SK Telecom has advanced to the third phase of the government's proprietary AI foundation model project with its proprietary AI model 'A.X K2,' which boasts 6.88 billion parameters. Yoon Kyung-sang, head of SK Telecom's AI CIC, revealed plans to expand beyond this model to provide AI services that the general public can experience in their daily lives.
In a column on SK Telecom's newsroom on August 24, Yoon emphasized that the 'intelligence improvement loop,' which combines model performance, inference economy, service completeness, and trust, will determine future AI competitiveness. He noted that while competition for frontier models remains fierce, the key differentiator will be how efficiently the acquired intelligence is deployed and transformed into real-world problem-solving.
Yoon highlighted the significance of A.X K2's entry into the third phase of the proprietary AI project, stating that it has been recognized not only for its technical performance but also for its applicability in industries such as defense, manufacturing, law, and taxation. He pointed out that possessing a good model and operating a service that millions trust and use daily are two separate challenges, stressing the importance of full-stack competitiveness that encompasses 'infrastructure-model-service.'
To secure inference economy, Yoon proposed several strategies: activating a portion of parameters for each request using a mixture of experts (MoE) structure, varying inference depth based on question difficulty, employing lightweight and acceleration technologies to maintain quality while enhancing serving efficiency, and distributing tasks so that lightweight models handle simple queries while large models tackle complex tasks. He explained that system and semiconductor-level optimizations, including memory hierarchy management and computational resource allocation, are essential for achieving sustainable economy at a national scale.
Yoon defined the core of future AI competition as 'the fastest intelligence improvement loop.' He stated that the speed of the cycle, which involves identifying failure points during actual use, generating evaluation and learning signals, and improving and redistributing models and services, will determine competitiveness.
He identified the next stage of AI services as a transition from merely answering questions to executing tasks on behalf of users and delivering results. Yoon noted that in the agent era, it is insufficient to explain service quality solely by the accuracy of individual models, as small errors can accumulate in processes involving planning, searching, tool invocation, and execution. Therefore, he emphasized the importance of 'end-to-end success rate,' which ensures that user intent translates into actual outcomes.
'Everyone's AI' is an initiative led by SK Telecom in collaboration with specialized companies across various sectors, including finance, mobility, media, education, healthcare, public services, taxation, and caregiving. Yoon explained that SK Telecom's role is not to create all services directly but to act as an 'orchestrator' that connects user intent with optimal models, technologies, and specialized services.
The plan for service expansion includes starting in everyday areas such as finance, mobility, media, education, and health, and then broadening to specialized domains for young entrepreneurs, small business owners, and seniors, including tax administration and caregiving. Regarding reliability, Yoon stated that SK Telecom will ensure response speed through model optimization and serving acceleration technologies, while also establishing a security system that continuously checks risks through objective evaluations of model quality, ongoing AI red teaming, real-time safety monitoring, and guardrails to safely handle personal and sensitive information.
Yoon remarked, 'Trust should be defined not as a state free of risk but as the ability to manage risk.' He added that the essence of responsible AI operation lies in the ability to continuously identify risks, limit their impact, quickly recover from issues, and improve to prevent the same errors from recurring.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.