NVIDIA has chosen the Korean language as the starting point for developing global artificial intelligence (AI) learning technologies. The company plans to generate and learn from Korean language data before expanding its efforts to other countries and languages. This move is seen as a strategy to leverage Korea's technological competitiveness to enhance its influence in the global AI ecosystem. However, NVIDIA has provided only a general explanation regarding the specific business benefits it hopes to gain from its Korean language research.
Kim Hyun-woo, NVIDIA's research manager and head of the NVIDIA-KAIST joint AI research institute, explained during a press conference on November 10 at NVIDIA Korea's office in Gangnam, Seoul, that the establishment of the joint research institute was motivated by KAIST's growing global reputation.
Kim noted, "KAIST's outstanding research achievements and the unique position of the Korean language as a mid-resource language have created a golden opportunity, especially given the high interest in AI among Koreans."
The value of the Korean language, according to NVIDIA, lies in its potential as a research foundation that can be expanded into the global AI market. While Korean does not have as rich a learning resource as languages like English or Chinese, it is classified as a mid-resource language, meaning it has more data than minority languages. This makes it a favorable condition for developing and validating new data learning technologies.
Kim stated, "Our goal is to develop data learning methods that can be applied to other languages starting from Korean. We aim to research how to create a forest rather than just a single good tree and share that knowledge with the world."
When asked about the specific business benefits of Korean language research for NVIDIA, Kim did not present a concrete revenue model. He responded, "As a company, we dream of an open AI ecosystem. Our goal is to research methodologies that can be used globally, starting with Korean, so that each country can develop its own AI in its language."
The joint research institute, announced in July, is NVIDIA's first collaboration with a university in Asia. Over the next five years, the two parties plan to invest a total of $300 million (approximately 400 billion won) in the partnership, with $250 million allocated for computing resources. The specifics of how the remaining funds will be utilized are still under discussion.
Kim emphasized the significance of a major AI tech company establishing a research institute and hiring researchers in Korea, stating, "This demonstrates the exceptional capabilities of Korean researchers and we hope to encourage more global companies to set up research institutes in the country."
He also opened the possibility of expanding collaborations with other universities, saying, "Our goal is to excel so that we can open up opportunities. It doesn't have to be limited to NVIDIA; other companies may also come in after seeing Korea's capabilities."
He clarified that the collaboration with Seoul National University is different in nature. The AI Technology Center (NVAITC) being developed with Seoul National University focuses on joint research and education, while the KAIST joint research institute will directly hire NVIDIA full-time researchers and provide computing resources.
The operation of the research institute is still in its early stages. Kim mentioned, "We have started the hiring process for researchers and plan to hire several by the end of the year." Discussions are ongoing with domestic NVIDIA cloud partners regarding securing GPU resources, although the specific number of devices has not yet been finalized.
In response to questions about whether the research institute's operations would continue despite a change in leadership, Kim stated, "I don't think I'm the right person to comment on that, but my goal as the head of the institute is to fulfill my responsibilities without any issues."
The core task of the joint research institute is to address the shortage of Korean language learning data. According to presentation materials, there are approximately 24 million Korean documents in the multilingual learning dataset mC4, which is only 1/163 the number of English documents, totaling 3.93 billion.
Kim pointed out, "It is really difficult to find high-quality public datasets in Korean. This shortage spans all fields, not just specific areas." He added, "It is challenging to close the gap using only human-generated data, so we will research technologies for generating synthetic data and efficiently learning from existing data."
Progress is already being made. Kim previously announced that the Korean synthetic dataset 'NemoTron Persona Korea' ranked first among popular datasets on Hugging Face for ten days, with approximately 140,000 downloads.
There are also examples of domestic companies utilizing this data. Kim noted, "The dataset was used in 'K-Exa One 2.5' released by LG AI Research the day before, showcasing a virtuous cycle of creating better models by leveraging each other's achievements."
Copyright issues related to the use of synthetic data are also identified as a research topic. Kim explained that it is unclear how existing AI models obtained their original data. The research institute plans to focus on generating new data using the knowledge from already trained models.
The research outputs will be made publicly available to prevent any single company from monopolizing them. Kim stated, "The NemoTron Persona dataset has also been made available for commercial use, and we will develop data and learning methods that other companies can utilize."
The first research results are expected to be released in the second half of next year. Kim expressed optimism, saying, "We have interesting projects in the pipeline that could be unveiled around that time, and we aim for research initiated in Korea to contribute to the global AI ecosystem."
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.