The second phase evaluation of the Independent AI Foundation Model (Dukpamo) project, led by the Ministry of Science and ICT, has been completed. All four teams participating in the evaluation have been recognized as 'notable models' by global AI research institutions, increasing interest in which teams will advance to the next stage.
According to industry sources on August 11, the results of the second phase evaluation are expected to be released as early as this week. The evaluation included LG AI Research Center, SK Telecom, Upstage, and Motif Technologies, which was selected through an additional application process, resulting in a four-way competition.
In the first phase evaluation, LG AI Research Center ranked first in both benchmark and expert/user assessments. Subsequently, LG AI Research Center unveiled 'K-Exa One 2.0,' SK Telecom introduced 'A.X K2,' and Upstage released 'Solar Open 2' last month on Hugging Face. Motif also launched 'Motif-3' earlier this month.
According to the Ministry of Science and ICT, all four teams' models developed in the first half of this year have been listed as notable models by Epoch AI, a U.S. nonprofit AI research organization. Epoch AI selects models that have significantly impacted AI development based on factors such as computational load and performance.
Last year, the U.S. had the highest number of notable models at 59, followed by China with 35 and South Korea with 8. This year, the countries producing notable models include the U.S., China, and South Korea.
In the past month, download statistics from Hugging Face show that Upstage's 'Solar Open 2' recorded 25,870 downloads. Motif's beta version, 'Motif-3 Beta,' had 5,062 downloads, while SK Telecom's 'A.X K2' had 2,844 downloads, and LG AI Research Center's 'K-Exa One 2.0' had 1,462 downloads.
There is keen interest in how evaluations from global research institutions and responses from the open-source market will influence the results of the second phase evaluation. Attention is focused on whether LG AI Research Center, which ranked first in all categories in the first phase, will maintain its lead, whether SK Telecom and Upstage, which have released new models, will catch up, or whether Motif, which joined through the additional application process, will advance to the next stage.
Notably, for the first time, 200 citizens will directly use the models to assess their performance. If the benchmark scores among the models are close, the perceived performance experienced by general users in various situations could become a decisive factor in the evaluation results.
Bon Kang-ho, a researcher at the Software Policy Research Institute, stated, "Just because a model scores high in benchmarks does not necessarily mean it offers good user benefits or convenience. If around 200 users apply the models to various cases that suit them and reflect the benefits they perceive in the process, it can yield valid results."
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.