"Just as the steam engine was the driving force of the new industry in the past, artificial intelligence (AI) will be the driving force of future industry."
During a visit to South Korea in 2017, Zhang Yaqin, then president of China’s largest internet search company Baidu, remarked on the coming era of AI popularization. Now, nearly 7 years later, his prediction is coming true. The Chinese AI startup DeepSeek is paving the way for popularization after successfully implementing low-cost, high-performance AI.
Now, as the director of the Artificial Intelligence Industry Research Institute (AiR) at Tsinghua University, how does Zhang view the emergence of DeepSeek? He is known as 'China's IT genius.' He entered the University of Science and Technology of China at age 12 and obtained a Ph.D. in electrical engineering from George Washington University at age 23. He has served as the president of Microsoft China and the head of Microsoft Asia’s research and development, transforming Baidu from a 'search company' into China's leading 'AI corporation.' I asked Zhang about the implications of DeepSeek for the global AI industry and how he envisions changes in the industry.
Since the beginning of the year, DeepSeek has amazed the world.
"I believe that DeepSeek has achieved the most significant innovation in the AI industry since the release of OpenAI's ChatGPT in November 2022. Although DeepSeek built its AI model based on research achievements in large language models (LLM) from leading companies like OpenAI, Google, Microsoft, and Meta, it has achieved practical innovation in various sectors, including algorithms, architecture, engineering, products, and open-source business models. It gained 100 million users worldwide just a week after its launch, the fastest for any consumer-targeted AI product. With DeepSeek's success, a new era of LLMs and generative AI characterized by higher efficiency, high performance, and open-source solutions is unfolding. This change will reduce the overall cost of LLM and accelerate the popularization of AI."
About 7 years ago, you emphasized that AI would become the steam engine of the future in Korea.
"That's right. I have always believed in the enormous transformative power of AI. I also see AI as the key engine of the Fourth Industrial Revolution. I just didn't know exactly what technology would emerge and when. I believe that the LLMs introduced by OpenAI and DeepSeek have played a significant role in ushering in the true era of AI."
There is an analysis that the 'scaling law' which suggests that more data is needed to create stronger AI models has hit a crisis due to DeepSeek.
"I don't believe that the scaling law has ended. I think it will continue in different forms. While the effect of the scaling law can diminish in the pre-training stage, I believe new scaling laws may emerge in the testing and inference stages, particularly for models with higher levels of reasoning abilities."
There is also a prospect that DeepSeek will increase AI accessibility, leading to a surge in usage.
"There is a good chance of that. Even if an AI model optimizes and reduces the computing resources required for a single operation, as the fields where AI is used and the frequency of its use increase, overall demand will inevitably rise. For example, if AI utilization expands explosively in various industries such as healthcare, agriculture, transportation, and science, the demand for not only high-performance semiconductor chips but also low-spec chips used for edge devices will also increase accordingly."
What changes will the AI industry face in the future?
"Overall, computing demand in the AI industry will continue to rise, but the way it is met will gradually shift to a 'decentralized' model. It seems that a hybrid structure will emerge, where data centers use ultra-high-performance semiconductor chips, while edge devices use low-cost chips. With increased application of AI to physical devices such as robots, autonomous vehicles, and drones, this decentralized structure is expected to become more realistic."
Do you think DeepSeek has accelerated the arrival of artificial general intelligence (AGI)?
"Yes. I believe that DeepSeek's emergence will accelerate the advancement of both informational AI and physical AI. Especially since countries or corporations that do not possess advanced AI models can still use efficient AI at low costs through open-source ecosystems, I predict that the spread of AI will accelerate globally."
So when do you think the AGI era will begin?
"It will take 15 to 20 years for a complete AGI era to be realized. This is because there is a need for a process that progresses from informational AI to physical AI and then to biological AI."
Watching DeepSeek's success, hope is spreading among startups that 'we can also create great AI models.'
"DeepSeek has opened a path for innovative startups to compete with big tech corporations. Amid a trend towards democratizing the AI industry, which allows anyone to participate, even if DeepSeek does not completely replace big tech, it seems likely to significantly change the industrial landscape. Big tech will continue to provide cutting-edge models through large-scale infrastructure and cloud services, and startups or app developers are expected to optimize specific functions or reasoning abilities using these models to offer customized AI that fits various price-performance conditions."
Many Korean corporations have pursued a 'fast follower' strategy rather than a 'first mover' strategy until now. Will this strategy still be valid in the AI era?
"South Korea already possesses world-class technology in the fields of LLM and generative AI. The High Bandwidth Memory (HBM) technology led by SK hynix is essential for implementing AI models. Not only fast computing power is important for LLM, but also the capability for high-speed data access and storage. South Korea has a definite competitive advantage in this area. Additionally, various infrastructures such as ultra-high-speed internet, wireless networks, data infrastructure, and the semiconductor industry have advanced. This means South Korea is well-equipped to nurture the AI industry. However, to maximize this competitiveness, South Korea should not act alone. I would advise that close cooperation with major corporate entities in countries such as China and the United States is necessary. A strategy that collectively realizes the immense potential of the AI industry will be more effective."
What is the secret behind the continued emergence of outstanding AI corporations in China?
"Ultimately, it's about talent. China possesses the largest pool of young AI talent, including many of the best. Furthermore, China is at the forefront of internet, mobility, e-commerce, and advanced infrastructure sectors. It occupies an advantageous position in digitalization and vast data accumulation. This serves as the foundation for a new wave of AI."
Plus Point
It's not just DeepSeek: China's startup AI assistant 'Manus' rises.
"It's the second DeepSeek moment."
On March 8 (local time), the American economic magazine Forbes reported on the Chinese AI startup Monica. Recently, Monica launched an AI model called 'Manus,' which showed superior performance in GAIA benchmark tests compared to OpenAI's deep inference model 'DeepResearch' released in February this year. GAIA is an indicator that evaluates AI's problem-solving capabilities, measuring not just simple language model tests but how effectively AI can perform actual tasks. Manus is an 'AI agent' that can judge and act independently without human commands. For instance, when given a vague request like 'Find an apartment in San Francisco,' it shows the optimal listings considering not just property listings but also crime rates, rental trends, weather, etc. Forbes evaluated, "Manus is surpassing the limitations of existing AI and shifting the paradigm centered on 'autonomous actions,'" adding that "Silicon Valley should now seriously consider the possibility that China will take the lead in the AI industry."