【Financial Analysis】DeepSeek breaks through the siege of AI Computing Power: Is it the "singularity moment" for the banking industry?

Recently, a salon with the theme of 'How does DeepSeek change the rules of AI? The high threshold of AGI is disappearing?' was successfully held. The salon was jointly organized by New Web Bank and the School of Management Science and Engineering, Southwestern University of Finance and Economics. Li Xiusheng, Vice President of New Web Bank, Wang Jun, Director of the Department of Computational Finance, School of Management Science and Engineering, Southwestern University of Finance and Economics, and Wei Hao, head of the Risk Science Department of New Web Bank, jointly discussed the technical secrets behind DeepSeek and its application prospects in the banking industry.

How does open source evolve compared to closed source?

When OpenAI's deepening collaboration with Microsoft sparked controversy over 'ecosystem monopoly,' and when NVIDIA is facing the strictest AI chip export control in U.S. history, DeepSeek's open-source strategy unexpectedly opened another door. Different from traditional closed AI models, DeepSeek's openness allows enterprises to use more advanced large models at a lower cost, enhancing the capabilities of intelligent assistants in multiple scenarios.

In the software industry, both open source and closed source models coexist with successful cases. Li Xiusheng uses Linux and Android as examples, stating that these two, as representatives of open source software, have greatly promoted the development of the operating system field. He also points out that Apple, as a model of the closed source model, has always maintained a leading position in high-end mobile applications. Although these institutions take different paths, they have achieved significant accomplishments.

From the perspective of absorbing contributors from around the world, I personally prefer the open-source model because it can bring together more wisdom and strength to jointly promote the advancement and innovation of technology. In the future, open source and closed source may continue to develop in parallel, but the potential of open source is worth looking forward to. Li Xiusheng expressed.

Wang Jun believes that open source and closed source are mutually integrated and competitive. Open source technology is open to the public, which can attract numerous developers to participate, promote rapid technological iteration, but its profit-making ability and business model are uncertain. Closed source, on the other hand, focuses on building its own moat and barriers to entry, requiring huge investment, and having a unique but less diverse business model. Both have their own advantages and disadvantages, so they may learn from and integrate with each other in practice, and form a competitive situation in certain areas.

From the perspective of market entities, DeepSeek, as an open-source, low-cost, and efficient large model, has had a significant impact on the leading technology companies in the market. "For closed-source large model companies like OpenAI, DeepSeek's pricing strategy forces them to rethink their business models and technological optimization directions. As for chip companies like NVIDIA, the release of DeepSeek proves that top-level reasoning can be done without necessarily relying on high-end GPUs, prompting such companies to consider how to adjust their investment logic and development models for AI infrastructure," said Wang Jun.

However, it is worth noting that general artificial intelligence large models face challenges in solving digital risk control problems. Wei Hao said, 'Although large models have wide-ranging capabilities, such as understanding issues, performing mathematical operations, and generating code, their performance in the vertical field of risk control is not satisfactory.' The reason is that the training of large models mainly relies on public internet data and code, lacking specialized data corpus training for the risk control field. Therefore, their logic may not fully align with the actual requirements of risk control.

Small and medium-sized banks counterattack with DeepSeek?

According to the research report of Zheshang Securities, the entire training process of DeepSeek-V3 used less than 2.8 million GPU (graphics processor) hours. In comparison, the training time of Llama3-405B released by Meta, the American Internet giant Metaverse platform company, is 30.8 million GPU hours. The training cost of DeepSeek-V3 is about 5.576 million US dollars, while the training cost of the language model GPT-4 released by OpenAI (the American Open Artificial Intelligence Research Center) for chat robot ChatGPT is as high as several hundred million US dollars.

Compared with the traditional large models that require tens of millions to billions of investments, the cost of DeepSeek's localized deployment can be as low as less than a million RMB. According to the latest news from the Ministry of Industry and Information Technology, three basic telecommunications companies have fully connected to the DeepSeek open-source large model. Currently, in the financial sector, from banks, funds to securities, many institutions are actively deploying DeepSeek. Since May 2024, the new online bank has applied the DeepSeek large model in the system development scenario, respectively building research knowledge question-answering assistants and code continuation assistants, reducing the time frontline engineers spend on looking up technical information during the development process.

Li Xiusheng believes that in the field of artificial intelligence, the emergence of DeepSeek has brought about two major conceptual changes. First, the emergence of DeepSeek has broken the obsession with 'great power comes great responsibility', no longer blindly pursuing extreme computing power. In the past, it was often believed that breakthroughs could only be achieved by stacking huge computing power, but DeepSeek has proven that efficient performance can also be achieved with lower computing power through algorithm and model optimization. Second, DeepSeek has further intensified the debate between open source and closed source. OpenAI popularized the concept of large models through ChatGPT, but its closed-source strategy limited the dissemination of technology. The emergence of open-source models like DeepSeek has lowered the technical threshold, allowing more institutions to apply large models. This change has profound implications for industries such as banking and other financial institutions.

"For the future, with the continuous advancement of technology and further cost reduction, large models will no longer be the exclusive luxury of large banks, but will be widely used in small and medium-sized banks and other financial institutions. This will bring important technological trends to commercial banks, driving them to develop more intelligently and efficiently." Li Xiusheng said.

In the digital risk control field of the banking industry, large model technologies such as DeepSeek have a wide range of applications. According to Wei Hao, the release of DeepSeek has made technicians very excited because it can rival the top reasoning model level of OpenAI, and the weights are open source, with a lenient license, and can be used locally and controllably.

Weihao described the practical use experience, "When dealing with unstructured data, large models like DeepSeek can enhance semantic understanding and text processing capabilities, allowing us to extract information from a wider range of data. In addition, the technology of general intelligent models can also be borrowed by risk control models to improve the accuracy of customer evaluation and make better decisions."

Wei Hao pointed out that the deep thinking ability of DeepSeek R1 can improve the ability to understand intent and semantics through the thinking chain training mode. This ability is not limited to Chinese, and it can also perform well in handling long contexts and complex intentions.

As a highly informatized industry, the banking sector has undergone several major changes in its computer systems. From replacing manual operations with computer systems to the emergence of mobile internet, banks have continuously restructured their business processes. Now, with the rapid development of artificial intelligence, banks are facing the challenge and opportunity of the fourth round of information system evolution. In this era of large-scale models, how should banks build adaptive intelligent technology application capabilities that suit themselves?

Li Xiusheng believes that the advent of the era of large models requires banks to rethink how to reshape their business management and processes from the perspective of the full application of artificial intelligence. Banks need to first consider how to build applications, and then consider how to organize data, improve data quality, perform label annotations, and apply external data. Overall, commercial banks need to think from a strategic level, taking into account factors such as computing power, data, algorithms, and applications.

According to his introduction, since its establishment, XWBank has fully applied artificial intelligence technology in anti-fraud and credit risk control, realizing efficient and large-scale loan processing. However, with the emergence of large models, banks have begun to consider exploring and trying in more areas. Currently, XWBank has applied large models in the customer service field, successfully replacing some manual customer service, and is trying the application of large models in marketing and post-loan management.

In addition to the banking sector, Wang Jun predicts that in the fields of manufacturing, climate risk prediction, computer, education, media and entertainment, etc., intelligent applications related to large models will have a significant improvement. Wang Jun pointed out, "In the manufacturing industry, large models can monitor the reliability of parts or batteries, predict their lifespan; in climate risk prediction, artificial intelligence algorithms interpret future weather conditions, providing warnings and route optimization for highways, etc.; in the computer field, large models can assist in code completion, code understanding, and construction; in the education field, based on students' learning habits and behaviors, personalized large models can be built to assist students' learning; in the media and entertainment field, large models can be used for content generation, model construction, and scene creation, such as animation production, game design, and short video production, and can also synthesize digital people for e-commerce recommendations, etc.

What kind of AI talent do banks need in the future?

According to the "2024 Annual Report on the Development of China's Banking Industry" released by the China Banking Association, there is a natural convergence between finance and artificial intelligence. AI large model technology can fully tap into the massive data of the banking industry, and the banking industry has rich scenarios suitable for AI large model technology. Currently, AI large models are promoting comprehensive innovation in the service, marketing, and product fields of China's banking industry, catalyzing the accelerated arrival of the "bank of the future".

With the widespread application of large models by banks, higher skills are also required for technical personnel. Li Xiusheng believes that in the Internet application industry, Internet thinking has contributed to the success of Internet giants. With the advent of the era of artificial intelligence, the demand for talent has shifted to finance and technology compound talents with artificial intelligence thinking.

In recent years, Xinye Bank has emphasized Internet thinking and will focus on artificial intelligence thinking in the future. When designing business products, customer marketing, daily operations, and building a comprehensive management system, artificial intelligence thinking is being integrated. Therefore, the bank will assess whether employees have this ability, foundation, or potential to cultivate talents needed for future bank development.

"The continuous advancement of artificial intelligence technology brings challenges to banking practitioners, but also provides new opportunities. Facing these changes, practitioners need to stay calm, keep learning, keep up with the times, and find their position in society and the business world," Li Xiusheng encouraged. "Technical personnel need to adjust themselves and apply artificial intelligence technology to enhance their abilities. Business personnel do not need to worry too much about being replaced because the application threshold of artificial intelligence technology has been lowered. Even people who do not understand computers can use AI tools to build processes and applications and demonstrate their value. Therefore, as long as they are willing to learn and keep up with technological changes, banking practitioners will not be eliminated but will be better able to adapt to the advancement of technology in the era."

From the perspective of risk control business, Wei Hao pointed out that practical experience is the key to mastering artificial intelligence. In the field of risk control, the application of artificial intelligence technology requires higher talent requirements. It not only requires a deep understanding of the technical principles, but also a full understanding of the advantages, capabilities boundaries, and risks of the model to ensure the correct application of the technology. Therefore, risk control personnel need to have a solid technical foundation and extensive knowledge.

Wang Jun also said that colleges and universities are also committed to cultivating compound talents in the field of AI + expertise. "We have optimized the curriculum to include courses such as data analysis, data mining, machine learning, deep learning, and multimodal data, so that students can be exposed to artificial intelligence at the undergraduate level. Practical training projects and experimental courses have been added, and students are encouraged to participate in competitions such as financial technology competitions to transform knowledge into practical ability. In addition, we hope to strengthen industry-university-research collaboration with the industry, and provide students with a deeper understanding of the needs and businesses of the industry through joint laboratories and expert lectures, so as to stimulate their learning motivation and cultivate talents that meet the needs of the industry. ”

When it comes to the next development trend of AI technology in the banking industry, Li Xiusheng stated that with the development of AI and large-scale model technology, commercial banks are ushering in a new round of reshaping. This not only involves system upgrades, but will also profoundly change the business processes, product forms, decision-making mechanisms, personnel composition, and job settings of banks. "Although the essence of financial risk management remains unchanged, the service methods, product forms, and operation mechanisms will undergo tremendous changes. This process may be gradual, but it is expected that the face of commercial banks will be completely renewed in three to five years."

(Article Source: Xinhua Finance)

Source: East Money

Author: Xinhua Finance