On July 31, Reuters published a review of more than 80 Chinese research papers and patents, concluding that researchers linked to the Chinese People's Liberation Army had trained domestic military systems on the outputs of American artificial intelligence models. The method described in the review is known as "model distillation," which is simpler than its name suggests.
Hot Topics
The human body ages in leaps and bounds Scientists identify two critical ages
From your phone screen to your data How do malicious apps sneak in and steal your secrets?
This method involves asking a large, expensive model thousands of questions, collecting the answers, and then training a small, inexpensive model until it masters a specific task. The small model doesn't become a copy of the large one, but it improves on the part it was trained on, and it works at a very low cost.
Among the findings of the review was a 2024 research paper from the National University of Defense Technology, a key research university affiliated with the Central Military Commission, describing an image processing model that had been scaled down and compressed to fit the computing hardware of a drone. The aim was to enable the drone to continue analyzing what it sees, navigating, and targeting even when communication with its operator is lost. In other words, the drone would decide for itself what to look at.
"Distillation technology does not require advanced equipment or penetration of a closed network, but rather a subscription to the service that anyone can pay for."
China publishes enough information to test this conclusion: its military lists its procurement requests on a public platform, and its researchers file patents in open registries. Here, we attempt to examine that record over the past four years, seeking a narrower answer than simply a question of success or failure: What has the US embargo actually done to Chinese artificial intelligence?
The story began five months before the Reuters report. Last February, Anthropic, the US-based developer of the Cloud model, announced that three Chinese commercial AI labs—DeepSec, Moonshot AI, and Minimax—had attempted to exploit its capabilities. These labs had created approximately 24,000 fake accounts and conducted over 16 million conversations through them. That same month, OpenAI, the developer of the ChatGBT model, informed a committee in the US House of Representatives that Chinese entities were circumventing access restrictions through technical intermediaries.
Tags:
technology