SK Telecom will enter its “AI for everyone” business by highlighting full-stack capabilities that link its proprietary AI model, infrastructure and public-facing services. It plans to build an “intelligence improvement loop” that uses experience and data secured from real services to improve its model.

Ryu Kyung-sang (유경상), head of SKT’s AI CIC, outlined the strategic direction in an SKT Newsroom column on Sunday. He said the focus of AI competition is expanding from bigger models and higher inference performance to execution capabilities through real services and the speed of improvement.

SKT recently unveiled its proprietary AI foundation model “A.X K2” with 688 billion parameters. A.X K2 strengthens capabilities such as math and science reasoning, Korean-language knowledge, long-form understanding and agent functions compared with the previous model. An elite SKT team using A.X K2 passed the second-stage evaluation of the government’s project to develop an independent AI foundation model and advanced to the third stage with Upstage and LG AI Research.

◆ Competition over the “intelligence improvement loop” beyond model performance

Ryu said it will be difficult to assess future AI competitiveness based on model performance alone. He said factors such as a model’s intelligence and inference costs, service completeness, the speed of retraining and improvement based on real data, and reliability must be combined into a single system. He said that in an environment where millions of people use large-scale models simultaneously, it is hard to secure economic viability by processing all requests with the highest-performance model.

To address this, SKT will use a mixture-of-experts (MoE) architecture that selectively activates some parameters per request, technology that adjusts the inference level depending on question difficulty, and model slimming and acceleration. It will route simple questions to a lightweight model and complex tasks to a large model to improve service operating efficiency.

It is also pursuing optimisation at the system and semiconductor level, including memory management and allocation of computing resources. It aims to lower inference costs across the entire process from model to service by leveraging capabilities to design and operate AI infrastructure directly.

Ryu cited “speed of improvement” as a key factor in AI competition. He said competitiveness depends on how quickly a company can repeatedly identify where a model fails in real services, reflect that in evaluation and learning, and provide the improved model back to users.

SKT plans to deliver this through full-stack AI capabilities spanning model, service and infrastructure, and experience in operating large-scale services. It will use operating experience from its AI service “A. (A.)”, which has more than 10,000,000 monthly active users, to link feedback secured from services to improvements in models and systems.

◆ Consortium of 20 organisations...from AI that answers to AI that executes

In its AI-for-everyone business, SKT is focusing on implementing “AI that executes” that goes beyond simply answering users’ questions and connects requests to real actions and results.

Ryu said that in the era of AI agents, not only a model’s accuracy but also the “end-to-end success rate” that links user intent to real results through multiple stages such as planning, search, tool calling and execution will become important.

SKT has formed a consortium with companies in areas including finance, mobility, media, education, medical care, the public sector, tax services and caregiving. About 20 companies, institutions and organisations, including SKT, are participating in the consortium. Separately, it has signed memorandums of understanding with 11 companies including Hana Bank, NH Nonghyup Bank, Tmoney Mobility and Bucketplace and is pursuing additional cooperation.

SKT will serve as an “orchestrator” in the consortium, analysing user requests and linking them to suitable AI models, technologies and specialised services. The AI-for-everyone service will start in everyday areas such as finance, mobility, media, education and health, then expand to specialised areas such as tax, administration and caregiving for young people, small business owners and seniors.

◆ From performance to trust...aiming for everyday AI used by the public

SKT said reliability and safety will become more important as AI penetrates deeper into people’s lives and the execution of services such as finance and medical care, and it is strengthening related technology.

It will raise response speed through model optimisation and serving acceleration, while applying objective evaluation of AI model quality, ongoing red-teaming, real-time safety monitoring, guardrails and other measures. It will also build a security system to protect personal data and sensitive information.

Ryu said the ultimate goal is to build a virtuous cycle in Korea that leads from AI model development and service expansion to real usage experience and training of next-generation models. SKT plans to combine A.X K2, AI infrastructure, large-scale operating experience from A., and the capabilities of specialised companies by sector to make AI for everyone an AI service that the public can use in daily life.

Ryu said, “Future AI competition will not be about making the best model once, but about building the fastest ‘intelligence improvement loop.’”