[Edaily Reporter Yun Junghoon ] “In defense artificial intelligence (AI), what matters most is not creating new strategies that humans couldn’t have conceived, but how accurately and quickly it executes a commander’s orders according to established procedures.”

Kang Min-jun, Head of the G.AI Team at SKTelecom(017670), made these remarks regarding the development direction of defense-specific AI models at the “5th Conan Defense AI Summit” held on the 13th at the Nine Tree Locaus in Yongsan, Seoul. He explained that in the defense sector, the core competitive advantage will lie not in AI’s ability to provide creative responses, but in its ability to swiftly and accurately carry out assigned missions while adhering to military doctrine and operational procedures.

Team Leader Kang emphasized, “When performing duties in the military, there are established procedures and doctrines,” adding, “AI should be developed not to generate new ideas, but to accurately and quickly carry out instructions in accordance with those procedures.”

To this end, SKT is focusing on developing a defense-specific foundation model that understands actual work processes and procedures, rather than simply a language model (LM) with a vast store of knowledge.

Team Leader Kang stated, “It is crucial not just to teach the model knowledge, but to teach it how to work.”

This means that defense-specific AI, upon receiving orders from a commander, must be able to plan all necessary steps without omission in accordance with standard operating procedures (SOPs), military doctrine, and manuals, and select the appropriate tools based on relevant data to carry out the task.

He explained that the criteria for evaluating AI model performance have also been changing recently. While in the past, the amount of knowledge a model possessed was key, recent analyses indicate that the ability to understand complex instructions and formulate execution plans, as well as the ability to complete actual tasks from start to finish, have emerged as major evaluation factors.

Team Leader Kang stated, “I believe a model that can plan effectively to ensure instructions are carried out accurately and without omission, in accordance with established procedures, is the very definition of a defense-specific model,” adding, “SKT is also developing models in this direction.”

He also emphasized the importance of data structuring to enhance the performance of defense AI. He explained that while vast amounts of data—such as video, audio, and satellite imagery—are generated on the battlefield, simply collecting this data is insufficient for AI to accurately understand the information.

Team Leader Kang identified ontology technology—which connects the meanings and relationships of different data types—as a key element. He explained that structuring data through ontology enables language models to analyze information and provide responses with greater accuracy.

He also mentioned “Conan ArgoX,” a battlefield information fusion and analysis platform unveiled that day by KONAN TECHNOLOGY INCORPORATION(402030). Team Leader Kang expressed his expectation that data fusion platforms like ArgoX will play a crucial role in structuring various battlefield information.

He also provided an update on the advancement of the ultra-large AI model currently under development by SKT Corporation.

During the first phase of development, SKT focused on securing the technical capabilities needed to develop a large-scale AI model with 500 billion parameters (500B). In the second phase, the company concentrated on improving model performance, achieving a composite score of 70.6 points and ranking first in the overall evaluation among the four participating companies.

In the third development phase, currently underway, the company is focusing on strengthening coding capabilities and the ability to follow instructions, as well as improving the operational efficiency of large-scale models. In particular, it is incorporating publicly available defense data into the model’s training to enhance its understanding of the defense sector.

Team Leader Kang explained, “The third-phase model incorporates publicly available defense data as training data,” adding, “Once the model is complete, it will be possible to ask questions about defense-related data and receive answers.”

The company also outlined the development direction for its Agent AI platform. Whereas previously humans had to design workflows step by step and utilize AI for individual stages, the outlook is that the system will evolve so that, in the future, humans will simply provide guidelines and necessary tools, and the AI will autonomously devise execution plans and review the results.

SKT is standardizing a collaborative approach in which multiple AI agents divide tasks among themselves and synthesize the results, and is applying this to relevant platforms.

Team Leader Kang stated, “AI models should evolve toward accurately and quickly planning and executing human instructions rather than generating new knowledge,” adding, “SKT is preparing defense-specific models capable of implementing this.”