The Ministry of Science and ICT is advancing its 'Everyone's AI' initiative, aiming to provide every citizen with a personal AI agent by the end of the year.
On July 21, the Ministry and the National IT Industry Promotion Agency (NIPA) held a briefing on the 'Everyone's AI' project at the Front One building in Mapo, Seoul. Over 100 representatives from companies including LG Uplus, Saltlux, Rebellion, Eastsoft, and AXZ attended the event. Selected companies will develop general-purpose AI chatbots and public AI agents using 512 B200 GPUs supported by the government.
Below is a Q&A from the 'Everyone's AI' briefing.
- Will the beta service be available to all citizens from the start?
The beta service is set to launch by the end of December, targeting all citizens. However, during the beta phase in September, companies will need to design their operational plans according to the service characteristics. Companies ready to launch can operate from the start for all citizens, while others may gradually expand their user base. There are no restrictions on including public institutions as demand sources or consortium participants, with the main goal being to provide a nationwide B2C service by year-end.
- Will domestic AI semiconductors (NPU) be considered in evaluations, and can multiple domestic NPU companies participate in consortia?
While the announcement specifies GPU support, domestic AI semiconductors may positively influence qualitative evaluations for ecosystem activation. Detailed criteria will be provided after internal review. Currently, participation in one consortium is the principle, but simultaneous participation in different consortia by the same institution will require further internal review. Additional guidance will be provided considering the possibility of multiple consortia being selected with the same company.
- Will the government support the integration of public services like Government 24 and Public MyData?
The government is coordinating with relevant departments to facilitate the implementation of public AI agents, including cooperation with Government 24 and Public MyData. If companies encounter issues requiring public institution collaboration during the project, the government plans to assist in discussions with relevant departments and agencies.
- Can companies participate by only applying AI models to existing services?
Simply replacing AI models or adding features to existing services does not align with the project's intent and is unlikely to score well in evaluations. However, integrating with existing platforms is permitted. The principle is to establish a separate 'Everyone's AI' service, and companies can autonomously propose service formats such as apps or websites. The key is to implement new services that are easily accessible to all citizens.
- How will the GPU support be provided?
The government will provide the B200 GPUs through cloud service providers (CSPs). Support will be based on a Kubernetes environment, and GPUs must be used within the project's scope. However, using idle GPUs for service development and enhancement will be allowed.
- How will the allocation of 128 and 256 GPUs be evaluated?
If two companies are selected, each will receive 256 GPUs. If three companies are selected, the allocation will vary based on evaluation rankings, with 256 and 128 GPUs assigned accordingly. Companies applying for 256 GPUs will compete among themselves, while those applying for 128 GPUs will compete separately. If the GPU allocation changes after final selection, the self-funding amount may also be adjusted.
- Is it possible to participate without a proprietary AI model?
Having a proprietary AI model is not a mandatory requirement. The stipulation of utilizing over 50% domestic AI models refers to meeting the criteria for independent AI foundation models, not necessarily proprietary models. Companies without their own models can use domestic open-source models or models from other domestic companies. Currently, open-source models from participating companies in the independent AI foundation model are also available, and securing usage rights through collaboration with other domestic companies is permitted.
- How will the 50% domestic AI model requirement be evaluated?
During the evaluation phase, the focus will be on whether the service design and model routing methods are reasonable. In the operational phase, the ratio of domestic model usage will be regularly checked using output tokens and system logs. The method for calculating the usage ratio is currently being finalized.
- To what extent can foreign AI models be used?
The use of foreign AI models is permitted only for minimal functions that cannot be implemented with domestic models. The portion of foreign model usage will not be eligible for government support, and excessive use beyond the necessary minimum may be reflected in evaluations.
- If 2-3 companies are selected, will all services be launched?
The evaluation committee will comprehensively review the business plans and select either two or three companies. The final selected companies will aim to provide services with a target launch in December.
- How will performance be evaluated?
The focus will be on user-perceived performance rather than benchmark figures. An annual evaluation will consider user numbers and service activation levels, and if performance is deemed insufficient, the possibility of changing support targets will be reviewed.
- What will be the support for operating costs and revenue models next year?
Starting next year, discussions are ongoing with the finance authorities regarding support for not only GPUs but also service operation and enhancement costs. Government support will focus on essential functions such as general-purpose AI chatbots and public AI agents. Companies will have the discretion to decide whether specialized services are free or paid, and various business models to attract users will be permitted.
* This article has been translated by AI.
Copyright ⓒ Aju Press All rights reserved.