The "100-model war" is intensifying, and application innovation in the vertical field is the key to breaking the situation

Reporter Li Yucheng

Ding Rong, a trainee reporter

Image source: Generated by Unbounded AI tool

Since the beginning of this year, my country's large-scale model industry has developed vigorously, and the popularity of the track has continued to rise. At the Artificial Intelligence Summit Forum of the 2023 Global Digital Economy Conference held recently, Jiang Guangzhi, secretary of the party group and director of the Beijing Municipal Bureau of Economy and Information Technology, said that more than 80 large models have been released nationwide.

The large-scale model industry has entered the era of "Hundred Model Wars", and there are constant speculations and debates. "Is there redundant construction in the 'Hundred Models Competition'?" "How to promote application innovation in vertical fields?" has become the focus of the industry.

A number of industry insiders interviewed by a reporter from the Securities Daily said that the large-scale model industry is one of the key contests in future technological competition. For most enterprises, exploring application innovation based on large models for vertical scenarios, vertical industries, and vertical fields will be the key direction to overcome. Partial success will drive the industry's overall prosperity, innovation enthusiasm and creative vitality.

Enterprises competing for layout

In the field of my country's large-scale model industry, companies such as Ali, Huawei, and Tencent have deployed relatively early. In 2019, Ali carried out large-scale model research and development, and released the "Tongyi" large-scale model series in 2022. Tencent also disclosed the research and development progress of its "Hunyuan" large model in 2022.

Entering 2023, ChatGPT set off a wave of large models. Baidu released the big language model "Wen Xin Yi Yan" in March, becoming my country's first ChatGPT-like product. Afterwards, large-scale models released by many enterprises in our country rushed to appear.

The "Securities Daily" reporter sorted out and found that as of July 3, there were more than 80 large-scale models with parameters above 1 billion in my country. In addition to Internet giants, listed companies on the artificial intelligence track such as SenseTime, Yuncong Technology, and iFLYTEK, start-ups such as Light Years and Baichuan Intelligent, Shanghai Artificial Intelligence Laboratory, Harbin Institute of Technology and other scientific research institutes, Large models have been released successively.

"We divide large models and related products into three categories. The first category is general-purpose large models, such as OpenAI's ChatGPT, Baidu's 'Wenxin Yiyan', HKUST Xunfei's 'Xinghuo', etc. are all general-purpose large models; The second type is the large industry model; the third type is the application service based on the general large model or the industry large model. Most of the products that have been released so far are concentrated in the first and second types.” Some insiders said, “From the parameters Looking at it, the parameters of general large-scale models such as 'Wen Xin Yi Yan' are at the level of 100 billion, and the parameters of large models of other enterprises or start-up companies are usually at the level of 10 billion or 1 billion."

Is there duplication of construction in the "Hundred Models War"? Chen Duan, director of the Digital Economy Integration Innovation Development Center of the Central University of Finance and Economics, said in an interview with a reporter from the Securities Daily: "In the PC era and the mobile Internet era, China has grown a number of Internet companies that were once world-leading. It is a good thing that the competition for the next round of survival competition has actively joined the early R&D competition of large-scale models. However, the competition in the field of general-purpose large-scale models will have a Matthew effect in the future, and only a few strong ones can win."

Zhang Xiaorong, Dean of Deepin Technology Research Institute, also said in an interview with a reporter from Securities Daily: "Large model training has a high threshold, subject to the limitations of data and computing power, the number of companies that can really run through the commercialization of large models will There will not be many. At present, domestic large-scale models are still in the laboratory stage, and manufacturers need more time and resources to train models and debug parameters, etc. The market competition environment is conducive to industrial development, and in the end it will inevitably be washed away by big waves.”

Landing application is the key

"Judging from the large models that have been released, there are more general-purpose large models and fewer application-oriented large models. At present, there is a gap between domestic products and ChatGPT. We must face up to this gap and make key breakthroughs through scene-based traction to create Benchmarking has formed a demonstration, effectively and quickly narrowing the gap.” Wang Peng, a researcher at the Beijing Academy of Social Sciences, said in an interview with a reporter from the Securities Daily.

Chen Duan told reporters: "The market demand for realizing the value of large models in specific fields and scenarios is very large, and it will also have good market compatibility and accommodate more competitors. Therefore, for more For many participants, exploring application innovation based on large models will be the breakthrough direction."

This view has also been highly recognized by track companies. Robin Li, founder, chairman and CEO of Baidu, said recently: "More important than the number of large models is the application, which is a breakthrough in the application of vertical fields. The key point of the new international competition strategy is not how many large models there are, but the number of large models. How many native applications are there on the large model, and to what extent these applications have improved production efficiency."

At the press conference of the "Jiadu Zhixing Traffic Model" of the listed company Jiadu Technology, Liu Wei, the chairman of the company, said: "The birth of a general-purpose large model is only a starting point, and it will eventually settle down to a specific application scenario and solve specific problems in the industry. Above. With the industry model as the key driver, empowering the industry to improve production efficiency and service quality will bring about profound changes in economic and social development and the industry. The data and feedback brought by the advantages of application scenarios will further accelerate the industry's large-scale development. Iteration of model technology will also become the greatest competitiveness of domestic large models.”

It is understood that in the fields of government affairs, public security, and medical care, large-scale application-oriented models in vertical fields are gradually being implemented. The listed company TRS has created a professional large-scale model of government affairs based on its own official documents, policy documents, government affairs guidelines and other data as professional training data; the MYAI large-scale model independently developed by Meiya Pico focuses on vertical field applications, and has been used in public security It has been applied in the construction of big data projects in many industries, such as government affairs, taxation, and enterprise digital transformation.