Government reorganizes independent AI foundation model business... Frontier AI development parallel two-track strategy announced

Regarding the rumor of discontinuation of the independent AI foundation model (Dokpamo) development project and the direction of Frontier AI development, the government stated its position that "Dokpamo is not being discontinued but is evolving in full."

The Ministry of Science and ICT held a press conference at the Seoul office of the Korea Intelligence Agency (NIA) on the 2nd and announced the direction of the Dokpamo business reorganization.

Ryu Je-myeong, Second Vice Minister of Science and ICT, announced at the meeting that the Dokpamo project will continue to be promoted, but a separate frontier AI model development project will also be carried out starting next year.

He explained that DokpaMo will evolve in a two-track manner to expand the ecosystem so that it can be used more widely in the industrial field, while Frontier AI will evolve on a two-track basis to break through technical limitations and secure top-level performance.

Vice Minister Ryu said that the performance of the global AI model has developed more rapidly than expected since the release of the 'Mythos Preview' in April, and the government collected opinions from industry and experts in May and June.

As a result, a consensus was formed that it would be difficult to catch up with the frontier model with the existing Dokpamo level of support, and to reflect this, it was announced that an investment project worth 4.7 trillion won for the development of the frontier AI model was organized in next year's government budget.

The following is a Q&A with Ryu Je-myeong, Second Vice Minister of Science and ICT, and Choi Dong-won, Director of AI Infrastructure Policy.

◇ Dokpamo, redesigned as a two-track for ‘frontier development and industrial expansion’

-- If the Dokpamo and Frontier development policies are implemented in parallel, please explain in an understandable way what is fundamentally different between the two and what will change in the future, and why separating the two policies is an efficient strategy. ▲ (Vice Minister Ryu Je-myeong) As a result of directly asking companies and experts, the common opinion was that new investment that is different from the current monopoly is needed. Companies also agreed that it is not enough to just give more GPUs while maintaining the current method of dividing resources among companies to compete, and that they must gather and focus all resources to catch up with the gap with the frontier model. Therefore, the existing Dokpamo business follows two tracks: one in which model development evolves and strengthens towards frontier model development, and the other in which the ecosystem expands to be more widely used in the industrial field.

-- Some of the companies participating in DokpaMo were eliminated through the contest, and the remaining companies invested resources with the expectation that they would receive follow-up support if they reached the final stage. What is your stance on these expectations and complaints?

▲ (Vice Minister Ryu Je-myeong) We were aware from the beginning that the primary election method would have a significant stigma effect on companies that were eliminated, and we also knew that it would cause social controversy. However, I believe it is true that through fierce competition, much greater results have been achieved than the typical appearance business method. In addition, last year's public offering guide already clearly stated that support and scale may be adjusted or discontinued depending on policy and budget situations.

-- If the monopoly continues, will the current method of one out of three companies failing and two surviving companies be maintained as is? If not, will it actually come to a halt?

▲ (Vice Minister Ryu Je-myeong) We will conduct a third evaluation to determine rankings and select two companies. However, the method and scale of support that the two companies will receive next year may change. The same was true in stages 1 and 2, when the support method was internally adjusted while the goal did not change. Similar to the complete redesign of the software-centered university business into an AI-centered university project, the achievements made so far are not lost, but serve as a stepping stone to the next stage.

-- Please confirm whether it is correct to provide incentives to the remaining two teams during the 4th evaluation of Dokpamo and the Frontier new business contest. I am also curious about the extent of the new investment, whether it is the establishment of an SPC, and whether there are plans to receive overseas investment as well.

▲ (Vice Minister Ryu Je-myeong) I am cautious about saying anything because many things have not been confirmed. Several choices are possible for the two companies remaining after the third evaluation. You can just let people participate in a new game without prioritization or incentives, or on the other hand, there is also a way to give consideration to existing expected benefits and trust. However, we recognize that there is a need to consider the expected profits and trust of existing participating companies, but we do not think it is desirable to choose a method that distorts the overall resource allocation due to such considerations.

-- Do you intend to receive financial investment from policy banks or other companies in the Frontier AI business?

▲ (Vice Minister Ryu Je-myeong) The criteria for public offering and selection, such as which entity will be the anchor company and how many companies will join forces at an equal level, have not yet been confirmed. I know there is a lot of discussion about how to structure a strategic partnership. Although it is unusual for the government to take action before the budget is finalized and approved by the National Assembly, as time is urgent, they are rushing to establish public offering standards as quickly as possible to ensure that the market can be predicted. -- An explanation to the effect that “there is no mistake since it was announced in the notice that it could be changed” can be read as meaning that the government cannot guarantee policy reliability. To summarize, next year, Dokpamo will be absorbed and reorganized into the frontier business, and Dokpamo can be understood as going into two tracks: the frontier track and the industrial expansion track.

▲ (Vice Minister Ryu Je-myeong) If it was seen as having that intention, I would like to withdraw that expression. This means that even if there is a change in policy, there must be a reasonable reason, and this does not mean that “there is no mistake because it was announced.” We will do our best to make no changes so as not to cause problems with government consistency or trust, but we ask for your understanding that there are limits as the changes in the AI ​​field are so large that they exceed expectations. And rather than 'absorption', please understand that Dokpamo's existing model development business will evolve and strengthen through frontier model development, and the existing Dokpamo business will evolve in a two-track manner that will expand towards greater use in the ecosystem.

◇ Two-track conversion starting next year… Selected as a frontier business in February-March

-- When is the exact end of the Dokpamo project? I am curious as to whether the project will end with the selection of two companies in December of this year, whether it will continue next year, and if it continues, whether a new frontier operator will be selected separately.

▲ (Vice Minister Ryu Je-myeong) In the original public offering plan, it was planned to select two out of three companies through a third evaluation at the end of this year and support these two companies to achieve 95% of the frontier level by the end of next year. However, it was announced that this plan could be discontinued or the scale of support could be adjusted at any time. Now, instead of two companies receiving support for next year as originally planned, the business is being redesigned into two tracks: a frontier model development track and an industrial field expansion track. Of the two companies selected in the third evaluation, one will be able to participate in the frontier model development track, and the other company will receive the existing level of support and participate in the expansion track. For these two companies, we are considering how to provide incentives as the existing support level changes.

▲ (Policy Director Choi Dong-won) The goal we pursue is to secure an excellent AI model. To achieve that goal, we are currently using various methods, and we are continuing to consult with Dokpamo companies in the process of creating the 3rd evaluation standard.

-- You said that you are aiming for a model that can compete with Mythos. Will security-specific AI also be integrated into Frontier AI in the future? In addition, while overseas competing models are known to use more than 100,000 GPUs, the government's holdings are about 39,000 even with the addition of 10,000 Verarubin units. Isn't the goal set too high?

▲ (Vice Minister Ryu Je-myeong) The security-specialized model was promoted first because the situation was urgent. The key is that Dokpamo must evolve to a level that industrial sites can choose from. Even if it is not necessarily a 100,000-piece model, there are cases where small-sized models show higher benchmark performance in certain areas. I think it is unrealistic to aim for a level equal to that of models with powerful capital in the United States. However, I believe that if private investment is added to government support, a goal comparable to China's top open model can be presented.

-- I am curious when the third evaluation and final two site selection will be confirmed, when the specific plan for the two-track direction will be confirmed, and whether the two-track process will begin in 2027.

▲ (Vice Minister Ryu Je-myeong) The understanding that the two-track process will begin next year is correct. Considering last year's experience with the AI ​​National Data Center Special Purpose Corporation (SPC), the Frontier Model project may require a lot of time for financial screening. Therefore, it is difficult to hold a contest before the budget approval of the National Assembly in December of this year, and our goal is to prepare to conduct a contest to select the project execution entity as soon as the budget is approved. The 3rd evaluation of Dokpamo will begin in September and run until the end of January next year, and the evaluation will be completed around mid-February. In connection with the results of this evaluation, we plan to finalize the selection of next year's frontier model investment target within February at the earliest or March at the latest.