[Reporter’s Notebook] One Year After the DeepSeek Shock: South Korea Scrambles for Originality

Artificial Intelligence (AI). Pixabay

One year ago, the Chinese artificial intelligence (AI) startup DeepSeek shocked the world by releasing an open-source inference model. DeepSeek became synonymous with cost-effective AI models by achieving high-performance AI at low cost and efficiency, without advanced hardware infrastructure.

South Korea quickly sprang into action. The government determined that securing domestic technological capabilities was necessary and launched its own AI foundation model project. With a determination to establish a "top three in AI" system in a short period, the government has been leading efforts such as supporting graphics processing units (GPUs).

However, whether the project would be developed "from scratch" became a controversy from the outset. The government's emphasis on "originality" led to a negative perception of open-source utilization spreading throughout the industry. Due to the lack of clear guidelines, confusion arose over what could be considered an original model. As a result, Naver Cloud failed to pass the first round of evaluations for not meeting the criteria, and major companies subsequently announced they would not participate in the re-bidding process.

In contrast, DeepSeek actively leveraged open source to maximize efficiency. By incorporating modules from Google and OpenAI in the model development process, DeepSeek accelerated its progress, later unveiling ultra-fast long-form training and inference technologies that surpassed existing mechanisms, and introducing innovations such as sparse attention technology to reduce inference costs.

This is why industry experts point out that fully utilizing South Korea’s outstanding processing capabilities would be more advantageous in terms of time and cost. There is also skepticism toward developing entirely original AI models. Some are concerned that focusing solely on domestic applications could make it difficult to secure global competitiveness. At the same time, there are calls to focus more on vertical (specialized) models rather than general-purpose ones.

There are certainly areas where data sovereignty is required. However, we are now in an era where companies are actively adopting ChatGPT, and individuals are using ChatGPT or Gemini and sharing their experiences. The two AI models from Google and OpenAI, which are locked in fierce development and update competition, have already become deeply embedded in our lives. The key for AI models is how widely they can be used. While originality is important, it is more crucial to skillfully utilize open source to improve performance and ensure control.