How china's AI catch-up is being driven by talent and innovation

According to The Wall Street Journal, China's AI models have significantly narrowed the gap with U.S. counterparts in recent years. This progress is attributed to a combination of long-term talent networks, open-source technologies, returning researchers, and improved computational efficiency. Key figures such as Tang Jie, co-founder of Zhipu AI; Yang Zhi Lin, founder of Moonshot; and Liang Wenfeng, founder of DeepSeek, are highlighted as central to this advancement. The article traces the development of China's AI talent pool back to the early 2000s, with the establishment of the 'Yao Class' at Tsinghua University by Turing Award winner Yao Qizhi. Tang Jie, who studied under Yao, later became a key figure in AI research at Tsinghua, while Yang Zhi Lin pursued his PhD at Carnegie Mellon University before returning to China to found Moonshot. This movement of talent between academia and startups has created a robust network supporting AI innovation. In 2018, China relaxed regulations on researchers starting businesses, leading to the formation of companies like Zhipu AI, which set out to develop models comparable to GPT-3. Despite challenges such as limited funding and access to advanced chips, Chinese firms have focused on improving computational efficiency. DeepSeek, for example, introduced the Multi-Head Potential Attention (MLA) mechanism and adopted the Mixture of Experts (MoE) approach early, which contributed to its successful open-source model launch in early 2025. Collaboration among Chinese AI firms has also accelerated progress. Moonshot's subsequent models incorporated techniques validated by DeepSeek, while DeepSeek adopted optimizations from Moonshot. This exchange of knowledge through papers, open-source models, and technical practices has created a different diffusion mechanism compared to the closed model development seen in the U.S. Meanwhile, talent returning from abroad has further bolstered China's AI landscape. Tencent recruited two Tsinghua graduates who previously worked at OpenAI, while ByteDance brought in researchers from Google. However, access to high-end chips remains a critical challenge. Liang Wenfeng noted that chip availability, not talent, is the key disparity between Chinese and U.S. AI companies. Researchers at Alibaba and Zhipu reported receiving only about one-fifth the number of advanced chips compared to their counterparts at OpenAI and Google.