What Is the New Strategy?
On January 9th, the U.S. Department of Defense (DoD) released its Artificial Intelligence Acceleration Strategy. The Strategy promotes a familiar outcome: for the U.S. military to become “AI-first.”
According to the Secretary of Defense Pete Hegseth’s memorandum on the Strategy, this AI-first status is to be achieved through four, broad aims:
- Incentivizing internal DoD experimentation with AI models.
- Identifying and eliminating bureaucratic obstacles in the way of model integration.
- Focusing the U.S.’s military investment to shore up the U.S.’s “asymmetric advantages” in areas including AI computing, model innovation, entrepreneurial dynamism, capital markets, and operational data.
- The initiation of Pace-Setting Projects, or PSPs, that will “serve as tangible, outcome-oriented vehicles for rapidly completing” AI-related scaffolding of infrastructure, data, models, policies, and talent.
What Are the Mission Areas of the PSPs?
The initial set of seven PSPs will coalesce around seven specific mission areas grouped into three categories:
Warfighting
- Swarm Forge: a “competitive mechanism” between U.S. forces and elite tech innovators.
- Agent Network: Development and experimentation with AI Agents for battle management and decision-support.
- Ender’s Foundry: AI-enabled simulation capabilities, including simulation-development (sim-dev) and simulation-operations (sim-ops) feedback loops.
Intelligence
- Open Arsenal: Technical Intelligence (TechINT)-to-capability development pipeline acceleration.
- Project Grant: Transforming deterrence from “static postures” to “dynamic pressure with interpretable results.”
Enterprise
- mil: Democratizing AI experimentation at all classification levels.
- Enterprise Agents: Building a “playbook” for AI agent development and deployment for enterprise workflows.
How Will PSP Progress be Evaluated?
PSPs will have program leaders. Progress is to be “demonstrated” by these program leaders on a monthly basis to the Deputy Secretary of War and Under Secretary of War for Research and Engineering in each mission area listed above. Initial demonstrations occur six months from the memorandum (i.e., July 2026).
The Chief Digital and Artificial Intelligence Office (CDAO) will be responsible for “enabling” the initial set of seven PSPs. (The CDAO likewise plays a critical role throughout the memorandum’s various directives.)
The memorandum emphasizes that progress must keep pace with these PSPs in projects under the DoD’s six Critical Technology Areas.
Each military department, combatant command, and defense agency and field activity is tasked with identifying at least three projects for prioritization in order to “fast-follow” the PSPs within 30 days.
(Where) Will DoD Resources be Allocated?
The memorandum instructs the CDAO to devote efforts and resources to eleven areas for the purpose of firming each up. An inexhaustive list of these areas includes: AI compute (e.g., infrastructure, like data centers), access to data across the DoD, and talent acquisition and engagement.
Particular attention should be paid by defense contractors to the memorandum’s emphasis on Modular Open Systems Architecture (MOSA). It notes that system architectures “must enable component replacement at commercial velocity to maintain overmatch” – a relatively innocuous continuation of existing MOSA efforts by the DoD.
The memorandum also notes that the need for “modular interfaces and associated documentation” is to be attained “without prime contractor support” – a qualifier likely of interest to some contractors.
Strikingly, the memorandum directs the CDAO to ensure that AI vendors be able to deploy the “latest models” within 30 days of those models’ public release. Emphasizing its importance: “This shall be a primary procurement criterion for future model acquisition.”
What’s New, and What’s Not
As Defense One’s Patrick Tucker notes, this AI Acceleration Strategy is the DoD’s third of its kind in as little as four years. Perhaps marking a difference this time around, Tucker also notes that the release of the Strategy and the DoD’s welcoming of Elon Musk’s AI chatbot, Grok, comes as the latter undergoes scrutiny by European allies and other state actors.
The memorandum’s directive to the CDAO to make available vendors’ “latest models” within 30 days of their public release as a “primary procurement criterion” is striking, too, and represents a sharp break from the days when it took Microsoft 18 months to make GPT-4 available in the Azure Government Top Secret cloud (and air-gapped at that).
The PSPs are new as well, though it is unclear how different these will be in practice when compared to a litany of other DoD efforts to push the Department, as it were, into a new era.
Still, continuities are detectable from the past four years.
For one thing, the capabilities of certain classes of AI models have continued to improve in the interim in some respects, a fact reflected in the memorandum’s bullish tenor on the capabilities gleaned from AI. Likewise, the public sector’s expectation that private sector AI basic research & development will continue apace is a fascinating cultural aspect of AI policy – a technology that, whatever its many shortcomings, has primarily been developed by commercial actors.
Moreover, the memorandum’s call for reducing bureaucratic barriers to accessing data for models – whether for training new models or for prompting chatbots with DoD data – is a familiar one, with efforts ongoing. The models that are the apparent subject of the memorandum are fundamentally reliant on data, and de-classification or otherwise arranging for classified data to be shared with them is a recurring DoD theme.
The emphasis on MOSA is, likewise, a continuation of efforts to widen the pool of available suppliers for DoD acquisition needs, including the ability to swap out components in a “plug-and-play” manner.
Possible Redundancies
Several goals outlined in the memorandum – including internal experimentation with AI models, agent development for battle management and decision-support, and AI-enabled simulation capabilities – are already being pursued across the DoD in various forms and could be points of redundancy.
Indeed, seemingly unrelated efforts like the U.S. Air Forces’ and U.S. Navy’s Collaborative Combat Aircraft (CCA) efforts, the testing of Leidos’ AlphaMosaic battle management system by the U.S. Air Force, and new means of transferring simulated autonomy to real-world deployments at the Defense Advanced Research Projects Agency (DARPA) are each underway (and have been for some time).
More than this, several of the directives are formulated at a high level of generality, sufficient to make their real-world implementation a matter of individual personnel decision-making in the course of their workflows. This applies, in particular, to internal DoD experimentation with chatbots that can interact with DoD data (a vague and somewhat unspecified instruction).
Interestingly, this mirrors the deployment of Enterprise Large Language Model (LLM) applications in the commercial domain, where uses are sometimes directed, and sometimes not, though the manners in which the models are used depends in large part on the employee’s preferences. It remains to be seen how this shakes out within the DoD, and likewise remains to be seen – for both defense and commercial domains – how and to what extent this impacts overall productivity.
Following vs. Leading
The memorandum is notable in that it embraces the private sector-to-public sector dynamic, in which the former develops the fundamental techniques driving applications taken to be of revolutionary significance (like chatbots or agents) and the latter adopts them without haste.
As alluded to above, this marks something of a disjuncture between AI and other powerful technologies, where many of the latter were developed largely or primarily within the confines of U.S. military research & development (e.g., GPS, drone technology). The memorandum seemingly embraces and extends this dynamic, adopting a stance in which the DoD is to shore up the peripheries of privately developed technologies (e.g., constructing AI and autonomous system evaluation standards) and identify use-cases for them within the DoD (e.g., internal experimentation, AI agents for enterprise workflows).
Space prohibits a more extensive commentary, but as noted elsewhere, this ‘following’ rather than ‘leading’ stance of U.S. government AI policy intersects with its bumpy rollout across commercial sectors.
Model Objectivity and Evaluation Standards
Secretary Hegseth’s memorandum includes a section – notably, at the very end – that ‘clarifies’ the meaning of “Responsible AI.” Specifically, it includes an opposition to “Diversity, Equity, and Inclusion and social ideology” in AI models, including of “tuning” of models that is “ideological” in nature, such that it interferes with their ability to provide objectively truthful response to user prompts.”
Along with this, the memorandum instructs the CDAO to “establish benchmarks for model objectivity as a primary procurement criterion within 90 days…”
To be sure, that this was inserted at the end of the Defense Secretary’s memorandum is an indication of its lesser importance relative to the primary directives set forth for the PSPs.
Moreover, regular observers will recognize this as a continuation of a broader political dispute over the proper development of generative AI models, a theme which is present within the second Trump administration’s Executive Orders on AI and, generally, bungled commercial deployment by companies like Google.
All that said, the technical reality is more limited than it appears: building AI models that “provide objectively truthful response to user prompts” is an extremely fraught endeavor, particularly when applied to open-ended domains.
The models that are implicitly the subject of this inclusion are so-called generative AI models. These are models that generate new outputs that bear resemblance to their training data. The model during training learns how each data point – say, each human-generated sentence – relates to one another statistically. The improvements in generative AI models over time are primarily an improvement in their surface-level alignment with human judgments: most humans accept that the Earth revolves around the Sun and, lo and behold, this is ‘discovered’ by the model during training (should the internet be populated with the reverse idea, the model would ‘discover’ geocentricism). So, too, for even the most advanced examples, including mathematical abilities.
There is, then, no means of constructing models that adhere to any standard other than that which they are held to by humans.
This presents an unwieldiness regarding model procurement: AI vendors will be expected to meet standards that can, at best, only be approximated rather than met with some specified level of precision. Contractors should be aware of the risks involved with tailoring generative models to such ends, as their steerability is a fraught matter in sufficiently open-ended domains of application. Use-specific (or category of use-specific) may be the most fruitful opportunities, so far as generative models are concerned.
Similarly, should DoD data be made available for certain types of model training, this will go a significant way towards sharpening its likely outputs and decrease the burden on vendors to meet the CDAO’s standards.
Of course, it is plausible that the CDAO’s eventual benchmarks will be grounded in these models’ technical realities, and an expectation of some level of unwieldiness will be baked in. How this intersects with Secretary Hegseth’s (separate) directive to the CDAO to develop “AI system usage and mission impact metrics for evaluating the success of these AI acceleration efforts” is an open question, and one to be monitored.
Finally, note that, although generative models of the type popularized by LLMs are the clear target of the Strategy, there does not appear to be an explicit mandate to underpin the various directives with this class of models. In principle, a variety of techniques could underpin “agents” and the like.
Vincent Carchidi has a background in defense and policy analysis, specializing in critical and emerging technologies. He is currently a Defense Industry Analyst with Forecast International. He also maintains a background in cognitive science, with an interest in artificial intelligence.
