Some problems with the Future of Defense Task Force report

I. Using the Manhattan Project as a model, the United States must undertake and win the artificial intelligence race by leading in the invention and deployment of AI while establishing the standards for its public and private use.

 

IV. To compete against 21st century adversaries, Congress and the Department of Defense must identify, replace, and retire costly and ineffective legacy platforms. The Task Force recommends that Congress commission the RAND Corporation (or similar entity) and the Government Accountability Office to study legacy platforms within the Department of Defense and determine their relevance and resiliency to emerging threats over the next 50 years.

Those were two important recommendations (#1 and #4) from the Congressional Future of Defense Task Force Report 2020. It was released with some fanfare on Sep. 23, 2020. Read the whole thing if you haven’t already. Here are some belated thoughts:

Revolutionary changes?

They want the DoD to do some “revolutionary” things. But FY21 budget is already on the Hill. FY22 is already going up the chain. So FY23 is what they hope to affect? Or by the time they study it enough, perhaps FY24 or FY25?

I was pretty underwhelmed to see that the Task Force asked RAND or a similar outfit with the GAO to go identify and replace legacy systems. RAND is great but they’re pretty institutional. And there was no target or anything, like reprogram $X billion in the FYDP to emerging tech priorities by FY22 submission. They’ll likely undershoot or do more “innovation theater.”

If Congress was real about that, then they would not have plussed up legacy shipbuilding (+$2.4B), F-35s (+$1.2B), MQ-9s (+$344M), etc. Congress still (sort of) has the opportunity to basically rejigger the entire FY21 budget, or create a sizable portfolio of innovation fund. But they didn’t put their money where their mouth is.

This must all be taken with a grain of salt. The members on the Future of Defense Task Force may be earnest in their pronouncements, but cannot convince the rest of the members to go the same way (particularly the Appropriations Committee). This portends badly for military service leaders who want to reprioritize funds to new programs and emerging tech. They have the power to shape proposals, but intermediate layers in Comptroller and OMB might flip on them if there isn’t Congressional support.

Manhattan Project for Artificial Intelligence

On the Manhattan Project for AI, which is an odd thought, that translates to perhaps $5B per year (inflation adjust a $2B project between 1942-45). It seems that some DoD folks are planning to recommend like a $200M AI scaling fund — they wanted to go larger but think $1B+ is just out of the realm of political reality, even though Congress itself used the term Manhattan Project.

In any case, the Manhattan project had basically a single technical objective (though they took multiple paths on fissionable material and the gun). What’s the equivalent for AI/ML? Is it curated data and enterprise workflows/tooling that would enable competition and a variety of applications? As I argued here, the “narrowness” of AI means that there will be a wide variety of applications, each potentially with their own data, training, algorithms, etc.

One way of thinking about it is that the AI Manhattan Project would fund “AI-native” platforms and capabilities rather than being a source of funds for existing programs to “sprinkle” AI onto.