Si vis pacem, simula bellum

Today, we begin a series on AI-driven simulation and defense spanning posts, blog articles, and related formats. And, frankly, it feels a bit déjà vu. At Scensei, we’ve been working in this space for more than fifteen years, going back to our time at the Center for Computational Social Science at George Mason University. Back then, under Donald Rumsfeld and the era of Department of Defense-led human, social, cultural, and behavioral modeling the atmosphere felt thick with mission relevance, technological advantage, and willingness to succeed. If any of this sounds familiar today, it should. 

The rhetoric has changed; the logic hasn’t. So rather than pretending we know where this is all headed, we’ll do something more modest: prompt some reflections on how AI-driven simulation is reshaping warfare – and why that matters. Because one thing is certain: si vis pacem, para bellum. Or rather: si vis pacem, simula bellum. 

We’ll start with what we call simulation as a service. At its simplest, simulation as a service means outsourcing data generation via simulation to enable certain tasks, say to create insights, facilitate training, and provide decision support. 

A client brings a problem – force design, operational risk, procurement trade-offs – and an external team builds and runs simulations to generate insights. We at Scensei built OPTIMA to address exactly these kinds of questions: Is it smarter to scale cheap drone production or invest in fewer, high-end systems? Will speed matter more than armor? What breaks first under pressure? Indeed, discussions with serving NATO officers suggest that wargaming is presently experiencing a revival. 

Another task is to use simulation as a training environment, for humans as well as for AI-driven systems. Yet another task for simulations is to serve as hunch engines that externalize intuition, explore plausible future scenarios and support judgment. A further objective is to leverage multi-agent simulations to support, enhance and control battle management systems. 

The point here is not to rebrand simulation as a service as agentic system. The point is that simulation, the machine itself, is changing – and with it simulation as a service. Simulations are already becoming not only more expressive but more precise and accurate. Eventually we will be beamed up to the holodeck. 

The long-standing gap between human and machine cognition is narrowing. Sparring with simulations will become routine – and eventually, mandatory. Aviation figured this out decades ago: no aircraft flies before it’s been simulated. The same logic is creeping into military operations. Soon, no J9 will sign off on a mission that hasn’t been run, re-run, stress-tested, and broken, by its loyal wingman “sim” first. 

Finally, learning and agency. Our multi-agent system OPTIMA already is an adaptive agentic system that can be nested in even larger agentic systems that are capable of recreating themselves and make decisions. Call it autopoiesis, if you like. The result is a system that doesn’t just represent the cognitive battlefield but increasingly re-constitutes it. 

In OPTIMA’s case, this means the AI doesn’t just initialize scenarios – it creates them. It assigns behaviors, evolves agents, and generates outcomes that shape the cognitive battlespace. The old distinctions – constructive versus virtual versus live; analytics versus training versus operational – start to blur. Eventually, they disappear. “Shall we play a game?” 

There is a certain cruelty baked into evolutionary processes, a fact best confronted with a sober kind of realism. Systems that survive do so because they adapt. Simulation as a service, if it is to remain a reliable instrument rather than a decorative one, must internalize that logic. 

For simulation as a service to get better, it must submit to standards. Explainable AI, for example, is one of them. Simply bolting large language models onto simulations and calling the result “intelligent” won’t suffice. Textual fluency is not cognition, and eloquent output is not understanding. Alternative AI architectures – ones that more closely resemble human cognition – are already beginning to challenge the prevailing orthodoxy. 

The delivery model must evolve as well. The idea that meaningful simulation requires bespoke federations of models running on opaque infrastructure buried deep inside institutional basements belongs to another era. While kids plug in game consoles and NVIDIA ships increasingly powerful small edge AI modules, defense organizations still act as if speed were negotiable left of boom. It isn’t. 

At the same time, humans should resist the temptation to outsource judgment without due consideration. Experience-based decision-making remains a comparative advantage we (still) possess. Simulation (as a service) is not a conveyor belt that brute-forces its way through the decision space. It is, and remains, partly an art. Exhaustive search is not only inelegant; it’s inefficient. Principles like satisficing still matter. So does Ockham’s razor. Or, to put it in more familiar military terms: simplicity of action. 

In the end, we want AI’s generative force to augment our own. Imagine, test, adapt to win today’s and tomorrow’s battles. The technology is already here. What’s missing is urgency.