The global landscape for AI chips is rapidly evolving, and the emergence of the domestic large model DeepSeek is poised to redefine the foundational logic of technological evolutionNow, with a skyrocketing increase in usage, DeepSeek has become the fastest application to exceed 30 million daily active usersAs an inquiry into the AI frenzy continues, discerning the genuine potential of DeepSeek's technology has become paramount.

The debate around open-source versus closed-source models has intensifiedWhile the partnership between OpenAI and Microsoft has sparked discussions regarding "ecological monopoly," and as Nvidia faces the strictest AI chip export controls in U.S. history, DeepSeek's open-source approach has unexpectedly opened another avenueIn contrast to traditional closed AI models, DeepSeek's accessibility allows businesses to utilize advanced large models at a lower cost, enhancing capabilities across multiple scenarios of intelligent assistance.

In the software industry, both open-source and closed-source models coexist, each boasting its success storiesLi Xiusheng cites Linux and Android as prime examples of open-source software that have significantly propelled the development of operating systemsConversely, Apple exemplifies a closed-source model, continually leading high-end smartphone applicationsDespite their divergent paths, both have achieved remarkable success.

“From the perspective of absorbing global contributors, I personally favor the open-source model, as it harnesses collective intelligence to drive technological advancement and innovation forwardThe future will likely see open-source sharing the stage with closed-sourceHowever, the potential of open-source remains highly promising,” Li expresses.

Wang Jun argues that open-source and closed-source are interwoven, with elements of competition and collaborationOpen-source technology attracts many developers and accelerates rapid iterations, but it comes with uncertainties surrounding profitability and business models

Advertisements

Closed-source focuses on building its protective barriers and requires significant investmentEach model presents advantages and disadvantages; hence, they may borrow from one another in practice and may manifest competitive dynamics in specific fields.

From a market perspective, DeepSeek acts as an open-source, cost-effective, and efficient large model, shaking up major tech companies. "For closed-source model firms like OpenAI, DeepSeek's pricing compels them to reevaluate their business models and technical optimizationsFor chip companies like Nvidia, DeepSeek's arrival demonstrates that top-tier inference is achievable without relying solely on high-end GPUs, forcing such companies to rethink their investment logic and growth strategies in AI infrastructure," Wang admits.

Nonetheless, it is crucial to recognize the challenges faced by general-purpose AI large models concerning digital risk controlWei Hao mentions, “While these large models exhibit expansive capabilities like understanding inquiries, performing mathematical calculations, and coding, their performance in the specific area of risk control is not satisfactory.” The root cause lies in the fact that the training of large models primarily relies on publicly available internet data and code, which lacks specialized datasets for risk control, potentially leading to a mismatch with the actual needs of the field.

Can small to medium-sized banks turn the tide using DeepSeek? How can they cultivate intelligent technological application capabilities?

According to Zhejiang Merchants Securities' research report, the entire training process for DeepSeek-V3 required fewer than 2.8 million GPU hoursFor context, the training duration for Meta's Llama3-405B model reached 30.8 million GPU hoursThe training cost for DeepSeek-V3 is approximately $5.576 million, whereas OpenAI's GPT-4 digital model for ChatGPT amounted to hundreds of millions of dollars.

In stark contrast to the traditional huge investments required for large models, localized deployment of DeepSeek can incur costs under one million yuan

Advertisements

Recent announcements from the Ministry of Industry and Information Technology indicate that three major telecom companies have fully integrated DeepSeek's open-source large modelPresently, within the financial sector, multiple institutions ranging from banks to funds and securities are rapidly adapting to DeepSeekStarting from May 2024, New Internet Bank has begun employing the DeepSeek large model in system development scenarios, creating both a knowledge question-and-answer assistant for research and a code completion assistant, shortening the time it takes for engineers to retrieve technical material.

Li Xiusheng believes that DeepSeek's entry into the artificial intelligence domain has ushered in two significant conceptual shiftsFirst, it dismantles the fixation on the concept that “great power yields miracles,” shifting the focus from blindly pursuing extreme computing powerPreviously, there was a prevailing belief that only by stacking vast computing resources could breakthroughs be achievedHowever, DeepSeek illustrates that efficient performance can be obtained through algorithm and model optimization, even with lower power usageSecondly, DeepSeek further intensifies the open-source vs. closed-source rivalryOpenAI popularized the large-model concept via ChatGPT, yet its closed-source approach limits technological disseminationThe rise of open-source models like DeepSeek has lowered technological barriers, allowing more institutions to apply large modelsThis shift significantly impacts financial institutions such as banks.

“In the future, as technology progresses and costs drop further, large models will no longer remain exclusive luxuries of major banks but will be broadly applicable to small and medium-sized banks and other financial institutionsThis will herald a crucial trend of technological transformation for commercial banks, propelling their movement toward greater intelligence and efficiency,” Li asserts.

In the domain of digital risk control within banking, the application prospects for DeepSeek and similar large model technologies seem extensive

Advertisements

Wei Hao shares, “The release of DeepSeek has generated significant excitement among technicians because it can compete with the top inference models from OpenAI, all while being open source, having a flexible license, and allowing for controllable, local usage.”

He elaborates on practical experiences, stating, “When dealing with unstructured data, models like DeepSeek enhance semantic understanding and text processing capabilities, enabling us to glean information from a broader set of dataFurthermore, the technology behind general intelligent models can serve as a reference for risk control models, enhancing customer assessment accuracy and facilitating better decision-making.”

Wei highlights that DeepSeek R1's deep thinking capabilities can be enhanced through its chain of thought training mode, yielding improved intent and semantic comprehensionThis competency transcends language boundaries, also exhibiting remarkable performance when engaging significant context and complex intentions.

As an industry with a high degree of information technology dependence, banking has undergone several transformative milestones, from deploying computer systems to replace manual operations to the advent of the mobile internet, continuously reinventing their operational processesIn light of the rapid advancements in artificial intelligence, banks are now navigating a fourth wave of information system evolution and face challenges and opportunitiesHow, then, should banks cultivate intelligent technological application capabilities suited to their needs within the era of large models?

Li Xiusheng asserts that the arrival of the large model era necessitates that banks rethink from the standpoint of comprehensive AI application how to remodel their operations and management processesBanks should first consider how to establish applications, followed by organizing data, enhancing data quality, labeling, and utilizing external dataOverall, commercial banks must consider the strategic layer while attending to multiple dimensions, including computing power, data, algorithms, and applications.

As he elaborates, New Internet Bank has extensively applied AI technology in fraud detection and credit risk control since its inception, achieving efficient and large-scale loan processing

However, with the emergence of large models, banks are considering further exploration across various domainsPresently, New Internet Bank has successfully adopted large models in customer service, replacing certain roles previously held by human customer service representatives, and is trialing large model applications in marketing and post-loan management.

Beyond the realm of banking, Wang Jun forecasts that large model-related intelligent applications will markedly enhance in fields such as manufacturing, climate risk prediction, computing, education, and media entertainmentHe argues that “in manufacturing, large models can monitor the reliability of parts or batteries, predicting their lifespan; in climate risk forecasting, AI algorithms can interpret future weather patterns, providing warnings and route optimization for highways; in computing, large models can assist in code completion, understanding, and creation; in education, personalized large models can support students' learning based on their habits and behaviors; and in media and entertainment, large models can generate content, build models, and create scenarios, such as animation production, game design, and short video production, while also synthesizing digital personas for e-commerce recommendations.”

As AI threatens to disrupt job markets, will it create new opportunities instead? What type of AI talent will banks need in the future?

A report published by the China Banking Association indicates that finance possesses an inherent coherence with AI, showcasing that large AI model technology can effectively unearth the vast amount of data in the banking sector, which offers rich scenarios well-suited for AI applicationCurrently, large AI models are driving a significant transformation across service, marketing, and product domains within the Chinese banking industry, accelerating the arrival of the “future bank.”

As banks increasingly adopt large models, there are also mounting expectations regarding the capacity and skills required from technology professionals

Li Xiusheng notes that in the internet application industry, internet-thinking has propelled the success of large internet firmsAs we transition into the era of artificial intelligence, the societal demand for talent is shifting toward those who possess an AI-centric mindset, capable of melding expertise in finance and technology.

In recent years, New Internet Bank has emphasized internet-thinking and plans to prioritize AI-thinking moving forwardIn product design, customer marketing, daily operational activities, and the establishment of an overall management system, AI-thinking is being integratedHence, banks will assess whether employees possess this capability, foundation, or potential to cultivate the necessary talent for future banking advancements.

“The continuous progression of AI technology presents challenges for banking practitioners but also opens up new opportunitiesIn confronting these transformations, practitioners must remain calm, continually learn, stay abreast of trends, and find their place within society and enterprises,” Li encourages. “Technical professionals need to adapt themselves, employing AI technology to elevate their capabilitiesMeanwhile, business personnel need not overly fear replacement, as the lowered barriers to applying AI technology now enable even those without a computing background to utilize AI tools for crafting processes and applications, thus showcasing their valueTherefore, anyone willing to learn and keep pace with technological advances in banking will not only adapt but thrive amidst the rapid evolution of technology.”

From the perspective of risk control, Wei Hao emphasizes that practical hands-on experience is crucial in mastering artificial intelligenceThe application of AI technology in risk control necessitates more advanced talent requirementsAdequate technical understanding must be paired with a deep awareness of model advantages, capability boundaries, and associated risks to ensure proper application

Advertisements

Advertisements