Financial big model, listen to the wind in the distance

As an important application of AI technology in the financial industry, financial big models are facing the challenge of transformation from high-speed coverage to optimal value. This article deeply analyzes the development status, difficulties and future development direction of financial big models, and provides valuable insights for professionals in the field of financial technology. It is recommended for friends who are interested in financial AI applications, technological innovation and industry trends.

The financial big model is considered to be the first stop for the industrialization of the AI ​​big model.

From 2023 to now, the development of China's financial big model market has indeed confirmed this judgment. According to relevant data, there are currently nearly 20 financial big models in the Chinese market, and leading financial institutions have either applied AI big models or demonstrated their research and development and application plans for AI big models.

It should be noted that in the context of reducing costs and increasing efficiency, financial institutions currently need to take into account multiple digital investment goals such as mobile finance and autonomous replacement of digital systems. The resources that can be allocated to AI big models are not abundant, and the return on investment brought by AI to financial institutions is still very limited. Is it really worthwhile for the financial industry to efficiently launch big models?

I remember that in 2023, when I attended a fintech summit, I communicated with a bank representative. He said that the most fundamental motivation for financial institutions to build big AI models is worry. Before the advent of the mobile Internet era, the global financial industry generally underestimated the impact of new technologies. Subsequently, mobile Internet platforms such as Apple Pay, Alipay, and WeChat shared the dividends of the era. Preventing a similar situation from happening again is the underlying motivation that drives financial institutions to increase investment in big AI models.

If this is the case, it is not enough for the financial industry to simply move quickly towards a big model. It also needs to listen to the wind in the distance and be able to trace back medium- and short-term action plans from long-term goals.

Today, we will start from this point of view and talk about where the financial industry will go next after quickly using big models.

1. Financial big model, from high-speed coverage 1.0 to value optimization 2.0

From the global to the Chinese market, the big model changes brought about by generative AI have set off a wave of technological innovation in the financial industry in more than a year.

Internationally, OpenAI regards the financial industry as the first stop for the implementation of GPT technology. For example, it has cooperated with Morgan Stanley to launch an investment advisory robot based on GPT-4.

In China's financial industry, AI big models have been widely used at an unprecedented rate. In just over a year, leading banks, securities firms, and insurance companies have all completed the implementation of financial big models.

For example, the Industrial and Commercial Bank of China announced that it had built the first full-stack, self-controlled, 100 billion-level AI large model technology system in the industry, and achieved innovative applications in multiple financial business fields. For example, in branch operations, it launched an intelligent assistant for branch employees based on large models. Agricultural Bank of China AI launched ChatABC, an AI large model application similar to ChatGPT, and conducted an internal pilot in the technology question-and-answer scenario. Postal Savings Bank of China connected to Baidu's "Wenxin Yiyan" to apply large language models in scenarios such as intelligent risk control, intelligent operations, intelligent investment research, and intelligent marketing.

At the private bank level, MyBank applies big model technology to industrial finance to identify the credit profiles of small and micro enterprises. In the insurance industry, ZhongAn Insurance has built the "Zhongyou Lingxi" system, which brings big models into business scenarios such as intelligent customer service, expiration reminders, and intelligent operations.

From the perspective of technology and solution providers, the Chinese market has already shown a variety of financial big model technology supply channels. Tencent Cloud and Ant Financial have both launched financial big model solutions. The financial industry requires both the implementation of big models and the update of infrastructure. In 2023, Huawei launched a financial-grade PaaS solution based on the Pangu big model, and released financial big model solutions for 10 scenarios in three categories: AI for Data, AI for Business, and AI for IT.

At the level of open source big models, Du Xiaoman open-sourced the Xuanyuan big model, which was trained based on the Bloom big model with 176 billion parameters. It has been applied to various business scenarios of Du Xiaoman, covering scenarios from marketing, customer service, risk control, office to R&D, and has achieved a leading position in a series of big model evaluations.

It can be seen that the financial big model has quickly gone through the 1.0 era characterized by high-speed coverage. In the first stage, the financial big model-related technologies and solutions were quickly launched, and leading financial institutions competed to try them out, laying a good foundation for the development of China's financial big model.

The findings of the first phase of the financial big model are obvious. For example, the technology is developing rapidly and users are highly motivated. At the same time, the financial big model covers a very complete range of industries, from banking to insurance, securities and other fields. In addition, the relevant technical capability supply chain has been improved. Both closed-source and open-source models are taken into account, and multiple deployment methods are complete. The software and hardware infrastructure supporting the financial big model is already relatively complete.

But in the future, the financial big model needs to move from the exploratory 1.0 stage to the 2.0 stage which must require a rate of return and reflect the value of long-term development.

During this stage, the challenges faced by big financial models will become more complex, and issues of strategic methodology will also surface.

2. Challenges that have been exposed

Looking at the implementation process of the financial big model, we can find that the logic of advantages and disadvantages is exactly the same as the previous implementation of AI technology in the financial industry. In the first stage, the industry will intuitively feel that AI is very useful. But in the future, whether AI can bring sufficient return on investment and whether it can penetrate into the core of the business will be a greater challenge.

At present, financial big models have also encountered similar problems. First and foremost, the problems of intelligence illusion and data pollution brought by big models are difficult to match with the extremely high requirements of the financial industry for professionalism and security.

Secondly, due to the extremely high confidentiality level and security compliance requirements of the financial industry, privatization is often required for large model deployment, and the model is prohibited from learning large amounts of data and excessive data flow is prohibited. This has given rise to two problems. One is that privatization deployment has brought great pressure on R&D and operation and maintenance costs to financial institutions. The other is that the high security level and restrictions on data use have brought about problems such as poor performance of large financial models.

For financial institutions, not only is the cost of developing and training large models too high, but the cost pressure of deploying models in scenarios is also great. Since financial large models are still in the exploratory stage and difficult to bring actual business feedback, they are often deployed within the institution or in marginal businesses. This leads to the continuous increase in deployment costs, but the commercial value has not been released.

Finally, there is a gap between the high security of the financial industry and the immaturity of big model technology. Financial big models are generally not deployed in core financial businesses. For example, intelligent customer service supported by big models not only cannot replace customer service personnel, but may also require frequent wake-up of manual customer service, and require customer service to re-understand user needs. On the industrial application side such as risk control, although big models have shown great potential, they are still unable to truly handle more complex risk anomalies.

In this case, the big financial model is likely to be marginalized like many financial technologies after it has been running for a period of time and its popularity has faded.

How can we break through the long-standing barriers between technology and industry? This may require the financial big model to hear some wind from afar.

3. Looking back at the present from the future

In the famous book "Banking 4.0", Brett King discusses the return to first principles in the first chapter. That is, we need to return to the origin of the problem and the essence of design. In the financial context, we first need to understand why banks were designed, and then discuss how banks should develop.

If we want to break the barriers that financial big models may face, we must also have such an awareness: to think, to judge, and even to assume what the future big models can bring to finance, and then work backwards from the future to the present. Even if this future is relatively far away, at least the wind in the distance will not lead us into a dead end.

So, what can AI big models bring to finance? This question can be understood from two dimensions: one is the technical dimension and the other is the application dimension.

From a technical perspective, the technical upgrade path for large language models has been relatively determined, and the industry has reached a clear consensus on its technical development paradigm and engineering practice.

For example, we recently read "Big Language Model: Principles and Engineering Practice" written by Yang Qing, Executive Chairman of Du Xiaoman Financial Technology Committee and General Manager of Data Intelligence Application Department. In addition to clearly proposing a number of key technologies of the big language model, the book also points out its characteristics in emergence and reasoning capabilities, as well as the widely-watched large language model scaling law. As the model training and inference capabilities improve, the big language model will continue to show intelligent emergence effects. This technical anchor is the long-term value pursued by the financial industry, and it is also the value concern of financial technology suppliers such as Du Xiaoman in exploring big language models.

From the application dimension, there are many application scenarios of financial big models, but they can be summarized into three aspects:

1. Intelligent customer service for the general public - Intelligent Advisor.
2. Intelligent credit identification for industrial users - credit system construction.
3. To enhance the capabilities of employees within financial institutions.

The ultimate of these three capabilities is the long-term value that the financial big model may bring. For example, in the future, intelligent customer service driven by a big language model may become an intelligent consultant or a one-to-one financial service expert. It can not only complete basic functions such as business docking and information notification, but also provide customized financial solutions based on user needs, realizing the transformation of financial customization capabilities from large users and corporate users to inclusive users.

Combining the technical and application dimensions, we can anchor the future of the financial big model in three aspects:

1. Intelligent and ubiquitous corporate and personal credit services.
2. Financial customer service is less staffed or even unmanned.
3. The ultimate in comprehensive intelligent capabilities of financial institutions.

If mobile finance is about bringing banks and securities firms directly into the hands of users, bringing finance closer to users, then AI big model finance allows users to go further, allowing all of their demands and assets to be connected to financial services.

4. Use the distant wind to measure the present road

For some time, there has been a debate between theoretical and practical thinking in the field of financial big models.

The so-called "empty talk" means that financial institutions should pay more attention to the algorithm itself, pay attention to the performance of the model, and first launch the big model, and use the hammer of the big model to find the nail of the financial business. The practical approach requires starting from the application scenarios of the financial industry, giving priority to the security, compliance and cost control of intelligent financial applications, and then adapting and tailoring the capabilities of the big model on this basis, so that the big model can serve financial services.

To some extent, both have correct factors, but perhaps both can also add new ideas. Because both focus more on short-term choices and ignore the starting point of the financial big model, which is the cross-era upgrade of the industry and the pursuit of long-term changes like mobile finance, or even more profound.

The idea that needs to be supplemented in the financial big model is that in addition to being pragmatic and idealistic, it is necessary to focus on the future and deduce the current work step by step from the idealized goals that may eventually be achieved. Let the wind from afar blow the current steps. From this perspective, financial institutions need to take into account the efficiency of launching big models while taking into account more AI technology innovations that may appear at any time in the future. Make their own digital architecture and software and hardware infrastructure adapt to the AI ​​technology iterations that may come at any time.

Specifically, the financial model of "serving the future" may include the following three levels of actions:

1. Lay a solid foundation for the financial big model, including its own R&D system and open and effective cooperation with technology suppliers to avoid technology stranding due to over-conservatism.
2. Take into account the engineering and imagination of financial big models. Facing the long-term development of big language models, the financial industry cannot be limited to having big models, but needs to hone its engineering control over big models, so that it can refine big models to meet its own needs at any time, and actively start creative big model exploration. It is brewing changes internally and not letting go of external opportunities.
3. Build long-term and clear intelligent goals. For the financial big model, we must face up to its short-term limitations and long-term possibilities. Strategically return to the first priority principle of the big model itself, and then use this as a goal to reverse the progress of each step.

When big models appear in front of the financial industry, our first feelings are surprise and reverie. However, when big models are actually applied, people tend to focus too much on the present, such as whether the big model is used, the business scenarios covered, and the return on investment brought. At this time, the emerging challenges and unclear values ​​often breed contradictions, leaving financial institutions in a dilemma. As long as we let the wind from afar blow in and let the long-termism of financial intelligence reflect the present, most problems will be solved.

Author: Feng Ciyuan; Source public account: Brain Extreme (ID: 341401)