Still casting big models？ The venture capital circle is waiting for Li Yanhong’s AI native application.

　

With the continuous development of artificial intelligence technology, entrepreneurs and investors in the AI community are looking for the next explosion point. However, they found that although the big model has made remarkable progress in technology, it is still out of reach in commercial realization. At the same time, the big model is a "gold-swallowing beast", the development and maintenance costs are extremely high, and people have become accustomed to the phenomenon of burning up hundreds of millions of yuan in financing in three months.

On the contrary, "AI native application" is the "golden beast" in the AI era. As Li Yanhong, founder, chairman and CEO of Baidu, has repeatedly emphasized, "AI native application is the real opportunity for entrepreneurs and entrepreneurs. The application has more obvious promotion effect on business, more room for innovation and more direct liquidity."

So facing the AI wave, what is the current investment attitude of the domestic market towards the big model? What is your attitude towards the innovation and development of application layer? How to do a good job of AI native application?

01

The big model is a "golden beast", and the investment is getting cold.

Since the release of ChatGPT in early 2023, it has ignited the enthusiasm of large models at home and abroad, and various funds have flocked. According to the data of IT Orange, as of September 25th, 2023, there were 73 domestic AIGC investment events with an investment amount of 8.295 billion yuan.

According to relevant media reports, the number of pre-training models on Hugging Face, the world’s largest large model open source community, has increased from 100,000 to more than 300,000.

As far as China market is concerned, according to the data released by Beijing Economic and Information Bureau, as of the beginning of October 2023, the number of large AI models published in China has reached 238.

According to the slow-motion observation, after the popularity of the first three quarters, it actually entered the fourth quarter, and the grand occasion of "everyone investing in a big model" in the past is no longer there. The media and venture capital circles frequently come up with prejudging conclusions such as "big model investment is getting cold", "just watching but not investing" and "no company can invest".

Founder Hao, the managing partner of Xiaomiao Langcheng, once said in an interview with the media: "At present, the investment opportunities at the model level can only continue the game among some players with strong capital. For investment institutions whose management scale is not particularly large, if they are not laid out as soon as possible before the industry is hot, they will not participate in the investment of large models at the current time. "

Why "no one throws a big model"? There may be three reasons behind the analysis.

First, investors began to realize that there are high technical barriers and cost thresholds to develop large models that are easy to use and available.

In a word, if you want to make a big model, you have to spend money like water. As a "pathfinder", OpenAI began to "burn money for a long time" without thinking about how to commercialize it. In April this year, according to media reports, the GPT model has 175 billion parameters, the training cost is 12 million dollars, and OpenAI burned 1 billion dollars. Seven months later, the input cost is even more incalculable.

The reason why Baidu can make a big model of Wenxin is also because it has invested almost 150 billion yuan in the chip layer, framework layer, model layer and application layer of artificial intelligence in the past ten years.

In addition, objectively speaking, under the hot appearance of the big model, there are still some technical gaps at home and abroad. Judging from the parameter scales of billions and tens of billions of language-based models of major domestic manufacturers, and the fact that most start-up companies are still making parameter models of billions and tens of billions, there is still a distance of 2-3 years compared with ChatGPT.

Second, the law of commercial competition is that whoever is efficient wins. What investors want to see is, under the same effect, whose reasoning cost is lower? Or at the same cost, whose effect is better?

Li Yanhong said: "’Working hard to make a miracle’ refers more to the process of exploring a large model from 0 to 1. Going back, this game is not this game, and it is more opposite. What is the opposite? It is the same as all commercial competition laws, that is, whoever is efficient wins. You can raise money, and I can also raise money. In the end, I use 10 yuan to make 100 points, and you use 10 yuan to make 120 points. Over time, you win. In other words, in order to make the effect of 100 points, I use 100 yuan, you use 80 yuan, and you win. "

Take Baidu as an example. Since it was released in ERNIE Bot in March, Baidu has reduced its reasoning cost to 1% of the original. It only dared to call 10,000 times, but now it dares to call 1 million times a day. This is because Baidu has layout in chip layer, framework layer, model layer and application layer, so it can be optimized end to end.

Third, the basis for investors to invest money is to see if there is any income, and the big model itself does not directly generate value.

First of all, as a general tool or platform, the big model provides basic ability and technical support, rather than a solution to a specific problem directly. Just like a hammer can be used to beat all kinds of things, it does not produce any specific products or services.

Secondly, the value of the big model lies in that it can be applied to actual scenes to solve specific problems. Only by combining the large model with specific application scenarios, through customized development and optimization, can its real value be brought into play.

Finally, the application of large-scale model requires professional skills and knowledge, including data preparation, model training, optimization and deployment. These tasks need to be completed by professional teams, so the application of large models also involves the input of manpower and resources.

In short, the big model itself does not directly generate value, but needs to be combined with application scenarios and professional skills input to play its real value.

So how to judge the quality of a big model?

"To be application-oriented, running points and brushing the list don’t count. Now many teams don’t know what is good and what is bad, but it is very unreliable to rely on third parties to score. There are more than 200 large models in China, many of which are on this list and that ranking, but there is actually no usage. I am afraid that the number of calls of Wenxin Big Model is greater than that of all the 200 combined. " Li yanhong said.

02

The native application of AI is the "golden beast" in the AI era.

There are many big models in China, but few AI native applications are developed based on them.

As mentioned above, as of October, 238 large models have been released in China, compared with 79 in June. In other words, the number of four-month-old models has tripled. But few people can name one or two native AI applications in China.

On the other hand, in foreign countries, besides dozens of basic models, there are thousands of AI native applications, which are not available in China market.

"The symbol of human beings entering the AI ​ ​ era is not to produce many large models, but to produce many AI native applications." Li Yanhong introduced that the big model itself is a basic base, similar to the operating system, so the final developers have to rely on a few big models to develop a variety of native applications. Therefore, it is a great waste of social resources to repeatedly develop the basic big model. "In the AI ​ ​ native era, we need 1 million AI native applications, but we don’t need 100 big models."

Why should China focus on the native application of AI? Why is the AI native application the "golden beast" in the AI era?

First, it can directly promote the business, especially the original biochemical transformation of AI for existing products; The AI native application built on the basic big model can affect the key indicators of the business and bring revenue and profit.

Microsoft is a very worthy example. It is a company that doesn’t make a big model, but it has made Copilot, the most successful AI native application at this stage, and has started to charge users at a price of $30 a month. At present, the market value of Microsoft has reached 2.8 trillion, five times that of OPEN AI.

Even domestic investors have set their eyes on the investment in application-layer products. For example, Bai Zeren, vice president of linear capital investment, once told the media: "We are also very concerned about the progress and changes of the big model itself. Considering the current market competition pattern and capital threshold, we will tend to invest in the application layer and new infra and other opportunities. I am more concerned about how the new technology landing industry can solve industrial problems more effectively and bring great commercial value to the industry. "

The second is a more direct way to start a business. Unlike big models that burn money, they have their own liquidity. Most of the generated AI products are paid.

In early November, Baidu launched version 4.0 of Wenxin Big Model, and tried to charge a subscription fee. At present, the number of paying users keeps a relatively high growth every day, and many users are willing to pay for Wenxin.

The same business model has also run abroad, including Microsoft’s copilot, and the annual revenue of three kinds of products abroad has exceeded 100 million US dollars: image generation, copywriting and code writing. In addition, in February this year, OpenAI also launched ——ChatGPT Plus, a paid subscription version of ChatGPT, with a monthly fee of $20.

Third, the real prosperity of new technologies must be the prosperity of applications, and the era of PC and mobile Internet is the best example.

For example, in the PC era, there was basically only one operating system, but there were many softwares developed based on Windows. In the era of mobile Internet, there are only two mainstream operating systems, Android and iOS, and there are 8 million mobile applications.

The big model era is very similar to the development of the above two eras. The big model itself is a basic base, similar to the operating system, so the final developer will rely on a few big models to develop various native applications. Therefore, it is a great waste of social resources to repeatedly develop the basic big model.

"In the AI ​ ​ native era, we need 1 million AI native applications, but we don’t need 100 big models." Li yanhong said.

03

In 2024, the year of opportunity for AI native applications?

Although domestic investors are concerned about AI at present and dare not invest in the application layer, it is impossible for them to let go of this trend opportunity, so the AI application layer will also be laid out in the future. It is worth noting that after a year of development and concentrated research, more and more investors have the ability to identify large models, and the living space of "shell-type" applications will be less and less.

Some people in the market have judged that 2024 must be the year of opportunity for the native application of AI. The author also agrees with this view, and analyzes the factors behind it mainly in three aspects.

First of all, China has a leading basic model, which provides a solid foundation for the development of AI native applications. For example, at the World Congress, Baidu Qian Fan launched its own "App Store"-AI Native App Store, which provided a trading platform for applications developed based on Wenxin Big Model. At present, the application store has launched AI applications developed by Baidu and customers for five scenarios: smart office, marketing services, industry functions, production efficiency improvement, and analysis and decision-making.

Secondly, similar to the emergence of mobile native applications such as WeChat, Tik Tok and Uber in the mobile era, excellent AI native applications will also be born in the AI native era. These applications will be developed based on the basic big model, with unlimited innovation space.

Finally, with the continuous progress of technology and the continuous innovation of applications, AI native applications will be applied in more fields, such as intelligent customer service, smart home and autonomous driving. These applications will bring more convenience and intelligence to people’s lives, and will also promote the further development of AI technology.

There are three key factors to develop good AI native applications: First, there are relevant industrial policies to encourage the development of AI native applications based on large models.

"China’s leading industries are often prescient in industrial policies. For example, solar photovoltaic, power batteries, new energy vehicles and other industries have been promoted by favorable policies. From January to August 2023, the global sales of new energy vehicles reached 8.23 million, and China’s new energy passenger cars accounted for 61% of the global new energy. If China is a fuel vehicle, it will curb demand, and it will also pay vehicle purchase tax if it is restricted and restricted. However, new energy vehicles have no such restrictions and naturally develop better. " Li Yanhong explained.

Second, the existing enterprises use the big model to have a positive effect on its business core key indicators. In this regard, Li Yanhong’s idea is: "This degree of attention needs to be raised. It’s easier said than done. In fact, the response of big companies is very slow. Even when big companies represent backward productivity, don’t look at what big companies are doing? "

Third, when and in which field can super applications appear? We need more startups and VCs to try together and make more efforts in this respect.

At the just-concluded geek park innovation conference 2024, Li Yanhong said that "the native application of AI is valuable, and the progress of large models is not an opportunity for most people". He encouraged everyone: "The real value lies in native applications, and native applications are great opportunities for large factories, small and medium-sized enterprises and entrepreneurs. I hope everyone can grasp them as soon as possible and try as much as possible. We will be able to find a path that suits our own development. "

tag

In the era of big model, practitioners, entrepreneurs and investors seem to have unlimited scenery. In fact, it took only half a year to cook oil, from rushing to the track to hitting the wall.

Although the big model era provides many opportunities, not every track is smooth sailing, only a few companies can stand out and are doomed to a narrow escape. Instead of competing for the distant overlord throne in the field of big models, it is better to innovate in the field of AI native applications, and perhaps create the next Super APP.

References:

{1} 36Kr, "Big model investment is getting cold? Who is "creating" the winter? 》

{2} The first new voice, "After more than 200 days of boiling AIGC, investors reached three major consensuses"

{3} China Business Intelligence Network, "AIGC’s Rapid Development and Broad Future Application Scenarios"