Chinese AI assistant copies American developments: scandal around Kimi K3

The recent revelation that the Chinese company Moonshot AI, which caused a buzz with its new artificial intelligence model Kimi K3, had in fact copied it from the American company Anthropic has raised serious concerns, especially in U.S. political circles. While the tech sector has always been characterized by rapid exchange and borrowing of ideas, the scale and methods used to “replicate” Kimi K3 raise questions about ethical norms and competition in the artificial intelligence industry.

Paths of borrowing: hard drives and friendly countries

According to representatives of the U.S. presidential administration, Chinese AI developers resort to complex schemes to gain access to advanced American technologies. One method that raises concern is transporting hard drives with AI models to countries that have friendly relations with the United States. This allows them to study and analyze American developments without directly violating U.S. laws. Further use of reinforcement learning techniques and state-of-the-art NVIDIA graphics processors then allows them to create their own, albeit derivative, versions of the software.

Exposure through self-identification

The fact that Kimi K3 was created on the basis of Anthropic’s Claude model became obvious almost immediately after its presentation. During one test conversation with Kimi K3, the model introduced itself as “Claude, an AI assistant created by Anthropic.” This is a simple, yet highly significant piece of evidence that immediately drew the attention of experts and regulators.

Official reaction and accusations

Michael Kratsios, chief science and technology adviser and director of the White House Office of Science and Technology Policy, officially confirmed that Moonshot AI copied Kimi K3 from the Fable model developed by Anthropic. He also added that Moonshot had developed “a sophisticated internal platform for conducting large-scale distillation against American models.” According to him, this platform allows them to quickly switch between different access methods to avoid detection. This suggests a deliberate and systematic approach to copying rather than accidental borrowing.

Technological arsenal and methods of avoiding detection

Moonshot AI’s high-tech lab likely has access to several NVIDIA GB300 servers, which are among the most powerful graphics processors on the market. In addition, there is a possibility that they may have obtained additional resources through servers located in Thailand. This allows them to perform the compute-intensive tasks required for the “distillation” process (that is, transferring the functionality of a large model to a smaller one).

An important element of Moonshot’s strategy is the rapid switching between different types of access and the constant querying of data from Anthropic’s models. This makes it harder to detect their activity, creating the impression of legitimate interaction.

High efficiency and not just “theft”

Despite the accusations of copying, it cannot be ignored that the Kimi K3 model, built on Claude Fable, was significantly improved. It demonstrates high efficiency for use in data centers thanks to reduced “key-value cache” and built-in memory-use optimizations for each token. This suggests that Moonshot did not simply “compile” someone else’s model, but also carried out some work on optimizing and adapting it. However, it was Anthropic’s original development that became the platform on which Kimi K3 was built, casting doubt on the originality of the Chinese development.

AI market context and geopolitical risks

This incident takes place against the backdrop of growing global interest in artificial intelligence and intense competition between countries for technological leadership. The United States, as one of the leaders in AI development, is concerned that its technologies may be used for commercial or military purposes by other states without proper oversight.

“Distillation” is a process in which data and functionality from a large, powerful model (a so-called “teacher”) are transferred to a smaller, more efficient model (a “student”). This method is used to create more accessible and cost-effective solutions. However, when the “teacher” model is someone else’s intellectual property and the distillation process is carried out without permission, it may be considered copyright infringement and industrial espionage.

The consequences of such actions can be serious. On the one hand, they can stimulate innovation by forcing companies to compete at new levels. On the other hand, they can reduce incentives to invest in original research and development if the results can be copied so easily. In addition, this creates additional geopolitical risks, as technological superiority in AI can be used to strengthen military power or espionage.

In this situation, questions arise about the effectiveness of international mechanisms for protecting intellectual property and regulating the development of artificial intelligence. Companies that invest billions in developing advanced AI models need confidence that their achievements will be protected. At the same time, countries seeking to catch up with the leaders will look for ways to accelerate their development. The story of Kimi K3 is a vivid example of how difficult it is to strike a balance between innovation, competition, and ethical norms on a global scale.