In 2018, the rise of pre-trained models began, and by 2022, the explosive popularity of ChatGPT brought people the dawn of approaching AGI through large-scale pre-training paradigms. According to the government work report in 2024, the development and application of big data, AI, and others have been deepened, and the "AI+" action has been carried out to create a digital industry cluster with international competitiveness. It is worth noting that this is the first time AI has been written into the government work report, which also means that AI has entered a new era of widespread recognition and full-speed acceleration.

With the continuous development of AI technology, large language models have gradually highlighted the advantages of strong language capabilities and high accuracy and low cost in processing unstructured data. In recent years, during the digital transformation process of enterprises, they usually face the practical problem of high costs in establishing data platforms or knowledge graphs. Domain large models provide new solutions for enterprises to build knowledge systems, promote digital transformation, and promote the development of new quality of production forces.

Advertisement

The development of policies and technology has made people pay high attention to large models. In fact, as early as 2014, Zhongguancun Kejin has focused on the enterprise service track of marketing services, operation training, accompaniment, quality inspection, etc., and provided AI dialogue scene services for them, and has accumulated PB-level data in more than a hundred dialogue scenes. At the end of 2023, this conversational AI technology solution provider launched a series of large model product services, and the domain large model technology successfully passed the national network information department's algorithm filing.

Zhang Jie, Vice President of Technology at Zhongguancun Kejin, said: "Large models are facing the impossible triangle of generality, professionalism, and economy. The domestic general large model track has just emerged, and everyone is still at the starting stage on the domain large model track. Domain large models, as the 'last mile' of AI services, will exist for a long time. We combine the language ability of domain large models and the factual knowledge ability of domain knowledge bases to help high knowledge density enterprises apply large model technology, build a unified knowledge platform, and efficiently create all-media interactive applications."

By delivering high-performance domain large models and scene co-creation to customers, Zhongguancun Kejin helps enterprises achieve digital and intelligent transformation and upgrading and significant cost reduction and efficiency increase. In addition, it also solves the current limitations of large models, such as insufficient professionalism, questionable compliance, and low economy, through core technical capabilities such as composite AI, continuous learning, model security and trustworthiness, and platform-based services (MaaS).With the development and in-depth application of large model technology, Zhongguancun Kejin has gradually formed a new concept of large model application: focusing on the "three major directions" of knowledge equity, decision-making equity, and experience equity, and deploying the "eight major applications" of large model +, namely large model knowledge assistant, large model platform, large model + human-computer interaction, large model + BI intelligent decision-making, large model + insight analysis, large model + anti-counterfeiting security, large model + consumer protection compliance, and large model + intelligent office and digital management. A variety of large model applications have been implemented in the business scenarios of many leading financial institutions, central enterprises, government service departments, and public security departments in various provinces and cities.

Promoting the "last mile" landing of AI in enterprises

As a representative of new quality productivity, AI is expected to achieve a leap in productivity by empowering various scenarios in enterprises. In 2023, more than 200 large models emerged in China, known as the year of large models, and a "hundred model war" was triggered. The popularity of large models has also opened up the market for generative AI, and multimodal interaction capabilities such as voice and video, in addition to text, have gradually become the key to competition among AI manufacturers.

"Zhongguancun Kejin has a clear positioning from the early stage of development: to be a domain large model rather than a basic large model, to integrate multiple technical capabilities, and to promote the landing of composite AI in the 'last mile' of enterprise scenarios. Our domain large model has a wide range of compatibility and can be flexibly adapted to different basic large models. The key to this is the accumulation of high-quality domain data and efficient fine-tuning technology, and the basic large model can be flexibly replaced," Zhang Jie said. The reason why he can be deployed before the arrival of the technological wave is closely related to the resource advantages and technical accumulation of Zhongguancun Kejin over the years.He mentioned that even before the emergence of ChatGPT, Zhongguancun Kejin had already established a full-stack self-developed technical system, including natural language processing, speech recognition, multimodal biometric anti-fraud, OCR, and other AI atomic capabilities, as well as technical capabilities such as real-time audio and video communication, big data, etc.

In addition, in the field of domain-specific model fine-tuning technology, Zhongguancun Kejin has a "unique skill". First, partition positioning is performed for different tasks, then the synergistic gain relationship between different tasks is found, and finally, hierarchical fine-tuning is performed for specific tasks. Ultimately, it can achieve less dependence on data volume, lower computational power consumption, and higher inference performance. It should be understood that when large models perform different natural language processing tasks, they will use different working areas, which is similar to the existence of different brain areas in the human brain.

Zhongguancun Kejin can use partition positioning technology and hierarchical fine-tuning technology to perform "CT scans" of tasks in different working areas to achieve customized rapid enhancement effects. By accumulating evaluation datasets, the evaluation model is good at extracting, generating, multilingual, etc., injecting domain knowledge into general large models quickly and at low cost, and then delivering high-performance domain-specific large models to customers.

Zhang Jie explained: "Fine-tuning domain-specific large models is a technical task, and it is not as simple as just putting on a shell and pouring in some data for fine-tuning. It is necessary to open the 'black box' of the large model. Our unique hierarchical fine-tuning technology can accurately enhance more than 100 different NLP tasks such as intent recognition, sentiment classification, or anaphora resolution. Among them, under the premise of achieving the same improvement effect, the domain-annotated data can be reduced to 1/6 of the original."

It is reported that through half a year of research and development, Zhongguancun Kejin has spent 100,000 card days of computing power, sorted out the synergistic gain relationship map between atomic tasks, used for mixed data, to improve the efficiency of model training, and achieve single-card inference, single-week iteration, single-week completion of model fine-tuning, and avoid catastrophic forgetting of the model. This technology enables general large models to be efficiently trained into domain-specific large models, and combined with domain knowledge bases, to meet the specific needs of enterprises in various scenarios.Taking the financial industry's enterprise knowledge base as an example, in the customer service scenario of cooperation with a leading wealth management company, Zhongguancun Science and Finance has independently developed an enterprise knowledge base model and intelligent customer service, and other artificial intelligence technologies to create an intelligent knowledge base for customers. It has capabilities such as multimodal document analysis, automatic extraction of QA question and answer pairs, and automatic tagging of knowledge content. By integrating the customer's enterprise WeChat and business expansion APP applications, it helps customers significantly improve the accuracy of the customer service system's question intent recognition and reply, reducing the operational workload of text customer service by more than 70%, and improving the answer effect by more than 50%.

While accumulating resources and technology, Zhongguancun Science and Finance's various technical capabilities have gradually been recognized by the industry and have participated in the industry's standard recognition. Specifically, its model capabilities have been certified by the authoritative institution of the China Academy of Information and Communications Technology, with comprehensive semantic, voice, and visual basic capabilities. In the evaluation of 5 capability domains and 46 capability items, the accuracy has reached more than 95%, and it has reached the highest capability level of the current domestic large model 4+, accelerating the efficient commercial landing of domain large models.

In addition, Zhongguancun Science and Finance also participated in the formulation of the first authoritative certification standards for large models in the industry, and fully participated in the formulation of industry standards for other product technical capabilities, including the first authoritative certification standard for intelligent marketing products in China, the "Trusted Face Recognition Guard Plan" of the China Academy of Information and Communications Technology, the basic capability requirements for intelligent dual recording systems, and the basic capability requirements and assessment methods for real-time audio and video services.

So far, Zhongguancun Science and Finance has applied for and accepted a total of 438 patents, obtained 329 software copyrights and CMMI5 international certification. In terms of data resources, Zhongguancun Science and Finance has served more than 200 digital transformation scenarios of more than 1600 enterprises, serving more than 500 million users, and has accumulated a large amount of industry know-how and high-quality data in various fields.

It is better to teach a man to fish than to give him a fish: Enhance the enterprise's independent operation and maintenance capabilities of large models.Large model technology is expected to yield a decade of application dividends. However, many companies still have some misconceptions about large models, especially the belief that the larger the model, the better the effect. In fact, although large models with stronger human-computer interaction capabilities are more outstanding in performance, they are not substitutes for traditional models.

In reality, the innovative combination of large and small models is the future development trend of AI. This combination can decompose complex tasks based on cognitive reasoning, plan precisely, and call various models. In this way, more accurate risk analysis, prediction, and decision-making can be achieved. For example, large models plus small models, decision engines, keywords, rule settings, etc.

On the other hand, from a technical perspective, although the development of large models has been rapid in recent years, they are more like a growing child, so large models are not a one-step solution but need continuous learning.

In addition, it cannot be ignored that the current large models have limitations such as insufficient professionalism, questionable compliance, and low cost-effectiveness. Only by overcoming these problems can the "last mile" of large models be solved.

To solve the above problems, Zhongguancun Kejin has built four core technical capabilities: combined AI, continuous learning, model security and trustworthiness, and platform-as-a-service (MaaS).Zhang Jie stated that the essence of large models is a knowledge base of language and common sense about the world. Zhongguancun Kejin has adopted a dual approach of knowledge injection and external knowledge base to enable enterprises to form a practical "technology + system" framework. This not only ensures that the behavioral data or performance data of human employees form a closed loop but also allows the large models to become "smarter" the more they are used.

It is reported that to reduce the operational and maintenance costs for users, Zhongguancun Kejin will also provide an easy-to-operate toolkit when delivering products, allowing users to maintain operations independently, thereby further enhancing the product experience and efficiency. At the same time, the application development platform can develop large model applications through a zero-code approach, fully meeting enterprise needs and effectively reducing the innovation costs of internal enterprise applications and shortening the trial and error cycle.

In terms of product security and privacy protection, Zhongguancun Technology has taken comprehensive and meticulous measures. In the pre-event stage, precise data screening and strict data purification are implemented to ensure the purity, security, and comprehensiveness of the data. In the event stage, a wealth of legal and administrative case knowledge has been accumulated for compliance training. In the post-event stage, in-depth security testing and risk assessment are carried out through post-processing techniques such as "illusion detection" to ensure the stability of the product and the privacy security of users.

It should be understood that since the large model is a new technological revolution, it still requires a certain period to be fully applied in various enterprise scenarios.

In the short term, the advantages of good memory, strong understanding, and the ability to read unstructured documents can be utilized as assistant applications, such as knowledge assistants, writing assistants, and training assistants, to empower internal employees of enterprises.At present, many enterprises have a vast amount of unstructured documents. In the past, to utilize them, it was necessary to train small models or manually annotate data, and then process them into a knowledge graph, making the cost of parsing very high.

In knowledge-intensive industry fields, corporate employees have to deal with a large amount of documents and information every day, including various types of research reports, reports, notices, official documents, etc. Zhang Jie believes that targeted writing assistants are expected to become a business demand that will explode in the near future.

In the medium term, based on the unified knowledge base that has been built, more domain knowledge will be combined for use in dialogue scenarios. For example, making phone calls to recall old users, automatically revisiting users who have made transactions, and empowering customer service seats, etc.

Starting from the first half of 2024, Zhongguancun Kejin has successively co-created with several leading securities firms, trying in the fields of intelligent investment research and intelligent investment advice. Due to the more targeted and diverse content of the copy written by large models, the proportion of complex financial product sales has increased to 30%. At the same time, co-creating the application of large models in compliance scenarios with a certain bank, deeply exploring the field knowledge related to internal and external regulations and laws and regulations, and empowering the compliance control of various links in financial business.

Zhongguancun Kejin has co-created with a certain provincial public security department in the anti-fraud police situation in the public security field. Based on the fine-tuning of the anti-fraud knowledge base data domain, an anti-fraud intelligent brain has been created and has achieved significant results. It has greatly improved the work efficiency of police officers and reduced the workload of police officers, and greatly reduced the property losses of the people. In addition, Zhongguancun Kejin has also developed a maintenance assistant in the field of industrial operation inspection, providing the best practice guidance for maintenance personnel, and the maintenance efficiency of a single work order has increased by 30%, reducing the cost of manual operation.In the long term, large models need to enhance their reasoning and decision-making capabilities on the basis of fundamental language abilities by effectively integrating factual knowledge and procedural knowledge, such as in operations and maintenance monitoring, financial risk control, etc.

Zhang Jie said that he hopes that in the future, generative AI technology can be used to build a new type of collaborative production relationship between humans and machines, helping enterprises to create "super employees" and achieve strong AI applications with analytical decision-making capabilities.

Conclusion

With the rapid and continuous development of large model technology, the field will become increasingly crowded, and the field of large models is also facing a trend of technological reshuffling.

Therefore, for domain large model product application service providers, on the one hand, they need to have strong technical capabilities to quickly and at low cost fine-tune domain large models; on the other hand, they also need to enhance the large model's ability to judge facts and values based on existing language capabilities and rich industry know-how accumulation on the path of AGI development.The future of domain-specific large models is promising for solving more problems in industries/enterprises, applying in more scenarios, and achieving transformative productivity improvements.