As the first domestic AAA game masterpiece "Black Myth: Wukong" has become popular all over the world, the space computing power market that realizes the seamless connection between the physical world and the virtual world behind it has quickly become the new favorite of the capital market. Under the wave of large models, huge demands for computing power have emerged.
At the same time, infrastructure construction to support the improvement of computing power has also attracted industry attention. Since 2024, AI large models have entered the first year of application. More and more commercial banks have continued to increase investment in the infrastructure construction of large models. Multi-modal model architecture capabilities are becoming a key direction for the layout of financial institutions.
The trend of multi-modal large models is highlighted
With the continuous improvement of large model capabilities, since 2024, banks and other financial institutions have increased the scenario application and ecological construction of AI large models.
Recently, a number of listed banks have disclosed the progress of building AI large model platforms in their annual reports. China Merchants Bank’s semi-annual report shows that the bank has strengthened the systematic construction of large language models and made comprehensive efforts in areas such as infrastructure, reasoning and training platforms, algorithms and models, application development frameworks and scenario applications. Continuously improve the construction of the internal large model experience platform, strengthen in-depth communication with more than 100 large model ecological chain companies, promote the internal and external ecological construction of large models, and accelerate the application of cutting-edge technologies such as AI large models in the company.
China Construction Bank made it clear in its semi-annual report that it will continue to promote the construction and application of large-scale financial models and fully empower 79 in-bank business scenarios in six major sectors: corporate finance, personal finance, capital asset management, risk management, technology channels, and comprehensive management.
Ping An Bank's semi-annual report pointed out that in the first half of the year, it independently developed a large model open platform and strengthened the construction of basic capabilities such as computing power platform, large model base, large model development and operation integration (Ops), agent, and application development platform.
At the 2024 CMB Pujiang Digital Financial Ecology Conference, Zhou Tianhong, general manager of the Information Technology Department and chief information officer-designate of China Merchants Bank, said that large models will become the largest factor affecting human society and profoundly change people's economy, society and life in the future. methods; banks will also actively explore the application of large model scenarios and continue to accelerate the implementation of cutting-edge technology applications such as AI large models.
AI large model infrastructure construction and scenario application exploration are becoming key focuses of banks’ financial technology layout.
Yu Wujie, deputy general manager of the Information Technology Department of China Merchants Bank Head Office, revealed that the bank established a laboratory at the end of 2017 and began to conduct research in various aspects such as speech, language, vision, and images through the study of traditional technologies. "Since the launch of ChatGPT at the end of 2022, banks have invested more resources in the field of large models. Now China Merchants Bank has made infrastructure construction and scenario applications a key investment direction."
At the same time, Yu Wujie also pointed out that the current large language model has the ability to understand, a certain generation ability and preliminary logical reasoning ability, but has not yet reached the stage of complex logical reasoning and principle derivation. At the level of scenario application in the financial industry, Yu Wujie believes that the development of large models has gone through three stages: in the first stage, many digital products were produced, and large model capabilities were superimposed on the products to improve the efficiency of existing business processes; In the second stage, the capabilities of AI are naturally integrated into the generated applications, optimizing the customer service system process, allowing customers to obtain financial services by themselves; in the third stage, the large model will reshape everything, including the underlying operating system, organizational model, process division of labor, etc., bringing to have a more profound and essential impact.
Judging from the application practice of large models in financial institutions, the current industry is generally in the first and second stages. It is worth noting that as we enter the development stage of AI large models, the application of large models has put forward higher requirements for the infrastructure construction of enterprise large models.
Liu Zhaoyang, senior algorithm expert of Alibaba Cloud Bailian Large Model Platform, said that there are several directions that can be explored in the development of large model technology. Among them, One for all multi-modal models that support language, voice, and image input at the same time are a major trend. This is a set of A technical paradigm capable of processing multi-modal input and output such as text, images, and videos, including image understanding and generation.
According to the latest "Artificial Intelligence Large Language Model Technology Development Research Report (2024)" released, future large models will pay more attention to the fusion and processing of multi-modal data, and will tend to improve adaptive and transfer learning capabilities. Interpretability algorithms are used to improve transparency, allowing large language models to better understand and adapt to complex and changeable practical application environments.
However, Wang Guangrun, chief scientist of Tuoyuan Intelligence, pointed out that most of the current multi-modal models are based on the technical architecture of 7 years ago. Although these models have made certain progress, they still have many shortcomings, such as high training and inference costs, Prone to hallucinations, not good at long-term planning, and unable to complete complex tasks independently.
Wang Guangrun revealed that in response to these problems, Tuoyuan Intelligence proposed innovative ideas to reshape the foundation of multi-modal large models through a new technical architecture. “This innovative architecture not only significantly reduces the training and testing costs of large models, but also significantly lowers the threshold for small and medium-sized enterprises to enter the era of large models, thus promoting the equalization of technology.”
Computing infrastructure construction speeds up
The development and application of large models are highly dependent on powerful computing power support. Liu Zhaoyang said that computing power is the scarcest resource in this era. In today's era, computing power will basically become the biggest cornerstone for the development of every enterprise or the development of artificial intelligence.
Gao Wen, academician of the Chinese Academy of Engineering, director of Pengcheng Laboratory, and Boya Chair Professor at Peking University, emphasized that the development of models such as GPT relies on big data, big models and large computing power; the scale of computing power is the core element of national competitiveness, and the construction of computing power The power network is crucial and needs to solve challenges such as core computing power supply, communication connections and computing power scheduling to promote the development and application of AI.
Against this background, more and more leading companies continue to increase investment in large-scale infrastructure construction.
Chen Xi, deputy general manager of the Information Technology Department of China Merchants Bank Head Office, revealed that the bank is also currently accelerating the construction of an AI cloud platform to provide the basic capabilities and model service platform required for AI business applications, focusing on the training cluster and the inference cluster. Computing infrastructure construction.
Chen Xi said that the "three years to the cloud" mentioned before was a comprehensive cloud. With the emergence of large models, the proportion of intelligent computing will become larger and larger, and it is not only the upgrading of infrastructure, but also involves higher-level development. Paradigm changes.
Liu Zhaoyang pointed out that after GPT starts to reserve Transformer large models in 2020 or 2021, both the number and scale of large models, as well as the demand for computing power and data of the large models behind them, will show significant exponential growth. trend.
Under such a trend, large models also bring greater challenges to computing power support.
Zhou Wei, chief architect of Kunlun Core Financial, pointed out that the growth rate of computing power requirements for large models is much greater than the growth rate of the hardware itself, which is the so-called Moore's Law; at the same time, in the context of competition between China and the United States, especially domestic chips are still There will be problems with the neck being stuck. "So generally speaking, the global computing power supply is not satisfied with the current software demand."
In addition, Zhou Wei also said that how to evaluate whether a certain computing power can meet the demand depends not only on the computing power itself, but also on comprehensive indicators. In Zhou Wei's view, it is now generally accepted that computing power is not a simple indicator such as a simple calculation of floating point numbers or main frequency or core number. It is a comprehensive value of different hardware capabilities in computing, storage and communication.
Zhou Wei emphasized that in order to meet the needs of large model pre-training or fine-tuning of computing power, different computing power must be used as a heterogeneous mixed computing resource pool, and training tasks, inference tasks, and Agent, Rag.
While strengthening investment in the construction of computing power infrastructure, some financial institutions have also begun to pay attention to the improvement of the innovation capabilities of financial technology talents to further assist the construction of AI large models and application practice exploration.
Zhou Tianhong said that looking to the future, following the steam age, electrical age, and information age, human society is about to enter the intelligent age; only the flourishing of technology applications can promote the overall development of "AI + finance".
Gao Xulei, director of the Digital Finance Development Office of the China Merchants Bank Head Office, also revealed that the bank launched the Pujiang Digital Finance Learning Program to increase the frequency and density of exchanges and strive to create more innovative directions. In Gao Xulei's view, innovation does not occur in isolation, but blooms at the intersection of ideas, experience and culture in a suitable environment. "In the wave of digital finance, everyone is a witness, participant and creator. I hope that all financial institutions can jointly study the development laws of digital finance, try innovative models and methods, and jointly promote the development and application of cutting-edge digital finance technologies."