The Defense Department wants to see data as a product as it looks to industry for help in identifying use cases for large language models and generative AI, according to its AI technology chief who spoke about its new AI adoption strategy.
“We’ve talked about data as an asset, but what do you do with assets? You hoard them,” said William Streilein, CTO at DOD’s Chief Digital and AI Office (CDAO), at the AFCEA International and AFCEA Hawaii TechNet Indo-Pacific conference in Honolulu Tuesday. “This focus on data products is one of our key areas of emphasis as we move out with this strategy.”
The message comes a week after President Biden’s executive order calling on agencies to develop safe and trustworthy AI, and weeks after the office’s adoption strategy that outlines an agile approach to AI that integrates speed, learning and responsibility. Foundational to this strategy is quality data.
“We are plagued by hype that stands in our way, not only for adoption, specifically in the case of CDAO, but it keeps us from seeing it as a technology,” Streilein said. “It is a technology that makes sense of data — that's really what it is. And the technology is getting better and better to make better sense of that data.”
Streilein noted the importance of this work toward broader efforts at the department like CJADC2 and multi-domain operations.
This focus on quality data is why the office is reaching out to industry and other partners through efforts like Task Force Lima and the Tradewind acquisition vehicle. Through Tradewind, for example, the office is seeking help to develop a maturity model for large language models. This maturity model would be used for assessing AI solutions within use cases and workflows.
“Within the DOD, it's a difficult thing to engage with a large language model that isn't inside the fence, if you will. How do we engage with a model that is so powerful, but we can't share our data with it?” he said.
Streilein sees a lot of promise in efforts like the Army’s Project Linchpin program in which CDAO is trying building out “AI scaffolding.”
“They are major partners to CDAO and one of the first pilot cases, if you will, for us to prototype our services for what's known as AI scaffolding,” he said of the project, which aims to “put pipelines of AI capability throughout the Army — and we're totally behind it.”
Moving forward, Streilein noted the biggest — and most expensive — challenge ahead is data labeling.
"There's identifying which data you have from a sensor. There's bringing it in. And then putting labels on it that say to the algorithm, this unit of data represents this aspect of my mission," he said. "Right now, the best methods are done by humans."