Logo of the Chinese AI corporation DeepSeek./Courtesy of DeepSeek

The shockwaves from China's artificial intelligence (AI) startup DeepCheek's generative AI 'DeepCheek R1' continue to ripple. With comparable performance to global big tech companies' generative AI like ChatGPT and LLaMa at a low expense, calls have even arisen questioning the necessity of high-performance graphics processing units (GPUs). Following the release of DeepCheek R1, NVIDIA's stock price plummeted on the New York Stock Exchange, and in the domestic market, which opened on the 31st after the Lunar New Year holidays, SK hynix's stock has dropped nearly 10%.

DeepCheek's strength lies not in its remarkable performance but in its low expense. By delivering similar performance to generative AI developed by OpenAI or Meta at a lower expense, it has fundamentally changed the dynamics of generative AI development. Yanbo Wang, a professor at the University of Hong Kong, noted in an interview with Nature that "DeepCheek presented a blueprint showing that it is possible to develop large language models without the sizable capital and hardware typically invested in Silicon Valley," adding that "this has opened the door for numerous new large language models to emerge."

DeepCheek revealed that it utilized 2,048 GPUs on its training model 'V1' developed prior to the launch of DeepCheek R1. The total purchase expense for the GPUs was only about 8 billion won. In contrast, the training expense for OpenAI's GPT-4 is known to be around 144.7 billion won, making DeepCheek's expense only 1/18th of that. DeepCheek V3 used more than 2,000 NVIDIA H800 chips, significantly lowering expenses compared to Meta's Llama 3.1 405B, which used around 16,000 H100 chips released last July.

The H800 chip has half the performance of the latest H100 chip, which also means the price is lower. While the IT industry is racing to develop high-performance GPUs to enhance AI capabilities, DeepCheek has shown that it can achieve similar performance with affordable and lower-performance chips.

Because it was developed at a low expense, its expansion potential is also significant. DeepCheek has also released for free 'DeepCheek Coder,' a generative AI that creates programming code. DeepCheek's free coder reportedly performs better than GPT-4.

DeepCheek's ability to deliver outstanding performance at a low expense has been attributed to reinforcement learning and cold-start data training. The DeepCheek team introduced the development process of DeepCheek R1 on the preprint site arXiv on the 22nd and stated that "we trained the model purely through reinforcement learning without any initial supervised learning, achieving strong reasoning abilities."

Existing large language model (LLM) approaches utilized supervised learning prior to reinforcement learning to enhance efficiency. In supervised learning, developers provide correct answers, and the model learns reasoning based on those answers. In contrast, DeepCheek used reinforcement learning alone, searching for and learning data independently rather than relying on curated data. Without supervised learning, the likelihood of making errors during reinforcement learning increases. The DeepCheek team addressed this drawback by rewarding correct reasoning during reinforcement learning.

Professor Gong Deok-jo of Gwangju Institute of Science and Technology (GIST) emphasized that "how effectively the existing technologies have been optimized will be key to the DeepCheek algorithm," and added, "If the research team’s claims are accurate, it seems DeepCheek has successfully implemented challenging optimization problems."

However, experts are uncertain whether the model is actually learning data as efficiently and achieving outstanding performance as the research team claims. In reality, many researchers express skepticism that DeepCheek achieves the performance stated in the paper or even surpasses GPT.

Professor Gong noted, "Even if the research team's claims are true, it is highly likely that they are receiving significant support from a considerable-scale data center behind the scenes," and added, "Given that they reference previously developed LLMs, it would be difficult to conclude that they completed it at that level of expense from scratch."

There are analyses suggesting that the foundation of DeepCheek's success lies in the Chinese government's proactive investments in AI talent. In 2017, the Chinese government announced its goal to become a frontrunner in the AI sector by 2030. To achieve this, it established AI-specialized departments in over 440 universities and concentrated various support measures, such as scholarships, research funds, and industry-academia collaboration, in the AI field. DeepCheek's founder, Liang Wenpeng, is known to have appointed many domestic talents.

There are voices urging Korea to pay attention to China's AI talent policies that fostered DeepCheek. Professor Moon Hyung-nam of Sookmyung Women’s University (President of Korea AI Education Association) stated, "The emergence of DeepCheek seems likely to bring significant changes to the global AI market," noting that "this could be both a crisis and an opportunity for our corporations."

He also added, "The success of DeepCheek highlights the importance of high-tech talent," stating that "discussions regarding data protection and technological ethics are also becoming increasingly important."

Professor Gong remarked, "Unlike China, where data can be collected indiscriminately, most countries, including Korea, face significant challenges from the outset when it comes to data collection," adding that "being able to optimize using large amounts of data was largely possible thanks to the presence of outstanding talents."

He continued, "Korea is in an environment where nurturing talent is becoming increasingly difficult due to the 52-hour workweek and increasingly easier science and math education," suggesting that "in light of the DeepCheek shock, there is a need to reassess how effective domestic AI talent cultivation and research culture are."