《中國移動&KT：2024大語言模型驅動的AI新時代中電信運營商的思考與實踐（英文版）（22頁）.pdf》由會員分享，可在線閱讀，更多相關《中國移動&KT：2024大語言模型驅動的AI新時代中電信運營商的思考與實踐（英文版）（22頁）.pdf（22頁珍藏版）》請在三個皮匠報告上搜索。

1、1 2 Contents Acknowledgement.3 Abstract.4 Preface.4 KTs AI Transformation utilizing Agent and Data.4 NTT DOCOMOs Strategic Journey towards Digital Transformation and Enhanced Customer Experience.5 China Mobiles Transition to AI+to Amplify Scale Empowerment.5 1 LLM Adoption Strategies in Industry.6 2

2、 Emerging Challenges and Technical Foresights.7 2.1 AI Application Perspective.7 2.2 Data Fueling Perspective.9 3 Application Tooling Platforms.11 3.1 China Mobile Jiutian Large Language Model Application Platform.11 3.2 DOCOMO LLM Value-Added Platform.12 3.3 KT SLM/LLM Platform.13 4 Generative AI A

3、pplication Cases.14 4.1 Generative AI for Network O&M.14 4.2 Generative AI for Customer Service.17 5 Future Outlook and Industry Suggestions.21 6 Abbreviations.22 3 Acknowledgement SCFA was established in 2011 by China Mobile,Koreas KT,and Japans NTT DOCOMO,aiming to promote a tripartite cooperation

4、 framework for global technology standards and industry ecosystems.In 2022,the AI Workgroup was established,focusing on the development and application of AI technology,promoting technical exchanges among member companies,and guiding and facilitating the application and cooperation of AI technology

5、within the industry.This White Paper has been produced as a collective effort within the SCFA AI WG,and on its behalf the following editing team(listed in alphabetical order):China Mobile:Lingli Deng,Bo Yuan,Xuefeng Zhao,Xiangyang Yuan,Di Jin KT:Jiyoung Kim,Jaeho Oh NTT DOCOMO:Issei Nakamura,Kuanyin

6、 Liu,Aogu Yamada,Satomi Kura,Takeshi Kato SCFA AI WG China Mobile Contact: KT Contact: NTT DOCOMO Contact: 4 Abstract This document analyzes the challenges of scale adoption of Large Language Models(LLMs)into industrial applications,highlighting the problem of reinventing the wheel of common capabil

7、ities,the performance bottleneck of network communication,the improvement of productivity by utilizing work-oriented SLM/LLM based AI agents,and proposes technological development trends such as innovation in fundamental algorithms,standardization of application tool platforms,and Cloud-Edge collabo

8、ration.It showcases contributing CSPs strategic layout in AI technology,data integration,application tooling platforms,as well as a variety of generative AI applications,and looks forward to the future development of AI technology,data integration and industry collaboration recommendations.Preface K

9、Ts AI Transformation utilizing Agent and Data With the rapid advancement of AI HW and SW technologies,generative AI models are evolving into various versions.Alongside this,generative AI Agents are swiftly permeating our daily lives.The paradigm shifts to a practical AI Agent competition,reflecting

10、users Gen AI demands,is closely related to the handling and accommodation of extensive customer data.As AI advances,the importance of data in corporate activities has become even greater,and Data-driven AI Agents based on customers and companies are at the center of Corporate Transformation Using AI

11、.To succeed in AX,it is essential to collect and utilize data from corporate activities effectively,and the primary innovation of AI companies must be driven by Data-driven AX.In the Era of AI Agents,where AI is becoming central to corporate and personal daily services,KT is pursuing the enhancement

12、 of AI competitiveness using AI Agents as one of its successful transformation directions into an AICT company.Under the multi-model line-up strategy,which combines its self-developed AI language model Mi:dm with models based on open-source,KT aims to provide a variety of customer/industry-specific

13、models and AI Agents to the market,based on high-quality data learning and utilization.KT is moving forward with the goal of enhancing productivity by utilizing work AI Agents for its employees,and it also plans to spread new AI experiences to customers by applying them to its Genie TV.By developing

14、 these AI Agents and launching services,KT expects to secure customer AI data and conceive specific AI business models utilizing the data.Strengthening AI MSP competitiveness by providing Model as a Service comprehensively and through global AI Agent technology/business cooperation,KT will lead the

15、AI market and ecosystem construction.5 NTT DOCOMOs Strategic Journey towards Digital Transformation and Enhanced Customer Experience NTT DOCOMO(DOCOMO)set the goal of improving customer experience and reforming business structure with digitalization of business management,and promotion and execution

16、 of data utilization as our medium-term strategy toward 2025.Initiatives in digital transformation at DOCOMO include network optimization through data utilization,AI and human resource training,and the promotion of digital marketing.AI platforms for image recognition,voice recognition,and customer a

17、nalysis are being offered to enhance DOCOMOs competitiveness by applying these technologies to its services.Since 2014,DOCOMO has been building a big data infrastructure that collects data such as user information,usage history,network traffic and payment history from almost 100 million users and mo

18、re than 270,000 base stations as an effort to promote digitalization of business management and data utilization.The platform incorporates external data from business partners and AI technologies to create value across various business fields,such as Mobility as a Service,retail,banking,and the meta

19、verse.Leveraging new technologies like generative AI to find new revenue streams and grow the business is not an easy task.It requires strategic planning,including training personnel,and a lot of trial and error.DOCOMO is not only focusing on developing the foundational technologies for generative A

20、I but is also actively working on various initiatives to create use cases and train personnel through continuous experimentation and refinement.China Mobiles Transition to AI+to Amplify Scale Empowerment In the face of the wave of change,China Mobile,as the largest mobile communication operator in t

21、he world,has always anchored its strategic positioning of world-class information service technology innovation company.In terms of network computing infrastructure,a communication network with the widest coverage and the largest user scale in the world has been built,with more than 1.9 million 5G b

22、ase stations accounting for 30%of the worlds total,over 90 land and sea cable systems connecting 78 countries,and the largest single intelligent computing center of global operators with 18000 GPU cards.Jiutian,a series of large foundation models of language,vision,voice,structured data and multi-mo

23、dality have been constructed,on top of which more than 40 large industry models are launched,forming a comprehensive AI portfolio including platforms,capabilities,and large-scale applications.Over 10,000 AI+projects have been launched to promote the intelligent and green development of various indus

24、tries,such as energy,manufacturing,medical caring,transportation and others.Along the way,it is noticed that the transition to AI+signifies the shift of AI technology from a mere technical application to a comprehensive empowerment deeply integrated into industrial development.The challenges faced i

25、n this process include the 6 limitations of LLMs in critical task execution,the waste of resources caused by the repetitive development of common capabilities,and the bottleneck effect of network communication.To address these challenges,China Mobile calls on all parties in the industry to work toge

26、ther in building a comprehensive AI+industry ecosystem to promote innovations at the fundamental algorithm level,standardization of application tooling platforms,and new models of Cloud-Edge collaboration 1 LLM Adoption Strategies in Industry Artificial intelligence,representing the new generation o

27、f information technology,is rapidly emerging as a significant driving force for new quality productivity.Among these,generative AI technology based on LLMs is significantly empowering various industries,leading to an explosive growth in the application of AI models across industries,heralding the ar

28、rival of a technological and industrial revolution,where the information service system and the economic and social operation systems are deeply integrated,profoundly changing peoples lifestyles and modes of production.LLMs have demonstrated extensive and profound impacts on current industrial appli

29、cations,emerging as pivotal tools in the digital transformation of enterprises.From knowledge management to handling complex tasks,LLMs are progressively integrating into core business processes.One notable application is retrieval-augmented generation(RAG),which combines external knowledge bases wi

30、th generative capabilities to effectively address complex queries.This approach is particularly effective in customer service,where LLMs assist companies in extracting precise answers from massive internal documents,thereby enhancing service efficiency.Moreover,LLMs play a significant role in buildi

31、ng and managing enterprise knowledge bases,facilitating intelligent querying and updating through natural language understanding and knowledge extraction.In handling complex tasks,LLMs exhibit powerful capabilities such as automated report writing,marketing copy generation,and code generation,signif

32、icantly boosting productivity and automating business processes.LLMs have also found widespread use in automated customer service systems,where their deep understanding of natural language allows them to handle complex customer intentions and contextual interactions beyond the reach of traditional c

33、hatbots.Additionally,LLMs contribute to personalized recommendations by generating customized content,offering precise suggestions that help businesses achieve higher customer satisfaction.To realize these applications,LLMs leverage various techniques to optimize their performance in specific scenar

34、ios.The adoption of LLMs in industry can proceed in different ways,depending on the technological requirements and application context.For applications with lower technical barriers,enterprises can quickly deploy L0 and L1 models by integrating domain-specific knowledge bases,making this approach su

35、itable for scenarios that require rapid implementation without intensive model optimization.In scenarios requiring domain-specific customization,L0 models can be fine-tuned by uploading customized datasets and applying low-code configuration to produce L1 models adapted to specific tasks.This method

36、 suits situations where data 7 accumulation and model adaptability are needed,allowing for more precise responses to particular business requirements.For applications with higher technical demands and more complex contexts,enterprises can adopt a comprehensive model development process,encompassing

37、data collection,processing,pre-training,and fine-tuning,ensuring model performance and stability in intricate applications and meeting the needs of high-precision,high-reliability operations.Furthermore,LLM deployment can be realized through multi-model convergence platforms,enabling broader collabo

38、rative applications.Enterprises can utilize modular plugins and centralized agents to build complex business systems that integrate multiple models,thereby facilitating cross-industry application expansion and fulfilling the requirements of sophisticated application ecosystems.In conclusion,the indu

39、strial deployment of LLMs spans from basic knowledge base integration to full-scale model customization and multi-model management,creating a multi-layered application system that ranges from low technical barriers to highly customized implementations.Through these diverse approaches,LLMs are drivin

40、g the development of intelligent industries,providing flexible and personalized solutions across sectors,and empowering enterprises with efficient operations and intelligent decision-making capabilities.2 Emerging Challenges and Technical Foresights With the in-depth development of the fourth indust

41、rial revolution characterized by digital intelligence,there is a foreseeable trend of the mutual embrace between traditional industries and AI technology to address emerging challenges for LLM scale adoption:on the one hand,the deepening integration of industry information resources and data governa

42、nce empowers the innovation of LLM applications by providing desired raw data materials;on the other hand,continuous innovation in LLM algorithms and engineering tools addresses the applicability and economic issues of large-scale production environment applications.2.1 AI Application Perspective Ch

43、allenge:Large language models currently do not possess the capability to be directly applied in key decision-making processes in production environments.Foresight:Innovation in basic theories for reasoning acceleration,full-process autonomous control at the fundamental algorithm level,to realize aut

44、onomous cognition,autonomous evolution,and autonomous breakthrough of AI agents.Currently,LLMs serve as powerful information processing tools capable of executing tasks such as natural language processing,image recognition,language translation,text generation,and image recognition.However,large lang

45、uage models themselves lack environmental perception capabilities and do not possess autonomy and proactive decision-making abilities,usually requiring human input or triggering to process 8 information in a preset manner.Therefore,they face difficulties in executing dynamic and complex tasks,as the

46、se tasks typically require perception and understanding of the real world,the ability to adapt to environmental changes,and making decisions that align with the goals.Hence future innovation at the basic algorithm level will focus on the following areas:Autonomous cognition Future algorithms will pl

47、ace greater emphasis on the autonomous cognitive capabilities of intelligent agents,enabling them to better understand and predict their environment,with enhanced perception,reasoning,and decision-making capabilities of the environment,as well as adaptability in complex environments.Autonomous evolu

48、tion Algorithms will be designed to evolve on their own,continuously optimizing their performance through machine learning.Intelligent agents will be able to learn from experience,automatically adjust their behavior to adapt to new tasks and environments,thereby improving their generalization capabi

49、lities.Autonomous breakthrough To achieve a higher level of intelligence,algorithms need to be able to achieve breakthroughs on their own without human intervention.This involves innovative algorithm design,enabling AI agents to discover new solutions and even surpass the performance of human expert

50、s in some cases.Moreover,to support the development of the above capabilities,algorithms and AI agent operation optimization and control technology also need iterative innovation,including reasoning acceleration technology to improve the responsiveness and efficiency of AI agents for complex tasks,a

51、nd full-process autonomous controllable algorithms to ensure their stability and reliability.Challenge:The vertical repetitive development of a large number of common capabilities leads to resource waste and slows updates and upgrades.Foresight:The rise of application tooling platforms serving as LL

52、Ms plus domain specific knowledge bases,with plugins,tools,enhancing professional capabilities while not losing basic capabilities for AI agent customization development.In the current field of artificial intelligence,we face a significant challenge,that is,the vertical repetitive development of a l

53、arge number of common capabilities,which not only leads to resource waste but also makes the process of updates and upgrades slow.This phenomenon is particularly prominent in the rapidly developing AI technology because it involves a large amount of research and application development.To address th

54、is challenge,it is foreseen that an important direction for future technological development is the innovation of application tool platforms.In particular,AI agent customization and development platforms will be key,which can provide low-code solutions to enable non-technical users to create office

55、agents,financial agents,and other professional tools easily.Such platforms provide basic LLMs combined with professional knowledge bases,as well as plugins and tools,which can enhance professional capabilities while keeping basic capabilities.Through such platforms,one may not only reduce resource w

56、aste but also accelerate the advancement of AI technology,thereby promoting the healthy development of the 9 entire industry.Challenge:The bottleneck effect of network in connecting data and cloud computing infrastructure is highlighted as the last mile of LLM deployment and user empowerment.Foresig

57、ht:Cloud-Edge collaboration is leveraged to enable premise(network edge,home terminal)personalized AI agent services.In todays digital era,the bottleneck effect of network communication has become the restricting last mile for LLMs to reach and empower users.To solve this problem,it is foreseeable t

58、hat the new model of Cloud-Edge collaboration will become mainstream,especially on the end-side of the network edge and home terminal,by providing personalized intelligent agent services as a solution.The network edge and home terminal on the end-side are key links in the Cloud-Edge collaboration,an

59、d AI agent services can be deployed at these endpoints to reduce the dependence on centralized cloud computing resources.In this way,data pre-processing,analysis,and response can be executed closer to the user,reducing data transmission latency and bandwidth requirements.e.g.,by deploying intelligen

60、t gateways at home terminals,functions like home automation control and security monitoring can be realized with improved responsiveness and reduced network load.In addition,based on the AI agent customization and development platform,personalized AI agent services can be customized according to the

61、 specific needs and usage habits of users,providing more accurate and efficient services.This not only includes applications in professional fields such as office agents and financial agents but can also be extended to various aspects of life such as personal health management,education,and entertai

62、nment.By calling on the LLMs and professional knowledge bases distributed in the end-to-end network on demand,integrating plugins and tools,etc.,personalized AI agents can enhance their professional capabilities while not losing responsiveness or customer experience.In summary,through the developmen

63、t of Cloud-Edge collaboration and personalized AI agent services,the bottleneck problem of network communication can be effectively solved,promoting the widespread application of LLMs in various fields and achieving a true intelligent transformation.2.2 Data Fueling Perspective Challenge:The lack of

64、 standardization of scattered data hinders the starting point for data-driven AX.Foresight:Data Governance for data classification,data standardization and systematization,and grade management of data.Data governance is a series of processes related to data standardization for AI,to ensure consisten

65、cy in data names,data descriptions,and data formats.The following three stages are necessary to implement data governance successfully.Meaningful classification of company-wide data It is crucial to systematically 10 classify various types of company-wide data,such as enterprise data,customer data,m

66、anagement data,and infrastructure data,according to their types and purposes.Systematic classification of data is the starting point for efficient management,utilization,and execution of AX in the near future.Standardization and systematization of classified data It is necessary to manage and unify

67、standards so that customers can understand from the same perspective at any contact point with the possibility of connections between company-wide data.Additionally,to improve the readability of business data by applying data standardization and secure AI utilization is needed.Managing data grades a

68、nd constructing grade-based clouds connected with the appropriate security systems It is essential to establish a grading system by creating management indicators(quality,utilization,and cost)for data and accordingly configuring grade-based clouds.From the security enhancement perspective,it should

69、be available to choose access control,monitoring,and log management according to the data grade.Challenge:Data integration is required to manage data that makes unfragmented in one place.Foresight:Cloud-based integrated platform for data centralization,analysis,and modeling.It is required to build a

70、 cloud-based ML data platform that can centralize company-wide data to resolve existing data issues.Building an integrated data platform helps centralize the data and gradually resolve the issues caused by data silos.To continuously manage the data integration effectively,it is necessary to consiste

71、ntly align a modernization of AI,Data,and IT infrastructure so that the process of data accumulation by the alignment between AI and Data and availability of assets by the alignment between Data and IT continues to circulate.Through the direction of data collection and availability of assets,it is e

72、xpected to achieve the effects such as improving decision-making,and predicting issues by utilizing customer data,management data,and infrastructure data.Challenge:Data Serving should be prepared to integrate and distribute the data appropriately.Foresight:Company-wide collaboration,secure and accum

73、ulation of capabilities,data monetization.Even if the process of integrated data governance and management is carried out properly,it cannot be said that data-driven AX has been fully realized.To effectively integrate the accumulated data and distribute it as needed,a dedicated organization that lea

74、ds data planning and execution must be established as well as a collaborative system based on domain-specific ML Ops.An expertise in data governance and domain-specific data can be secured through such a collaborative system.Additionally,it is necessary to expand data utilization businesses based on

75、 the acquired 11 data operation and management capabilities and to convert this experience into external business capabilities.3 Application Tooling Platforms In response to numerous challenges that greatly limit the efficiency of users in building intelligent agents during the development process,s

76、uch as high technical barriers,long development cycles,difficulties in improving model performance,complex deployment and maintenance,insufficient customization and flexibility,difficulties in team collaboration,and ensuring security compliance,both China Mobiles Jiutian Large Language Model Applica

77、tion Platform and DOCOMOs LLM Value-Added Platform enable one-stop intelligent agent application development.3.1 China Mobile Jiutian Large Language Model Application Platform China Mobiles Jiutian Large Language Model Application Platform has capabilities such as application construction,plugin int

78、egration,model playground,and inference services,offering a full-process,one-stop production tool for LLM applications.It provides a combination of autonomous planning and scheduling with controllable manual scheduling to improve scheduling accuracy and reduce model hallucinations,achieves enhanced

79、management of private domain knowledge bases to improve the accuracy and professionalism of answers,integrates a rich set of official plugins to facilitate the construction of a broader range of application capabilities,integrates various memory capabilities to personalize model responses and integr

80、ates with third-party applications to provide access to APIs and other inference services,which helps individual and enterprise customers to develop their own AI applications at a low cost and in a timely fashion,promoting the application and implementation of LLMs in various industries.Figure 1 Ill

81、ustrative Workflow of Jiutian Large Language Model Application Platform 12 As shown in Figure 1,the Jiutian Large Language Model Application Platform provides one-stop intelligent agent services for individual and enterprise customers,in supporting more than 100,000 users to quickly build more than

82、1,500 customized intelligent agent applications,covering multiple scenarios such as office,social,entertainment,and daily life,helping AI to empower various industries.Looking to the future,consumers needs are becoming increasingly complex,and higher requirements will be proposed for the quality,sta

83、bility,and refinement of services.To empower users to build diverse and complex applications,the platform will focus on standardizing processes,supporting multimodal data,low-code workflows,and optimizing the core capabilities of intelligent agents.By comprehensively upgrading intelligent agent serv

84、ices,it ensures excellent quality,stability,and reliability,enriches the plugin ecosystem,and provides an efficient,intelligent,and comprehensive construction experience,in order to help its customers seize the initiative in digital transformation,accelerate the pace of innovation,and achieve a leap

85、 in business value.3.2 DOCOMO LLM Value-Added Platform Since August 2023,DOCOMO have been developing the LLM Value-Added Platform to promote digital transformation within our internal operations and provide new services using LLMs.This platform is utilized within the DOCOMO Group,boasting approximat

86、ely 7,000 monthly active users and around 1,000,000 calls per month.The major features available on the platform include:LLM There are various LLMs available as open-source software(OSS)or software as a service(SaaS).These LLMs differ in terms of cost,input/output token limitations,proficient langua

87、ges,tasks,and specialized domains.Multiple LLMs are provided in this platform so that users can select the most suitable LLM for their needs.Currently,the available LLMs include the GPT-4o and GPT-4o mini models provided by Microsoft Azure OpenAI Service,and the tsuzumi developed by NTT Laboratories

88、.We plan to add more LLMs in the future.RAG Since the data used to train the aforementioned LLMs is mostly publicly available information from the internet,RAG is necessary for use cases that require domain-specific knowledge,such as internal policies.The platforms RAG is built on the Microsoft Azur

89、e and employs a unique architecture that improves search precision and response quality through enhanced preprocessing of registered documents.Users can utilize RAG by placing the target documents in a specific folder within Microsoft SharePoint used internally.The documents are registered through b

90、atch processing,after which users can select the document ID and input their queries via a WebUI.Currently,there are over 100 projects using RAG within DOCOMO.AI Content Moderation Given that LLMs can generate text freely based on input prompts,there is a risk of unintentionally producing unethical

91、content,especially in business contexts such as advertising,press releases,and email communications.Although manual checks are conducted,an automated system to detect unethical content was needed internally as a safeguard.Thus,a custom machine learning 13 model is created that categorizes and detect

92、s unethical expressions based on DOCOMOs internal standards.On the platform,users can press a check button to automatically detect if the LLM output contains any unethical content.Figure 2 Components of DOCOMO LLM Value-Added Platform Moving forward,alongside promoting internal use of the LLM Value-

93、Added Platform,DOCOMO aims to implement additional features requested by internal users,enhance RAG precision,integrate with internal databases and AI systems to develop agent functionalities,and add more LLMs.These efforts aim to further improve the platforms convenience and utility.3.3 KT SLM/LLM

94、Platform KTs LLM platform provides all services from AI infrastructure and hardware to various service applications in Multi Model,and manages safety for all AI products,making it a full stack AI platform.The platform pursues Customization,Reliability&Responsibility,Convenience,and Efficiency in a w

95、ide range of areas.The main features of the platform can be divided into the following five categories:Governance An E2E Governance system has been established that evaluates and monitors the safety and reliability of AI-applied products using KTs unique Responsible AI Framework and tools.Infra and

96、Data A High-quality Sovereign Cloud has been built that securely and conveniently integrates LaaS,PaaS,and SaaS in a multi-cloud system,including KTs own cloud.Platform An AI studio platform has been built that enables AI Ops,including SLM/LLM Ops,data training,and integrated management.Model It pro

97、vides various language models,including KTs self-developed language model,Mi:dm,and various open source-based models,allowing customers to choose according to their situation and needs.Agent Domain-specific AI Agents are developed,provided and optimized for customers needs and context.14 Figure 3 KT

98、s SLM/LLM Platform KTs LLM platform can be utilized in various areas,and major use cases where the platform has been applied include AICC,AI consultation assist,Genie TV copywriting,AI spam filtering,image generation object book,and AI newsletter,contributing significantly to productivity improvemen

99、t and new customer experiences.4 Generative AI Application Cases China Mobile,DOCOMO and KT,as leading telecommunications operators,have become the backbone of participating in the construction of intelligent computing infrastructure and promoting the generative AI application cases.In particular,le

100、veraging our rich operational experience in telecommunications networks,we are internally building autonomous networks with Network OAM agents to ensure cost reduction and efficiency enhancement,improving customer experience with consultant AI assistant,as well as externally actively exploring and d

101、eploying new types of innovative AI services.4.1 Generative AI for Network O&M China Mobile AI+Network Native Intelligent Agents Bridge the Digital Divide in Operations and Maintenance Systems,Achieving Efficient Network Operation The field of network operations and maintenance(O&M)is characterized

102、by high system complexity,diverse expert knowledge,and significant human resource investment.To further enhance the scheduling and orchestration capabilities between 15 systems,as well as the integrated analysis capabilities of network data across various domains,leveraging the understanding,selecti

103、on,thinking,action,and memory characteristics of LLM intelligent agents can accurately comprehend the on-site needs of network O&M personnel,provide 24/7 online services such as precise Q&A for home broadband business knowledge,immediate verification of current network data,root cause analysis for i

104、ssues,generation of processing solutions,scheduling execution,and result verification,which assists experts in various fields to handle current network issues,offering suggestions and optimization directions.Figure 4 Jiutian Agent for Home Broadband Installation and Maintenance As shown in Figure 4,

105、in handling home broadband comprehensive scheduling scenarios,China Mobiles Jiutian Network LLM identifies installation and maintenance intents through multimodal interaction methods.Combining its own abilities of memory,observation,thinking,and execution,it formulates plans and breaks down tasks.Ba

106、sed on the results of each step,it determines the next steps transition and solution output,completing the main process control and intelligent flow.During task execution,it confirms the tools and parameters to be leveraged,completes result analysis,and outputs solutions.Ultimately,it achieves unifi

107、ed scheduling and orchestration of IT capabilities,data,and small intelligent capabilities of various domain-specific network management systems,realizes collaborative innovation between large and small models,reshapes the integration model,and provides new momentum for solving autonomous network te

108、chnical challenges.China Mobiles Jiutian Home Broadband Comprehensive Adjustment Scenario Intelligent Agent has currently docked with five comprehensive scheduling solution providers,covering the comprehensive scheduling systems of 27 provinces,and is expected to complete full network rollout before

109、 the end of the year.Taking Henan Province as an example,with an installation and maintenance team of 9,000 people and a daily workload of about 2030 tickets,the intelligent processing ratio of ticket dispatch reaches 60%,achieving a reduction in costs and increase in efficiency of 7.3 million RMB/y

110、ear within the province.16 To further enhance the intelligent agents processing capabilities for complex network tasks,intelligent agent technology is expected to evolve from single intelligent agents to hybrid intelligent agents.Through multi-agent negotiation,it will achieve autonomous planning of

111、 network tasks,self-optimization of capabilities,and breakthroughs,ultimately realizing the transition from assisted decision-making to autonomous decision-making in the field of network operations and maintenance.DOCOMO Social Media Based Analysis using LLM for Improving Mobile Network Performance

112、DOCOMO has been identifying low-network quality areas and detecting signs of customer complaints by combining network traffic data,customer complaint information,and location information.In addition to these efforts,from August 2023,DOCOMO began utilizing the DOCOMO LLM Value-Added Platform to analy

113、ze network quality-related content posted by customers on social media.This allows for a more specific and rapid identification of areas requiring intervention to enhance network quality.We collect 22,000 posts related to DOCOMOs network quality from X per month.By executing two prompts on these pos

114、ts,we extract(1)location information of low-network quality areas,(2)the degree of dissatisfaction,and key keywords related to network quality.Prompt1:Extracting Location Information-Output:Location name(facility name),prefecture name,and location category Prompt2:Estimating Dissatisfaction Category

115、 and Extracting Keywords-Output:Dissatisfaction category(dissatisfied,satisfied,neutral,or others)and keywords related to mobile network quality For each post,we use LLM to extract information as follows:Figure 5 Samples of Output We then combine the LLM analysis results with our own mobile network

116、data and plot the location points on a map.By leveraging this map information,we can identify areas with poor network quality and use it to improve network quality.We manually labeled approximately 300 posts from X as either satisfied or dissatisfied,17 and the classification accuracy of the LLM was

117、 over 75%.Tasks that used to take several days with manual classification can now process 20,000 data entries in a day by using LLM.Since the system began operating in August 2023,we have used LLM classification results to select and implement targeted network improvement measures at 2,000 locations

118、 nationwide,including railway lines.In the future,DOCOMO will continuously improve its network quality by conducting a more comprehensive analysis of network quality across a broader area.This will be accomplished by integrating data collected from X with feedback on network quality gathered from cu

119、stomers through the contact center.4.2 Generative AI for Customer Service DOCOMOs Communication AI for Customer Support For customer service,such as call center and chatbot,existing dialogue AI and other technologies have not been able to provide personalized responses.To achieve a higher level of c

120、ustomer satisfaction,DOCOMOs Communication AI understands each individuals situation and attribute,and reflects that information to customer support.Figure 6 Technology assets that realize AI customer service Figure 6 shows the system configuration diagram of Communication AI,which combines several

121、AI assets of DOCOMO.The component surrounded in yellow line is Voice DX Platform,which is DOCOMOs AI asset for speech recognition.The component in the green rectangle is CX Analysis Technology and docomo Sense,which analyze big data of DOCOMOs users,and estimate customers needs.The results obtained

122、from these AI assets are input as prompts into DOCOMO LLM Value-Added Platform in the blue box,and it will generate answers to the customers.Solutions provided by Communication AI include engaging in conversations that empathize with customers using conversation contexts,making proposals that exceed

123、s 18 customer expectations by understanding customers latent needs from their characteristics and behavioral data.Figure 7 How different AI technologies can work together to provide customer support Figure 7 shows how Communication AI can work as a customer assistant.(i)shows its ability to estimate

124、 customer speech rate and adjust the conversation to match their pace.For example,when a customer speaks slowly,then the AI also speaks slowly.In(ii),after the AI assistant resolved the customers problem,it proactively suggests DOCOMOs another service that might be of interest to the customer based

125、on the amount of data used by the customer.(iii)shows that Communication AI can have more flexible conversations compared to conventional chatbots that use fixed sentences by utilizing generative AI.Communication AI is under verification for commercial release after 2024.KT AI Consultant Assistant I

126、n the rapidly accelerating business competition for productivity,improving corporate productivity has become a crucial issue.To enhance productivity,it is important to consider various factors that can reduce employee work efficiency,costs,and time.The needs for Employee AI agents are growing to red

127、uce costs and time,automate tasks to prevent errors,and quickly process and solve problems that arise from human contributions.One issue is the efficiency of customer service representatives performing repetitive tasks.As all tasks are handled manually by people,efficiency decreases,and a considerab

128、le amount of time is consumed,leading to resource wastage.KTs AI Customer Assist,which applies generative AI,can utilize the amount of customer service data to summarize and analyze consultation content,thereby reducing the workload of customer service representatives.This AI Consultant Assist is an

129、 AI agent solution applied with Mi:dm SLM,which supports the efficient management of customer responses by utilizing real-time STT technology,curation and summary features.This service can analyze real-time voice and conversation contents during or after the consultation.It classifies the consultati

130、on 19 contents and recommends necessary information to individuals.Additionally,using this AI model,it can collect consultation data,learn from that data,and provide knowledge-based content recommendations and summaries of actual consultation contents.Figure 8 AI Consultation Assist Service Flow We

131、can achieve hyper-automation by rapidly processing repetitive tasks through AI agents so that it can improve work efficiency and productivity,reduce time and costs,and address the issue of declining productivity.The speed of operations can be increased by hyper-automating from standardized corporate

132、 tasks.Furthermore,work productivity can be greatly improved by preventing human errors or mistakes.Additionally,stress from monotonous work can be reduced,leading to significant time savings and increased employee satisfaction by replacing repetitive and manually performed tasks with automated oper

133、ations.Through this service,approximately 80%of the consultation content was able to summarize,reducing the average consultation time by up to 22 seconds as well as saving about 6 billion won in total costs.KT Genie Copywriting and Genie Themes Cases of generative AI application in building personal

134、ized marketing strategies based on individual customer interests or behavior patterns are expanding.Previously,marketers had to generate promotional copy phrases and other tasks by referencing external resources,with most of the work being done manually.Naturally,applying standardized marketing meth

135、ods to proceed with tasks led to many limitations in fully satisfying customer needs.To address this,there is a growing trend to expand services centered on personalized marketing strategies that match the interests and behavior patterns of individual customers using generative AI agents.KTs AI agen

136、t services,such as Genie TV and Genie Theme,which leverage generative AI,can quickly learn and analyze amounts of customer data to understand individual customer interests,preferences,and behavior patterns.Based on the information,they can automatically generate and recommend optimized content for i

137、ndividual customers.Companies can now effectively conduct real-time,customer-centric marketing strategies and campaigns,thereby focusing on attracting and managing customers by 20 meeting their hyper-personalized needs.Genie Copywriting is an AI agent that generates curation phrases needed for the h

138、ome portal of Genie TV,utilizing generative AI to create new marketing phrases according to the desired keyword length.Genie Themes is an agent that generates promotional phrases visible on VOD or movie content of Genie TV,allowing users to select preferred tags and create personalized phrases using

139、 the resources.Figure 9 Genie Copywriting and Genie Themes Service Flow Marketers can easily create personalized messages for posters,movies,etc.,using the AI agent based on accumulated data through the curation platform.The service was implemented by utilizing KTs Mi:dm AI model,and it goes through

140、 the process of automatically generating,verifying,and applying the phrases.Through this process,it is possible to generate new phrases of various lengths and create personalized curation phrases for different categories by combining the tags preferred by the user,utilizing minimal resources.It redu

141、ced production resources and improved the productivity of marketers tasks that were previously done manually with AI,enabling them to quickly recommend the desired content in a customized manner for each customer.This allowed for the fulfillment of hyper-personalized customer needs in a timely manne

142、r.KT AI Spam Filtering Most corporate infrastructure/system operations were previously executed in a rule-based method,classifying and addressing already analyzed types of problems according to predefined rules.As technology advances and the complexity of system configurations and operational elemen

143、ts increases,the issue of risk management has become increasingly important.To overcome new types of disruptions and error issues in telecommunication networks,we are being made to analyze network conditions and preemptively block problems using AI agents.SPAM is also part of the network issue,and K

144、T is promoting services that apply AI models to preemptively block users risks from SPAM messages.KTs AI spam filtering solution,developed with the SLM model,is an AI agent that 21 applies an action model to recognize,judge,predict,and process results,capable of analyzing and judging its patterns by

145、 self-learning various SPAM message data,and can also judge new forms of SPAM messages,playing a crucial role in protecting customers safety and enhancing the efficiency of network operations.This agent solution was developed by using Mi:dm SLM and it was designed to combat illegal SPAM in Korea.Thi

146、s solution works by pre-processing data from a database containing illegal SPAM datasets.The AI model then performs the context recognition and the intention analysis to determine comprehensively whether the messages are SPAM or not.Figure 10 AI Clean Messaging Service Flow When users receive ambigu

147、ous messages containing URLs among the texts,these messages are stored in the database.Subsequently,when similar types of messages are received,pre-processing is performed to utilize the KT AI model,which recognizes the phrases and analyzes the intention of the messages.Based on the analysis results

148、,the solution determines whether the messages are SPAM,their risk levels,and the degree of maliciousness.By utilizing small language models instead of large language models,the time spent can be reduced to determine and infer SPAM messages.Additionally,it is possible to preemptively block the messag

149、es related to illegal loans or fraudulent investments by reducing the time of AI training to apply new types of SPAM messages quickly.5 Future Outlook and Industry Suggestions Looking forward,China Mobile,Korean Telecom and DOCOMO will further strengthen the exploration in Generative AI agents and f

150、oster a collaborative standardized ecosystem together so as to accelerate the further development of advanced theoretical innovations and value-adding integration platforms for the intelligent agent industry.To start with,successful transformation using AI agents will be effective in improving the e

151、fficiency of work productivity so that makes development available in various business fields,including customer AI agents that enable hyper-personalized customer experience,business AI agents offering executive decision support,employee AI agents 22 for improving employee work productivity,and infr

152、astructure AI agents assisting both predictive and proactive network operations.Secondly,basic algorithm research needs to be improved to provide a solid theoretical foundation for the autonomy and reasoning acceleration of intelligent agents in order to cover the above application scenarios.At the

153、current stage,focus on the ability enhancement of single AI agent,such as visual work-flow orchestration,memory,and plugin customization;the second stage plans to achieve the application of multi-modal multi-agent in industrial scenarios;in the long run,the evolution from AGI(Artificial General Inte

154、lligence)to ASI(Artificial Super Intelligence)requires controllable mechanisms to effectively supervise/align AI to ensure its conformity to human values.Last but not least,we are willing to work with all parties in the industry to reach consensus on the target vision of and evolving path to a more

155、intelligent,efficient,and convenient future,which will be guiding us to reasonably position,orderly work,and efficiently collaborate to provide strong technical support for the sustainable development of the industry in the following areas:AI agent grading standard to specify long-term development m

156、ilestones and capability construction paths,guiding the corresponding intelligent agent R&D planning;Multi-agent collaboration standards to ensure the effective representation of atomic capabilities(e.g.LLMs)and the realization of responsible AI,including communication protocols,technical architectu

157、res,general agent framework;and Collaboration mode between humans and AI agents to optimize human-computer interaction,and improve collaboration efficiency.6 Abbreviations AI Artificial Intelligence AICT AI&ICT AX AI Transformation HW Hardware LLM Large Language Model MLOps Machine Learning Operations O&M Operation and Maintenance Q&A Question and Answer RAG Retrieval Augmented Generation R&D Research and Development SLM Small Language Model SW Software