Deputy Defense Secretary Kathleen Hicks specifically called out the “so-called valley of death” as a barrier to innovation in a virtual speech at the National War College on Friday, saying that the DOD needs to be able to take concepts and theories and develop them into new capabilities that can be fielded by the military.
Part of the problem is the DOD’s rigid budgeting process, whereby acquisition programs, in addition to the multiple years they often take to establish, are usually expected to have clearly defined long-term plans and outputs, said Eric Lofgren, a research fellow at George Mason University’s Center for Government Contracting.
Technology programs like software development or artificial intelligence, by contrast, are often developed iteratively, can change significantly in relatively short periods of time, and don’t necessarily fit into that budget model.
“Software is never done, it’s agile,” he said. “You have a road map, but you don’t have a waterfall plan. Then AI just blows it out of the water — how could I know what the outcome of this algorithm will be before I collect the data, before I train the data? You have to justify [the project] using your ability to predict and plan, but you can’t do that when the whole problem is the knowledge and the data problem.”
There’s a risk that if the DOD continues to struggle to convert that initial interest from tech firms into long-term success, “this generation gets disillusioned” and those firms become reluctant to put further effort into working with the department, Lofgren said.
That was a nice article from Daniel Wilson, DOD Tech Outreach Undermined By Contracting Barriers. I think the prediction-based model of acquisition is really running into problems.
When you know the Laws of Nature and mathematics — such as Newton, Kepler, and Schrodinger’s equations — engineers can in theory design the space of possible solutions for systems. Designing a program approval process based on precise planning and costing makes sense. The world is complicated, but smart people are smart enough to find optimal solutions using their engineering models.
That process really doesn’t makes sense for AI/ML. Consider automating target recognition using drone data. You know the requirement for the application is to identify and track enemy combatants. But is the data you have in bits and pieces all over the department sufficient to train an algorithm? How could you know until you’ve brought it together, tagged it, and tried?
And so DoD might have a requirement, yet it has no basis for creating a linear schedule of tasks that can be costed out. Like software, an AI/ML application is never done. This is a huge challenge for processes based on 20th century hardware. Obviously, it is not the commercial approach to AI/ML that is at fault. It is the Pentagon’s outdated processes that need transformation.
Leave a Reply