The Humble Task of Implementation is the Key to AI Dominance
Our dynasty’s majestic virtue has penetrated unto every country under Heaven, and Kings of all nations have offered their costly tribute by land and sea. As your Ambassador can see for himself, we possess all things.
In 1793 China was the Middle Kingdom, the center of the world that received tribute from all others. Its dominant posture was based in part on its historic mastery of technology, from gunpowder to the compass and moveable type. However, centuries of unchallenged dominance blinded China to the growing technological prowess of the West and stifled innovation. It went on to endure a century of military, economic, and societal humiliation at the hands of Western powers, who leap-frogged China during the Industrial Revolution.
After decades of unprecedented growth, China today possesses, by some measures, the largest economy in the world. By the middle of the century, Beijing aspires to become “a global leader in terms of composite national strength and international influence.” To do this, President Xi Jinping believes China must “develop and control” artificial intelligence (AI). As Chris Brose mentions in his book, The Kill Chain, a focus on emergent technology, namely AI, is core to China’s attempt to leapfrog U.S. dominance.
For now, the United States appears to hold an advantage over China in terms of AI. If the United States is to maintain its advantage in AI, which is diminishing in the face of Chinese advances in development, the current conversation about AI development needs to shift. Instead of focusing on fantastical claims about how AI development will usher in an era of exponential efficiency, fully AI-enabled autonomous systems, and artificial general intelligence (i.e., AI that can understand any task that a human can), today’s commercial and military AI developers need to focus on the humble mandate of implementation. Implementation in the Defense Department is a colossal challenge as warfighters and acquisition professionals struggle to understand AI, let alone engineer the proper environment for AI development and integration amidst stultifying bureaucracy.
To overcome these barriers, the Defense Department needs a common platform for AI development and application. A development platform will bring data, AI developers, and the military into a single ecosystem to develop AI at speed and scale. The Pentagon should also make significant changes to its acquisition approach by overhauling intellectual property ownership, modernizing contractor solicitation and payment, and rebalancing a mix of relationships with large, medium, and small-sized AI development companies.
AI Challenges: Data, the Development Environment, and Integration
The first challenge to implement AI is finding the data required to build an algorithm. It is common in the Defense Department to assume that there is a wealth of data “out there” waiting to be pulled through commercial algorithms. Based on personal experience, I assure you that is not the case. The reality is that the data is either mired in legacy weapon systems, with proprietary code restricting access to would-be developers, or it is not thoroughly cleaned, organized, and wrangled (transforming the data format to another useable format) for a specific use case.
For example, while working on the Humanitarian Assistance Mission Initiative in the Joint Artificial Intelligence Center, wildfire data from full motion-video from an MQ-9 Reaper drone had position data and telemetry burned into the imagery. This meant an algorithm would be continually confused by the irrelevant numbers and letters etched into the images, making the data significantly more difficult to use for AI development. That is an example of the type of data that is “out there” for the Defense Department.
Another barrier to AI development in the Defense Department is the lack of a vendor-friendly software platform that brings together data, AI code languages and their libraries, and testing software required before deploying code. Without this development environment, it is impossible to support the experimental nature of algorithm development or provide mechanisms to respond to user feedback. For the Defense Department’s platform to be useful, there must be a well-defined and established process that elicits and incorporates user evaluation and input — something that happens with depressing irregularity today. In my experience, solving one user requirement typically begets another, unforeseen requirement. Identifying a user’s needs demands a software platform in which developers and users sit virtually or literally side-by-side. Once modifications are made to the algorithms the next critical function to the development environment must be the ability to dynamically retrain the algorithm and to conduct test and evaluation. Ultimately, an AI model that is fielded once and never updated is useless; in those cases, it is better not to use AI solutions at all.
Additionally, the Defense Department should empower its vendors to use the cutting-edge commercial AI tools they are accustomed to having at their fingertips in the private sector. Specifically, the Defense Department should replicate the DevSecOps ecosystems commercial developers use so there is no seam between the government solution and the best commercial AI tools.
During my experience on the team responding to wildfires, we faced a relatively simple yet debilitating problem: We were not allowed to bring in the latest versions of vendor-preferred tools like PyTorch, which was vital to our model development. Instead, we were forced to make our vendors use an older version, since that was the only information assurance-approved tool suite available. By developing an ecosystem with updated, industry-standard containers, the Defense Department can architect various sandboxes or domains where developers can have access to the tools they need to develop models.
Finally, successful AI implementation depends on the delivery of AI-enabled insights to the user. The key is to filter those insights into whatever legacy weapon system that the user is most familiar with. In the case of optimizing wildfire discovery, we used a National Guard innovation network on commercial internet that was secured through common, commercial practices. Once we were able to develop algorithms, we could demonstrate relevant capability by improving the intelligence analyst workflow. Here, the team inspired by the Strangler pattern philosophy that corporations like Google and now the U.S. Air Force recognize as a best practice. The Strangler pattern can be a useful approach for the Defense Department because it edges out an old system while replacing it with a new system. The difference in many cases around the Defense Department is that it relies on human effort, such as staring at intelligence sensors and flagging activity.
The Answer to AI Implementation: Joint Common Foundation
The Defense Department needs a new development environment for its AI systems. This would enable developers, scientists, acquirers, and warfighters to test prototypes and explore new operating models for their domain. The Joint Artificial Intelligence Center’s Joint Common Foundation can play a central role as the cloud-based development environment that brings data, libraries, cutting-edge AI development tools, and open system architecture standards under a single technical solution to the Defense Department.
With extensive experience as both an acquisition and intelligence officer, I am attuned to the criticism about the sluggish pace of system upgrades in the Defense Department. This criticism is entirely warranted. As an intelligence officer (and deployed to Combined Joint Task Force Operation Inherent Resolve as a targeting officer) I engaged in intelligence, surveillance, and reconnaissance operations during the counter-ISIS campaign from 2015 to 2018. While serving in these roles I was acutely aware of the voluminous data our sensors generate and especially the amount of potentially valuable information that ended up unviewed on the cutting-room floor during the deliberate targeting process.
Yet, the vision of an AI solution that could reduce the analyst’s workload and facilitate ground force commanders’ decision-making was overly ambitious — analysts could not even accomplish basic tasks such as enhancing imagery to make specific annotations. A networked development environment would have allowed us to innovate and improve our workflows with the help of contractor support. Operators know the mission and AI engineers know how to build algorithms.
Fully building and deploying the Joint Common Foundation would require the Joint Artificial Intelligence Center to tackle three daunting challenges — cybersecurity for AI, data sharing, and cloud engineering expertise. With a clear vision of the challenges, the Joint Artificial Intelligence Center can solve for its development environment and subsequently create the catalyst for AI delivery at scale.
Establishing a better development environment is key to AI implementation at the Defense Department. First, the Joint Artificial Intelligence Center will have to focus less on networks and more on software — AI is software code at its base — to secure its cyber networks. Next, the Pentagon has a data sharing problem. In order to make the Joint Common Foundation effective, the Joint Artificial Intelligence Center must create a fast lane to data access that is staunchly bounded and controlled. Finally, the Joint Common Foundation will require a steady cadre of cloud computing experts. In order to attract these experts the Defense Department must continue to highlight its unique mission while simultaneously creating a familiar development environment for the cloud engineers. These engineers are integral to building the development environment so that the top-talent AI engineers can get to work. However, all this would mean little without commensurate improvements to Department-wide acquisition and contracting processes and procedures.
AI Development Requires a New Approach to Acquisition
Today, the defense acquisition system is so difficult to navigate that it makes little business sense for the best AI companies to enter the defense market. This is due in large part to misalignment of incentive structures between the Defense Department and small AI startups from three perspectives. This requires addressing three acquisition-related problems: outdated intellectual property strategies; slow solicitation and payment processes; and reimagining how big defense contractors, acting as system integrators, and small AI startups work together to make the AI whole more than the sum of its parts.
The Defense Department tends to oversimplify its intellectual property needs by purchasing it outright from commercial companies. Instead, and as long as the Defense Department’s data is protected, the Defense Department should favor a licensing or royalty approach to AI development, especially for smaller AI startups. Insofar as intellectual property for AI is concerned, trained AI algorithms could be made almost irrelevant in months, weeks, or days as better techniques are rapidly developed and fielded. Requiring the intellectual property purchase of every trained model is a bad investment. Instead, the true value stems from the continuous improvement and adaptability of the feedback loop. When it comes to AI, we should consider the lessons learned regarding data management and parameter tuning — inherent in the skillset of the AI developer’s workforce — as the true value proposition to the government.
Another major barrier to entry for small and medium AI developers is the time to be placed on contract coupled with the glacial pace of the Defense Department’s payment system. Typical solicitation processes can take months of development and deliberation, with considerable efforts on the company’s part to respond to government requests for information. When added to the financial toll of the technical evaluation process that precedes contract award activities, many small and medium companies simply do not have a deep financial well to endure the pain. It would not be unreasonable to expect that the process I just described could take 12 to 18 months to yield first payment. It would be an understatement to say that small and medium sized businesses are discouraged from seeking government contracts that use confusing and circuitous paths to contract award and compensation. To accelerate the process, it is imperative to leverage acquisition tools like the Commercial Solutions Opening, Other Transaction Authority, or the recently released recently released Acquisition Pathway Interim Policy and Procedures launched by the Under Secretary of Defense for Acquisition and Sustainment Ellen Lord.
Finally, many small and medium companies are not postured to endure the Defense Department’s byzantine information assurance requirements or the integration process to function on top of legacy defense systems. As an example, the Pentagon can benefit by a contractual arrangement with a system integrator, who then chooses sub-contractors to meet the military’s specific AI requirements. However, in those cases, the Defense Department must consider whether or not the system integrator, as the prime contractor, could also be competing with its own sub-contracted AI developers in any way. If the system integrator is also in the AI development business and is controlling the data engineering and developing the user interface, there is a strong incentive for the system integrator to favor its own algorithm developers instead of third-party subcontractors.
Defense Department AI acquisition requires a number of new entrants into the acquisition arena, forming a new National Security Innovation Base. Yet, they will still rely on elements of the defense industrial base that developed and continue to sustain many legacy weapons systems. AI acquisition is a team sport, requiring participation from both small startups and more established defense contractors. These new partnerships will be vital to drive AI development and implementation across the technology valley of death.
The Defense Department Needs to Move Faster
Many say that the Defense Department must move faster. I agree. But haste alone is not enough. Speed and scale for AI implementation requires a multi-layered foundation built on the right data, algorithms, and integration. It also needs a technical solution in the form of a shared development environment that enables swift collaboration and iteration. And it requires a practical, deliberate approach to acquisition that invites collaboration from the world’s leading AI developers.
AI is forcing a paradoxical shift in how we view weapon systems and industrial acquisition. Performance, speed, and agility are today’s currency both in warfare and in business. Unfortunately, the current military industrial complex is consistently trading military advantage for the familiarity of hardware and weapon systems with big price tags and exorbitant downstream sustainment costs. What feels like nostalgia and comfort continues to lull the United States into greater disadvantages against competitors like China and Russia, who are headlong into new technology frontiers. Implementing AI will keep the United States technologically ahead of its adversaries but it will also serve as a forcing function to shift defense acquisition to be more rapid, affordable, modular, and operationally relevant.
Absent major structural changes to how the Defense Department acquires, fields, and sustains emerging and disruptive technologies like AI, the military risks losing its edge over China and Russia. Losing the edge could make the difference between the United States winning or losing a major-power conflict. The United States maintains certain inherent advantages in emerging technology, but it cannot afford to slow down. China, in particular, is invested in competition with the United States over the future of AI, with a direct bearing on its ability to project military power. The Pentagon needs to thoroughly embrace a new environment, blending the best of the Defense Department with the best of the commercial tech industry in a way designed to maintain the country’s technological competitive advantage for decades to come.
Maj. Matthew Cook is an active duty Air Force acquisition and intelligence officer currently stationed at the Joint Artificial Intelligence Center, Washington, D.C., as the military Assistant to the director. Before that role he spent his time in a variety of acquisition and intelligence roles ranging from acquiring intelligence systems to deploying as a targeting officer overseas in the war against ISIL. He is a graduate of the U.S. Air Force Academy (BS) and Tufts University (MS).
Image: Department of Defense