At the Xilihu Forum held today, Robin Lee said that too many big models in China would cause waste, and the AI native developed based on the big models would have great development prospects.
As of October, 238 large models have been released in China, compared to 79 in June. There are already thousands of AI native applications overseas, but there are few native applications in the Chinese market, regardless of the size of the model.
Robin Lee said that the symbol of human entering the AI era is not to produce many large models, but to produce many AI native applications. In the PC era, there was basically only one operating system, Windows, but there were many software developed based on the Windows system; In the era of mobile internet, the mainstream operating systems are only Android and iOS, and there are as many as 8 million mobile applications. The era of big models is also similar. Big models themselves are a foundation, similar to operating systems, and developers need to rely on a few large models to develop various native applications. Therefore, constantly repeating the development of basic large models is a great waste of social resources.
In the AI native era, there is a need for 1 million AI native applications, but not 100 large models. Robin Lee said that if China's industrial policies can more encourage AI native applications based on the big model, it will certainly build a prosperous AI ecosystem and promote a new round of economic growth.
The AI technology stack is divided into four layers, namely chip layer, framework layer, model layer, and application layer. The top layer is native applications, and chips, frameworks, and models all require AI application drivers. Only by implementing more scenarios can a larger data flywheel be formed, which can make the chips sufficient and user-friendly. Globally, AI native applications are becoming the main trend. The most successful AI native application, Copilot for Office365, which Microsoft collaborates with OpenAI, requires users to pay $30 per month.
Baidu has launched a large model plugin platform, through which individuals and businesses can turn their data and abilities into plugins, such as legal assistants, resume assistants, decoration assistants, etc. Plugins are a special type of AI native application that is not only easy to get started, but also allows enterprises to more conveniently use the ability of large models without the risk of private data leakage, reducing the threshold for developers, and is conducive to building an AI native application ecosystem. Baidu currently has over 8 million AI developers.
API is the main way for AI native applications to call large models. In physical industries such as manufacturing, energy, electricity, chemicals, and transportation, they will become extremely important footholds for big models and AI native applications, and will also become the main battlefield for promoting the integration of data and reality. On the Baidu AI Cloud Qianfan model platform, more than 17000 enterprises have developed industrial models and solutions.
If there is no intelligent emergence ability, the value of specialized large models is very limited. To train specialized large models, it is necessary to have sufficient parameter scale, sufficient training data volume, and continuous iteration to generate intelligent emergence, and large models can have the ability to bypass analogy.
Robin Lee believes that the industrialization model of the big model should combine the general ability of the basic model with the professional knowledge in the industry field. That is to say, the large model sets the small model, and the specialized small model has fast response, low cost, and the large model is more intelligent, which can be used to cover the bottom.
Robin Lee said that since its opening on August 31, the API call volume of Wenxin Grand Model has shown an exponential growth. The call volume of Wenxin Grand Model alone is more than that of more than 200 other big models combined.
Whether in China or the United States, Robin Lee believes that the best AI native applications have yet to emerge. Just as "mobile native" applications such as WeChat, Tiktok and Uber were born in the mobile era, there will be excellent AI native applications developed based on these big models in the AI native era.
Baidu has undergone AI native refactoring of its various product lines. The new search has the characteristics of extreme satisfaction, recommendation stimulation, and multiple rounds of interaction. Writing an article of 3000 words can be completed in just one minute, and Baidu Wenku can generate over 20 pages of PPT in one minute, including chart generation, formatting, etc., with almost zero cost. The new library has also achieved a transformation from a content tool to a productivity tool.
The greater space for imagination lies in the fact that large models have spawned AI native applications that have never been before. Baidu is also incubating new AI native applications. For example, Comate, an intelligent code assistant. Baidu has tens of thousands of engineers, and currently, for every 100 new lines of code added, 20 are generated by AI, which is still growing rapidly. Through human-machine collaboration, it can help significantly improve research and development efficiency.