[비즈한국] Naver035420, which has been emphasizing 'Sovereign AI,' has failed in the government-led selection process for a national representative AI. In contrast, LG003550 passed the first hurdle with the highest score, proving its technological prowess. Although the government originally planned to select four teams in the first round of evaluation, Naver Cloud and NC AI were eliminated. There are plans to select one additional team through a 'repechage' round later, but Naver Cloud has stated that it is not considering re-applying.
LG at the Forefront… Recognized for 'Practical AI' Ready for Field Deployment
The Ministry of Science and ICT (MSIT) and the Institute for Information & Communications Technology Planning & Evaluation (IITP) announced on the 15th that among the five consortia (Naver Cloud, LG AI Research, SK Telecom017670, Upstage, and NCSoft036570) in the 'Independent AI Foundation Model Project,' three companies—LG AI Research, Upstage, and SK Telecom—have advanced to the second stage. Leading candidates Naver Cloud and NCSoft (NC AI) were eliminated.
This first-stage evaluation was conducted with weighted scoring criteria: 40 points for benchmarks, 35 points for expert review, and 25 points for user assessment. The evaluation comprehensively assessed AI model performance (AI Frontier Index), potential for practical field application, cost efficiency including model size, and usability and ripple effects, including plans for impact on domestic and international AI ecosystems.
In particular, LG AI Research made significant strides, achieving the highest scores in all three evaluation categories. 'EXAONE,' developed by LG AI Research, has aimed for a 'domain-specific' model that can be immediately deployed in industrial settings rather than a general-purpose chatbot since its early development stages. In this evaluation as well, LG earned high scores for its 'specialized performance' based on expert data training in fields such as bio, chemistry, and manufacturing.
On the other hand, for Naver Cloud, the 'independence' evaluation was the deciding factor. Naver was eliminated for failing to meet the criteria of being free from external control or interference due to the use of open source. Previously, a controversy over independence arose when it was revealed that Naver used Alibaba's open-source AI model 'Qwen' during the development of its model encoder technology.
However, to prevent unnecessary controversy, the Ministry of Science and ICT did not disclose the specific scores received by each of the five consortia for each item. The Ministry stated, "Naver Cloud's elite team's AI model was judged not to have met the independence criteria," and added, "Many expert evaluation committee members raised issues regarding the limitations of its independence."
Ryu Je-myung, 2nd Vice Minister of Science and ICT, explained, "Using pre-existing weights is akin to 'free-riding' on someone else's learning experience. While using open source according to global licensing standards is a given, this project's purpose is to design and train models ourselves."
Although the use of open source is widespread in the global AI ecosystem, the view is that for AI foundation models, the basic condition for ensuring model independence is to initialize weights and carry out training and development, even if validated open-source software is used strategically.
Naver's Elimination: What Held Them Back?
The elimination of Naver Cloud, which has emphasized independent Korean AI capabilities, has dealt a significant shock to the industry. Naver built the country's first independent large-scale AI ecosystem through 'HyperCLOVA X.' However, in this evaluation, it is interpreted that they failed to highlight clear advantages over competitors in terms of the 'in-depth problem-solving ability in specific domains' and 'cost efficiency' requested by the government.
The survival strategies of SK Telecom and Upstage, which passed alongside LG, were also evaluated as clear and distinct.
SK Telecom received positive evaluations for its vast infrastructure unique to a telecommunications company and its B2C (Business-to-Consumer) data processing know-how accumulated through operating the 'A.' service. Its ability to show hybrid operations, utilizing not only its own models but also global models like Anthropic's through a multi-Large Language Model (LLM) strategy, acted as a realistic alternative.
AI startup Upstage also fared well in the foundation model race, which typically requires massive capital. Upstage employed a strategy to capture both lightweight design and high performance through its 'Solar' model. It proved that it could achieve efficiency comparable to big tech models despite having a lower parameter count, fitting the evaluation's objective of prioritizing cost-effectiveness.
The government plans to select a total of two elite teams by the end of the year. Furthermore, since two teams were eliminated contrary to expectations, the government intends to open up re-application opportunities to both the previously eliminated firms and new entrants. Through this, the plan is to form a total of four elite teams by the first half of this year. However, through its recent statement, Naver has effectively made it official that it will not participate in additional open calls.
Vice Minister Ryu stated, "Four and a half months after starting the project, all five elite teams have achieved the success of having their models listed as 'Notable AI Models' by the non-profit research institution Epoch AI," and added, "We will continue to provide opportunities for as many companies as possible to utilize GPUs."