A smart ‘Telco LLM’ that thoroughly understands communication services will be launched as early as this June.

On the 30th, SK Telecom announced that it is developing the ‘Telco LLM,’ which has learned internal guidelines of telecom operators, such as South Korea’s specialized terms like 5G plans, T Membership, and public support funds, as well as AI ethical values. The development is expected to be completed by June.

‘Telco LLM’ refers to a language model specialized in the telecommunications industry, unlike general models like GPT and Claude.

Through partnerships with OpenAI and Anthropic, SKT is collecting vast amounts of data on telecom services, products, membership benefits, and customer consultation patterns, sorting through structured and unstructured data to train it on A.X, ‘GPT,’ and ‘Claude’ to create an LLM specialized in telecommunications.

The Telco LLM is fine-tuned based on various general models such as SKT’s A.X, OpenAI’s GPT, and Anthropic’s Claude. This is part of SKT’s multi-LLM strategy.

Eric Davis, head of AI Tech Collaboration at SKT, explained, “It’s not easy for telecom companies to solve various services and problems with a single general LLM,” adding that “creating multiple Telco LLMs tailored to telecom data and domain know-how through fine-tuning and benchmarking is SKT’s unique multi-LLM strategy.”

This strategy allows telecom companies to optimize LLM usage for various applications such as AI contact centers (AICC), distribution networks, network operations, and in-house tasks.

SKT is developing various Telco LLMs by incorporating Korean telecommunication-related data into its LLM A.X, OpenAI’s GPT-4, and Anthropic’s Claude through collaboration.

With a diverse lineup of specialized LLMs for telecommunications, they aim to diversify performance and cost-saving measures.

On the 5th, OpenAI announced via its official blog that it has launched custom models to help companies train and optimize AI models for specific domains. They mentioned that they improved the dialogue performance related to telecommunications in Korean through fine-tuning GPT-4 in collaboration with SKT, citing it as a representative case of industry-specific fine-tuning.

Compared to general LLMs, the Telco LLM can carry out a high level of generative AI tasks within the telecommunications domain, making it highly applicable, according to SKT.

To build the Telco LLM, SKT first collects telecom data and sorts and refines structured and unstructured data. Based on this, they fine-tune the general LLM specifically for telecom firms and conduct reinforcement learning with human feedback (RLHF) followed by final benchmarking (model evaluation).

General LLMs struggle to address customer requests effectively, such as recommendations for plans, due to a lack of understanding of telecom knowledge like number portability processes and procedures. The fine-tuning process of the Telco LLM aims to solve this issue by learning from additional telecommunication-related data.

The fine-tuned Telco LLM undergoes a reinforcement learning process based on human feedback.

Consultants evaluate the content answered by the Telco LLM based on quality and satisfaction. This involves scoring how useful the answers were to customer inquiries, how well the context was understood, etc., through evaluations made by a human.

Finally, they benchmark the Telco LLM’s language skills, reasoning abilities, and task performance specific to telecommunications.

For example, if the Telco LLM receives low scores in areas like assisting customers with plan inquiries or service modifications, additional relevant data is gathered, and learning continues through the fine-tuning process. This entire cycle is repeatedly conducted to make the Telco LLM smarter.

Currently, a customer service consultant takes approximately three minutes to handle a customer call and over thirty seconds for post-consultation processing. With the introduction of the Telco LLM, it is expected to significantly reduce the time by providing solutions to consultants during customer calls and summarizing consultations afterward.

Previously, customer service consultants required extensive experience and training to become adept at organizing customer inquiries, searching and summarizing necessary documents, and recording the consultation process. The Telco LLM streamlines this entire process.

Moreover, among the Telco LLMs, the Claude version, which incorporates telecommunication-related data, is rigorously learning ethical principles that AI must follow and can accurately grasp the nuances of emerging slang and Korean insults or threats in the context.

The Telco LLM is also useful in the operational management of telecom network infrastructure.

When infrastructure operators encounter issues during network monitoring, they can input questions into the Telco LLM in real-time and receive answers to solve the problems.

Since the Telco LLM quickly provides relevant answers based on information from equipment manuals and response cases, it can shorten response times compared to manually searching for information.

SKT plans to expand the use of the Telco LLM for data analysis generated during infrastructure management and information retrieval based on accumulated data.

Jung Min-young, in charge of SKT’s AI platform, stated, “The Telco LLM will enhance operational efficiency in various areas of telecom operations, from customer service and infrastructure to marketing/network distribution points, legal affairs, and HR,” adding that “we plan to continuously increase use cases utilizing the Telco LLM.”

SKT has also unveiled an “Intelligence Platform” that allows telecom companies to efficiently build and develop generative AI applications.

This is a package for enterprise AI development and operation that encompasses everything from multi-LLM to multimodal, orchestration, and retrieval-augmented generation (RAG).

SKT’s Intelligence Platform does not rely on a single LLM. General LLMs can incur high inference costs proportional to performance and may have a low understanding of specific domains. In applying this to telecom services, it is inefficient, so SKT believes using a multi-LLM specialized in telecommunications is necessary.

SKT is currently applying this easily and efficiently usable Intelligence Platform in services such as A.(A-dot) and plans to continually expand applicable use cases.

There’s an endless range of services that telecom companies can implement, such as customer service call bots, chatbots, distribution channel assistants, infrastructure assistants, and internal business innovations.

SK Telecom stated that “telecom companies, both in South Korea and globally, as well as enterprises with similar job characteristics, will likely save costs and time on developing large-scale platforms through the Telco LLM Intelligence Platform.”

Lee Sang-jin daedusj@autodiary.kr