*Editor’s note: K-VIBE invites experts from various K-culture sectors to share their extraordinary discovery about the Korean culture.
Matthew Lim's AI Innovation Story: Necessity of a Korean-Language LLM
By Matthew Lim, AI expert and director of the Korean Association of AI Management (Former Head of Digital Strategy Research at Shinhan DS)
In May, a WikiDocs member known as "Jeonnae Hacker" released a ranking of Large Language Models (LLM) based on their usability in Korean. "I created my own problems to test various LLMs, evaluating how user-friendly they were in Korean, and converted the results into a score out of 100," he explained. However, he acknowledged limitations, saying, "The benchmark design is simple, with some calculation flaws. I didn't conduct multiple tests per model on the same questions to derive an average, and the number of attempts, timing, and frequency varied for each LLM, which may have led to inconsistent scoring."
In the results, Claude 3 Opus claimed first place, while the widely recognized ChatGPT-4 was ranked third. Though the experiment has clear limitations as it represents just one power user's findings, it draws attention to the need for deeper consideration of LLM usability in Korean.
Relying on multinational LLMs that "support" Korean is not sufficient. Korean's complex system of honorifics, subtle nuances, and the unique cultural context formed over millennia cannot be addressed through simple "support." Expecting a foreign AI to fully grasp the cultural essence embedded in a phrase like, "Have you had your meal, sir?" is overly optimistic.
This challenge becomes even more serious when discussing Korea's history and social issues. Terms like "comfort women," "East Sea," "Dokdo," and "Sewol Ferry" represent more than just events or concepts; they encapsulate the pain, conflict, and healing process of Korean society. Entrusting foreign AI with these complex contexts risks misinterpretation or diminishment of our historical and social issues—an unacceptable outcome.
The need for a Korean-specific LLM goes beyond cultural matters and extends to economic impacts. According to a report from the Korea Foundation, the economic ripple effect of Hallyu (Korean Wave) amounted to around 65 trillion won in 2021. A Korean-language LLM could further accelerate the production, distribution, and global spread of Hallyu content. Additionally, considering the growth of AI-based service industries, developing a Korean LLM could become a new driver of economic growth.
Of course, developing a Korean LLM requires significant investment in time and resources. However, this should be viewed as an investment rather than a mere cost. As AI technology becomes a core component of national competitiveness, securing an independent LLM model is essential, even from a national security standpoint.
How do things stand for us today? A closer look at the current state of LLM development by domestic companies is not just disappointing—it is infuriating. The behavior of major corporations, in particular, is shocking. Take Naver's "HyperClovaX," for example. Despite being the only entity able to fend off Google's dominance in global search engines, Naver's LLM performance is abysmal. It frequently fails to grasp the context of conversations and often gives irrelevant answers to even simple questions. This situation not only disappoints but exposes the stark reality of the current state of AI technology in Korea.
I still remember the awe I felt in March last year when OpenAI announced GPT-4. I expressed my hope that we, too, could develop such a model and shared my high expectations for Naver on social media. Given that training an LLM requires extensive data, I believed Naver had an advantage over Kakao, which has data rich in youthful expressions, due to its vast trove of blog, cafe, and KnowledgeiN data. The disappointment I felt on the day of the HyperClovaX launch seemed surmountable at the time. However, seeing little to no improvement nearly a year later leaves me feeling not only disappointed but hollow.
What is even more infuriating is Naver's attitude despite its poor performance and quality. Instead of focusing on performance improvements, they have been heavily invested in exaggerated marketing and promotion. It feels like someone posing in front of a mirror without properly dressing. This behavior is closer to manipulative acts aimed at managing stock prices and corporate image rather than genuine efforts toward technological progress.
In contrast, the startup Upstage’s "SOLAR" model has shown relatively strong performance. Last December, it dominated the Hugging Face’s open-source LLM ranking board, occupying seven out of the top ten spots, including sweeping the top six positions. I still vividly remember the sense of excitement and thrill from that time.
This stark difference between major corporations and emerging startups highlights the urgency of developing a domestic LLM that not only competes globally but also addresses Korea's unique linguistic, cultural, and economic needs. Such a model would play a crucial role in safeguarding the accurate narrative of Korea’s history while positioning the country as a leader in the AI-driven future.
However, due to the nature of their scale, it is inevitable that domestic companies face limitations when competing with global giants. This is another tragic aspect of Korea’s AI ecosystem. While capable startups strive against all odds, they find themselves marginalized without adequate support, forced to focus on projects that generate immediate revenue rather than on further research and development.
This reality starkly underscores the urgency of developing a Korean-language Large Language Model (LLM). Korea can no longer afford to remain idle. The state of LLM development among domestic companies has gone beyond being merely embarrassing—it now poses a serious threat to national competitiveness. The notion that Korea is lagging in the global AI race is no longer a mere concern but a harsh reality.
Thus, it is time for the government, major corporations, startups, and academia to come together and take immediate action. The government should lay the groundwork for LLM development through large-scale investments and regulatory reforms. Major corporations must leverage their financial resources and infrastructure to engage in long-term research and development. Additionally, there should be bold investment and support for competent startups, allowing companies like Upstage to compete on the global stage.
Furthermore, we need to strengthen the collaboration framework between academia, industry, and research institutions. It is essential to ensure that talented individuals from universities and research centers are directly involved in LLM development. By fostering open-source communities, we can create an ecosystem where diverse ideas and technologies are shared and advanced. When such a cooperative system is in place, Korea will finally be able to possess a world-class LLM.
A Korean LLM is not simply about developing an AI that speaks Korean fluently. It is about creating an AI that can fully embody Korea’s language, culture, history, and values. It presents an opportunity to preserve our cultural sovereignty in the digital era while further promoting Korea's soft power on the global stage.
Of course, there are many factors to consider carefully during the development of a Korean LLM. Issues like resource concentration, the risk of data bias, concerns about technological dependence on certain companies, and privacy challenges are all valid. However, these potential risks should not deter Korea from pursuing the development of a Korean LLM. Rather, they underscore the need for a more cautious and responsible approach. If the government, corporations, academia, and civil society work together to address these challenges, we can develop a safe and effective Korean LLM.
The development of a Korean LLM is not merely a technological challenge. It is a crucial opportunity to preserve our cultural identity and assert our voice in the global AI era. Acknowledging the risks and overcoming them wisely, while creating an AI that embodies our unique language and culture—this is the path we must pursue. Developing a Korean LLM is no longer an option but a necessity. An AI that understands our language and sentiments is the future we must shape. The time for action is now. Let’s shape our future with our own hands.
※ External contributor articles may not necessarily reflect the editorial stance of this publication.
(C) Yonhap News Agency. All Rights Reserved
