Dong-A Ilbo IT Science Team journalists introduce noteworthy technologies, trends, and companies in IT, science, space, and bio fields. “What is this company?” The behind-the-scenes stories of tech companies changing the world with technology! From ideas that surprised the world to the current concerns of founders, they delve into everything about them.
On the 7th, Cho Dong-yeon, Vice President of SK Telecom's Innovative Model Division, explains the 'Independent AI Foundation Model'. Provided by SK Telecom
It is considered a consortium equipped with a full stack from semiconductors and data to services centered around Large Language Models (LLM). Additionally, with Seoul National University and KAIST conducting preliminary research necessary for technology, it can cover all aspects, which is a strength.
On the afternoon of the 7th, Cho Dong-yeon, Vice President of SK Telecom's Innovative Model Division, was met. SK Telecom is one of the five elite teams selected for the 'Independent AI Foundation Model' project supported by the Ministry of Science and ICT. The Independent AI Foundation Model project aims to nurture a 'national representative' AI at a level comparable to big tech. Fifteen teams applied for the project, which was narrowed down to ten through a document review, and further reduced to five through a presentation evaluation. It overcame a competition ratio of 3:1 among prominent companies.
When asked about the strengths of the SK Telecom consortium compared to other consortia, Vice President Cho showed modesty by saying, “It is difficult to compare with other consortia,” but confidence was felt in his explanation of the strengths of the consortium formed by SK Telecom.
SK Telecom participated in this project by forming a consortium (team) with △Krafton (domestic game listed company) △42dot (mobility AI) △Rebellion (inference AI semiconductor (NPU) manufacturer) △Liner (AI agent company) △SelectStar (AI data specialized company) △Seoul National University Industry-Academic Cooperation Foundation △KAIST.
In this project, SK Telecom plans to apply 'omnimodal' technology that can integratively process various forms of data such as text, images, voice, and video. Furthermore, it plans to develop AI agents that all citizens can easily utilize and lead AI innovation and transformation in the office, manufacturing, automotive, gaming, and robotics sectors.
Q. There are opinions that SK Telecom is lacking compared to other companies like Naver or LG AI Research Institute.
For example, SK Telecom's Korean-specialized LLM, A.X 4.0, is implemented using Alibaba's open-source model, which seems to lead to such evaluations.
What do you think about this?
“I think having models implemented in two ways is rather a strength. At first glance, it might seem easier to create by bringing in an external model. However, technically, this method is more difficult. External models are already trained, so if retrained, they may forget what they learned before. The technology to keep learning without forgetting is more challenging.
Moreover, SK Telecom's A.X 3.0 was created from scratch (built from the very first stage of the model). In other words, SK Telecom has more, more diverse, and more challenging technologies.
Furthermore, we aim to create an LLM with better performance than now. The technology commonly used in current LLMs is called 'Transformer'. Simply put, a Transformer creates a sentence sequentially by generating the next word that matches the created character. However, in the case of 'Diffusion LLM', words pop out as if they appear suddenly, creating sentences much faster.
We are not necessarily committed to Diffusion LLM. Of course, we will start based on Transformers initially, but we aim to develop technology beyond Transformers and implement LLM based on 'Post-Transformer'.”
Q. You have stated that you will develop a super-large AI that surpasses the existing domestic LLM scale in this project. What does it mean to surpass the existing scale?
“For example, the 'Kimi K2' recently unveiled by Chinese AI startup Moonshot has 1 trillion (1000B) parameters. Currently, 1000B level is emerging globally. The standard set by the government for this project was '95% of the global level'. While we cannot specifically say what size model we will create, it would be difficult to meet the government standard without aiming for that level (1000B). Based on the current situation, we think it is possible to achieve that scale (1000B) during the project period.”
On the 7th, Cho Dong-yeon, Vice President of SK Telecom's Innovative Model Division, explains the 'Independent AI Foundation Model'. Provided by SK Telecom
Q. Although included in the five teams initially supported, one team will be eliminated through evaluations every six months in this Independent AI Foundation Model project. The Ministry of Science and ICT plans to narrow it down to a total of two teams by the first half of 2027.
“Our consortium aims to be among the two teams, and we will strive to achieve that. We see possibilities in various aspects.
The biggest is technology. SK Telecom did not start LLM suddenly. It has been continuously researching the latest technologies since before the term LLM existed, so the technology has been accumulating.
Also, the companies forming the SK Telecom consortium deal with not only NPU but also data and service aspects. No matter how high a benchmark score an LLM receives, it is meaningless without service experience. By providing services, we can identify issues with the LLM and the necessary technologies, thereby gaining global competitiveness. Our consortium has experience in both technology and service aspects, which is a significant strength.”
Q. The SK Telecom consortium is said to prioritize applications in manufacturing, automotive, gaming, and robotics in the B2B sector. Does this mean it will also develop into physical AI?
“Agents have two aspects: agents on computers and physical agents. In fact, we are pursuing both directions. We are considering the application of physical AI that connects to 'knowledge-based' models that assist in manufacturing sites and can actually manipulate.”
Q. Among the companies forming the consortium, the game company 'Krafton' and the autonomous driving mobility company '42dot' stand out.
“Krafton may appear to be a game company, but it has a significant number of advanced AI personnel. It is among the top in Korea. In addition to the technology to create LLM, it possesses abundant AI technology. As a game company, it also has strengths in the visual field of AI. As a company providing very large services, it is expected to provide much help within the consortium.
Moreover, the space where AI can be most tangibly experienced outside of computers is thought to be automobiles. Whether driving or using public transportation, everyone rides in a car. Therefore, 42dot holds significant meaning as a member in terms of both physical AI and application aspects.”
Q. Rebellion, a manufacturer of inference AI semiconductors (NPU), is also included in the SK Telecom consortium. What role does it play?
“Currently, when a model is created, it is adjusted to fit the NPU. To put it simply, it feels like squeezing a model into a fixed NPU. In this case, the model's performance may not be fully realized. However, creating an NPU tailored to the model can result in a more efficient and faster-operating LLM.
Of course, there are physical limitations to NPU, so it cannot be made freely. Nevertheless, Rebellion can provide insights on how the model should change to optimally design the NPU, creating synergy in the process.”
Q. Other elite teams also have Seoul National University and KAIST as members. What strengths do the Seoul National University and KAIST researchers participating in the SK Telecom consortium have?
“There are many researchers who have studied various technologies involved in LLM or LLM itself. Professor Kim Geon-hee from Seoul National University has outstanding technology in the multimodal field. Professor Do Jae-young has worked extensively on language and agent technologies at Amazon, such as Alexa. Professor Yoon Sung-ro served as the Chairman of the Fourth Industrial Revolution Committee and has researched various AI technologies in academia. Professor Hwang Seung-won has conducted a lot of research in language technology. Additionally, Professor Lee Ki-min from KAIST possesses technology that extends to agents and physical AI.
We will do our best with the companies in the SK Telecom consortium and the researchers from Seoul National University and KAIST to be among the final two teams.”
ⓒ dongA.com. All rights reserved. Reproduction, redistribution, or use for AI training prohibited.
Popular News
