On August 26th, the Wall Street Journal reported thatChina AI In recent years, the model has rapidly narrowed the gap with the head model of the United States, not behind a short-term and sudden technological breakthrough, but as a result of a long-standing network of talent in higher education institutions such as the University of Tsinghua, as well as a combination of open-source technology, brain gain and higher computing efficiency. According to the report, a group of researchers represented by the co-founders Tang Jie, Yang Shizuku, the founder of the dark side of the moon, and Leung Wenbei, the founder of DeepSeek, constitute an important force in China ' s pursuit of AI。
According to Wall Street Journal, many Chinese and American industry figures believe that the gap between the best AI models in China and the most advanced models in the United States has now narrowed to months. In June of this year, Mask predicted that China would not be able to catch up to the top of the Anthropic model until the first quarter of 2027, and Tang Jie then responded, "It won't take that long."。
THE REPORT TRACED THE FORMATION OF THE CHINESE AI TALENT SYSTEM TO MORE THAN 20 YEARS AGO. IN 2005, THE TURING PRIZE WINNER, YUO ZHI, FOUNDED THE “YO CLASS” AT QINGHUA UNIVERSITY TO TRAIN TOP TALENT IN THE COMPUTER FIELD. AT THE SAME TIME, TANG JIE HAS LONG BEEN ENGAGED IN DATA MINING AND MACHINE LEARNING RESEARCH IN QINGHUA, AND HIS LABORATORY LATER BECAME ONE OF THE MAJOR SOURCES OF CHINESE AI TALENT. HE STUDIED ALGORITHMS IN THE QINGHUA MACHINE FROM TANG JIE, THEN WENT TO THE UNIVERSITY OF CARNEGIE MELLON TO STUDY FOR A DOCTORAL DEGREE AND, ON HIS RETURN, TO CREATE THE DARK SIDE OF THE MOON。
This movement of people from higher education and start-ups has evolved into an AI talent network. In 2018, China further eased restrictions on the entrepreneurship of researchers and the commercialization of research results. The following year, Tang Jie hatched from the Qinghua laboratory and later developed into a smart AI company. Since OpenAI published GPT-3 in 2020, the development of a homogeneity model has been targeted。
HOWEVER, COMPARED TO THE UNITED STATES AI ENTERPRISE, CHINESE COMPANIES HAVE LONG FACED A SHORTAGE OF FUNDS AND ADVANCED CHIPS. JEFFREY CALCULATED THAT DURING THE RELEVANT PERIOD, THE INPUT OF CHINA SCIENCE AND TECHNOLOGY WAS LESS THAN ONE FIFTH OF THAT OF ITS AMERICAN COUNTERPART. AGAINST THIS BACKGROUND, INCREASED EFFICIENCY IN THE USE OF ENERGY BECAME AN IMPORTANT PATH FOR CHINESE AI FIRMS TO CATCH UP。
DeepSeek is the representative. It was mentioned that DeepSeek reduced memory consumption in model operation through multiple potential attention mechanisms (MLAs) and earlier introduced the Mixed Expert Model (MoE) in China to reduce the need for chip computing. These technologies formed an important basis for the DeepSeek shock at the beginning of 2025, when the low-cost open-source model that it launched triggered a shock in US science and technology and forced Chinese counterparts to re-examine technology routes。
The exchange of technology between Chinese AI enterprises has further accelerated the iterative pace. The moon's dark side follow-up model uses MoE and MLA variants used or validated by DeepSeek, while DeepSeek uses training techniques optimized by the dark side of the moon。Different teams learn from each other through papers, open-source models and technical practices, creating a technology diffusion mechanism different from the closed model route。
At the same time, the return of talent has become another technological path. Two Qinghua graduates who worked in OpenAI were recruited, and byte beats were introduced to Chinese researchers who worked in Google. According to reports, the technological advances of US AI, the return of talent from abroad and China's indigenous research network have together accelerated the catch-up by China AI。
However, calculus remains the main constraint facing Chinese AI enterprises. Liang Wenfeng said in May this year that chips, not talent, are the most critical gap between Central America and AI. According to researchers in the head of the Ari Baba, the brain spectrum, etc., they often receive only one fifth of the high-end chips received by OpenAI, Google colleagues。