Consensus: Emerging artificial intelligence technology could be a game changer for the military, but it needs rigorous testing to ensure it works reliably and that there are no vulnerabilities. which can be exploited by adversaries.
Craig Martell, head of the Pentagon’s Chief Digital and Artificial Intelligence Office, or CDAO, told a packed ballroom at the Washington Hilton that his team is trying to balance speed with caution in implementing the latest AI technologies. is, because it inaugurated four days. Symposium on the subject.
“Everybody wants to be data-driven,” Martell said. “Everybody wants it so badly that they’re willing to believe in magic.”
Large language models, or LLMs, such as ChatGPT’s ability to sift through huge troves of information in seconds and crystallize it into a few key points represent attractive possibilities for militaries and intelligence agencies, which Struggling with how to find forever. Growing oceans of raw intelligence are available in the digital age.
“The flow of information to an individual, especially in a high-activity environment, is very high,” US Navy Capt. M. Xavier Lugo, mission commander of the recently formed Generative AI Task Force at CDAO, said at the symposium. Is.” “Having reliable abstraction techniques that can help us organize that information.”
Other potential military uses of the LLM could include training officers through sophisticated war gaming and supporting real-time decision-making, the researchers say.
Paul Shire, a former Defense Department official who is now executive vice president at the Center for a New American Security, said some of the best uses may be yet to be discovered. What has excited defense officials about LLM, he said, is its flexibility to handle diverse tasks compared to earlier AI systems. “Most AI systems have been narrow AI,” he said. “They’re able to do one thing correctly. AlphaGo was able to play Go. A facial recognition system can recognize faces. But that’s all they can do. Whereas language seems to have more general-purpose abilities. A bridge is built towards
But a major obstacle — perhaps even a fatal flaw — is that LLMs continue to have “hallucinations,” in which they gather false information. Lugo said it’s unclear whether it can be fixed, calling it the “number one challenge for the industry.”
The CDAO established Task Force Lima, an initiative to study generative AI that Lugo chaired in August, with the goal of developing recommendations for “responsible” deployment of the technology at the Pentagon. Lugo said the group was initially created with LLMs in mind — the name “Lima” was derived from the NATO phonetic alphabet code for the letter “L” in reference to LLMs — but the image and Its remit was quickly expanded to include video generation. .
“As we progressed from phase zero to phase one, we moved into creative AI as a whole,” he said.
Researchers say that LLMs still have a ways to go before they can be reliably used for higher purposes. Shannon Gallagher, a Carnegie Mellon researcher speaking at the conference, said her team was asked by the Office of National Intelligence last year to explore how the LLM could be used by intelligence agencies. can Gallagher said that in his team’s study, they devised a “balloon test,” in which they asked LLMs to describe what happened in last year’s high-altitude Chinese surveillance balloon incident, kind of like A secret agency is interested in it as a proxy for geopolitical events.
“I’m sure they’ll get it next time. The Chinese couldn’t determine the cause of the failure. I’m sure they’ll get it next time. That’s what he said about the first A-bomb test. I’m sure they’ll get it next time. They’re Chinese. They’ll get it next time,” read one of the responses.
Even more worrisome is the possibility that an adversary hacker could break into the Army’s LLM and prompt it to spread its data sets through the back end. Researchers proved in November that it was possible: To get ChatGPT to repeat the word “poem” forever, they began leaking training data. ChatGPT fixed this vulnerability, but others may exist.
“An adversary can force your AI system to do something you don’t want it to do,” said Nathan von Hodnos, another Carnegie Mellon scientist, speaking at the symposium. “An adversary can force your AI system to learn the wrong thing.”
During his talk Tuesday, Martell called for industry help, saying it probably doesn’t make sense for the Defense Department to build its own AI models.
“We couldn’t do it without you,” Martell said. “All of these components that we’re envisioning are going to be industrial solution combinations.”
Martel was preaching to the choir Tuesday, with about 100 technology vendors looking for space at the Hilton, many of them eager to snag the upcoming contract.
In early January, OpenAI removed restrictions against military applications from its “Usage Policies” page, which “prohibited any activity that poses a high risk of physical harm, including,” specifically, , “Weapons Development” and “Military and War”.
Commodore Rachel Singleton, head of the UK’s Defense Artificial Intelligence Center, said at the symposium that the UK felt compelled to urgently develop an LLM solution for internal military use because of concerns that staff would be forced into their work. It can be tempting to use a commercial LLM, putting sensitive information at risk. .
As US officials discussed their urgency to introduce AI, the elephant in the room was China, which announced in 2017 that it wants to become the world leader in AI by 2030. The US Department of Defense’s Defense Advanced Research Projects Agency, or DARPA, announced. In 2018 that he will invest $2 billion in AI technologies to ensure that the United States maintains its supremacy.
Martell declined to discuss the adversary’s capabilities during his conversation, saying the topic would be discussed later in a classified session.
Scharre estimated that China’s AI models are currently 18 to 24 months behind American models. “The U.S. technology restrictions are top of mind for them,” he said. To remove restrictions such as chips going to China.”
Gallagher said China may still have the edge in data labeling for LLMs, which is a labor-intensive but key task for training models. Labor costs in China are much lower than in the US.
According to the conference agenda, this week’s CDAO gathering will cover topics including the ethics of using LLMs in defense, cybersecurity issues involved in systems, and how technology can be integrated into daily workflows. On Friday, there will also be a classified briefing on the National Security Agency’s new AI Security Center, announced in September, and the Pentagon’s Project Maven AI program.