“What one Predator drone pilot described of his experience fighting in the Iraq war while never leaving Nevada: ‘You’re going to war for 12 hours, shooting weapons at targets, directing kills on enemy combatants. Then you get in the car and you drive home, and within 20 minutes you’re sitting at the dinner table talking to your kids about their homework.’”
-P.W. Singer
October 29, 2025 — Semi-autonomous weapons have arrived, and Ukraine is the test bed.
- Ukraine is burning through 10,000 drones a month. Russia is spending $100 m per month on drone production. Both are rapidly innovating — from first-person view (FPV – drones flown by remote pilots) to fiber optic drones (unspooled fiber optic cables provide the drone with a secure high-bandwidth connection).
- On the battlefield and against civilian targets, ~70% of the casualties in the war are currently from drone attacks. Drones, using zero-day tactics and systems disruption, are also conducting strategic attacks against national infrastructure and strategic military assets.
- In contrast, the US just announced its drone strategy. Its focus is on mass-producing drones and buying off-the-shelf systems in bulk (without any Chinese parts). The problem is that the US doesn’t have a domestic drone industry. China does. It outproduces the US a thousand to one, and the DoD desperately throwing a few billion at defense contractors won’t change that.
When facing a deficit like this, the best strategy isn’t to play catch-up; it’s to reinvent the game. Autonomous weapons provide that opportunity. Semi-autonomous weapons are merely a waypoint on the path to autonomous weapons, so catching up there shouldn’t be the objective. It’s time to vault to the endgame.
This first step is building a domestic commercial autonomy industry (see the report “American Autonomy” for more) in air, land, sea, and space. The next step, and likely the hardest, is inventing an autonomous weapon system from scratch. Let’s dig in.
Autonomous Weapons
Autonomous weapons can;
- Maneuver and navigate on their own, traversing complex environments without human support.
- Counter and penetrate defenses to find a suitable target and attack it successfully.
- Understand mission orders (orders designed to allow decentralized execution per the commander’s intent) provided verbally or through written instructions, and execute these mission orders while adapting to complex conditions.
From the description, it appears there will be a need for three AIs.
- An AI that will enable it to interact with and navigate a complex environment (this is already in motion for commercial applications).
- A military-specific AI designed to survive in a lethal environment, counter defenses, defend itself, and attack suitable targets it identifies. Think of this as muscle memory and reflexive decision-making based on training.
- An AI (modified LLM running a complete OODA loop) that can understand and execute mission orders. An AI that can be conversed with and trained. Think of this as orientation and decision-making based on experience. This perspective will be the most challenging part to develop. Autonomous commercial vehicles will eventually have a layer like this, but it won’t be as advanced (i.e., a trained Grok + a Tesla robotaxi).
Assumptions and Cognitive Filters
While it is impossible to provide specifics on how these AIs will develop, we can gain some insight into their direction and timing by making some assumptions. Here are some assumptions developed for AI (primarily for LLMs, but it also applies to other artificial cognitive capabilities).
- AI will increasingly become a commodity. There are multiple alternatives. Some are even open source. Soon, the older models in wide circulation will have, with external refinements, all of the capabilities required to make this possible.
- Advances in hardware will make it possible to easily run multiple commodity AIs locally, on the weapon system itself.
- There’s no need to wait for AGI (artificial general intelligence). What we currently have is enough. It’s also potentially a cul-de-sac (dead end). Regardless, AGI will likely be far better suited for deep research than real-world applications.
Of course, running commodity models locally won’t be enough. It will take some manipulation of the model to yield the results needed. Here are some of the likely modifications we’ll see (this may be tough to understand if you see AI as a tool);
- AIs need a persistent mental model (for the mission order layer). A model that can remain intact across sessions and improve through training and experience (like a human being). The company that builds this capability will likely be far more successful than the big AI companies we see today.
- AIs and human beings work better together than alone — it grounds them. Human beings introduce constraints (time, resources, etc.), values (this is good/bad), and a vector for improvement (orientation). AIs provide unbounded potential. Think of crew/team members rather than tools.
- The more human-like an AI is — from language use to personality — the easier it is to integrate it into human-run/built environments. We already know how to train a recruit, how to measure acceptable performance, and how to build trust through dialogue and interaction. The more human-like an AI is, the better able it is to access these highly evolved social interactions. Interestingly, companies developing robots are finding that bipedal robots are the optimal form factor, since the human shape is both highly evolved and best suited for most tasks.
What it Means
Here’s some thinking on what this means;
- In the past, weapon systems took decades to build and changed slowly. Autonomy changes this. For example, new capabilities developed by field tests or simulation (testing scenarios in full physics simulators depicting actual environments) could be downloaded to existing weapon systems, making it possible to upgrade a weapon system significantly without any meaningful hardware changes. A process of improvement that used to take many years would shrink to weeks and, in time, days.
- One of the best ways to test, improve, and train a weapon system’s AI is through simulation. The bigger, faster, and more realistic the simulator is, the better the results. A national initiative that builds these simulators would be decisive in accelerating the commercial development of autonomy and the success of autonomous weapon systems.
- The standardized ways that military members interact with each other, both generally and within teams/crews, will prove to be a significant benefit to human/AI teaming. This standardization means that AI will remain relatively portable and not tied to a specific human teammate. This portability will make upgrading autonomous weapon systems currently in the field easier.
John Robb
Global Guerrillas & Grey Swan Investment Fraternity
P.S. from Addison: Ahead of tomorrow’s Grey Swan Live! with John Robb, we’ll review some of his recent writings, such as this piece today on autonomous warfare making its debut in the Ukraine conflict.
John is a regular contributor to Grey Swan Investment Fraternity and the author of Brace New War. If you’re a paid-up reader, you can review his analysis of Antifa, Terrorism and Political Warfare in the October issue of the Grey Swan Bulletin.
Mr. Robb’s our go-to expert for exploring the geopolitics of the Trump administration’s tariff strategy, understanding Gaza, the global intifada and the ongoing standoff in Ukraine. We also look to Robb, in his work as a consultant to the military’s Joint Chiefs of Staff, to keep abreast of innovations at the forefront of the rapidly developing technical future of warfare.
With the markets rallying on positive news of a U.S.-China trade deal today, John can point us to the next global hotspots – and some of the more attractive investment opportunities that may be off the radar with traders happily blowing the AI bubble.
