Large-scale models become a new track for artificial intelligence ["2022 White Paper on the Development of China's Large-scale Models" Released——]

　　◎Reporter Liu Yan and He Peiqi

　　The popular ChatGPT around the world has brought large-scale model technology into more people's field of vision. Can the strength of domestic large-scale model manufacturers support ChatGPT-like applications?

The "2022 China Large-scale Model Development White Paper" (hereinafter referred to as the "White Paper") recently released by the international authoritative consulting organization IDC has become a reference for a glimpse of the domestic large-scale model industry.

　　As a conversational robot, ChatGPT's ability to "know astronomy from above and geography from below" comes from the ability support of large models. As Wu Lianfeng, vice president and chief analyst of IDC China, said, ChatGPT would not have been born without long-term investment in large models Such an application, and behind the big model lies a revolution in the artificial intelligence implementation model.

　　The "White Paper" pointed out that from a technical point of view, large models originated in the field of natural language processing, represented by Google's BERT, OpenAI's GPT and Baidu's Wenxin large model, and the scale of parameters has gradually increased to hundreds of billions and trillions. The amount of data used for training has also increased significantly, resulting in improved model capabilities.

　　As the demand for digital transformation grows, more and more applications of AI are used in enterprises. IDC predicts that the market size of China's artificial intelligence software and applications will reach 21.1 billion US dollars in 2026. Artificial intelligence has entered a critical period of large-scale application. However, how Solve the problems that have begun to emerge, such as high development threshold, complex and diverse application scenarios, and dependence on scene annotation data?

AI big models bring new hope.

　　The data shows that since 2020, the number of large-scale models in China has increased sharply. From 2020 to 2021 alone, the number of large-scale models in China has increased from 2 to 21, which is on the same level as the United States and significantly ahead of other countries.

The "White Paper" shows that Baidu Wenxin's large model has built a three-tier system of "large model + tool platform + product and community", which is widely used in energy, finance, aerospace, manufacturing, media, cities, social sciences, and film and television. The key path for the landing of the large-scale model industry.

　　Judging from the industry's first large-scale model evaluation framework proposed in the "White Paper", Baidu's Wenxin large-scale model is in the first echelon in the market structure, leading in product capabilities, ecological capabilities, and application capabilities, and is widely recognized by the industry.

Wu Lianfeng said: "Under the large-scale model evaluation framework proposed by IDC, Baidu's Wenxin large-scale model performed very well, which is a solid foundation for its large-scale language model Wenxin Yiyan."

　　It is understood that Baidu has released the pre-training large model ERNIE 1.0 in March 2019, and will start to apply the Wenxin large model to the search business in 2020, empowering search relevance, in-depth question and answer and content understanding, etc., and develop ChatGPT-like products in China Have a first-mover advantage.

　　Under the pressure of the emergence of ChatGPT, related companies have expressed their opinions one after another, and "big model" has become a key word without exception. In just a few lines of official announcements, Baidu used a paragraph to introduce its own AI four-layer architecture layout, focusing on the Wenxin large model; Google CEO Sundar Pichai said that its own AI conversational robot Bard (Bard) is supported by the large model LaMDA.

　　Talking about ChatGPT, Huang Tiejun, dean of Beijing Zhiyuan Artificial Intelligence Research Institute and professor of Peking University School of Computer Science, told the reporter of Science and Technology Daily: "Natural language interaction has crossed a hurdle and is accepted by the public, no matter from the history of artificial intelligence development. , or the history of computer development, is a milestone. Technically, China has no problem making similar applications, but there is a big difference between having technology and making products with good user experience. It is It is a matter of channel and operating experience for ecological construction and serving a large number of users."

　　Robin Li, founder, chairman and CEO of Baidu, also said that ChatGPT is a new opportunity after the development of AI technology to a certain stage. How to turn such a cool technology into a good product that everyone needs is actually the most difficult step. is also the greatest and most influential.

　　Can Chinese companies make a difference in ChatGPT-like applications?

The first to be tortured is Baidu, which claims to have a first-mover advantage. Judging from the list it announced recently, the list of well-known companies joining Baidu's Wenxin Yiyan ecosystem is getting longer and longer.

　　In the past few years, Baidu has publicly emphasized the important role of large models as a new type of AI infrastructure for many times, calling for industry attention.

The "White Paper" pointed out that the industrial chain with large models as the ecological base will become an infrastructure that can be reused on a large scale in the intelligent upgrade. Chinese large-scale model manufacturers are relatively complete in terms of model layout. Continue to explore the breadth and depth, constantly consolidate the product construction based on the large model, and promote the large-scale model technology from the laboratory to the large-scale implementation.

　　"Science and Technology Daily" Issue 03, February 21, 2023