The Zhitong Finance App learned that Galaxy Securities released a research report saying that recently, the fourth batch of large domestic models passed the filing. A total of 13 companies and 14 large domestic models were included in this batch. The bank believes that big model applications will enter the implementation period in 2024. On the one hand, the commercial application of large models in vertical fields is accelerating, and on the other hand, multi-modal large models are emerging, and application scenarios will be even richer. The bank is optimistic about the commercial development of AIGC. It is recommended to pay attention to iFLYTEK (002230.SZ), Jinshan Office (688111.SH), Tonghuashun (300033.SZ), Fluorite Network (688475.SH), Weining Health (300253.SZ), Zhongke Chuangda (300496.SZ), Tulsi (300229.SZ), etc.
Incident: Recently, the fourth batch of domestic big models passed the filing. A total of 13 companies and 14 domestic big models were included in this batch. Among the companies that have passed the registration of large model products, there are a number of artificial intelligence companies, such as Fourth Paradigm, 10,000 Things, Qianyuan Technology, and Zhiyin Intelligence, etc., as well as a number of manufacturers deeply involved in vertical fields, such as Xiaomi, Zhilian Recruitment, Boss Direct Recruitment, Pulse, What is worth buying, Microstep Online, Xinyi Technology, Chuangsiyuan, and Palm Reading.
The big domestic model is moving from GM to a vertical field, and commercialization is accelerating.
In August 2023, the first batch of large models from 8 companies passed the filing. Among them, Baidu, Douyin, Baichuan Intelligence, the Chinese Academy of Sciences, Shangtang, MiniMax, and the Shanghai Artificial Intelligence Laboratory all had large model parameters at the 100 billion level, all of which were general-purpose models.
In November 2023, the second batch of 10 companies passed the big model registration. There are 100 billion large-scale models such as MoonShotAI (MoonShotAI), Kunlun World Wide (Tiangong Big Model), and Facewall Intelligence (Luca Big Model), as well as 10 billion large-scale models such as Go Out and Ask (Big Sequential Monkey Model) and NetEase Youdao (Ziyue Big Model). However, most of the big models that passed the filing in the third and fourth batches had parameters below the 10 billion level, focusing on vertical fields. For example, an “AI question and answer robot” that is worth buying is a vertical product in the consumer sector. It will act as a personal shopping assistant for customers to search for links to good products around the world through dialogue. The Xinyi Video Big Model launched by Xinyi Technology is currently the first artificial intelligence model focusing on the video field in China. As a multi-modal video model, it can implement functions such as script generation, material matching, AI video generation, intelligent editing and dubbing, and digital human broadcasting.
“The main theme of 2023 is how to make a big model, and the main theme of 2024 is how to use a big model.
” The 100 billion general-purpose model has a variety of skills, but after specialized training of the large vertical model, its performance is comparable to that of the 100 billion model in individual skills, and the 10 billion parameter model is lower in cost, more cost-effective, and more suitable for enterprises to use. In solving specific problems in specific industries, large models in the vertical field have more advantages, and enterprises have also gradually realized the application potential of large vertical models, actively invested in R&D and application, and accelerated commercialization of vertical models.
With the passing of the fourth batch of big model filings, the supervision of big models continues to be normalized.
The “Interim Measures on the Management of Generative Artificial Intelligence Services” is China's first generative artificial intelligence (AIGC) regulatory document, and the world's first AIGC management measure. The document proposes adhering to the principles of equal emphasis on development and safety, promoting innovation and the rule of law, adopting effective measures to encourage the innovation and development of generative artificial intelligence, and implementing inclusive, prudential, classification and hierarchical supervision of generative artificial intelligence services. This not only shows the regulatory authorities' attention and support for the development of big models, but also shows the strict requirements for big model management and regulation.
This document provides a clear legal basis for the development and application of large models, helping to protect users' rights and interests while preventing misuse of technology. Up to now, almost all large model products of domestic core companies have been registered. Theoretically speaking, obtaining large model filings helps large model products from various manufacturers to form a data flywheel effect, which has contributed to the development of domestic big models.
Looking ahead to 2024, multi-modality will become a hot topic in the industry, and the commercialization process of AIGC will continue to advance.
According to the latest data from the CCID Research Institute of the Ministry of Industry and Information Technology, in 2023, the size of China's large model market will increase rapidly, and application scenarios will continue to be enriched. It is estimated that in 2023, China's large model market will reach 13.23 billion yuan, with a growth rate of 110%. Currently, the industry is paying more and more attention to multi-modal large models. The multi-modal large model is of great significance to the development of AI, because to achieve general artificial intelligence (AGI), it is necessary to analyze diverse data, even tactile perception. Google released the large model Gemini in December 2023, which performed better than GPT-4V in multi-modal tests. Gemini directly uses multi-modal data (including text, audio, images, videos, PDF files, etc.) for training, and can seamlessly understand and manipulate these different types of information, so it is a “native” multi-modal; GPT-4V trains plain text, pure vision, and pure audio models and then splices them separately, so the multi-modal performance is not as good as Gemini. Meanwhile, OpenAI will also release the GPT-5 multi-modal large model in early 2024. The emergence of multi-modal large models will enrich application scenarios, so that AI will no longer be limited to understanding text and images, but will expand into the field of video and advance the commercialization process of AIGC.
Risk warning: risk of global macroeconomic downturn; risk of weak downstream demand; risk of increased regulation; risk of AIGC commercialization falling short of expectations.