Since the advent of ChatGPT at the end of 2022, the upsurge of AI big model in the technology industry has swept the world. In 2023, global technology giants all increased their layout, aiming to occupy a leading position in this round of scientific and technological revolution initiated by AI big model.
In this context, China’s science and technology enterprises quickly follow up. According to the data of CCID Research Institute of the Ministry of Industry and Information Technology, there are more than 19 language model research and development manufacturers in China. Among them, the model products of 15 manufacturers have passed the record. It is estimated that the market size of large language models in China will reach 13.23 billion yuan this year, with a growth rate of 110%.
In fact, AI technology has been developed for many years, and the reason why the popularity has risen rapidly in 2023 lies in the changes brought about by the increase in computation.
When AI Go defeated human players, the parameter scale of Google Bert was 300 million; After 2020, GPT-3 was born, and the parameter order has crossed to 175 billion; Iterating to GPT-4, the parameters continue to multiply, reaching about 1.8 trillion.
When the amount of calculation reaches the 22 nd power of 10, the model capacity will complete a leap from quantitative change to qualitative change, showing an amazing explosive growth.
With the expansion of the parameter scale, since the second half of the year, the subversive influence of technology has gradually penetrated into the industry and reconstructed the traditional industrial form and mode. Gartner predicts that by 2026, more than 80% of enterprises will use generative artificial intelligence APIs and models in production or deploy generative artificial intelligence applications.
And 2024 is the year when the application of AI big model broke out. Recently, Fu Sheng, Chairman and CEO of Cheetah Mobile and Chairman of Orion Star, said at the 2023 Exploration Conference that there will be many applications in the future, such as Didi, Meituan Takeaway in the era of mobile smartphones, which have not appeared in the past.
Image source: vision china
Competition and elimination
IT is generally believed in the industry that the AI big model era is another new era full of opportunities and possibilities after the IT era and the mobile Internet era. From a global perspective, all parts of the world are actively promoting the development and application of large models, among which the total number of general large models released by the United States and China accounts for 80% of the global release.
Following the footsteps of ChatGPT, major enterprises and institutions have joined the big model competition in succession, including Internet giants such as Baidu, Alibaba and Tencent, AI field manufacturers such as Shangtang Technology and Defiance Technology, big model start-ups such as Zhipu Huazhang, Baichuan Intelligent and Daguan Data, and universities and institutions such as Fudan University and Tsinghua University. According to public information, as of October this year, 238 large models have been released in China.
In the new big model competition, startups and big manufacturers have their own advantages. With years of accumulation and a large number of users, large manufacturers can obtain massive user data and feedback, but first-line startups usually have innovative technologies and business models, and there is also the possibility of rapid technology iteration. As Wang Xiaochuan, founder and CEO of Baichuan Intelligent, said, "Small innovation depends on big factories, while big innovation still depends on small factories".
In the "SuperCLUE Chinese Big Model Benchmark Evaluation Report, 2023", the average score difference between the big model of the big factory and the startup company is about 1 point, which is almost the same.
However, the capability gap between domestic and foreign big models still exists. "The gap between us and OpenAI is widening, not narrowing." Xiao Yanghua, a professor at the School of Computer Science and Technology of Fudan University, said.
In SuperCLUE evaluation, GPT4-Turbo is far ahead with a total score of 89.79, which is higher than all domestic large models and foreign representative large models. The big model with the highest score in China is ERNIE Bot 4.0, but it is still 15.77 points away from GPT4-Turbo.
"At present, the ability level of the mainstream big models in China is basically around GPT3.5." Wu Wei, an extraordinary capital partner, told 21st century business herald that in his view, the gap was more than half a year.
The core barriers of the big model include computing power, data and algorithm, so it is also the aspect where the gap is concentrated. Xu Dongliang, CTO of Du Xiaoman, said at the annual meeting of 2023 Financial Street Forum that only a few enterprises can complete industrial-level R&D from beginning to end.
Computational power is the cornerstone of large model training. Nowadays, the model parameters increase exponentially, and the calculation power required for training is huge. It costs tens of millions of dollars to train a large general model of 100 billion levels. At present, among the large models published in China, there are only about 10 manufacturers whose parameter scale reaches 100 billion or more.
Under the latest round of blockade and sanctions imposed by the United States, the demand for domestic alternatives is more urgent. But for a long time, chip and computing power will still be a huge gap between domestic big models and ChatGPT.
"Every company in China has to do it, which will create a problem. Each company does its own, but each company has limited data and computing power, and there is not much money to support research and development, and it is doing some very basic and repetitive things." Qiu Xipeng, a professor at Fudan University School of Computer Science and the head of Moss system, said in an interview with 21st century business herald.
With the continuous catching-up of domestic big models, the competitive pattern of the industry will also change. At present, the big model industry is still in the bubble period, and companies with technical strength don’t want to be left behind by the tide of the times, trying to catch up by training their own big models, so there is a hundred-model war and a thousand-model war. When the industry enters a mature stage, only a few enterprises can truly empower the industry, and the value can be precipitated after the bubble is squeezed out.
"In the future, large models will gradually show the trend of oligarchy, because in the case of limited computing power and too homogeneous large models, computing power resources and data resources will be concentrated in the future, and some large models with low value will gradually be eliminated." Li Qing, director of Sullivan Greater China, told the 21st century business herald reporter.
It is generally believed in the industry that only a few giants will win in the general model.
"The eliminated big model enterprises will not disappear, and they may find their own opportunities, such as making multi-modal big models or big models in certain industries." Wu Wei told reporters.
Since the second half of the year, the market’s enthusiasm for big models has dropped significantly compared with that at the beginning of the year. In the industry’s view, this is nothing more than the embodiment that the market overestimates the short-term impact and underestimates the long-term potential of new technologies. From another point of view, when the new technology disenchants, various problems gradually appear in the process of landing, and the excessive expectation for the big model is also in the process of becoming rational.
Application and commercialization
When the competition in the industry intensifies, the application of large-scale model has become the focus of attention of all parties.
In the research report of Gen AI ARC Survey in August 2023, IDC pointed out that among enterprises with more than 5,000 employees, 80% believe that generative AI will subvert their business in the next 18 months.
"The big model should never stay in the alchemy stage. We should promote it to become a scientific big model. Only by deep integration with the industry can we truly achieve sustainable development." Xiao Yanghua said.
The first is the transformation from a general large model to a vertical large model. Some industry views believe that the future development of large models will tend to be both general and specialized.
In June this year, Tencent Cloud officially announced the research and development progress of the industry model for the first time, and released Tencent Cloud MaaS service solutions for B-end customers. In July this year, Huawei released Pangu Big Model 3.0, which is "Do not write poems but do things", and deeply cultivated government affairs, finance, manufacturing, coal mines, railways, pharmaceuticals, meteorology and other industries. In addition, the tourism-oriented "Ctrip Asking", the medical-oriented Baidu "Spiritual Doctor" model and the education-oriented Netease "Zi Yue" model were also released in the second half of the year.
Among them, the financial industry is rich in application scenarios, and it is the earliest institution to carry out digital transformation, and it has become one of the best scenarios for the application of AI big models. The financial industry has accumulated a huge amount of data, including financial transaction data and customer information, and a good data foundation provides conditions for the landing application of AI big model. At present, two financial models, generation and decision-making, have been implemented in banks, securities and other financial institutions.
"How to match your business and business scenarios with the logic of AI, and how to seek AI value innovation, not just from the perspective of efficiency, may be a problem that needs to be considered in the current promotion." Li Qing said, "If you can’t use the big model flexibly, or you can’t fully adapt to your own business scenarios, it may be difficult to fully achieve the purpose of reducing costs and increasing efficiency to a certain extent."
For the latecomers, the opportunity of the big model lies in the application. AI native application based on large model technology is regarded as a path that will really ignite the industry. Wu Wei told reporters that from the perspective of investors, it is also more optimistic about startups with large model applications.
Li Yanhong, founder, chairman and CEO of Baidu, recently said at a round-table event that the Hundred Models War is a waste of social resources, and more resources should be put on super applications. At the Xili Lake Forum in November, he also said, "In the AI native era, we need 1 million AI native applications, but we don’t need 100 large models." At the scene of Baidu World 2023, Baidu took the lead in throwing out more than ten AI native applications.
The prosperity of new technologies must be the prosperity of applications, such as the wonderful duck camera that uploads 20 photos to generate real photos, and the Pika that can generate high-quality videos by inputting a few keywords. These hot innovative products have opened a new door for AI entrepreneurship. It is generally believed in the industry that native applications can incite greater commercial value and are the symbol of human beings entering the AI era.
They have the ability to integrate large models based on the original products, and also have scene-based applications that re-load large models. "At present, we see that many applications have chosen the second one." Li Qing told reporters.
He believes that the regulatory level may give some space. "When the big model first came out, because its understanding and future applications were relatively vague, strict strategies were adopted at first. Later, with the birth of applications and differences at home and abroad, the regulatory level may give appropriate flexible space to support the development of the industry."
Unfortunately, until the end of the year, there was still no AI application that continued to be popular in China. Even the wonderful duck camera, which was a smash hit, gradually faded out of public view only two months after it was launched, due to repeated payment and insufficient user stickiness. In Wu Wei’s view, the business model of domestic large-scale model application still needs to be explored. In contrast, overseas markets have greater opportunities. "The monthly income of similar overseas applications has exceeded one million dollars."
However, such problems should not be rushed. The high use cost of large models, the illusion of their own existence, and the lack of willingness to pay in the market are all reasons that limit the development of applications, and the solutions of these problems are advancing in the ever-changing changes.
Some people in the industry believe that the prosperous AI native application ecology needs three elements: large model, intelligent computing power and new paradigm of AI native application research and development to complement each other. Others believe that the starting point of AI native application will be in 2024. To be sure, the outbreak will take time, but the future of AI application is worth looking forward to.