neither super model nor super application. If it goes well, there will be major upgrade changes this year. If it comes to super applications, I think the number of users will increase by several orders of magnitude. I think these two increases are not only our own, but also the situation of each industry today. From the perspective of super applications, they have not been seen yet, But we have this sense of image in our mind, especially after the year from April last year to May this year, the team itself has enough touch. We have done language AI and super applications before, and this time we have a hope for what the future will look like.
After issuing this model, we can really move towards the "two wheel drive" mode in rhythm. When we move forward, there will be milestones. We must always have this "1" step out, so that we have experience and today in the water test, we can integrate the two kinds of information together, so that the team can walk more healthily.
Question: Just now I mentioned super applications, and I mentioned the data of 1 million DAUs. Can this be defined as super applications?
Wang Xiaochuan: No, I think super applications should be improved by two orders of magnitude before they are called super applications. These two orders of magnitude change from 1 million to 100 million. If they rise or fall 3.3 times, they will get between 30 million and 300 million. I think this is called super application.
Question: Are there any now?
Wang Xiaochuan: Now one million people think it's very nice. It must not be.
Question: What is the breakthrough point of super application?
Wang Xiaochuan: I think the first thing that super applications can break through is in the profession. Add the data density and the thinking degree of the profession. Today we did some basic homework. We saw that we can access and search, and we are doing some component level work. If we talk more, we will know where our direction is.
Question: Is there a more specific scenario for "100 Primary School Students"? Will we have some specific scenarios?
Wang Xiaochuan: It's not a new concept to use it in the workplace today. When GPT was released, it found that there would be a lot of things in the daytime on weekdays, so it would be used more in the daytime. But I think that in today's stage, the next stage is to play as professional roles. Further down, more professional roles are required to cover more scenarios for each role. Today is not the end of the service scenario. In my opinion, the service direct scenario is a sentence without any information. Therefore, when we use it today, we think it will naturally become more professional. For the example just given, one is that career scenarios will be used, and I still think the future is in the three words I just mentioned: Creativity, health and happiness, so we think that creation may be related to your work in the workplace. Going towards health and happiness, I think it is a greater application direction.
Question: With regard to the commercialization of the launch of "Baixiaoying" today, the existing products on the market, such as ChatGPT and Wenxin Yiyan, have a paid version. Which model is used by "Baixiaoying", and will it take a similar commercialization path in the future?
Wang Xiaochuan: Not only are there other considerations, but neither of the two mentioned is under our consideration. Including the words of Wenxin you mentioned just now, I think it doesn't earn as much as Baidu's library. I write something to sell to you, which will earn a lot more. Today's API revenue and the revenue of the model itself are not called successful practices. It is not the case that we want to catch up and analogy. We think these two things are not the sexiest business models today. Neither Wenxin nor ChatGPT.
Question: What is the sexiest business model in your opinion?
Wang Xiaochuan: You have super applications that can solve the pain points of users. I am undoubtedly optimistic about the income, but it seems that it is really meaningless to collect money now.
Question: Just now, we talked a lot about super applications. Do you think ChatGPT is a super application?
Wang Xiaochuan: I don't think so.
Question: What do you think the gap is?
Wang Xiaochuan: I think ChatGPT will have a chance with its current technology if OpenAI is diligent enough to super apply it. But because of the company's own genes, it does not invest too much energy in the product direction.
2. Big model price war: Baichuan Intelligent does not follow up
Question: What do you think of the giant's price cut? Is this scene familiar in the previous business war? How does Baichuan deal with it?
Wang Xiaochuan.
Second, I think the core of price reduction depends on what your business model is, If you are doing ToB services, you will sell a whole set of cloud services instead of the model itself at the end of the price reduction, so cloud manufacturers are more traditional service models and enter a new battlefield. Therefore, the price reduction is limited to the actions of cloud service providers. This thing is also quite familiar. Like the previous four dragons when they were ToB, everyone became like this. The limited scenario of price war, in fact, for start-ups, has gone beyond their range. Why do you pay special attention to price reduction today? We are more sensitive to this. In our view, it is not to participate in the price war, but not to take such a business model as your key position. If the business model is based on the service concept of API disclosure, in this case, this is not the responsibility of a startup company itself, nor is it in our concept of two wheel drive.
So first, I think this price cut is too optimistic about the future. Second, I am really familiar with it. I think this price cut is the same as the previous one sound of dripping water Meituan It's not the same, because the price war or subsidies at that time had a network effect, which was a bilateral network. The business model was changing the production relationship, the relationship between drivers and passengers, and the relationship between delivery staff and users. So that is to establish a bilateral network effect, and this price reduction is more like the price reduction method of the Four Dragons on the B side than the price competition like Didi Meituan. I think this is not like the C side. This time, it is not a change in production relations, but a direct supply of productivity and AI to productivity
I think for us, this thing is to stay out of it.
Question: If B-end developers develop some new applications through the general big model after price reduction, will the competition on the application end be significantly intensified?
Wang Xiaochuan: Now neither company can compete like this. One is not within the range of the factory. Second, in today's situation, we quickly want to make money with this model, and some small applications will rise, but not in our main direction. We don't want to compete with these smaller start-ups now. Each has its own niche, because they are not super applications. They may want to be small applications in a vertical industry.
Eating melon to watch the activity is the same as everyone's mood.
Question: President Sichuan mentioned just now that the release rhythm of Baichuan's new models will be extended to every quarter. In Baichuan's view, whether domestic or foreign companies are facing greater challenges or bottlenecks in the technology of large models, and whether frontier exploration is becoming more difficult?
Wang Xiaochuan: Last year, we put forward a concept that the ideal is one step slower and three steps faster. This is a reference to the United States. In fact, in the process of model exploration, the breakthrough from 0 to 1, the capital reserves of the United States and the culture of their scientists will be much stronger than that in China. Therefore, the cost of exploration will be very high. It can be simply said that every major technological breakthrough is led by the United States, so they have achieved it. We will follow them again. In terms of models, there are GPT4 We will follow after the invention of GPT4o, which can greatly reduce the investment in domestic costs, so that we can focus more on how to improve the user end. So the difference in China is to balance.
My basic judgment is that, with the same generation model, Baichuan has a better chance to make super applications. If the model crosses generations, it is the same thing, but it is our competitive advantage in the same generation model. We have chosen our own track before, and next time we may talk about how we can get closer to such super applications from today's model, which can only be done after the model has been improved to a certain extent, so we have our own application side reserves, which I think is China's practice of moving forward from applications behind the United States.
The bottleneck of the model, as I heard, at least Sora and GPT4o do not represent the biggest breakthrough in my mind. The legendary GPT5 may hear some delay in the United States, but it can be seen by the end of the year.
Question: You also mentioned that Baichuan may pay more attention to text ability than multimodality.
Wang Xiaochuan: I think it's more about intelligence itself. I mean multimodality without intelligence.
Question: If the long text rolls to infinity, what direction can it develop?
Wang Xiaochuan: I also have the ability to reason. I can do math problems, or I can think more about them. Today's model is not self consistent enough. The rigor of model thinking and good enough logic are all more important things. I don't think model length is its only thing. Now China also knows that model length text is very important, but it is only a necessary condition for growth, and not a sufficient condition. In terms of sufficient conditions, I think the model itself should emphasize the internal consistency and its reasoning ability. These two things are not solved by long windows.
This is the second issue of NetEase Technology's series of dialogue column Attitude AGI. More content, follow us.