After more than two years of exploration, a significant consensus has emerged within the financial sector: while large language models (LLMs) can achieve rapid question-and-answer capabilities through vast datasets, they fall short of addressing the complex reasoning needs of industry applicationsWhat is essential for real-world application is a complementary approach focusing on "slow thinking," which emphasizes logical inference and reasoning over speed.

The release of the OpenAI-o1 model in September of last year triggered a massive surge in the global AI landscapeThis model boasted enhanced reasoning capabilities, leveraging reinforcement learning and dismantling complex problems into understandable componentsIts performance in tackling mathematical problems and intricate tasks notably outperformed its predecessors in the GPT series, which were primarily general-purpose large models.

However, the improved reasoning ability of the o1 model came at a costIts response time stretched to about ten seconds, a stark contrast to the instantaneous responses typically provided by GPT series models, and its access came with high overheads.

As the Lunar New Year approached, DeepSeek unveiled its DeepSeek-R1 model, making it the first model globally to replicate the capabilities of o1 successfullyNot only did it match the reasoning performance of o1, but it reduced related computational resource consumption to one-tenth of its counterpart's demandRemarkably, the cost of input tokens accessed via the official API was merely one-fiftieth that of the o1 model, revealing great economic potential for financial institutions.

A senior executive in the digital transformation department of a prominent banking institution expressed palpable excitement, stating, "It is not an exaggeration to say that DeepSeek has developed the most cost-effective large model among China’s open-source alternatives.” They highlighted that the cost-effectiveness included significant savings in reasoning costs alongside a dramatic leap in the applicability of business scenarios

Advertisements

Many AI projects requesting funding had previously been rejected due to either exorbitant costs or impractical implementation plans, but the new landscape could foster a wave of innovative initiatives.

Numerous interviewees noted that R1 model democratized AI access, enabling smaller financial institutions to unlock a wealth of imaginative applications, thus catalyzing more certain use casesHowever, the efficient employment of this high-performing model still hinges on the foundational elements provided by traditional general models, such as training on appropriate datasets and algorithm fine-tuning.

Significantly, the announcement by multiple financial institutions to integrate the DeepSeek-V3 and full-scale R1 models does not indicate a retreat by other general model vendorsPrior to the emergence of the next technological "singularity," the coexistence of general models, deep reasoning models, multimodal models, and smaller models will continue, effectively serving various application scenarios as needed.

One might say that R1 embodies an all-encompassing "master's degree" level of proficiency, solidly situating itself within this vibrant ecosystem.

“So, the user's question is…” This is often how DeepSeek begins to respondWhen a query is directed towards DeepSeek, it engages in deep contemplation for about ten seconds before generating a responseDuring this time, it articulates how it interprets the question, what aspects it encompasses, and what type of response the user might be anticipatingThis is the essence of the "slow thinking" process facilitated by the deep reasoning model.

The slow reasoning capability of R1 results from innovative algorithms at DeepSeekUnlike the conventional "instant response" models, R1 does not produce immediate answersInstead, it conducts extensive reasoning over the content of instructions, employing chains of thought, consensus, and retrieval to formulate the optimal answer

Advertisements

Each piece of content generated by the model requires repeated contemplation, which in turn creates a plethora of output markers, further enhancing the quality of the model.

Several interviewees emphasized that R1's introduction is vital in bridging the reasoning gap left by traditional general models.

“Overall, today’s large models possess capabilities akin to that of a master’s graduate,” remarked a technology lead from a mid-sized brokerage firmHe noted that a fully capable digital employee is needed to balance both "fast thinking" and "slow thinking," which R1 effectively supportsPrior to the open-sourcing of R1, institutions relying on deep reasoning models faced the sole option of integrating the o1 model's APIGiven the regulatory and data security issues prevalent in the financial sector, the application of the o1 model in production was often impracticalThe advent of R1 allows for the analysis of intricate data while simulating the human reasoning process via thought chains, with its performance being on par with that of the o1 model.

One executive from a city commercial bank shared insights with the press, stating, “Previously, we were limited to general models like Tongyi Qianwen, DeepSeek-V3, and never had access to deep reasoning models like o1. With the launch of R1, AI applications now genuinely possess the capability for in-depth thinking."

Consider an example within the realm of intelligent marketing: the marketing scripts generated by traditional general models are derived either from rule-based systems that organize corpus or swiftly respond to directives provided by usersHowever, when faced with skepticism from the directive party, they may fail to deliver effective marketing outcomesIn contrast, the deep understanding facilitated by the reasoning model enables far more intelligent and accurate interactions with users.

It's important to recognize that while the R1 model excels in reasoning, it must undergo the rigorous scrutiny that traditional models have faced in the finance sector.

Authoritative testing from Vectara HHEM, which assesses AI models for "hallucination" rates—instances where a model generates information that is incorrect or fabricated—revealed that DeepSeek-R1 presented a hallucination rate of 14.3%. In comparison, the V3 model from DeepSeek recorded a hallucinatory rate of only 3.9%, both of which exceed the overall industry average.

The age-old saying, "Learning without thought is labor lost; thought without learning is perilous," holds great relevance in the AI space.

R1 model’s strong reasoning abilities contribute to its outstanding performance in mathematical and logical reasoning; conversely, the potential over-exuberance in producing outputs in the humanities can lead to hallucinations

Advertisements

A chief information officer at a financial institution recounted that when utilizing DeepSeek to generate a personal resume, he often encountered nonsensical outputs, whereas other general models provided remarkably accurate information based on publicly available resources.

The aforementioned city commercial bank representative candidly acknowledged that R1 currently stands as the best in logical reasoning skills among China's large models but still shows deviations in understanding specialized knowledge until localized financial datasets are properly integrated into its training. “Knowledge put into the model cannot immediately manifest its logical reasoning capabilities,” he commented, highlighting that all big models face fundamental challenges regarding computational power and data when being deployed in finance.

On one hand, governance surrounding data needs triggered by AI strategies remain an essential issue for financial institutionsR1’s pronounced hallucination problems necessitate higher-quality datasets and knowledge basesOn the other hand, for institutions hoping to deploy large models on-site, the energy costs comprise two components: the fixed cost of computational resources necessary for local deployment and the dynamic energy expenses engaged during the reasoning process, where the latter significantly influences the performance of the model.

However, relative to general models, R1 stands as a beacon of efficiency, displaying substantial advantages in both reasoning and training costs, thus markedly decreasing the expenses associated with practical deployment.

It's noteworthy that while various financial institutions have publicly announced their integration of DeepSeek-R1 or V3 models, the scale and nature of model integration vary based on each institution's resource capabilities, expenditure priorities, and practical applications.

For example, Postal Savings Bank and Jiangsu Bank are integrating the lightweight version of DeepSeek-R1 while a representative from a major bank’s fintech arm revealed that they have just begun testing the 14B R1 model

Conversely, as previously mentioned, a mid-size brokerage firm has rolled out the full-sized R1 model.

Such variations indicate that institutions are strategically exploring starting with lighter models due to the considerable memory demands that the full-scale R1 model entails as compared to others like the Qwen2.5, whose maximum size is limited to 72BThis illustrates the practicality many institutions adopt to initiate their endeavors in AI deployment.

The transformation is undeniable: DeepSeek does not only bring excitement but also releases a cascade of innovative application possibilities within the finance realm.

Although the degree of integration with DeepSeek's models may differ by institution, R1’s robust reasoning abilities provide a significant impetus for innovation, heralding a profound shift in AI applications for the financial industry.

Interestingly, the financial sector is no stranger to DeepSeek; reports indicate that various financial institutions commenced their engagement with AI large models over a year ago by integrating DeepSeek-Coder-V2 for coding language tasksWithin the IT department of the seven financial institutions interviewed, five utilized the intelligent code assistant models based on Coder-V2 to facilitate their services.

“We have experimented with nearly all coding language models available, and perhaps due to foreign exchanges being well-versed in quantitative programming, the Coder-V2 excels in code generation,” one interviewee noted.

“DeepSeek's open-sourcing of the R1 model has democratized AI access,” asserted a senior executive at a brokerage firmPreviously, utilizing reasoning models demanded an outstanding technical background, including skills in deep learning and neural network reasoningHowever, through DeepSeek’s publicly available technical reports, all institutions can now replicate the reinforcement learning workflows used to construct the R1 model within their own large models, enabling effective transfer of capabilities from DeepSeek to financial application scenarios.

In the securities landscape, novel pathways for exploring "AI + brokerage," "AI + investment research," "AI + advisory," "AI + compliance," and "AI + document" methods will be revealed

Advertisements

Advertisements