After the launch of its new generative AI task force – dubbed Task Force Lima – earlier this year, the Department of Defense’s (DoD) Chief Digital and Artificial Intelligence Office (CDAO) is continuing to make progress on the adoption of generative AI technologies across the DoD.
According to the Chief Technology Officer at the CDAO, Bill Streilein, the office has collected nearly 200 generative AI use cases across the department.
“We are working across the department – with other agencies, other PSAs, and the services – to figure out [and] help the DoD chart its course for generative AI,” Streilein said during a GovExec webinar on Dec. 14. “We’ve collected nearly 200 use cases over the last couple of months from the department where we solicited, we did a data call, tell us about your favorite generative AI use case.”
“We’re assessing them. We’re trying to understand which ones would be appropriate in the state of the technology, which is important to acknowledge,” he said. “[There’s] still a lot to learn about it … Within the DoD the consequences are perhaps higher, and we need to be responsible in how we leverage it.”
DoD announced the establishment of Task Force Lima on Aug. 10. The CDAO said the task force will examine generative AI use cases across the Federal government and develop recommendations on how the DoD can responsibly use these powerful technologies. The Pentagon said it plans to utilize generative AI models to enhance its operations in areas such as warfighting, business affairs, health, readiness, and policy.
Streilein explained during the webinar that the department is “excited about this capability, because it enables us to deal with the huge amounts of data that we have” including textual data like policy and planning documents or imagery and “other things like that.”
When it comes to large language models specifically, Streilein said the CDAO is “trying to experiment with it.”
He explained that they’re looking into what it would take to host their own model – like a ChatGPT – at various classification levels.
“We don’t know what the right model is yet,” Streilein said. “The issue is the DoD has sensitive data. And when you interact with a model, that prompt goes to one of these large models which now, if you think of it that way, has basically been leaked. And so, the DoD has to be very careful about that.”
“We can certainly use unclassified data that is not sensitive and interact to understand its capabilities, but for many of the use cases, the data that is involved is sensitive,” he continued, adding, “We are looking into what it would take to host our own model at various classification levels.”
“There’s a lot that we don’t quite understand related to the size of the model. The largest models are incredible in their capability,” he said. “But it may be for some of those use cases we could use a smaller model, and exactly the parameterization of that is yet to be determined. There are efforts in the department to create smaller models for focused applications, and those are very promising.”
