Text |Observer Mission** New Economy Observer Mission
Since ChatGPT detonated the market at the end of 2022, the trend of large models has been blowing for a whole year.
As a naturally data-intensive and technology-driven industry, the financial industry has become one of the hottest testing grounds for large models, which have been flooded after the release of the world's first financial model, BloombergGPT.
In China, from financial institutions to Internet giants, to fintech companies such as Ant Group and Sina Digital, they have successively bet on their own financial models, and have shown their own tricks in scenario applications and algorithm models, exploring the implementation of large models in the whole financial chain.
However, when more and more financial models are pushed to the public, the attitude of the market and manufacturers has also changed from the initial fanatical pursuit to the rational thinking stage after "disenchantment".
The reason for this is the barrier between the ideal and the reality: although the particularity of the financial industry determines that it is one of the best scenarios for the implementation of large models, due to the very strict requirements for the accuracy and compliance of information and data, most financial large models only stay at the junior assistant level and are not deeply embedded in the core links of the business.
Therefore, in the second half of the financial model, the competition point of institutions has shifted to reducing the "illusion", that is, to truly realize large-scale application, deeply integrate with scenarios to solve practical problems, and bring about a jump in industrial value. At present, the participants have consciously moved towards ecological co-construction to face the landing problems together.
Burst
At the end of March this year, when the market was still in the midst of a carnival of large-scale models, Bloomberg's BloombergGPT was born, burning the AI boom detonated by ChatGPT to the financial circle. Subsequently, Morgan Stanley announced the adoption of GPT-4 to manage its vast internal knowledge base, adding fuel to the fire of the boiling financial industry.
The dynamics of overseas institutions are like the wings of a butterfly, quickly transmitting the heat wave to the other side of the ocean.
At that time, Yang Zeyuan pointed out in the research report of CITIC ** that overseas financial giants represented by Bloomberg and Morgan Stanley actively made efforts to develop and apply large models, forming a strong demonstration effect of large models + finance. It is expected that with the gradual breakthrough of domestic large models, generative AI applications in the domestic financial industry are expected to gradually open.
In the past few years, with the wave of financial digitalization, domestic financial institutions have deeply embedded AI technology capabilities in business processes such as customer acquisition, risk control, investment research, consumer protection, and customer service, and built a basic disk for intelligent transformation. However, based on powerful content generation, logical reasoning, rapid iteration, and even decision-making capabilities, large models are expected to reconstruct the original infrastructure and management system of financial institutions, and greatly expand the imagination space for financial digital and intelligent transformation.
Wang Xiaohang, vice president of Ant Group and head of the financial model, judged that the model is bringing experience changes to the financial industry, "Every key function in the financial business chain is worth redoing with the big model technology." ”
McKinsey calculates,In the future, the overall income of financial enterprises will be3% to 5%.Can be generativeaigenaiProvided. Overall, the GenAI use case has the greatest value potential for the banking industry's frontline distribution, customer operations, technology, and legal, risk, compliance, and fraud departments, accounting for 70% of the overall value pool.
As a result, large models in the domestic financial field have emerged one after another, and major business directions such as data analysis, risk control enhancement, intelligent customer service, and investment consulting are also emerging.
In this huge AI feast, participants can be roughly divided into three categories:
One is financial institutions. Industrial and Commercial Bank of China, Agricultural Bank of China, Bank of Communications, China Merchants Bank, Zheshang Bank and many other banks have put forward the relevant deployment of large modelsIn August, the consumer finance company immediately released the first retail financial model "Tianjing".
The second is a major Internet company. Huawei's Pangu model, Qianfan model, Tencent Cloud industry model, iFLYTEK Xinghuo model, 360 intelligent brain model and other general models all take finance as an important driving force.
Among them, on December 14, Volcano Engine and Zhipu AI jointly released a high-performance financial model and started testing in an all-round way.
In addition, in July this year, Tencent Cloud Industry Model announced its solution for financial risk control scenarios, releasing the financial risk control model for the first time, integrating Tencent Security's extensive experience in risk control modeling, as well as massive fraud knowledge and multi-scenario risk control model capabilities accumulated over the past 20 years. In November, Tencent Cloud officially released the financial industry model solution, helping each financial institution to have its own large model.
The third is fintech companies. Du Xiaoman Financial, Qifu Technology, Transwarp Technology, Ant Group, Sina Digital, Lexin, Hang Seng Electronics and other companies have launched their own financial models.
Further from the perspective of industry application landing, major modelsIt mainly focuses on content information, product introduction, research report generation, virtual customer service interaction, and anti-fraud and other fields.
In terms of Du Xiaoman, large model technology has been applied to various business scenarios, from marketing, customer service, risk control, office to research and development, and has achieved initial results. Among them, in the field of customer service, the large model has promoted the service efficiency by 25%;In the field of smart office, the current intent recognition accuracy rate of large models has reached 97%.
Sina Digital has actively promoted the application of cutting-edge technologies in the financial field, and has carried out the practice of large-scale model application in many fields such as intelligent customer service, marketing design and R&D efficiency improvement, and has achieved certain results.
Among them, in terms of intelligent customer service, Sina Digital has built its own intelligent customer service assistant by using the fine-tuning technology of general large language model, combined with the knowledge base data accumulated in the field of financial customer service for many years and professional customer service experience.
Through intent recognition combined with dedicated APIs, the assistant can directly face users, respond to user requests efficiently 24 hours a day, accurately understand user intentions, resolve problems, and retrieve knowledge bases and obtain user business information through advanced capabilities, so as to provide users with more personalized services. In addition, the intelligent customer service assistant can also collaborate with the human agent to support the human agent in summarizing historical communication content and providing relevant suggestions.
In the R&D efficiency improvement scenario, the R&D team of Sina Digital Technology has introduced a large-model-based Copilot auxiliary tool after stripping away the sensitive, which can effectively help R&D personnel improve efficiency and quality in the whole development process such as demand analysis, architecture design, writing and testing.
Fences
However, with the launch of the financial model, its implementation in actual business is much lower than the concept at the beginning of the year.
Liu Shufeng, chairman of Hang Seng Electronics, recently revealed that more than 70% of financial institutions are in the large-scale model research stage, 8% are in the project approval stage, and 17% are in the testing stage. Only a small number (less than 10%) of customers are in the process of actual application.
From the perspective of the market, the industry is generally onThere are many applications in generative scenariosHowever, it is difficult to implement financial scenarios involving decision-making. That is to say,There is still a long way to go before the large model is deeply integrated into all endpoints of the financial business, releases due value, and then reshapes the production relationship of the financial industry.
Lv Zhongtao, chief technology officer of ICBC, said that the current stage of large models is not mature, and there are still problems such as ethical risks in science and technology. Therefore,It is not recommended to use it directly with customers in the short term
The reason for this is that the data of financial institutions is highly sensitive, involving customer privacy and financial security, and must be under strong supervision, which requires more refined and demanding indicators such as security, stability, compliance, accuracy, and reliability, and also poses more challenges to manufacturers' data storage and analysis capabilities, compliance and legality, and financial strength.
Jiang Ning, chief information officer of immediate consumption, believes that the biggest difficulty in generating a large model is full of economy, and you can not take risks if you answer wrong, but the most important model of the financial model is discriminatory, and you need to make trading decisions, and 1% of the mistakes will cause customer losses, which is the biggest difference between the financial model and the traditional model.
Wang Xiaohang said, "Although the ability to understand and generate large models is strong, there are still many challenges when encountering professional and rigorous industries. In addition to solving the illusion of large models, it is also necessary to pay attention to the training of large models in financial compliance and industry value proposition. ”
And for a long time,The data of financial institutions is generally siloedThe "chimney" type of fragmentation is not highly liquid。Therefore, when deploying large models, most financial institutions tend to choose to train in a secure and confidential state in a private cloud to meet strict data security and compliance requirements.
Therefore, in the process of promoting the implementation of large models, institutions often face a number of pain points such as high computing power costs and hidden concerns about data security. Specifically, the technical capabilities of various types of participants are different, and they also need to carry out massive data governance and data cleaningPrivate cloud training by institutions also causes a waste of computing power, which is costly and unaffordable for small and medium-sized financial institutions.
Guosheng** has estimated that the cost of GPT-3 training is about $1.4 million, and for some larger LLM models, the training cost is between $2 million and $12 million.
Co-construction
In the face of the dilemma of the implementation of the financial model, the obvious trend in the industry is that more and more institutions are embracing cooperation and interoperability to achieve a win-win situation through ecological co-construction.
Yao Qian, director of the Science and Technology Supervision Bureau of the China Securities Regulatory Commission, pointed out in an article that large models require huge computing power support and strict data governance, and it is often difficult for ordinary institutions and application departments to support the operation and iterative upgrading of large models. For this, it is necessary to build oneThe ecology of healthy interaction and co-evolution of various modelsto ensure that the artificial intelligence industry related to large models can be successfully implemented in various application fields.
We have also seen that various participants in the financial model in the market have jointly explored the successful implementation of the model by strengthening cooperation.
In the battle of "model", standards come first, and Tencent Cloud has actively taken the lead in formulating standards and specifications. As early as July, it and the Chinese Academy of Information and Communications Technology jointly launched the joint promotion plan of industry model standards, and jointly led the compilation of China's first financial industry model standard.
On November 30, the IEEE Financial Risk Control Model Standard Kick-off Meeting was held. The standard was initiated by Tencent and is the world's first large-scale international standard in the field of financial risk control. The conference was held under the guidance of the international authoritative standards organization IEEE, and academic institutions such as the China Academy of Information and Communications Technology, as well as WeBank, Instant Consumption, Du Xiaoman, Zhongyuan Consumer Finance and other institutions attended and jointly participated in the formulation of standards.
In addition, the "strong alliance" between financial institutions and leading large model manufacturers is also accelerating.
In August, Bank of Communications, Huawei, Tencent Cloud, and iFLYTEK announced the establishment of three joint innovation labsIn September, Zheshang Bank and Huawei signed a deepening strategic cooperation agreement to deepen cooperation in comprehensive financial services and AIGC scenario applications, so as to achieve a new pattern of resource sharing, complementary advantages, and mutual benefitIn November, Tencent Cloud released the financial industry model solution, and the first batch of China UnionPay and 11 partners announced their participation in the ecological co-construction of Tencent's financial model.
For most small and medium-sized financial institutions and fintech companies, a more ideal path is to introduce the leading basic large model of third-party vendors, fine-tune it on the basis of their own samples, build their own professional large model, quickly empower business processes, and strive to achieve corner overtaking in this model war.
Xu Dongliang, CTO of Du Xiaoman, believes that the industry model will help small and medium-sized financial institutions that actively embrace the model and narrow the technical gap with the head institutions."Everyone is back on the same starting lineThis is an opportunity for small and medium-sized institutions to bridge the 'digital divide' and the 'smart divide'."
According to Tencent Cloud's data, Dongfeng Nissan Financial Leasing has completed customized risk control modeling with only a small number of samples with the help of Tencent Cloud's risk control model, saving 70% of the modeling time, so that the bottom risk control model has a solid risk control immunity to support the development of financial business.
In terms of Sina Digital, it is also actively exploring the use of general models combined with the best libraries in the financial field to privatize the deployment of auxiliary tools to further improve performance and information security.
According to Gartner's latest Top 10 Strategic Technology Trends for 2024**, by 2026, more than 80% of enterprises will use generative AI's APIs or models, or deploy generative AI-enabled applications in production, up from less than 5% at the start of 2023.
It is believed that in the near future, with the joint efforts of the participants in the financial model and the support of national policies, the "barrier" of the implementation of the large model will be crossed, and the implementation of multiple core business ends, and the exponential production potential will be released, and the financial business model will be comprehensively reconstructed.
Disclaimer: The publication of this article by the New Economy Observation Mission is for the purpose of conveying more information and does not constitute any advice. Original articles are not allowed without authorization**.