The US and China Are Teaming Up on AI More Than You Realize

The US and China stand as major competitors in the realm of artificial intelligence, with organizations striving to outperform each other in algorithms, models, and cutting-edge silicon. Nevertheless, these leading AI nations engage in a notable amount of collaboration in advanced research.

A recent analysis by WIRED, examining over 5,000 AI research papers showcased at the top industry conference, Neural Information Processing Systems (NeurIPS), indicates a considerable level of collaboration between laboratories in the US and China.

The findings reveal that 141 out of the 5,290 total papers (approximately 3 percent) involve joint efforts between authors from US institutions and those from China. This collaboration remains relatively stable, with 134 out of 4,497 papers in 2024 featuring authors from both nations.

Additionally, WIRED investigated how algorithms and models originating in one country are shared and modified on the other side of the Pacific. The widely utilized transformer architecture, developed by a team at Google, appears in 292 papers authored by Chinese institutions. Furthermore, Meta’s Llama model has been significant in 106 of these papers. Meanwhile, the increasingly popular large language model Qwen from Alibaba is present in 63 papers that include US collaborators.

Jeffrey Ding, an assistant professor at George Washington University who monitors China’s AI landscape, expresses little surprise at this level of collaboration. “Regardless of policymakers’ preferences, the US and Chinese AI ecosystems are intricately intertwined—and both parties reap benefits,” Ding remarks.

This analysis undoubtedly simplifies the complex nature of idea and talent exchange between the US and China, as many Chinese researchers pursue their studies in the US, cultivating lasting relationships with colleagues.

“NeurIPS exemplifies international collaboration and underscores its significance in our field,” said Katherine Gorman, a spokesperson for NeurIPS, in a statement. “Collaborative efforts among students and mentors often persist long after graduation. Such cooperation signals are evident in various professional networks and among former collaborators.”

The latest issue of WIRED delves into the numerous ways China is influencing the current century. However, as US politicians and tech leaders leverage concerns over China’s ascent as a rationale for rolling back regulations and encouraging significant investments, our analysis serves as a timely reminder that both AI superpowers have much to benefit from continued cooperation.

A Note on Methodology

I utilized Codex, OpenAI’s coding model, to assist in the analysis of NeurIPS papers. After creating a script to download all the papers, I employed the model to analyze each one. This process involved having Codex generate a script to identify US and Chinese institutions within the author field of the papers.

This experiment provided intriguing insights into the potential of coding models to automate useful tasks. While there’s considerable apprehension about AI replacing coding jobs, this is a project I typically wouldn’t have the time or resources to undertake. Initially, I wrote scripts and had Codex modify them before requesting that it conduct the analysis directly. This required the model to import Python libraries, test various tools, and create scripts before generating reports for my review. The approach involved a significant amount of trial and error, necessitating caution since AI models can make surprisingly simple mistakes despite exhibiting intelligence. I ensured that every report contained a way for me to review the results, and I manually checked as many as I could.

This is an edition of Will Knight’s AI Lab newsletter. Read previous newsletters here.