Pentagon sunsets generative AI task force, launches rapid capabilities cell

The Defense Department is sunsetting Task Force Lima and launching a new initiative focused on accelerating the delivery of new generative AI capabilities.

Task Force Lima was stood up last year to help the Defense Department better understand how it can effectively and responsibly leverage gen AI tools such as large language models. Officials have taken lessons learned from that effort and stood up the Artificial Intelligence Rapid Capabilities Cell (AI RCC).

“Over the course of 12 months, Task Force Lima analyzed hundreds of AI workflows and tasks that AI tools could make more efficient or more effective. And we categorized all of those use cases into a smaller set of 15 areas aligned into two big categories: warfighting functions — like command and control [and] decision support — and enterprise management functions like financial management and healthcare information management. Upon completing its work, Task Force Lima submitted a detailed report,” Radha Plumb, head of the department’s Chief Digital and AI Office (CDAO), told reporters at a Pentagon briefing Wednesday.

The cell is being stood up to implement those recommendations with the aim of accelerating the delivery of frontier models and next-generation AI capabilities across the department, she said.

The initiative will be led by the CDAO in partnership with the Silicon Valley-headquartered Defense Innovation Unit (DIU).

The cell will be charged with identifying and testing technology through rapid experimentation and prototyping, assessing the effectiveness of technology and determining whether it can be scaled and sustained.

“If warranted, we’ll use defined acquisition pathways to scale the technology across the DOD enterprise, and that can be within CDAO, with the military departments or with other key components,” Plumb said.

“This rapid experimentation approach will allow us to test and identify where these cutting-edge technologies can make our forces more lethal and our processes more effective, but equally critically, the AI RCC will define the requirements for enterprise infrastructure … and support scaled AI development that includes compute development environment and AI-ready data,” she added.

The warfighting use cases that the cell will focus on include command and control (C2) and decision support, operational planning, logistics, weapons development and testing, uncrewed and autonomous systems, intelligence activities, information operations and cyber operations, according to a DOD fact sheet.

Enterprise management use cases include financial systems, human resources, enterprise logistics and supply chain, health care information management, legal analysis and compliance, procurement processes, and software development and cybersecurity.

The Pentagon is planning to allocate $100 million from fiscal 2024 and 2025 research, development, test and evaluation funding toward some of the initial efforts, according to Plumb.

Part of the investment will include $35 million for four frontier AI pilots that will kick off “immediately,” according to Plumb. Those will be conducted in partnership with the combatant commands and other DOD organizations in 90-day increments, including via the Global Information Dominance (GIDE) series of experiments.

About $5 million will go toward “rapid user-centric experimentation,” according to a DOD fact sheet.

The CDAO plans to work with DIU to tee up additional pilots “in the near future,” Plumb noted.

In mid-January, the Pentagon also intends to award about $40 million in Small Business Innovation Research contracts to fund generative AI solutions, including from non-traditional vendors. The department is still in the source selection process.

“We’ve received hundreds of responses to our request for solutions to leverage generative AI in specific DOD ecosystems, everything from applying commercial applications to healthcare and financial management to solutions in critical warfighting areas like autonomy,” Plumb said.

Another $20 million will go toward boosting compute and creating digital “sandboxes” to facilitate development, experimentation and testing. The department is taking a multiple cloud approach and plans to lean on major providers working under the Joint Warfighting Cloud Capability enterprise on that effort. The four vendors with contract spots on the $9 billion JWCC program include Google, Oracle, Amazon Web Services and Microsoft.

“We will have a sandbox with each major cloud provider. We’ll start with two sandboxes that will be available in mid-January with two providers, and then fast-follow with two additional sandboxes on the other two cloud instances by the summer,” Plumb told reporters.

The CDAO chief said she couldn’t provide a specific timeline for when new tech shepherded by the rapid capabilities cell will be ready for deployment, noting that it will depend on the performance of the technologies in testing and experimentation.

“I think industry continues to innovate and improve both the quality and reliability … of their generative AI models, and we’re watching that very closely and in close partnership with our industry innovators. The second piece, though, is … the department has to have its own reliability standards. We talk a lot about responsible AI. What that really means is, do the models perform the way you want them to perform? And do they do they do the things you want them to do? And do they not do things you don’t want them to do?” Plumb said.

“That’s true for all of our platforms and capabilities. We have to do that in weapon systems, we have to do that in our digital solutions, and we have to do that in our hardware. We have a specific set of standards and applications that we apply in the generative AI context to bound the risk and ensure the performance meets the reliability. Part of the pilots, the test and evaluation, and the generative AI-specific responsible toolkit are creating the pathways for that,” she added. “To my mind, this is a really ‘better brakes make faster trains’ approach where we’ve got a toolkit, we’ve got to test the technology, and then we’ve got to rinse and repeat to get it to the reliability level that will allow us to deploy it. That’s going to vary use case by use case, but that’s the approach we’re taking here.”

The risk management framework includes things like authority to operate (ATO) processes — which the Defense Department is trying to streamline — and identity credential and management (ICAM) solutions.

“Those are the tools that let us as a department, continue to review and make sure digital solutions we bring in from the commercial sector meet our cyber requirements and don’t provide threats,” Plumb said.

“There’s a broader set of issues in which we have to think about how we deploy AI into our ecosystems, how we think about data security and the data use in our systems now. That is an ongoing part of our discussions and part of what we want to get after with these [rapid capabilities cell] pilots. How do we bring in the very best commercial technology, marry it with our sort of unique, often classified data, use that for our warfighters, and then be able to scale that? And that is explicitly one of the things we need to work through within the context of these pilots. That’s going to vary a lot depending on what kind of data you’re using. And you can imagine security risks that relate to health information look very different than security risks that relate to cyber information, which in turn look really different than the risks related to autonomous systems. So there is a workflow use case specificity to this. That’s part of the pilot effort,” she said.

As for the leaders of the disbanded Task Force Lima, some will be joining the rapid capabilities cell and others will be working on other priority projects for the CDAO, according to Plumb.