(updated on 8/11/24)
Exia Labs is an American defense technology company that specializes in AI-enhanced analysis, training, and intelligence tools. Our flagship product is AI Commander, an AI system that generates winning warfighting strategies that humans can't and simulates those strategies and accompanying tactics before troops execute them, saving time and potentially lives.
This idea is championed by some of most prominent defense innovators and former military leaders like Gilman Louie, CEO and co-founder of America's Frontier Fund and co-founder of In-Q-Tel, former Chairman of the Joint Chiefs of Staff Mark Milley, and former Google CEO and former Chairman of the Defense Innovation Board Eric Schmidt. Just last week, Milley and Schmidt published a piece in Foreign Affairs that highlighted the importance of AI Commander,
“The next major conflict will likely see the wholesale integration of AI into every aspect of military planning and execution. AI systems could, for instance, simulate different tactical and operational approaches thousands of times, drastically shortening the period between preparation and execution. The Chinese military has already created an AI commander that has supreme authority in large-scale virtual war games.”
We believe that only way to build AI Commander is through games and AI, but before we discuss why, let’s first discuss the role of strategy in America’s last two wars.
Background
The United States failed in Iraq and Afghanistan not because we didn’t have the best military, the best training, the best equipment, or the best technology. We failed because of intelligence failure and poor strategy.
That is not just our opinion, but also the findings of commissions established by the United States government. The Commission on the Intelligence Capabilities of the United States Regarding Weapons of Mass Destruction reported in 2005 that the “Intelligence Community was dead wrong in almost all of its pre-war judgments about Iraq's weapons of mass destruction” and that “this was a major intelligence failure.”
Examples of poor strategy in the context of the last two wars include:
Achieving the Opposite Goal of National Policy: The U.S. policy in the Middle East aimed to contain Iran, yet we strengthened Iran through the Iraq War.
Contested Military Strategy: There’s still a debate as to whether or not pursuing the military strategy of counterinsurgency (COIN) was the correct choice in Iraq and Afghanistan. Critics like retired Lieutenant General (LTG) Daniel Bolger argued that adopting COIN contributed to losing the war. Advocates like COIN architect John Nagl argued that the strategy was essential for fighting irregular warfare and that we didn’t stick with it long enough to see it succeed.
Strategic Withdrawal: Effective strategy involves not only planning for success but also being prepared to adapt and make difficult decisions, including withdrawal, when circumstances change or when the initial goals are no longer attainable. Regarding Afghanistan, LTG Bolger said, "The record to date shows that no senior officers argued for withdrawal. Instead, like Lee at Gettysburg, commander after commander, generals up and down the chain, kept right on going."
Certain services like the U.S. Army has recognized that it “and the U.S. military as a whole, faced serious strategic setbacks in places like Iraq and Afghanistan, and a coherent national strategy did not take shape to address those challenges.”
While we’ve been mired in two counterinsurgencies, the People’s Republic of China (PRC) has emerged as America’s greatest geopolitical competitor. US Ambassador to China Nicholas Burns calls the US - China relationship, “the most competitive, the most important, and most dangerous relationship in the world.”
The PRC has been preparing to challenge American hegemony for decades. They have demonstrated their long-term strategic thinking through initiatives like:
The Belt and Road Initiative: A global development strategy aimed at enhancing regional connectivity and economic integration by financing and constructing transportation, energy, and telecommunications infrastructure, linking these regions with China through infrastructure and investment projects across Asia, Europe, and Africa.
State-Directed Strategic Capital: The PRC sees strategic capital as a means to achieve its broader goals of economic dominance, technological supremacy, and geopolitical influence while maintaining the Party's control over key sectors of the economy and society. China uses public and private capital to subsidize and incentivize investment in areas vital to their holistic view of national security.
Today, China is the #1 global producer of drones, ships, and rare earth elements, which are vital to both consumer and military technologies.
The largest Chinese drone manufacturer controls 70% of the global drone market share. Chinese drones command 90% of the US consumer drone market, 70% of the enterprise market, and 92% of the state and local first responder market.
China builds 46.59% of ships measured by tonnage per year. The US produces 0.13%.
China produces 60% percent of global rare earths, but processes nearly 90%.
21st-Century Space Race: The US is currently the leading superpower in space, but that lead is being challenged by China. Just last month, China landed an uncrewed spacecraft on the far side of the moon to “retrieve the world's first rock and soil samples from the dark lunar hemisphere.” Meanwhile, the People's Liberation Army (PLA) owns half of the world’s space based Intelligence, Surveillance and Reconnaissance (ISR) satellites.1 What strategies can the US devise to maintain its lead in space?
Fabius
Introducing Fabius, a strategy simulation platform. It will operate like a live-service game that updates with insights from every conflict, regardless of our involvement.
Named after Quintus Fabius Maximus Verrucosus, the Roman statesman and general renowned as the Shield of Rome, famous for utilizing the Fabian strategy. He prevented Hannibal from defeating Rome by employing an unconventional, unpopular, and distinctly un-Roman strategy: delaying the enemy. Even though it was contentious, this strategy proved to be exactly what Rome needed at the time. Fabius will present all feasible strategies, whether they are unconventional, unpopular, or unpredictable.
Fabius is designed to navigate the complexities of modern warfare and adapt to the rapid pace of battlefield innovation. It will include every military domain (space, cyberspace, air, land, sea), every instrument of national power (diplomatic, informational, military, economic), and includes emerging weapons systems (hypersonic, directed energy, autonomous drones, etc.). Custom scenarios can be generated using any combination of these elements.
The ideal version of this would be a perfect real-life simulation, similar to advanced simulations in Ender’s Game Battle School. Since that technology doesn't exist yet, we'll use modern game design techniques to translate real-life scenarios into game mechanics.
What we have discussed so far isn’t enough to reach our goal. We need an AI to search through the state-space of the scenarios that Fabius produces to identify all possible strategies. Utilizing Reinforcement Learning (RL) techniques, we will train policies that know how to play within the scenario. These policies can learn, adapt, and develop strategies to maximize the reward.
Games and AI were made for each other. The progress of AI has been often measured by its ability to beat human champions of games. Early milestones included beating Chess and Go champions. More recently, AI has been able to beat individuals and teams of players of real-time multiplayer games like Starcraft and Dota 2. Inspired by these achievements, the PRC developed CASIA Prophet 1.0:
“The CASIA Prophet 1.0 system developed by the Institute of Automation of the Chinese Academy of Sciences defeated eight champion-level players from the military and civilian sectors with a 7:1 victory, using algorithms for mission task analysis, terrain analysis, force comparison, opponent behavior estimation, combat deployment, situation cognition and prediction, combat decision-making, and automatic generation of combat plans, achieving simultaneous improvement in confrontation and transfer capabilities.”2
In environments where there isn’t a single win condition like Chess or Go, RL can show many possible strategies. For example, among equal strategies with equal rewards, China may choose a strategy of investment while Russia may choose a strategy of annexing neighbors. That choice comes down to things like culture, leadership, political systems, historical actions and trends, etc. Intention is needed to understand how the red team might play.
Red
“It starts with the enemy. Paragraph 1 is paragraph 1 for a reason…
understanding the threat enables how we operate, understanding our enemy capabilities,
that can be applied to how we can fight as an Army.”
General Gary M. Brito, Commanding General, United States Army TRADOC
The Department of Defense and Intelligence Community are sitting on a “Scrooge McDuck” level of data. LLMs are a key technology that can help us unlock this data to understand enemy intention.
Our approach is to utilize a Retrieval-Augmented Generation (RAG) app to start, which we’re calling Red. Like many RAG apps, Red enhances the capabilities of an LLM by first retrieving relevant information from our curated dataset and then generating responses. Unlike regular LLMs that rely solely on their training data, Red ensures more accurate and up-to-date answers by combining retrieval with generation, providing well-cited responses. We’re starting off by testing Red on our database of public Chinese military strategy. Red will begin to use various DoD datasets once we gain access. Our ultimate goal is to develop domain-specific models.
There’s one last piece of the puzzle. While Fabius is searching the state-space for all possible strategies, and Red is understanding intention, we will combine these two into what we call Red Team AI.
Red Team AI is a policy that is goal-aligned to the intention of a red team opponent. To do this, we intend to research and prototype a novel usage of RL through AI feedback (RLAIF) that augments the RL loop and utilizes Red to do the goal-orienting. This can help tune the RL policies to choose strategies that are more in line with their intentions.
What’s Next
So that was a lot. To summarize, here’s our high-level plan:
Fabius (Strategy Simulator)
Digitize select wargames into simulation libraries
Train RL policies that can play these simulations
Integrate more domains, more instruments of national power, and more emerging weapons systems into Fabius
Develop ability to create strategy simulations within Fabius
Red (Red Team Insights)
Build RAG app focused on single country
Develop transfer learning
Build a domain-specific model
Red Team AI (Simulate Strategy as an Opponent)
Research RLAIF techniques to combine Fabius and Red
Develop Red Team AI using techniques learned
Advancing RL techniques, researching RLAIF integrations, developing custom domain models, doing game design for wargames, we need help on all of it. Please reach out!
Space Capital: The Race for Space Superiority Report
K. Huang, J. Xin, J. Zhang, et al., “Intelligent Technologies of Human-Computer Gaming," SCIENTIA SINICA Informationis, 2020, Vol.4, No.50, Pp:540-550