Home > Opinion > Columns

On DeepSeek's shock waves: A conversation with Ha Jung-woo, Korea's leading AI expert

Published: 12 Feb. 2025, 20:01 Updated: 13 Feb. 2025, 13:51

The shock waves from China’s DeepSeek continue to spread. The Chinese AI startup, which has been in existence for less than two years, unveiled its generative AI model R1 on Jan. 20. Reports suggest that its performance surpasses that of OpenAI’s latest model, o1. Given the stringent U.S. technological restrictions on China amid a fierce tech rivalry, the news has sent shock waves through American society. On Feb. 3, The New Yorker published an article titled "Is DeepSeek China’s Sputnik Moment?", drawing a parallel between DeepSeek and the Soviet Union’s launch of the world’s first artificial satellite in 1957, which profoundly shocked the United States. The U.S. government and Congress have swiftly moved to block access to DeepSeek, and Korean authorities and businesses are following suit. But what exactly is DeepSeek, and what does its rise mean for Korea?

On Feb. 6, at the Federation of Korean Industries building in Yeouido, western Seoul, JoongAng Ilbo sat down with Ha Jung-woo, 48, the head of Naver’s Future AI Center and one of Korea’s foremost AI experts. Ha is also actively involved in policy discussions, serving as chair of the AI-Data subcommittee of the Presidential Committee on Digital Platform Government and as co-representative of the National Alliance for the Realization of a Just Science and Technology Society. He had just finished attending the 2025 Digital Policy Forum at the National Assembly, where he discussed strategies for fostering an innovation ecosystem to make Korea a top-tier AI powerhouse.

최준호 Choi Joon-ho

Choi Joon-ho
The author is the editorial writer and the science journalist of the JoongAng Ilbo.

Ha Jung-woo, the head of Naver’s Future AI Center and one of Korea’s foremost AI experts, poses for a picture during an interview with the JoongAng Ilbo in Yeouido, western Seoul, on Feb. 6. [JOONGANG ILBO]

Q. What is your evaluation of DeepSeek?
A. A model of this caliber is on par with OpenAI’s latest o1. It is incredibly advanced. While we need more time to assess its real-world industrial applications and serviceability, based on existing AI evaluation benchmarks, DeepSeek's R1 performs at least as well as o1.

Have you personally tested it?

I was probably the first person in Korea to try it. The speed is impressive, and the way it logically structures responses is remarkable. It is very well made. However, when I asked about sensitive topics like the Tiananmen Square incident or the history of Goguryeo [an ancient Korean kingdom], the responses were predictably aligned with the Chinese government’s stance. That said, we cannot fault China for developing an AI that reflects its national ideology and values — after all, the United States does the same with its AI models.

What stood out to you the most?

Its inference capability. For example, when asked, 'How many people might be inside this building?' DeepSeek doesn’t simply regurgitate a memorized answer. Instead, it estimates based on factors like the number of floors, area per floor and spatial utilization. Existing AI models struggle with this level of logical reasoning. The step-by-step analytical thinking was particularly impressive.

DeepSeek's R1 is reportedly the first open-source inference model. Why do you think they chose an open-source approach?

Personally, I believe this goes beyond corporate strategy — it’s a battle for national security and technological dominance. Open-source AI enables the integration of diverse datasets for application in new industries and services. Since AI is a foundational technology, its widespread adoption can lead to ecosystem dominance, creating a strong 'lock-in' effect.

Ha Jung-woo, the head of Naver’s Future AI Center and one of Korea’s foremost AI experts, speaks with the JoongAng Ilbo in Yeouido, western Seoul, on Feb. 6. [JOONGANG ILBO]

Right now, Meta’s large language model, LLaMA, leads the open-source AI space. However, DeepSeek, as the first open-source inference model, is superior to LLaMA in certain aspects. By making it open-source, DeepSeek has presented the United States with a serious challenge. The battle between U.S. and Chinese open-source AI ecosystems will likely determine the global AI power structure.

How did a startup that’s not even two years old achieve this?

DeepSeek’s parent company, High-Flyer, is a hedge fund specializing in quantitative investing using mathematical models and computer algorithms. DeepSeek spun off from its AI research lab, meaning the research was already underway before its official founding.

Moreover, since High-Flyer is a hedge fund, DeepSeek has substantial financial backing and has not sought external investments. DeepSeek reportedly possesses as many GPUs as all of Korea combined. They also pioneered several cutting-edge AI techniques.

There is also speculation that DeepSeek leveraged existing AI models like ChatGPT for training, despite their denial. However, given the overwhelming evidence circulating on social media, it’s likely they did, which would have significantly reduced development time and costs compared to starting from scratch.

Has China already established an independent AI ecosystem?

There are numerous highly capable AI firms in China. Since the AlphaGo breakthrough in the 2010s, the Chinese government has invested heavily in AI. Tech giants like Baidu and Alibaba have significantly advanced their AI capabilities, especially after the emergence of OpenAI’s ChatGPT. They’ve also received strong backing from the Chinese government.

Some see DeepSeek as both a crisis and an opportunity for Korea. What do you think?

I agree. Korea already has a model at the foundational level of DeepSeek's R1. Moreover, DeepSeek has disclosed about 80 percent of its methodologies through open-source channels. With a few additional steps, we could catch up. However, Korea’s biggest challenge is the lack of GPUs necessary for training models at DeepSeek’s level. Training AI requires a massive number of GPUs running simultaneously.

Can Korea produce a startup like DeepSeek?

We must create the conditions for it. The biggest hurdle is providing developers with the time and resources to focus on R&D. In Korea’s corporate ecosystem, could we afford to invest 100 billion won ($68 million) over two years without requiring immediate profitability? Another issue is how we treat AI talent. Many top researchers at OpenAI are Korean. Can Korean companies offer the same competitive compensation and working environment?

DeepSeek benefits from the financial strength of its hedge fund parent and likely government support. We need similar backing if we want to compete.

The United States has blocked access to DeepSeek, and Korea is following suit.

Unlike OpenAI, DeepSeek does not seek user consent before collecting data. It captures everything users type, including keystroke patterns, and stores it on servers in China. Under Chinese law, the government has unrestricted access to such data. DeepSeek users need to be aware of this.

What should Korea’s AI strategy be to become a global leader?

I believe Korea should focus on ‘Sovereign AI.’ If we want to become the world’s third-largest AI power, we must excel beyond our domestic sphere. First, we need AI that accurately represents Korean values and perspectives. Once we achieve that, we can expand to provide customized AI solutions to countries in the Middle East and Southeast Asia, which also seek AI models that reflect their own values. To do this, Korea must invest in digitizing data, nurturing talent and expanding AI infrastructure.

Translated using generative AI and edited by Korea JoongAng Daily staff.