SK Telecom set to complete Korean language LLM in June

The LLM was co-developed with global AI companies such as OpenAI and Anthropic. Telcom-specific data such as its products, membership benefits and information from customer centers are gathered and trained on multiple LLMs.

Eric Davis, vice president of SK Telecom’s AI Tech Collaboration division, explains the company's telecom-tailored large language models at the company's headquarters in central Seoul on Tuesday. [SK TELECOM]

SK Telecom, Korea’s leading mobile carrier, will complete the development of a Korean language large language model (LLM) tailored for telecommunication services in June, the company said on Tuesday.

The LLM was co-developed with global AI companies such as OpenAI and Anthropic. Telcom-specific data such as its products, membership benefits and information from customer centers are gathered and trained on multiple LLMs — SK Telecom’s A.X, OpenAI’s GPT and Anthropic’s Claude.

Once the development is completed, the Korean language service will roll out first, to be followed by releases of global versions of the telco-specific models in the future.

SK Telecom hopes to find business models in sectors such as AI Customer Centers, network infrastructure and logistics, although the road map has not yet been clearly outlined.

Eric Davis, vice president of SK Telecom’s AI tech collaboration division, explained that the company chose to train on multiple LLMs because it is not easy for one general-purpose LLM to provide services and solutions tailored to the different interests of telecom companies.

“We fine-tuned and benchmarked various LLMs and fed telecom data and domain expertise, allowing them to be selectively used in purposes that each model is best equipped for,” Davis said.

The fine-tuned telco LLMs then undergo a reinforcement learning process based on human feedback (RLHF), which involves assessing the quality and satisfaction of real customer queries.

Finally, tests of the telco LLMs’ language abilities, reasoning abilities and performance of related tasks are conducted by third-party clients looking to adapt the models to their respective businesses.

SK Telecom emphasized that the telco LLMs will radically reduce the time spent on tasks at customer service centers. Currently, it takes about 3 minutes for an employee at a center to handle one call, and 30 seconds or more for post-consultation tasks. Telco LLMs can significantly reduce this time as they allow employees to perform tasks simultaneously during the call as it provides an automated summary of the call and suggests solutions to the customer demands.

BY LEE JAE-LIM [lee.jaelim@joongang.co.kr]