Originally published in the Herald Insight Collection, this article was written in both Korean and English. Republished here is the English translation.

Korea’s AI strategy aims to make it one of the world's top three AI powerhouses. To achieve this vision, the government prioritizes four policy directions: 

1. Innovating technology and infrastructure 

2. Fostering startups and talent

3. Creating a culture of inclusiveness and fairness 

4. Securing global leadership in AI governance 

To achieve these policy goals, the Korean government has emphasized the need for ‘sovereign AI,’ positioning it both as an economic competitor and a national security priority. Economic competition refers to becoming a global alternative to the American and Chinese AI foundational models, while the security element refers to the ability to protect the outflow of (especially national security-related) data outside of Korea.

The Korean government has emphasized the need for ‘sovereign AI,’ positioning it both as an economic competitor and a national security priority.

A few months ago, the government announced the allocation of significant funds towards a sovereign AI foundation model project. Over KRW 200 billion (USD 73.5 billion) of public funds, to be exact, will go to the eventual winner. The funding will include data acquisition and processing, Graphics Processing Unit (GPU) rentals, and labor costs.

While the selection process is still ongoing, two companies (Naver Cloud and NC AI) have been eliminated from the race. This development gives us an opportunity to debate the outcomes of this project. Now, it is important to step back and ask fundamental questions: What is necessary for this Korean alternative to be successful globally? Is such a huge investment from public funds necessary?

American commercial models are currently leading in performance. However, China’s push to release high-performing open models with rapid updates has put pressure on other countries. Recent reports from Hugging Face, analyzing Open-Source model downloads, show significant interest in these models, and a corresponding increase in the downloads of Chinese models, surging to 17.1% in the recent year, “surpassing the collective of American model developers for the first time.” 

In response, the American AI Action Plan strongly emphasizes the development and deployment of an Open-Source, open-weights model that could be adopted globally. The U.S. Administration is keen to counter the global diffusion of the Chinese models. Unfortunately, the ROK is not listed in the Hugging Face analysis among the top ten countries developing high-performing Open-Source models.

AI technology is rapidly evolving, and the adoption modes are shifting. Open-Source foundation models are increasingly performing better. Expert analysis shows that the capability gap between the best open and proprietary LLMs is around one year and is expected to narrow further. Why Open Source, you might ask? Or what would it take? Open Source benefits society at large, allowing researchers and small businesses to build without incurring significant subscription costs or dependencies. 

However, just like the term ‘sovereign’, 'Open Source’ also needs a clear definition to avoid misunderstanding and ensure accountability. 

Open Source Initiative (OSI) requires that for an AI system to be “Open Source”, it should be “made available under terms and in a way that grant the freedoms to: 

• Use the system for any purpose without having to ask for permission
• Study how the system works and inspect its components
• Modify the system for any purpose, including to change its output
• Share the system for others to use with or without modifications, for any purpose.” 

Whether we are talking about a fully functioning system, or the model(s), parameters (including weights), or any other structural components, these conditions must be met.

Korea’s publicly funded sovereign AI models will also become Open-Source. This is critical to advancing accountability and fostering the adoption and diffusion of AI technology across the economy. However, to compete with current alternatives, Korean models need to be state-of-the-art (SOTA) and continuously updated to remain so. This means Korea needs to commit not only to developing alternative models but also to further investment to keep the models at SOTA levels.

Korea needs to commit not only to developing alternative models but also invest further to keep the models at state-of-the-art levels.

Open-Source AI systems can be scrutinized for their data and training processes; the code and model parameters are transparent. Open-Source systems reduce dependence on vendors that can unilaterally decide to retire a product or change its features. Users can avoid situations where their established workflows are disrupted by vendor decisions, as was the case with the discontinuation of GPT-4o without proper notice.

Researchers can audit models for biases, understand failure modes, and verify safety claims. Using Open-Source software may require additional compute resources. However, such access and control over the models make Open-Source systems very attractive for individuals and companies.

Developers can also fine-tune open models for specific domains or languages, experiment with architecture, and build specialized applications without API limitations. This means that if the underlying model performs solidly, further enhancements can improve the overall quality, for example, for Korean.

Additionally, these models can be localized and secured as necessary. Since Open-Source systems can be deployed and secured locally, without any data flowing to the model’s original developer or the country of origin, this largely undermines the argument for ‘national security.’ Some argue that national security concerns are not simply data flows but also ensuring domestic engineering talent who can compete at the SOTA level of AI development and governance. 

Korea should definitely have an Open-Source Korean-language model with better context on Korean culture, heritage, laws, and business expectations. However, is an investment of 73.5 billion USD necessary for an LLM? Or are there other ways of achieving Korea’s strategic goals?

Is an investment of 73.5 billion USD necessary for an LLM? Or are there other ways of achieving Korea’s strategic goals?

There are several ways for Korean funds to contribute to the national economy and the global advancement of AI. One alternative could be for Korea to develop SOTA small language models (SMLs) trained on the Korean language, cultural, and legal context. High-performance, Open-Source, small models can be adopted more easily by domestic businesses and developers, with lower barriers to entry. Such models could also be used to fine-tune other globally available models for the Korean context.

The second alternative could be for Korea to invest in developing new methods to enhance both compute and energy efficiency. Global majority countries particularly have significant shortages in compute capacity and energy resource availability. Improving these efficiencies is both a domestic and global contribution.

A third alternative may be to invest in developing world-leading techniques in AI testing, evaluation, validation, and trustworthy AI governance at large. Although academic and government capabilities are improving, much of the testing and evaluation work is still driven by leading AI companies with vested interests. Raising the bar for these techniques helps build trust in AI technologies, which can drive greater adoption and investment.

These suggestions can occur concurrently and reflect deeper needs regarding the current state of AI globally. Korean stakeholders rightly suggest that American and Chinese models are biased in different ways and do not reflect Korean cultural, linguistic, or legal nuances. 

However, the blade cuts both ways. Such bias is inevitable for any user or developer when engaging with a foreign model. A Korean alternative may also be criticized for the same shortcomings on the global stage. While a Korean language model would significantly help the domestic market and possibly lower the digital trade deficit, it may not be ideal for foreign users. 

However, a Korean SML to augment Open-Source foundational models would be cheaper to develop and continuously improve. Such a model would improve the overall quality of outcomes generated for Korean purposes. Furthermore, if Korea’s AI investments were geared towards compute and energy efficiency, it would be a significant contributor to planetary well-being.

Investing in these outcomes would support Korea’s goals of creating a culture of inclusiveness and fairness and securing global leadership in AI governance. At the same time, leading in these domains would mean stronger domestic talent and better security of national and economic assets against foreign intrusions.

These alternatives may sound too high-level or intangible at first. Some may argue that such goals could be a distraction to Korea’s ‘sovereignty’ goals. However, the case is exactly the opposite. Many businesses, large and small, are conscious of the risks associated with LLMs (such as hallucinations, safety, privacy, and security issues). 

If Korea can improve testing and mitigation techniques to address these issues, more Korean companies would be able to adopt AI domestically, and more companies would use its methods internationally. 

Furthermore, inefficiencies in AI - such as inference costs, energy bottlenecks, and GPU shortages - prevent many businesses from adopting large models or investing in AI infrastructure. Korea’s investments in developing efficiencies would be a significant enabler for deeper adoption.

In an interdependent world of AI, the goal should not be Korean AI dominance but rather ensuring that Korea possesses sufficient capabilities to avoid complete dependence on any single foreign provider while contributing to the global Open-Source AI movement. 

In an interdependent world of AI, the goal should not be Korean AI dominance but rather ensuring that Korea possesses sufficient capabilities to avoid complete dependence on any single foreign provider while contributing to the global Open-Source AI movement. 

The Korean advantage could lie not in attempting to replicate the model of American companies with proprietary models, or Chinese models with censorship, but in building governance institutions that channel AI development toward broadly shared prosperity rather than concentrated private gain. If significant public funds are to be used, the expectation should be to raise the bar for everyone.