"AI Must Be Improved Quickly and Redeployed to Stay Competitive"

"The future competition in artificial intelligence (AI) will not be about creating the best model once, but about building the fastest 'intelligence improvement loop.'"
Kyungsang Yu, head of the AI CIC at SK Telecom, said in a column for the SK Telecom Newsroom on the 24th that "SK Telecom's A.X K2 is a large foundation model with 688 billion parameters, but it is not sustainable to handle every request with the highest-performance model in an environment where millions of people use it at the same time." He added that "while the cost per token keeps falling as technology advances, the tasks assigned to AI are becoming heavier even faster."
He said that what SK Telecom sees as most important is the speed of improvement. "AI's competitive edge is not determined by a single round of training," he noted. "What matters is how quickly we can repeat the cycle of finding where the model fails in real use, turning those failures into evaluation criteria and training signals, improving the model, inference, and service, and then redeploying it to users." He added that as a well-designed loop accumulates model performance and experience with user context, AI becomes not a fixed product created once, but a system that adapts to each user and evolves on its own as it is used. He also explained that the experience and data accumulated in this process feed into the training of the next generation of models, gradually blurring the line between training and inference.
Yu emphasized that in AI services, what matters is not the accuracy of a single model response, but the end-to-end success rate in which a user's intent leads to an actual result. "The Republic of Korea must now go beyond being a country that uses AI well and become a country that can choose the next direction of AI on its own," Yu said. "That power of choice comes from the cycle I mentioned earlier."