Chinese researchers adapt Meta's Llama model for use in military intelligence.
ChatBIT is responsible for managing specific dialogues in the military field and supports operational decision-making.
Researchers in China have adapted Meta's Llama model to create a military intelligence tool called ChatBIT. Although Llama is an open-source artificial intelligence model available for free, Meta has reiterated that its use should be limited to non-military applications, according to its licensing terms.
However, the development of ChatBIT by a group of six researchers from three institutions—including two affiliated with the Military Academy of the People's Liberation Army—indicates how open-source technologies can be leveraged for unexpected purposes. Using an early version of the Llama model, these academics have integrated its parameters into the Llama 2 13B model, aiming to design a tool focused on the military domain. Supplementary documents even suggest that ChatBIT can process military dialogues and support operational decision-making, seeking to achieve performance comparable to 90% of GPT-4's capability. Nonetheless, no clear testing methods or field applications have been published to support these claims.
Experts in the field of military artificial intelligence in China have reviewed the related documentation on ChatBIT, endorsing the model's development and functionality. However, they criticize the lack of transparent comparison methods and the limitation of datasets, making it difficult to verify the presented results.
A report also highlights that the Chinese military's efforts to adapt Meta's open-source model for strategy and intelligence tasks set a precedent for using Llama for defense purposes. This has reignited the debate over the security risks associated with access to publicly available technologies. Despite the restrictive licenses imposed by Meta, the nature of open-source makes it challenging to enforce such limits, allowing governments from other countries to modify the technology for their needs.
In light of this situation, calls for stricter export controls and restrictions on China's access to open-source technologies have intensified in the United States. To prevent U.S. technologies from facilitating China's military advancements, lawmakers are also considering limiting U.S. investments in China's artificial intelligence, semiconductor, and quantum computing sectors.
Despite the warnings about ChatBIT, some experts question its effectiveness, pointing out that the model was trained using a relatively small dataset of 100,000 military dialogue records. This could affect its ability to handle complex tasks compared to other more advanced models that have been trained with trillions of data points.
Meta has responded to these developments by clarifying that the Llama 2 13B model used for the development of ChatBIT is already outdated and that they are working on a more advanced version, Llama 4. The company has also distanced itself from any unauthorized use of its technology by the PLA, asserting that any application of its models by this military goes against its acceptable use policies.
