The inner workings of China’s top artificial intelligence laboratories are often shrouded in mystery, viewed from the outside through the lens of geopolitical competition and benchmark scores. However, a recent firsthand account by American AI researcher Nathan Lambert offers a rare and nuanced glimpse into the culture, structure, and daily realities of the teams building some of the world’s most advanced models.
Lambert, who publishes the widely read Interconnects newsletter and serves as senior research scientist at the Allen Institute for AI (Ai2 ), recently spent time visiting leading Chinese AI labs. His observations, published directly in his Interconnects post, challenge several Western assumptions about the Chinese AI ecosystem and highlight the unique factors driving its rapid progress.
Ecosystem Over Tribalism
One of Lambert’s most striking observations is the highly collaborative nature of the Chinese AI research community. In contrast to the intense, often secretive tribalism that characterizes the rivalry between U.S. labs like OpenAI, Anthropic, and Google, Lambert found a more fluid exchange of ideas in China. He noted that researchers frequently move between institutions and maintain open lines of communication, fostering an “ecosystem” approach rather than a zero-sum competition.
This collaborative spirit is partly driven by necessity. As we have previously reported, Chinese AI companies face significant constraints regarding access to advanced compute. By sharing insights and techniques, researchers can optimize their limited resources and accelerate collective progress.
The Student Engine
Another key differentiator is the demographic makeup of the research teams. Lambert observed that student researchers, often graduate students from top universities like Tsinghua and Peking University, form the core engine of many leading AI labs. These students are deeply integrated into the development of frontier models, taking on responsibilities that would typically be reserved for senior engineers in the West.
This reliance on academic talent creates a dynamic, fast-paced environment where theoretical research is rapidly translated into practical applications. It also helps explain the sheer volume of high-quality AI research papers originating from Chinese institutions. The tight integration between academia and industry ensures a constant pipeline of fresh talent and innovative ideas.
The “Build-Not-Buy” Data Culture
When it comes to training data, Lambert noted a distinct “build-not-buy” culture. Rather than relying heavily on purchasing proprietary datasets or scraping the open web, Chinese labs are investing heavily in generating their own high-quality, specialized data. This approach is particularly evident in the development of models tailored for specific industrial applications, a strategy that aligns with China’s broader focus on AI diffusion.
By controlling the entire data pipeline, these labs can ensure the quality and relevance of the information used to train their models. This meticulous approach to data curation is a critical factor in the impressive performance of Chinese models on both domestic and international benchmarks.
Government Support, Not Direct Control
Lambert’s account also provides a nuanced perspective on the role of the Chinese government. While state support, in the form of funding, infrastructure, and favorable policies, is undeniable, Lambert observed that the government does not exert direct technical influence over the day-to-day operations of the labs. Researchers are largely free to pursue their own technical directions, provided they operate within the broader strategic framework established by Beijing.
This dynamic contradicts the common Western narrative of a rigidly top-down, state-controlled AI sector. Instead, it suggests a model where the government provides the resources and sets the overarching goals, while leaving the technical execution to the experts.
Admiration for Claude
Perhaps the most striking observation Lambert shared is that Chinese AI developers are, in his word, “obsessed” with Anthropic’s Claude models, despite Claude being nominally banned in China. This admiration is not merely academic; it actively informs the research directions, benchmarking strategies, and software development practices of Chinese teams. The irony is sharp: the same U.S. model that Washington is trying to protect from distillation is the one that Chinese researchers most closely study and emulate.
This observation aligns with the recent revelation that Chinese users have been accessing Claude via “shadow APIs”, highlighting the strong demand for access to top-tier U.S. models within the Chinese developer community.
A Formidable Competitor
Lambert’s firsthand account paints a picture of a vibrant, highly motivated, and uniquely structured AI ecosystem. Driven by a collaborative culture, a deep pool of academic talent, and a meticulous approach to data, China’s AI labs are proving to be formidable competitors on the global stage. As the race for AI supremacy continues, understanding the internal dynamics of these institutions will be just as important as tracking their benchmark scores.