With the continuous development and application of domestic big model technology, the wave of big models has swept across various industries, and financial big models have also become an important tool for investment in the financial field. Major technology enterprises, including Du Xiaoman, Qifu Technology, and Instant Consumer Finance, have taken the lead in layout and have successively released and implemented multiple financial big models, further empowering financial institutions.
However, current domestic big models are still in the early stages of research and iteration. What can big models do in the financial field? What are the difficulties and challenges faced in the research and subsequent development of financial models? How to better utilize large models for mining, analyzing, and predicting financial data? Recently, a Caixin reporter exclusively interviewed Fei Haojun, Chief Algorithm Scientist of Qifu Technology, on the above-mentioned issues.
In Fei Haojun's view, the financial big model will inevitably become a financial expert in the future, but the bigger the model, the better. The higher the parameters, the greater the computational resources invested. How to improve the model's effectiveness within limited computing power and make practical business applications faster and more efficient is the first thing to balance.
Caixin News Agency: What are the specific aspects in which the financial model empowers financial institutions, and how does it play a role?
Fei Haojun: With the continuous evolution of model capabilities, future big models will continue to change the most essential core layer of finance
At the end of November 2022, ChatGPT emerged, sparking a wave of big models. Since then, general models such as ERNIE Bot and Synonym Qianwen, as well as industry models such as MedGPT, Qifu GPT and ChatLaw, have been released. Subsequently, the application of financial big models in the financial industry has accelerated.
At present, AI big models are divided into general big models and industry big models. The former has a wider range of applications and can play a role in multiple fields; The latter focuses on a specific field or industry such as law, healthcare, finance, etc., and has high professionalism and targeting in vertical fields.
In Fei Haojun's view, the empowerment of financial institutions by big models can be comprehensive. In the financial business process, big models can empower all aspects related to people and human-machine interaction. "At present, the role of financial big models is mostly to improve efficiency."
Taking the practice of Qifu Technology's financial model as an example, in the telemarketing system, semantic analysis and clue mining help improve the accuracy of telemarketing clue recognition by up to 98%, while increasing the conversion rate by over 5%; In the intelligent marketing process, approximately 70% of image materials are generated by AIGC, and it is planned to annotate and rate the materials in multiple dimensions through a large model to achieve optimization of advertising placement; In the scenario of speech robot script generation, the quality rate of generated scripts has reached 70%.
"Taking call quality inspection as an example, in the past, we used traditional models to conduct some quality inspections to ensure the effectiveness of phone calls. Now, with the use of large models, the coverage and efficiency of quality inspections will be greatly improved. Our first version of the large model achieved 100% coverage in automated quality inspection, and the detection rate has increased by 15% compared to previous models." Fei Haojun explained.
"The big model is a derivative state that will continue to develop from now to the future, so the empowerment of the entire financial industry by the financial big model is also a gradual process." Fei Haojun believes that in the future, as the model's capabilities continue to evolve, the big model will gradually be applied to the core business layer of finance, continuously empowering the most essential core layer of finance, and playing an increasingly important role.
Caixin News Agency: What difficulties have been encountered in the research and development process of the current financial big model?
Fei Haojun: Data and computing power are the key factors that affect the effectiveness of large models, and hallucinations are the most common problems with large models
Under the trend of digital transformation in the financial industry, more and more companies are participating in the layout of the financial industry's big model. In May of this year, Qifu Technology was the first to announce the launch of its self-developed financial model, "Qifu GPT," which is also known as the first general financial industry model in China; In the same month, Du Xiaoman launched the first Chinese financial model worth billions of yuan in China, "Xuan Yuan"; In June, Hang Seng Electronics released the financial industry model LightGPT; In August, Consumer Finance immediately released its first retail finance model, "Tianjing"; In September, Ant Group officially released industrial grade financial models such as AntFinGLM.
"The limitations of data and computing power are the most fundamental issues that affect the effectiveness of large models, and they are also the primary issues faced by the industry in the process of implementing large models." Fei Haojun admitted that the parameter quantities of large models are roughly divided into billions, tens, and hundreds of billions, and different parameter quantities represent different computational resources invested, which also means different research and development costs.
Fei Haojun believes that larger models are not necessarily better. How to improve model performance within limited computing power and make practical business applications faster and more efficient is the first thing to balance. "We hope to 'make the model bigger' and then 'make it smaller'. Making it bigger is to make its capabilities bigger. For example, models like Chat GPT with billions of parameters have strong capabilities. Keeping their capabilities unchanged while making the parameters smaller, combined with various scenarios, can achieve better results in vertical fields," he explained.
Regarding the application of big models in the financial industry, Fei Haojun believes that as a data intensive industry, the financial industry has accumulated a massive amount of data in various aspects such as financial transactions, customer information, market analysis, risk control, etc., and has given rise to a large amount of efficient processing technology needs. At the same time, the financial industry continues to undergo digital transformation, and has done well in data collection, processing, cleaning, and other work in the past, with a good data foundation. Therefore, the financial industry is a field that has developed rapidly in the process of big model evolution.
"Illusion is currently the most common problem in large models." In Fei Haojun's view, learning a large model with massive data can generate subsequent text based on understanding the previous content, and produce "fabricated" content that is "out of nothing". This directly affects the reliability of financial analysis results, especially when communicating directly with users in practice, the stability is insufficient, which can cause certain harm to the user experience. "The fault tolerance rate in the financial industry is relatively low, and in various scenarios, it is difficult to truly achieve the use of 'to C' if there is an illusion."
Caixin News Agency: What is the future trend of financial big models? What challenges will there be?
Fei Haojun: The financial big model will become a financial expert and undergo significant transformation in the next 5 years
The wave of big models is sweeping across various industries, and the financial industry, which continues to undergo digital transformation, has become one of the optimal scenarios for big models to take the lead in landing and promoting their development. Regarding the future development trend of financial big models, Fei Haojun believes that currently, financial big models are still in the instrumental stage, helping to artificially promote efficiency improvement. In its vision, with development, the financial big model will replace some manpower and eventually become a "financial expert".
Fei Haojun stated that from a trend perspective, based on the same technological path, the future development of financial big models will be in line with the pace of AGI (General Artificial Intelligence), and there will be a significant transformation in the next five years. For the companies that are laying out the financial big model track, they believe that it is far from the competitive stage, and "everyone should cooperate and work together to make the financial big model better."
"With the continuous evolution and ecological advancement of the entire big model, technology companies, fintech companies, and financial institutions will inevitably move towards universal collaboration." Fei Haojun analyzed that top technology companies will complete the construction of the universal model, and on this basis, top fintech companies will complete the construction of vertical financial industry big models. Financial institutions or small fintech companies will focus on empowering the application layer of business, forming a situation of industry co construction.
Of course, technology needs to be imaginative in order to make breakthroughs, but it is even more important to be down-to-earth. Fei Haojun emphasized that the technological development of financial big models still needs to be down-to-earth. The financial big model technology must be strongly integrated with the business, and every stage of the entire practice process must empower the financial business process to bring value, thereby proving its own value, in order to continue to develop.
"How to collect, manage, and use data securely and compliantly is a fundamental issue in the big model ecosystem." Speaking of the challenges faced by the development of financial big models, Fei Haojun bluntly stated that the entire industry is facing a major challenge of data security compliance and privacy prevention. The usual practice in the industry is to establish a data fence, strictly desensitize all data, and build a security fence at the instruction level to perform multi-layer verification on all input instructions to prevent malicious use of large models.
At the ecological level, the continuous upgrading of models and the self innovation in this process are also challenges. In his view, the ability of financial big models to reach the level of financial experts still requires continuous derivation of data and technology, but this process will inevitably conflict with past systems. How to choose and discard past things requires certain decision-making.