Warfare by Committee, a kind of offside report on using LLMs to support wargame participation
These are just a few notes on how I got on using LLM’s during Jaap’s game of “Warfare by Committee”, played in the middle of May 2025, as an exploration of politics in the French General Staff pre-1940.
This was set in the late 1930s, after a French wargame had demonstrated that the French military needed to rethink their plans for a German invasion. Each of the players was one of the generals at the time: Gamelin, Georges, Hering, or Pretelat.
I played André-Gaston Prételat, a veteran of the First World War, and pilot. I’m now part of the French Supreme WarCouncil (Conseil supérieur de la guerre, generally abbreviated CSG).
Overall the game was to role-play the different generals, and determine how the French should respond to any forthcoming invasion based on a recent wargame. From the introduction: “One of the tasks of the CSG is the organisation of exercises, manoeuvres and war games. In fact, at the time of the game, a large map exercise has just concluded, and the members of the CSG are preparing to discuss the results at the next meeting. Before presenting these results, though, we shall give a background of the French military situation halfway through the 1930s”.
A brief summary of Prételat was that he could foresee the utility of airpower, was defensively minded, and wasn’t against challenging conventional thinking – all of which suited me pretty well.
The setting
As with many wargames, especially the more conversational or conceptual kind, I mainly know how much I don’t know about the period, or how much I’ve forgotten since last time. This time period is especially difficult because it’s so involved,
Scanning the briefings and online resources was useful. We discussed the history of the Maginot Line, where I knoew that the common misconception of the French not considering that the Germans would try to go around it was incorrect. But the game went into more detail about how inaccessible some of the terrain was to the north, and the French expectation that Belgium would play more of a role.
I leant into my character’s history, and advocated for a vast increase of airpower – able to take on the function of artillery strikes through bombers and the function of cavalry for reconnaissance. This increase in airpower would also be required to protect French rail from the German Luftwaffe, which would be needed for transport.
As always, a quick skim of my character’s Wikipedia page was invaluable, along with the briefing provided by the game runner.
But beyond that, I was suddenly struck how useful an LLM might be. Rather than bothering team-mates with questions, or hoping to find quick answers online, I could see how useful they were for this kind of task. Especially as being correct would be very helpful, but not essential.
Through a project I’ve access to Perplexity.Ai, which seems to be widely well regarded as able to answers questions in more depth than other LLMs. So that’s the one I tried. I mean I also have an account with GreenPT but although I thought I asked it something, I can’t see anything in my chat history. It’s EU hosted, and using sustainable energy ( you can find out more at https://greenpt.ai/ ) so I’m interested in how that fares against the more exploitative models. Also I’ve accounts with Anthropic and OpenAI, and I’m used to going from the Bing search engine to Copilot, but I didn’t see a reason to use any of those for this exercise.
My experience of the LLM
A summary of my experience with them is as follows:
Hallucinations – or other errors. Of course I’m asking an LLM because I don’t know, so I’m not the best judge of their correctness. But Perplexity didn’t make any obvious errors, or say anything that clashed with the briefings or discussions in the game.
Summaries – As part of its answers Perplexity always includes a Summary Table and a short conclusion. This was particularly useful, as I could read that first and then bounce up to the more detailed sections if there was something I needed to read more on. Otherwise this format allowed me to skim a subject while also listening to my fellow generals at the same time.
Entangled queries – I tend to try out less obvious strategies for these kinds of games, it’s my natural inclination, but also it’s interesting to see what more conventional thinkers might have missed. In this case the French has particularly heavy tanks available, more so than anything the Germans had produced up to that point. Those tanks were very defensive in nature and not especially quick, but also might have been tasked with defending both the northern and southern end of the Maginot Line. Originally I asked about them transporting themselves from one end of the Maginot Line to the other, but them I think Jim mentioned railways. Perplexity.AI was able to give me details about how long it would take for them to travel. But also it was able to answer, relatively quickly, whether the standard French rolling stock would be able to transport all of the French tanks, includes the “Char 2C”, a “land ship” standing at 69 tons. I imagine that looking up the tank weights, and then trying to find out what French rail could support in the mid to late 1930s, would have taken me a lot lot longer.
Similarly it was able to give me more information on General André-Gaston Prételat’s preferred military strategy, as I was enjoying taking on that mindset and seeing where it took me. The briefing did provide that information, as did Wikipedia, but I was able to get more detail straight away – whereas I assume a more conventional online search would have just brought up references to biographies containing far too much detail, or online arguments where I’d have no idea who were the valid authorities.
Unusual queries – Also Perplexity provided a usable answer for a more obscure queries. As we were thinking about further use of the rail network, it was considered whether the French should build artillery trains. I was pretty sure they could they could only fire in the direction they’re facing, due to engineering requirements, and recoil. Perplexity.AI was able to confirm this – and unlike asking my colleagues, I wasn’t interrupting it, or asking something that might be obvious.
Centaurs, always.
While it leans into my common position on these matters, the combination of AI and human knowledge is what gave me the most information during the game, and gave me the most enjoyment in the game. And especially the combination of the two enabled me to take a relatively active part right from my relatively uninformed start.
For example AI highlighted that French airpower was behind technologically. But also Japp and/or Jim pointed out, first I think, that aircraft manufacturing in France was almost a cottage industry, rather than operating at the industrial scale that would be needed. Something perplexity.ai later confirmed. Being able to think about that, and research it, while the main game conversation covered other subjects, was a distinct advantage.
I mean, to be clear, I expect the assembled players could have answered all my questions, but then the game becomes “Nick learns about the period before WW2”, great for me, not so much for the others.
Two further notes:
Firstly – as always ( and this is thought provoking about my relationship with books ), I enjoyed the “just in time” approach to obtaining and using knowledge far more than any other method I’ve experienced in the past. I think that I, to a much greater extent than average, need a “hook” to put knowledge on.
Secondly – one of the interesting outcomes of the game was that the best way for France to respond would have been to invade northern Germany before the start of 1939, going through Belgium. This was politically impossible, and would have created its own set of problems, but was an interesting option to consider as we left the 1930s and brought more of our contemporary knowledge to the fore in the post game discussion.
Oh, and well, there was a third note… with a defensive mentality in mind, and looking for the French to defend the North of the Maginot Line, I tried to find a river that the French could use as a defensive line as it would slow down any German assault. In the end I determined that the River Aisne was in the “wrong place”, and would have been much more useful if it was more to the East and North. Unfortunately there was not time in the game to determine how easy it would have been to move… even with an LLM to hand… maybe next time…
Related Posts
Discover more from Military Muddling
Subscribe to get the latest posts sent to your email.
What does LLM mean?
In this context, Large Language Model – https://en.wikipedia.org/wiki/Large_language_model
What was a bit abrupt Andy, but I’ve had a coffee now, so let me try to be a bit more helpful. Especially as I’d assumed that the term was commonly known… but if the term was commonly known, then you wouldn’t have asked 😉
Here I mean “LLM” as in “Large Language Model”. An LLM is the type of AI that’s become popular in the last few years because of its ability to analyse data, to accept commands in free text, and to reply in free text as well.
ChatGPT is the most well known example of this, partly because it was the first on the scene, partly due to effective marketing.