Finance Associated Press, December 9 (researcher Zhang Yuhong, reporter Guo Songqiao).On the first anniversary of the release of ChatGPT, the discussion of AI large models is still in the ascendant. Driven by ChatGPT, domestic large-scale model companies have also gradually explored and tried to deepen the water area.
In the domestic "100 model war", the application of proprietary large models for vertical industries has been highly concerned. Among them, the application of large models in the financial field is one of the focuses.
So, what is the current development of the financial model?Which financial model enterprises can win in the "100 model war"?
In an exclusive interview with Cailian, Huang Aizhou, Head of Fintech at KPMG China, believes that the quality and scale of algorithms and data are the core of competition. "The future development direction of the financial model is in the form of general model + agent (agent)".
Finance Associated Press: **The financial work conference put forward five major articles: science and technology finance, green finance, inclusive finance, pension finance, and digital finance. In your opinion, what kind of help can the financial model give?
Huang Aizhou:The large model is a thing that is partial to infrastructure, so it helps with these five aspects. Nowadays, the financial model is usually customized for the financial industry on the basis of the general model, so it is called the "financial model". The financial model has some clear applications in the three fields of green finance, inclusive finance and pension finance. For example, in green finance, AI does ESG data collection and analysis.
Fintech refers more to how finance can better serve high-tech and high-growth companies. From this point of view, when financial institutions provide financing services for science and technology companies, they can use large models to improve the effect of intelligent risk control. Digital finance is more about the digitalization of financial institutions. Science and technology finance, green finance, inclusive finance, and pension finance are how financial institutions serve different scenarios and customer groups, and digital finance is how financial institutions improve their digitalization. It turns out that there may be many existing AI applications that are relatively independent, such as customer marketing, intelligent risk control, public opinion management, compliance management, etc., and the large model can be like a base to open up these individual AI applications.
To sum up, first of all, if we focus on the "five major finances", the foundation is digital finance, the digital transformation of financial institutions, or the improvement of digital capabilities. Large models can definitely play a huge role in the improvement of digital capabilities, and the improvement of the quality of the base can in turn better promote the development of science and technology finance, green finance, inclusive finance, and pension finance.
Finance Associated Press: What are the characteristics of the financial model compared with other industry models?
Huang Aizhou:Large models require a large amount of data, and there is a lot of corpus training at the bottom to emerge inference capabilities. The financial industry is one of the industries that attaches the most importance to data security and data privacy protection, so it may be difficult to explore large models.
At present, banks are still in the exploration stage of the application of large models. In the future, I think there are two ways to apply large models in financial institutions: one is to use the existing open source large models as the base, using the trained open source models, supplemented by their own corpus for new annotation and training;The second is to explore and develop its own large models. I judge that this may be difficult, although everyone knows the basic technical framework, but the cost of computing power is too high. At present, it is quite time-consuming and laborious to really complete your own large model, and the effect is not necessarily ideal, so the first method is relatively reliable.
Finance Associated Press: What are the main applications and landing scenarios of the current domestic financial model?Who are the main players?
Huang Aizhou:There are three main types of application situations and landing scenarios of financial models: one is to provide investment advisory guidance. Large models can replace a large part of people's work and generate targeted investment references for thousands of peopleThe second type is to do more accurate intelligent risk control. In terms of risk control, the large model can accurately identify market signals and put forward some risk control suggestionsThe third category is to do regulatory compliance. For regulatory requirements, compliance and other issues, large models can effectively organize and organize information. There are three types of participants: technology companies, financial institutions, and technology research institutions, including university laboratories. These application products are further developed by technology companies on the basis of open source models. Financial institutions are more likely to purchase solutions provided by technology companies, and some are exploring the design of some agents to use large models to solve the problem of their own R&D capabilities. Universities and other research institutions are working on their own open source models.
Finance Associated Press: What do you think are the more prominent values of the large model of this subdivision of finance?
Huang Aizhou:The first is to be more capable. Financial models have reasoning capabilities on the basis of professional knowledge, and can perform more adaptable in specific fields. For example, now some customer service **, chat a sentence or two to know that it is AI, in most cases it is difficult to solve specific problems, now with a large model, this kind of communication and problem-solving ability is stronger. The second is a significant increase in efficiency. Before the big model came out, financial institutions also continued to improve efficiency through RPA (robots). However, I found that the use of RPA by specific business personnel is not very extensive, because it is still relatively cumbersome to use, and there is Xi cost to learn. In the future, it is likely that the financial model will be able to clearly understand the meaning of what you say in a natural language interaction and perform the corresponding work. At the same time, the application of large models in the field of system development such as writing can also play a great role, greatly improving efficiency. The third is that the application scenarios are wider and cover more user needs.
Finance Associated Press: What is the core of the current competition in the financial model?
Huang Aizhou:I think the most important thing is two capabilities: the first core is the algorithm. The principle is the same, but a good algorithm can make fewer parameters required and produce better results, but this requires top AI scientists. The second core is the quality and scale of the data. Data quality means that there is a large amount of data that is well labeled, which will definitely improve the performance of the model and the results produced by the model. Labeling itself is not a simple labeling, but is done with your own understanding of the scenario, customer, and business. In terms of the scale of data, because the emergence of large models depends on the continuous training of a large number of learning Xi, too little data is definitely not possible. There is now a direction in foreign countries that is based on the general large model, and there are all kinds of small companies and small teams to do agents. That is, the general model + agent model. ChatGPT's Markets is actually similar to this model. The general model can give a seemingly passing answer to the problem you need to solve, but you don't feel that it is enough to quench your thirst and not really solve your actual problem, so you need a special agent to solve the specific problem. For example, there are now 100 AI applications in the financial industry, and it is likely that these 100 AIs will eventually become 100 agents based on large models, with one agent responsible for solving the problem of financial product sales, and the other agent responsible for solving the risk control problem of a certain product.