The difficulty of the financial big model lies in whether it can penetrate deeper into the industry; its subversiveness is also based on going deep into the industry, empowering the development of the long-tail scene of the financial industry, and regaining "financial trust".
Author | Sihang
Editor | Pi Ye
Produced | Industrialist
"From an economic point of view, the digitalization process of the entire financial industry is not at a uniform speed; from a technical point of view, the emergence of catalysts will accelerate the industry's entry into the deep water area of digitalization. And the big model is the strongest 'catalyst' in the past decade." Hang Seng Electronics Chief scientist Bai Shuo told us.
The big model is becoming the second wave to advance the digitization of the financial industry.
In 2013, Internet finance was born. In the following ten years, the financial industry has experienced two revolutions brought about by AI.
The protagonist of the first revolution is discriminative AI, such as helping financial institutions to conduct better intelligent analysis and decision-making. At that time, Internet finance was at the top of the wave. Paperless, online, mobile, and remote finance all promoted the transformation and innovation of the financial industry chain.
In the first wave of AI, the most significant change is that the financial industry represented by banks has undergone a transformation from old to new paradigms.
However, this wave of financial industry revolution was not complete. Although the clarion call of "financial trust" has long sounded, in the financial industry, digital acceptance is not high. The benefits of artificial intelligence have not been fully utilized in the financial industry.
Among them, there are various reasons such as technical problems, compliance factors, and industry barriers, which hinder the arrival of the financial industry revolution. Until 2023, the big model changes the situation a little.
Objectively speaking, with the arrival of generative AI, the industry is regaining "financial trust".
1. Is a large model just needed in the financial industry?
At present, financial institutions generally have a low degree of acceptance of digitalization, and it is more difficult to fully realize digitalization. But the digitization of the whole process is the prerequisite for financial institutions to introduce large models. If it is still only applied at the tool layer, the large model cannot better empower the development of the industry, and it is not very disruptive.
Hang Seng Electronics told us, "If the digital transformation maturity of financial institutions is divided into 0 to 5, where 0 represents the initial stage, and 5 represents a completely data-driven business model. Currently, most financial institutions are at level 2. And level 3, a small number of institutions have reached level 4, and some even partially reached level 5."
Among all financial institutions, "banks, especially leading banks, have the best performance in digital transformation, followed by securities firms."
The reason why banks have the highest degree of digital acceptance is because banks involve many customer service and risk monitoring scenarios. Brokers are different, and their more application scenarios are in intelligent decision-making. These two different scenarios are precisely the areas where discriminative AI and generative AI are good at.
Specifically, the discriminative AI directly maps the input to the output, and predicts the output label by learning the characteristics of the input data, while there is no enhanced learning process of the generative AI between the input and the output. Therefore, discriminative AI is more used for tasks such as classification and regression, such as image recognition and speech recognition.
Generative AI is different. Its biggest advantage lies in the process of enhancing learning. Generative AI can learn the statistical characteristics of samples from existing data and generate new data on this basis. Therefore, in financial scenarios, it is more suitable for intelligent decision-making, and through the financial knowledge and news input in the large model, it can give suggestions on business marketing and venture capital.
This means that with the blessing of the AI model, there will be some changes in the financial industry that did not exist before.
As Bai Shuo said, the large model is the most intuitive "catalyst" that has had an impact on the financial industry for many years. Compared with metaverse, blockchain and other technologies, the large model can go deeper into the vertical field, subvert the industry, and bring actual value . Among them, the most intuitive impact is to bring a new way of working to the original position.
"For example, the changes in the positions of data analysts are very prominent. In the field of investment research, data analysts need to conduct data analysis to form content based on public data such as financial statements, public information, and research reports. The performance of large models in such data processing capabilities Very good, it can replace part of the jobs." Bai Shuo told the industrialists.
However, due to the obvious shortcomings of large models in terms of accuracy, timeliness, and professionalism, it is currently difficult to realize deeper value in the financial industry. At present, what the large model can do more is to provide a very friendly interaction between man and machine, and it still needs professional manpower to complete the financial professional work.
It can be said that in addition to rich imagination, as far as the moment is concerned, the more practical value brought by the large model to the financial industry is more reflected in some more interactive scenarios.
Banks are already taking action. In March this year, ICBC released the first general model for the financial industry based on Ascend AI. At the press conference, ICBC announced that the model has been applied in customer service, risk prevention and control, and operation management. For example, ICBC uses this model to support intelligent customer service to answer customer calls; another example is to use the large financial model to monitor the progress of the construction of industrial engineering financing projects.
Or it can also be said that the significance of the large model to the financial industry, before accelerating digital intelligence and regaining "financial trust", the more obvious change is the landing of the long-tail scene.
2. Where has the big financial model gone?
In half a year, all major Internet companies have entered the game; banks, securities companies and other financial institutions have also exited.
The reason why the financial large model is called "spire technology" is not only difficult in technology and compliance, but also in data and domain experience. In other words, the construction of a large financial model cannot be accomplished overnight, but requires certain conditions.
Taking major Internet companies as examples, Baidu, Tencent, Ali, and 360 can be regarded as the most qualified financial model leaders with their years of experience in fighting black and gray industries and their deep cultivation in the AI field.
Du Xiaoman was the first to make a move. On May 26, Du Xiaoman officially open-sourced the large Chinese financial model "Xuanyuan". Different from Wenxin Yiyan, the Xuanyuan model is the result of Du Xiaoman's long-term cultivation in the financial field, and has more high-quality trainable data. For financial large models, the quality of data in the financial field directly determines the performance of all aspects of the Xuanyuan large model.
In addition, from the perspective of the number of parameters, according to the official introduction, the Xuanyuan large model is trained on the basis of the Bloom large model with 176 billion parameters, and Xuanyuan also integrates financial noun understanding, financial market commentary, financial data analysis and financial news understanding and other data.
The next thing that spread the wind was the Ant Group. On June 21, it was reported that the technology research and development team of Ant Group is developing its own language and multimodal large model, which is internally named "Zhenyi". In response, Ant Group responded that "the news is true."
On the one hand, Ant Group’s confidence comes from Alipay’s years of industry experience in the financial field; on the other hand, it comes from Ant Group’s investment in trusted AI technology research in 2015. In 2016, Ant Group fully launched the AI intelligent risk control and defense strategy, which has been implemented in multiple scenarios such as anti-fraud, anti-money laundering, anti-theft, joint enterprise risk control, and data privacy protection. In the past two years, Ant Group has stepped up its layout in the AI field.
As early as 2019, the Basic Theory Research Center of Tsinghua AI Research Institute was established. The center's chief scientist Zhu Jun and his team released the third-generation artificial intelligence platform RealAI at the same time, and it is deeply integrated with financial, industrial manufacturing and other industry applications. And just two days before Ant Group announced its self-developed "Zhenyi", the new team led by Zhu Jun completed an angel round of financing of nearly 100 million yuan, led by Ant Group.
Finally, Tencent and 360 have also recently joined forces with the Institute of Information and Communications Technology to compile domestic financial large-scale model standards. For Tencent, over the past 20 years of experience in confrontation between black and gray industries and thousands of real business scenarios, these have given Tencent the most authentic industry data. And 360, which has always had the title of "safety guard", is no exception.
In addition to Internet vendors, there are also database vendors, such as Transwarp Technology, that are deploying in the direction of large-scale models in the financial field.
For the construction of large financial models, database vendors and Internet vendors are taking two completely different routes. The advantages of Transwarp Technology are twofold.
The first is the process, that is, the data "cleaning" and other processing involved in the model training process. As a database manufacturer, Transwarp Technology has a rigorous methodology for data processing, especially for heterogeneous data unique to the financial field.
In this regard, while developing the financial large-scale model "Wuya Infinity", Transwarp Technology also provides a one-stop enterprise self-built large language model tool chain. The tool chain includes the vector database Hippo, which is closely connected with the application of large language models, and a series of underlying processing technologies for databases. Among them, the most notable is the vector database Hippo.
In the financial field, data timeliness is one of the challenges in the implementation of large models. How to input real-time data such as emergencies and financial information into the large model is directly related to whether the financial large model can accurately analyze and make decisions. The vector database is the key to solve this problem.
The second major advantage of Transwarp Technology is its field data and industry know-how accumulated in the financial field for a long time.
Although Internet vendors and database vendors each have advantages in industry experience and model technology, vertical vendors should be the most equipped to build financial models. Because such manufacturers have high training model data, such as "Hang Seng Electronics", a manufacturer dedicated to providing financial digital solutions.
On June 28, Hang Seng Electronics released LightGPT, a large model of the financial industry. It is understood that the model uses more than 400 billion tokens of financial data (including information, announcements, research reports, structured data, etc.) and more than 40 billion tokens of language-enhanced data (including financial textbooks, financial encyclopedias, government reports, regulations Regulations, etc.), and supports fine-tuning of more than 80+ financial-specific task instructions, thereby enhancing LightGPT’s ability to understand in the professional field.
Bai Shuo said that for large financial models, the most important thing is data quality, that is, the size and quality of data for large model training, because this is related to what the large model can output. When the underlying technologies are similar, data quality is key. The second is engineering capability and industry experience. Among them, engineering capabilities include the selection, cleaning and transformation of data. For example, when the performance of the large model is not satisfactory or there are problems, the manufacturer knows how to judge which data is missing and which data needs to be supplemented, so as to improve the quality of the large model data .
However, in the process of landing a large financial model, the most important challenge that cannot be ignored is the security issue, that is, the choice between public cloud and local deployment.
In the financial field, a lot of data involves compliance, privacy security, and even regulatory issues, which cannot be disclosed, so it is difficult to go to the cloud. For example, Industrial and Commercial Bank of China, Agricultural Bank of China, Postal Savings Bank of China, China CITIC Bank, Industrial Bank, Bank of Jiangsu, Bank of Suzhou and many other banks and brokerages have chosen to access the general-purpose large model, that is, to build a large-scale model in a dedicated field by local deployment.
If you choose the local deployment method, you will inevitably face some difficulties, such as computing power challenges and parameter quantity problems. Whether or not the locally deployed financial institutions have sufficient computing power is one aspect, and another is whether the number of parameters is large enough. If the number of parameters is not enough, even if high-quality data is input, large models cannot "emerge".
For various reasons, manufacturers entering the large financial model face many obstacles.
3. Explore the value in depth of the industry
But there are still many problems. Even in overseas countries where the financial industry is more developed, the landing of large models is still a big challenge.
Source: Atom Capital
It is not difficult to see from the above figure that the financing amount of start-up companies is generally small; and except for the more well-known YC, there are not many star capital.
In China, at least for now, accuracy, timeliness, and security are the three major challenges facing the implementation of large financial models.
In terms of accuracy, large models cannot give expert-level answers in professional fields, especially when it comes to people's livelihood and economic issues. Bai Shuo told the industrialists, "From a technical point of view, we don't think that AGI can develop professional capabilities in a certain field, and professional things need to be handed over to experts. But what the large model can provide is the ability to connect humans and machines. If the two Combining the two can play a greater role.”
Another major challenge is timeliness. The process of data generation itself is fluid, and the data in the market is about accuracy, quality, and timeliness. "In terms of data timeliness, the training cycle of a large model itself determines that it is impossible to have timeliness, so supplementing time-sensitive data is a necessary condition for large financial models." Now many self-developed financial large models Many manufacturers have used vector databases to solve this problem.
Finally, and the most important challenge faced by large-scale models in the current domain, that is, data security. Since the data collected by the large model comes from public data, what the industry large model needs is domain data, and even some proprietary data such as research reports and papers that are not available in public channels.
In this regard, some enterprises and institutions choose to disclose the data, but more choose to deploy the large model locally. And this leads to another question, whether technical problems such as computing power challenges, parameter quantity problems, and engineering algorithms can be solved.
According to Bai Shuo's observation, some gaps in language skills can be resolved within 2 to 3 years, and the gaps between different large-scale model capabilities can also be bridged. The rest of the question depends on whether the big model can be embedded in a deeper industry to provide value.
Judging from the application scenarios of the current financial large model, the value provided stays more at the tool level. Specifically, the financial big model goes a step further on the basis of the traditional AI model, using high-quality knowledge data and intelligent attributes, and applying it to highly interactive scenarios.
But from a larger perspective, with the implementation of standards for large financial models, issues such as data compliance, privacy security, and training techniques have been resolved one by one. value. After challenges such as accuracy, timeliness, and security are eliminated, the financial model will work with "experts" to solve problems that cannot be solved at the moment and bring greater industrial value.
The difficulty of the financial model lies in whether it can penetrate deeper into the industry; its subversiveness is also more based on going deep into the industry to empower the digital development of the financial industry.