Our reporter Yuan Chuanxi Trainee reporter Liang Aonan
Once upon a time, the ancients used the calculation to calculate thousands of miles of Wan Li. Today, the AI big model can accommodate time, space, sun and moon in minutes and seconds.
In 2023, it was the first year of the outbreak of AI big model, which came by the wind of AIGC (Generative Artificial Intelligence), and the wave of big model swept the world and continued to be hot. Technology giants, small and medium-sized enterprises, and scientific research institutes have entered the market one after another, trying to occupy a favorable position in this new blue ocean market that has become a "battleground for military strategists." In the future that will be reshaped by technology, will AI bring subversive innovation or gradual iteration? What new kinetic energy will it bring to enterprises, industries and capital markets?
The Securities Daily reporter combed the data of Eastern Fortune Choice and found that as of December 24, the total market value of 69 AIGC concept stocks in A-shares was 818.247 billion yuan, an increase of 219.449 billion yuan compared with the beginning of the year, of which 55 companies’ share prices rose and the number of companies fell was only 14. From the perspective of the capital market, the wind of the big model has already blown.
"The big model is everywhere, and the golden decade of artificial intelligence has been opened since then." Hu Xuecan, a senior investment consultant of East Hi-Tech, told the Securities Daily reporter, "This will greatly change the whole social economy and industrial ecology and become another productive force."
AI triggered a "hundred-mode war"
Looking back on 2023, since the launch of ChatGPT-4 in March, domestic technology companies have entered the market by running. Baidu’s ERNIE Bot, Alibaba’s Tongyi Qianwen, Huawei’s Pangu, 360′ Zhinao’, Kunlun Wanwei’s Tiangong, JD.COM’s Lingxi, Iflytek’s Spark, Tencent’s Hunyuan and Shang Tang’s Day-by-Day-by-day-by-day-by-day-by-day-by-day-by-day-by-day-by-day-by-day-by-day-and other large models have appeared By the beginning of October, 2023, the number of large AI models published in China had reached 238, which was upgraded from "one hundred models" to "two hundred models".
Zhao Gang, president of Saizhi Industry Research Institute, told the Securities Daily reporter that domestic artificial intelligence enterprises all want to seize the opportunity of the AI era, and have laid out in advance around the general model, industry model and expert field model, thus creating a "hundred-model war". At present, the large model industry as a whole is in an active period of technological innovation.
"Compared with the early artificial intelligence model, this year, China’s large model has made a qualitative leap in parameter quantity, the overall modeling ability of complex tasks has improved, the learning ability and generalization have become stronger, and it has a higher level of cognitive interaction ability." Chen Xiaohua, executive director of the Meta-Cosmic Industry Collaborative Innovation Center in the Science Park of Beihou National University, told the Securities Daily reporter.
Compared with the general big model, many small and medium-sized enterprises have successively released vertical big models with deep cultivation and professional ability in order to gain a place in the industry. Among them, Ctrip released "Ctrip asks" in the tourism industry, Honey released "Mi Wen Xiu" in the field of intelligent proofreading, Netease released "Zi Yue" based on education, JD Health released "Beijing Medical Thousand Inquiries" in the medical and health industry, and Ant Group released a financial model … to promote artificial intelligence from perception to cognition, from recognition to generation, and from general to industry.
"The slowdown in user growth is a fact, and the early explosive growth is unsustainable. The big model must be integrated into thousands of industries and let the industry harvest value from AI, which can become a widespread need. This is the next growth space. " Wang Xiaohang, vice president of Ant Group and head of financial model, said.
In the eyes of the industry, the key to the development of large-scale models in China is to seize the scene bonus. At present, the ability of domestic large-scale models is enough in the vertical field, and the biggest opportunity is to develop large-scale models in the direction of industrialization, industrialization, verticalization and deep customization.
According to the data of quantum bit think tank, in 2023, the market size of China AIGC industry will be about 17 billion yuan, and it is estimated that the market size will reach trillion yuan in 2030. "In recent years, China’s artificial intelligence industry has flourished, with the core industry reaching 500 billion yuan and the number of enterprises exceeding 4,300. Innovations such as smart chips and general models are constantly emerging. " Xu Xiaolan, vice minister of the Ministry of Industry and Information Technology, said.
The "Hundred Mode Wars" gradually entered the second half, and the "Group Mode Era" came.
Accelerated commercialization
It is worth noting that the current large-scale model training needs strong computational support, especially the larger the parameters, the higher the model cost. For example, the language model ChatGPT-3 of OpenAI (American Artificial Intelligence Research Company) costs about $700,000 per day.
With the continuous increase of large model parameters by Internet and technology enterprises, the demand of traditional enterprises for digital intelligence transformation has increased, which has led to a sharp increase in computing power demand. Moreover, due to the difficulty in obtaining high-end chips and the lack of public computing power construction, computing power leasing has entered a new round of price increase cycle. Among them, Heji Technology, Zhongbei Communication, Huina Technology and other companies have recently announced a sharp increase in the fees for computing services, which has also led to a further increase in the research and development costs of large models.
Under the huge financial pressure, large model enterprises have to step up their efforts to find a road to commercialization. In fact, the commercialization path of large models of some established Internet companies is now relatively clear.
Zhao Gang told reporters that there are three main modes for realizing the large-scale model business: First, enterprises realize the use fee through the large-scale model platform, and adopt charging according to the use time or charging according to the amount of information, mainly for individual users; Second, enterprises realize the realization of customized development and deployment fees through large model App applications, mainly for enterprise users, and the internal data of users are also used to train large models, and develop customized large model applications for business needs such as enterprise marketing, product design and customer service to help users achieve the goal of reducing costs and increasing efficiency; Third, enterprises realize value-added services such as advertising, and develop value-added services such as advertising with a large model platform with a large number of users to give full play to the advantages of traffic portal and learn from the Internet model.
For example, Baidu reconstructed several product lines based on Wenxin Big Model this year, and used generative AI as an auxiliary function to embed the original business to provide users with value-added services. "In the realization of ERNIE Bot App, new field, new search, etc., Baidu will gradually establish some payment methods, such as the library input method has a paid version." Wang Fengyang, vice president of Baidu Group and head of the mobile eco-business system, told the Securities Daily reporter.
"At present, the path of commercialization of large models is relatively clear, and domestic manufacturers are actively exploring it. However, the road of commercialization of large models can not be limited to the exploration of business models, but also lies in solving the underlying problems of the development of large models." Many insiders told the Securities Daily reporter.
Although the road to commercialization of large models is long and difficult, the dawn has already appeared. The semi-annual report of 360 in 2023 shows that the "360 Zhinao" model has begun to generate income, with an amount of nearly 20 million yuan; Shang Tang Group also announced that the related income of generative AI increased by 670% in the first half of the year; Meitu said that the big model boosted the company’s profit in the first half of the year by 3.2 times.
"In fact, it is not difficult to develop a big model. What is difficult is how to make enterprises pay for it willingly." Chen Da, managing director of Fuchunshan Capital (HK), told the Securities Daily reporter that only large models that can really solve the needs of enterprises will be favored. For example, enterprises such as Good Future and Face-to-Face Intelligence have completed the training, development and optimization of large models with the services of Baidu AI Cloud, which has improved the effect of their own large models, reduced costs and increased efficiency for enterprises, and indeed brought convenience, so the willingness of enterprises to pay is also very high.
"In the future, the big model will be embodied as an ubiquitous intelligent service, and the commercialization model will be very rich. It can be embedded in various software and hardware systems, such as smart cars, intelligent robots, etc., commercialized in the form of smart products, and can also be deployed in the cloud to reconstruct existing business logic and provide commercialized large-scale model cloud services and various smart App platform applications. " Zhao Gang thinks.
Heading for the multi-modal trend
Although various large-scale models emerge one after another, continuous optimization and upgrading to promote the progress of the industry is still the main direction of major Internet companies. Recently, the upgrade of large model from single mode to multi-mode has become a hot spot in the industry, and multi-mode AI of many companies has become popular.
In November, OpenAI released GPT-4 Turbo and opened GPTs, once again subverting the industry and opening the ecological prelude of AIGC application. "Top Ten Technology Trends of Dharma Institute 2023" shows that the establishment of a unified, cross-scene, multi-task multi-modal basic model will become one of the mainstream trends of artificial intelligence development.
According to soochow securities Research Report, multimodal is the only way to realize general artificial intelligence. Modal data input can help improve the model ability and user experience, and allowing multimodal data output is more in line with the needs of the real world. The requirements of data, algorithm and computing power are higher than that of single mode. This wave of natural language model development provides technical reference for other modes, and the industry is expected to accelerate its development.
Although the big model triggered changes in the Internet industry this year, the whole industry is still in the early stage of development, and there are still some shortcomings in accumulation and precipitation. For enterprises, it is still worth exploring how to gain a firm foothold in the industry, effectively solve the needs of users and realize commercialization.
Looking ahead, Zhang Yi, CEO and chief analyst of AI Media Consulting Group, told the Securities Daily reporter that enterprises should not blindly follow suit and need to solve the demands of their existing basic products and user groups in the AI field. However, there is little difference between the current big models, so it is difficult for users to distinguish and choose the products that suit them. In the future, enterprises can stand out only by optimizing large-scale products to improve users’ work efficiency, reduce production costs and solve problems for customers, which is also the real development direction of large-scale models.
Zhao Gang predicted that there will be three major trends in the development of large-scale models in the future. First, the performance improvement and optimization of general large-scale models will be further deepened. Large-scale models will enter trillions of parameters, and the performance will be further improved by using clusters above 10,000 calories. At the same time, model tuning will be carried out around different application scenarios; Second, the thinking and exploration ability of the large model will be further improved, and the application will move from simple intelligent application to complex intelligent application; Third, the super application of the big model will appear, which will integrate the abilities of content generation, semantic understanding, reasoning, memory and self-awareness, and provide intelligent services with a strong sense of experience. Intelligent digital people or robots will become the standard in all fields of economy and society.
From general large model to vertical large model, from single mode to multi-mode integration, AI generation technology is constantly iterative, accelerating application landing and business model innovation, which is the general trend and the future has come.
Source: Securities Daily
Reporting/feedback
关于作者