What Does the Department of War’s AI Strategy Really Say?
The memo says the risks of not moving fast enough outweigh the risks of imperfect alignment. That's the one line in this whole document that should give us all pause.
In this episode · 13 chapters
- Six weeks after
- More lethal and efficient
- The Anthropic history
- Race, and speed
- The agent network
- What the NGA director claimed
- Who is the human in the loop?
- Intel into weapons in hours
- How much data
- Accelerate
- The two most important paragraphs
- The line that should give you pause
- Out with utopian idealism
And buried somewhere in the bureaucratic grandiose, chest beating language of this memorandum, is a very serious conversation about who controls AI.
Mark and Jeremy read Artificial Intelligence Strategy for the Department of War, signed by Pete Hegseth and issued on 9 January 2026. They weigh its pace-setting projects, from Swarm Forge to an agent network for battle management and kill chain execution.
They ask who the human in the loop would be, why turning intel into weapons in hours worries Jeremy, what a 30-day data deadline would demand of any storage architect, and whether AI makes war more likely or less likely.
Artificial Intelligence Strategy for the Department of War, issued on 9 January 2026 and signed, Mark Fielding says, by Secretary of War Pete Hegseth, sets out a speed-first plan for an AI-first US military. Reading it on the Thinking On Paper podcast, he and Jeremy Gilbertson single out its line that “the risks of not moving fast enough outweigh the risks of imperfect alignment” and its call for “models free from usage policy constraints”.
What does the Department of War's AI strategy say?
The memo, issued on 9 January 2026, sets out a speed-first plan for an AI-first US military. It argues that the risk of moving too slowly outweighs the risk of imperfect alignment.
What does the AI strategy memo ask the US military to become?
The memo directs the Department of War to become an “AI first war fighting force across all components”, Mark reads, and it opens by citing President Trump's Executive Order 14179 on sustaining America's global AI dominance. Jeremy warns that the memo rewards AI-first approaches even where AI may not be the best tool.
What are the memo's pace-setting projects?
The memo sets up seven pace-setting projects (PSPs), which Mark reads as projects “that will demonstrate the accelerated pace of execution” across war fighting, intelligence and enterprise. They include Swarm Forge, Agent Network, Ender's Foundry and Open Arsenal, whose stated aim is turning intel into weapons “in hours, not years”.
What is Agent Network and does a human review its decisions?
Agent Network aims to use AI agents “from campaign planning to kill chain execution”. Mark cites the Project Maven Wikipedia page: as of September 2025, the director of the NGA claimed Maven would begin transmitting 100% machine-generated intelligence to combatant commanders by June 2026. Jeremy says he sees no language in the memo on where a human must review those decisions.
What data deadline does the memo set for every department?
Mark reads that the memo gives every department 30 days to publish federated data catalogues exposing systems and data across all classification levels. Jeremy, who says he has sat in data centre planning meetings, says data architects would call a 30-day re-architecture impossible and risky, and questions how much data these departments should access at all.
How is Anthropic linked to the Department of War AI memo?
The hosts link the memo to Anthropic, which Jeremy says was classed as a supply chain risk after, in Mark's words, it sought terms barring use of its technology for mass US surveillance and full automation of the kill chain. Jeremy then reads the memo's requirement that the department use “models free from usage policy constraints”.
Where Jeremy disagrees with the memo
Jeremy rejects the memo's “speed wins” doctrine, arguing that imperfect alignment could mean people dying from AI that has not been properly safety tested, and that the Department of Defense should not adopt Silicon Valley's move-fast-and-break-things ethos. He also doubts the need to turn intelligence into a weapon within an hour.
Links from this episode
- Artificial Intelligence Strategy for the Department of War: Memorandum, 9 January 2026 (Defense Management Institute, PDF)
- Executive Order 14179: Removing Barriers to American Leadership in Artificial Intelligence (Federal Register, January 2025)
- DOW To Become an AI-First Warfighting Force, naming the seven pace-setting projects (AFCEA Signal, January 2026)
Questions
What does the Department of War's AI strategy say?
The memo, issued on 9 January 2026, sets out a speed-first plan for an AI-first US military. It argues that the risk of moving too slowly outweighs the risk of imperfect alignment.
Who signed the Department of War AI strategy?
Artificial Intelligence Strategy for the Department of War, issued on 9 January 2026 and signed, Mark Fielding says, by Secretary of War Pete Hegseth, sets out a speed-first plan for an AI-first US military.
What lines in the AI strategy concern Mark and Jeremy?
Reading it on the Thinking On Paper podcast, he and Jeremy Gilbertson single out its line that “the risks of not moving fast enough outweigh the risks of imperfect alignment” and its call for “models free from usage policy constraints”.
What questions do Mark and Jeremy ask about the memo?
They ask who the human in the loop would be, why turning intel into weapons in hours worries Jeremy, what a 30-day data deadline would demand of any storage architect, and whether AI makes war more likely or less likely.
Transcript
Mark Fielding Disruptors and curious minds welcome to the show. What you're about to hear is not fake news. It is not AI generated. It is a memorandum published by the White House signed by Pete Hegseth, the Secretary of War. Issued by the Department of War on January the 9th, 2026. The memo is called Artificial Intelligence Strategy for the Department of War, America's Military AI Dominance.
Six weeks after
Mark Fielding And we're reading this because six weeks after it was signed, the U S is at war with Iran and AI is identifying targets. And buried somewhere in the bureaucratic grandiose, chest beating language of this memorandum, is a very serious conversation about who controls AI.
And does AI make war more likely, or less likely? The memo opens, President Trump makes clear in executive order 14179, it is the policy of the United States to sustain and enhance America's global artificial intelligence dominance in order to promote human flourishing, economic competitiveness and national security. And then AI
AI enabled warfare and AI enabled capability development will redefine the character of military affairs over the next decade. This transformation is a race fueled by the accelerating pace of commercial AI innovation coming out of America's private sector.
Races have winners and losers. Who's in the race, Jeremy? Who is in the seat two of this race?
Jeremy Gilbertson talked about this a lot, Mark, whenever something becomes right, well, whenever something becomes a race, there's an urgency to And we forget about a lot of the rules and human necessities outside of that. So yes, we're in a race. The race is arguably, against China. I think it's is it listed in this? Does it spell it out specifically in this?
Mark Fielding Just say China. Just say China.
out. No, it's not. So this war, this race for an AI led military, How is America going to win this race? This memo outlines the strategy to win the race. The United States military must build on its lead. So it's already winning the race over our adversaries in integrating this technology established during President Trump's first term.
More lethal and efficient
Mark Fielding to our war fighters more lethal and efficient. To this end, aligned with America's AI action plan, I direct the Department of War to accelerate America's military AI dominance by becoming an AI first war fighting force across all components from front to The department will achieve this objective by... So yeah, the language is very...
aggressive, it's very attack, attack, attack. This is the art of war. 2026 style. Let's go through the strategy. So objectives. What are the objectives of this? Number one, unleashing experimentation with America's leading AI models department wide.
and rewarding AI first reconceptions of legacy approaches. So reconceptions of legacy approaches.
rewriting the rule book, the rules of engagement, throwing out what existed before and redesigning that with AI in mind, how would it look?
Jeremy Gilbertson there's certain approaches I think that could be reasonable in this. The safety of people involved, the of the military, safety of the soldiers. You know, how could that, how could we improve kind of a good thing. The experimentation piece of this challenging, right? Because this again, this is jet fuel. We don't even fully understand it. Like these AI models, we don't know how they produce what they produce, why they produce what they produce. We don't even know much about how they're as individuals. ⁓ does hallucination play into the experimentation as well? If you're experimenting with life and death, I mean, it's like, it's all good and well used having an hallucination at work, but having an hallucination on the battlefield, the consequences are life and death.
Jeremy Gilbertson Yeah. And then, you know, incentives, incentives usually direct people to what you're asking them to achieve. Right. And you're asking AI first approaches, even when AI first may not be the best approach. So unleashing, God, that's a strong damn word. Right. Like the language in this is the language in this is, is pretty big. then, you know, this next one, aggressively identifying and eliminating bureaucratic barriers to deeper integration. So this is really interesting. This ties into some stuff that we talk about a lot.
Jeremy Gilbertson Think about democracy just in general. it's inefficient for good reason. it takes a longer time to have things happen. And some bureaucratic barriers may need to be there, especially if there are decisions, whether to pull the trigger and automate or automate the pulling of the trigger for, for weapons, which I think we are, we're, largely referring to in this, in this
Mark Fielding Yep. And you have vestiges of legacy information technology and modes of warfare. So they're talking about, again, out with the old in with the new, and then focusing our investment to leverage America's core asymmetric advantages in AI computing, model innovation, entrepreneurial dynamism, capital markets and combat proven operational data from two decades of military and intelligence operations that no other military can replicate. Okay, well, we have
Anthropic or we had Anthropic, have OpenAI, we have Grok, we have Anduril, we have Palantir, we have the America's finest entrepreneurs.
The Anthropic history
Jeremy Gilbertson so remember Anthropic, just some history guys, Anthropic Dario He left OpenAI he largely disagreed with the safety protocols related to releasing products and things in the market. He started Anthropic to do it a little bit differently. you know, currently in the news,
They're suing the US government, I think, because they were put ⁓ into a risk profile. what was the official categorization mark? Do you have that handy? They're an official supply chain risk, I think is what they were called.
Mark Fielding risk. Yeah. Because they wanted to have a little bit of small print in their agreement whereby the tech couldn't be used for mass US surveillance and couldn't be used for full automization of the kill chain. They're not light demands. Yeah. At least consider.
Jeremy Gilbertson Probably not bad things to consider. Like, I mean, that those seem pretty damn important.
Mark Fielding Those seem pretty damn important. And you got
Jeremy Gilbertson That's right, that's right. let's continue.
Race, and speed
Mark Fielding So we're talking about race, we're talking about speed. So now the memorandum moves into the PSPs, the pace setting the projects that will demonstrate the accelerated pace of execution, focus and ethos we need to stay ahead. The PSPs will also serve as tangible outcome orientated vehicles for rapidly completing our build out of the foundational AI enablers, infrastructure data models, policies and talent.
which is an interesting one. So there's seven PSPs. We start with again, the race, the acceleration approach. So these are outlined war fighting, intelligence and enterprise. So we'll start with war fighting more bellicose language, Jeremy. swarm forge: competitive mechanism to iteratively discover, test and scale novel ways of fighting with and against AI enabled capabilities, combining America's elite war fighting units with elite technology innovators. Here again, we have big tech and big government and big bombs and big drones and big swarms of drones. This has been acted out in Iran as we speak, is it not?
Jeremy Gilbertson That it is, that it is.
The agent network
Mark Fielding Agent network. We're going to unleash AI agent development and experimentation for AI enabled battle management and decision support from campaign planning to kill chain execution. So if I just want to read you something from the project Maven Wikipedia page.
Jeremy Gilbertson Is this the as of September 2025?
What the NGA director claimed
Mark Fielding Yeah, as of September 2025, the director of the NGA claimed that by June 2026, Maven will begin to transmit 100 % machine generated intelligence to combatant commanders using LLM technology. That doesn't mean that they're doing that on the ground now in a live war scenario. It means that that is possible.
Jeremy Gilbertson 100 % machine generated intelligence. the key here, I think, is 100 % machine generated intelligence. What is the review process and protocol to...
take that through the human lens or are we limiting that? So technology scales what we do, right? So do you have now one person managing, I don't know, four or five theaters per se or areas of battle and then they're relying on machine-generated intelligence to make decisions still?
Or is this just going to automate? Is there going to be a handoff process? And if so, where is the handoff that this decision needs to be generated or reviewed by a human? This decision does not. I don't see any language in here that really talks much about that.
Who is the human in the loop?
Mark Fielding And who is the human if there is the human in the loop? Who decides who that human is? Well, from that language, it means I understand it as they don't doesn't need, or there won't need to be a human in the loop. If there needs to be a human in the loop, that will be a decision made by the Department of War to keep a human in the loop, even though from a technological perspective, it might not be necessary.
Jeremy Gilbertson So you mentioned Project Maven, Anduril, Palmer Luckey, So Palmer Luckey was the guy that came up with Oculus, right? Was that his first piece? it to Meta this is marketing or whether this is, him pitching his I hear is his idea is to like use all of this to avoid
Mark Fielding Yeah, he sold it to Meta for a billion.
Jeremy Gilbertson wars and conflicts kind of before they happen, whether I believe that or not. I don't know. I haven't researched it enough, but this certainly, this memo doesn't talk about the avoidance of conflict. talks about putting jet fuel on the ability to generate
Mark Fielding Yep. So Ender's Foundry. haven't read the book that this is related to. accelerating AI enabled simulation capabilities and SIM dev and SIM ops feedback loops to ensure we stay ahead of AI enabled adversaries. There we are again. It's very opaque looked up what the book The Ender's Game is, but it's about a child unknowingly fighting a real war through what he thinks is a simulation. And it turns out it isn't a simulation. If you've read that anybody, please let us know.
Jeremy Gilbertson I'm not sure that's a reference I want to see happening in an official strategy document for the, I'm going to still call it the Department of Defense. don't think.
call it the Department of War. Intelligence, open arsenal, accelerating the tech int.
Intel into weapons in hours
Mark Fielding turning Intel into weapons in hours, not years, turning Intel into weapons in hours, not years. have your area 51 modern warfare equivalent. You find this strange foreign Chinese technology on the war field in the straits of wherever you might find it.
determines what it can and can't do, suck it into the AI machine and use it. And you can do that in a day, not in years, which it would take now.
Jeremy Gilbertson Sometimes, sometimes days are good. Sometimes having more days are good. I don't know that, I don't know that we need to have something turn Intel into a weapon in an hour. already as humans kind of limited in our ability to make great decisions on behalf of the greater good. think it's kind of like, it's kind of like that idea when you get really pissed off, Mark.
Jeremy Gilbertson And your limbic brain is activated and you get into this tornado of just like, wow. And you're chasing down this mad. What do you, what do you, what do good people tell you to do? Take a breath, take a walk, step away from stuff for a minute. We need to be able to step away from shit for a minute before we turn it into something unstoppable. That's dude. That's yeah.
Mark Fielding to turn into a weapon. Yeah, you're right.
There does seem to be everything in this is written with the optimistic hope that everything is correct. Although that's not actually true, is it? Because later on, we should probably get to that. They say that there will be, we'll make some mistakes, but collateral damage is part of war and speed is more important than errors.
So at least they say there's no breathing space and that might cause some problems. But what the hell? What the hell fire?
Jeremy Gilbertson Well, speaking of that too, just, just jumping down real wartime approach to blockers. think this is listed down here somewhere and there is something called an ATO and authorization to operate. this kind of helps remove some of these, some of these blockers. And this is specific to data So this talks about what,
Amodei probably highlighted and said, Hey, you know, there's a lot of access to a lot of data we could use that could be used to create a surveillance state. I don't want to be a part of it. Here's a blocker that or here's a block remover that is put into this document saying, Hey, we could just use this vehicle to do whatever we want to do. Because it's a race, because we have to be because we have to win.
And listen, man, I'm not downplaying national defense. Like I'm not downplaying this stuff. That stuff's important, but man, is moving it quickly and maybe like recklessly, maybe a bit recklessly.
Mark Fielding And I think there's any maybe about it. if you read. okay, the middle part of the document, we talk about enterprise. talk about creating a website where all military personnel can access these models and practice and train and do what they want. enterprise agents, they call them. But then it gets to AI compute, data access. Okay. So all of the departments, have to establish, maintain and update federated data catalogues exposing their systems interfaces, data assets and access mechanisms across all classification levels as mandated by the department's 2021 memorandum creating data advantage. They have 30 days to do that within the day of this memorandum. so that deadline's passed. They talk about talent.
so first of all, technologically, that is a very difficult thing to do from a data storage perspective, from a permissions, from an access, from a data hierarchy, like everything involved in that.
it? How do you know? they've got open AI on board, is it?
Jeremy Gilbertson No, but you still have to, you still have to look at your data archi, your storage architecture. And I'm not a storage architect. I'm not a data architect. I've been around a bunch of them. I've sat in meetings, planning data center programs with them at the table, teaching me, helping me understand their piece of the puzzle of that. And if we had one on today and we said, Hey, you have to look across your organization, re-architect everything in 30 days to support this memorandum. They would say no way. And if we did, it would, it would create tremendous amount of risk.
So the accelerated drive for this, again, we talk about ⁓ potentially reckless.
How much data
Mark Fielding much data is that? How many quadrillion terabytes of data is that? All of that information is mind blowing.
Jeremy Gilbertson Well, but also how much of that data should they have access to? Should all of these departments have access to? Smelling a little surveillance, Mark. Smelling a little surveillance.
Mark Fielding Well, yeah, all of it.
well, before we get into the mad rush, we have the talent, which I think is big tech in bed with big AI. There's a line in here which I want to read out. I believe the best American talent will see this accelerated posture of AI capability development adoption at the DOW and I expect each service and component to attract and retain this talent.
To that end, I direct use of special hiring and pay authorities department wide, as well as novel talent programs from the office of personnel and management and other partners to accelerate our pace of technical talent hiring into AI roles. seeing Facebook is paying 200 million for the best talent, how much? Because I don't know what the going rate is for an engineer.
at the DOW, but are they going to times a thousand that? Are they going to start paying the talent millions per year to get them on board? Is that what this means? Or are they willing to settle for the second best talent or the third best talent? Because the best talent, except for the the warmongers is going to Silicon Valley. how do they what does what does this mean?
Jeremy Gilbertson This statement tells us that they're committing to investing heavily in resources to hire really smart people. So that's all I really gather out of that.
Accelerate
Mark Fielding Okay, you'll like this language, Jeremy. We're on to acceleration expectations. This strategy will accelerate our advantage and we must implement it with the warrior ethos. As a non-American, Jeremy, could you explain to me the warrior ethos?
Jeremy Gilbertson I have no idea what that means, dude. I really don't. It's capitalized. maybe potentially, potentially within the branches of the military, maybe that's terminology within, within the army or within the Navy or something we don't have privy to. I don't know statements about technology and aligning with warrior ethos, you know, I get, yeah.
The two most important paragraphs
Mark Fielding two most important paragraphs in this memo for me, speed wins and the AI model parity. Speed wins. We must internalize that military AI is going to be a race for the foreseeable future and therefore speed wins. We must weaponize learning speed and measure and manage cycle time and adoption rates as decisive variables in the AI era. We must accept.
that the risks of not moving fast enough outweigh the risks of imperfect alignment. There it is.
The line that should give you pause
Jeremy Gilbertson That's the one line in this whole document that should give us all pause. We must accept the risks of not moving fast enough. Outweigh the risks of imperfect alignment. So imperfect alignment least aggressive language in this, probably on purpose. Imperfect alignment, you know what that could potentially mean? people dying because we use...
something that hasn't been properly vetted, properly safety tested, properly incorporated into a true understanding of what AI is capable of and where the limitations lie.
Silicon Valley philosophy of move fast and break shit. But this is like, I'm gonna call it again, the Department of Defense. we shouldn't move fast and break shit. In all cases.
Mark Fielding Yeah, break shit takes on a new meaning here.
memo goes on for another page. Some of the headers below wartime approach to blockers, competition, centralized planning, AI native war fighting, modular open architectures. And the final one, clarifying responsible AI at the Department of War.
Out with utopian idealism
Mark Fielding out with utopian idealism in with hard-nosed realism.
Jeremy Gilbertson I'm going to echo just something right out of that section out with utopian idealism and with hard nose realism. department must also utilize models free from usage policy constraints that may limit lawful military
So essentially anyone who doesn't play with the vision in this document is going to be called a supply chain risk, just like Anthropo.
Mark Fielding And we say this from the comfort of our podcast studio. We are not in anywhere remotely dangerous. We're not on the war front. We're not in Iran. We're not military personnel.
but I will end the show with the same question I started it with. Does AI make war more likely or less likely?
Jeremy Gilbertson We'll leave that for you to think about and read about.
Mark Fielding Until next time, be disruptive, stay curious.
Jeremy Gilbertson Keep thinking on paper.




