.png){kind=logo}
Recent academic papers and analysts exposed that China is using Meta's publicly available Llama model to come up with an AI tool perhaps for military use. It was found that the Chinese research institutions which are affiliated with the People's Liberation Army (PLA) are developing this AI.
According to a June study, as reported by Reuters, six researchers-based institutions, including two affiliated with the PLA's primary research organization called the Academy of Military Science (AMS). It had developed, "ChatBIT" using an early variant of Meta's Llama.
The researchers used the Llama 2 13B large language model, released by Meta in February 2023, and took its parameters to develop a military-specific AI tool for gathering and processing intelligence. This AI tool is developed to generate accurate and reliable information that can be used for decision-making in operations.
It explained that ChatBIT had been specifically optimized for military-oriented dialogue and question-answers, reportedly surpassing other AI models.
The model is nearly on the same level as the highly-touted OpenAI ChatGPT-4. However, the researchers failed to give insight on how they defined performance, or whether ChatBIT has already been deployed in some form.
“This is the first clear evidence of PLA experts who, in a systematic way, study and exploit open-source LLMs like Meta's for military use." Sunny Cheung, Associate Fellow at the Jamestown Foundation, specializing in China's emerging technologies.
The members of the Chinese research team include Geng Guotong and Li Weiwei working at the AMS's Military Science Information Research Center along with Beijing Institute of Technology and Minzu University.
Reuters was unable to confirm the abilities and processing power of ChatBIT. Researchers said that the model had only been trained on 100,000 military dialogue records which were far fewer in number compared to most LLMs which train primarily on trillions of tokens. Joelle Pineau of Meta described it as unlikely that it's able to do this "with 100k dialogues versus trillions on most others".
Meta has continued to make a practice of publicly releasing most of its AI models, including Llama while putting conditions on their application. For example, a license application is required for services that have more than 700 million users. The models are also barred from military warfare or espionage applications.
However, Meta still faces problems of enforcement with regard to its policies due to the fact that its models are in the public.
Meta director of public policy Molly Montgomery put emphasis on the fact that "any use of their models by the PLA is unauthorized and violates their acceptable use policy."
The publication of this research coincides with the debates that have been ongoing between the US national security and technology sectors regarding the implications of public AI models.
In October 2023, President Biden signed an executive order aimed at governing AI development. According to him, there's a need to balance benefits from innovation with significant risks to security.
The US sealed rules for the past couple of weeks that limit investments into the AI and technology sectors of China, which adversely affects national security.
Defense Department spokesman John Supple said open-source models bring along both advantages and risks in their wake, reasserting that the DoD "remains vigilant regarding these capabilities".
Some analysts argue that China's success in indigenous AI development, and the establishment of multiple research laboratories has complicated its ability to maintain a technological lead over the US.
In another academic paper, researchers from the Aviation Industry Corporation of China (AVIC), a company tied to the PLA, report on using Llama 2 for training airborne electronic warfare strategies.
Some other topics that have been published in the PLA Daily on state-run forums include AI hastening weapons, and how the simulation and training efficiency of combat can be improved.
"Excluding China from these technological developments becomes ever more unfeasible as collaboration between Chinese and American AI scientists continues to flourish," according to William Hannas of Georgetown University's Center for Security and Emerging Technology.
