SKT will launch 'Telco LLM' specialized in telecommunications services in June. Telco LLM is an AI model developed by training general LLM with Korean telecommunications-related data. Through this, telecommunications companies are expected to be able to secure telecommunications expertise and cost efficiency by applying AI models optimized for each service and technology.

SKT held a press conference at the SKT Tower Supex Hall on the 30th and introduced Telco LLM, which was created based on general models such as Open AI and Antropic. SKT created Telco LLM after 8 months of development and will release the domestic version as early as June. After that, it plans to focus on releasing the global version.

The advantages of a telecommunications specialized LLM include improved performance, speed, and cost efficiency. SKT's Jeong Min-yeong, in charge of AI platform, said, "Compared to general LLM, Telco LLM can perform high-level generative AI tasks in the telecommunications field, so it has high usability and will be able to measure quantitative indicators when applying telecommunications-related tasks and services."

Telco LLM is part of SKT's 'multi LLM' strategy. It is explained that telecommunications companies can utilize Telco LLM for specific tasks such as AICC, network operation, distribution network, marketing, and internal work. Eric Davis, head of SKT AI Tech Collaboration, explained, “It will not be easy to solve the various services and problems that telecommunications companies want to provide with a single general LLM.” He added, “We have created various telco LLMs tailored to telecommunications data and domain know-how, and have enabled them to be selected and used according to the situation.”

■ General LLM + Telecommunications Data Learning = Telco LLM

SKT collected a vast amount of data on telecommunications companies' services, products, benefits, and customer consultation patterns based on its own AX (A.X), OpenAI's GPT-4, and Antropic's Claude, and selected structured and unstructured data to build a Telco LLM by training it on a global general foundation model.

On the 5th, Open AI announced its collaboration with SKT by releasing a customized model that supports training and optimizing AI models to fit a company’s specific domain. Jeong Min-young, in charge of SKT’s AI platform, also evaluated, “Compared to GPT-3.5, GPT-4 has significantly reduced hallucinations and improved response accuracy.”

In order to build a Telco LLM, the following processes are performed: △Selection and purification of unstructured and structured data, △Fine-tuning of general-purpose LLM for telecommunications companies based on this, △Reinforcement learning based on human feedback (RLHF), and △Final benchmarking (model evaluation). Eric SKT representative said, “The model is continuously improved through a fine-tuning process and is continuously updated on a monthly basis.”

Telco LLM is expected to be used as an AI agent to support counselors at customer centers. If Telco LLM is introduced to customer centers, it is expected to help counselors with their work and shorten the time required to handle consultations. Telco LLM helps counselors understand customer intentions and topics of consultation during calls instead of counselors, and helps counselors search for new information that is continuously updated based on internal information search (RAG) and enable marketing knowledge search (MRC). It also provides a summary of consultations after consultations.

Telco LLM is also useful for the operation of the telecommunications company's network infrastructure. When a problem occurs during network monitoring, the operator can ask Telco LLM a question in real time using natural language to receive a solution, and receive an appropriate response based on information such as equipment manuals and response cases. SKT plans to expand the application of Telco LLM to data analysis generated during infrastructure operation and information search based on accumulated data in the future.

■ ‘Intelligence Platform’ that Helps Build Generative AI Apps

SKT also unveiled an ‘Intelligence Platform’ that allows telecommunications companies to efficiently build and develop generative AI applications. This is a ‘full-stack package for enterprise AI development and operation’ that includes multi-LLM, multi-modal, orchestration, and augmented search generation (RAG).

SKT's intelligence platform makes it easy to use multi-LLM rather than single-LLM. SKT is currently applying it to services such as A. (Adot). The intelligence platform is expected to support automation in customer center call bots, chatbots, distribution channel assistants, infrastructure assistants, and internal work.

Jeong Min-young, head of SKT’s AI platform, explained, “We expect that domestic and global telecommunications companies, as well as companies with similar business characteristics, such as customer centers, infrastructure, marketing, distribution networks, legal affairs, HR, and other internal work, will be able to reduce the cost and time required for developing large platforms through the Telco LLM Intelligence Platform. We plan to continuously expand the use cases of Telco LLM.”