Image source: topwar.ru
The US Military is holding a four-day symposium on the topic of military use of large language models of artificial intelligence, such as ChatGPT. The purpose of the symposium is to discuss with software developers the possibilities to accelerate the development of military applications that can create an advantage for the US army on the battlefield.
The main speaker at the event was Craig Martell, Doctor of Computer Science, Director of the Pentagon's General Directorate of Digital Technologies and Artificial Intelligence (CDAO). Opening the symposium in the crowded ballroom of the Washington Hilton Hotel, he stated in particular:
After initial excitement around the world about the appearance of ChatGPT and deep fake image generators, Pentagon officials became concerned that AI could be used against US interests. The main danger is that the emerging artificial intelligence technology could change the rules of the game for the military. And therefore, it needs intensive testing for emerging vulnerabilities that could be exploited by opponents.
- US Navy Captain M. said at the symposium. Xavier Lugo, commander of the newly formed Generative AI task Force at CDAO.
Another possible military application of LLM could include training officers through sophisticated war games, as well as real-time decision-making assistance.
Paul Scharre, a former Defense Department official and vice president of the Center for a New American Security, said:
According to M. Xavier Lugo, the critical drawback of LLM systems so far is that they can have “hallucinations” during which they conjure up inaccurate information. It is not yet clear if this can be fixed, so it can be called the number one problem for industrial AI use.
Shannon Gallagher, a researcher at Carnegie Mellon, talked about how LLMs can be used by intelligence agencies using the example of the "Chinese balloon test" (the incident with the Chinese observation balloon occurred last year):
In general, the symposium paid great attention to the development of AI in China. Charre estimated that Chinese artificial intelligence models are currently 18-24 months behind American ones. However, this gap can be overcome due to the cheapness of labor in China, since labeling data for LLM is a time-consuming but key task in training AI models. China has said it wants to become a world leader in artificial intelligence by 2030. Therefore, US technological sanctions will not be lifted from China in order to prevent the accelerated development of the PRC.