AI for Grunts: An Intro to Artificial Intelligence (Part I)

by Major Ryan Shannon

Sure the sleeves are still cuffed, but the discipline it takes to maintain boot blouses like that in the jungle is impressive. Image Source

Editor’s Note: This is the first of a two-part submission on AI and its impact on the Rifle Company, and it is the third installment in our series about emerging technology. Part I provides a relevant and focused background on AI and Data for small unit leaders and Part II will focus on potential applicability in both garrison and deployed environments. You can read our previous articles on data sciences here and here.

---

So your boss wants you to use AI to make your company better. Now what? What does this even mean? Where do you get AI? What do you use it for? How can you integrate it?

These articles explore these questions, trying to describe what Artificial Intelligence (AI) is, practical implications of AI for a rifle company right now, and implications for a rifle company in the future. Part I will help you wrap your mind around AI when it comes to pictures, software, and big data. Additionally, it will preview AI’s applicability for junior leaders at the company level.

Bottom Line Up Front

The potential for your company to leverage AI is huge… as an end-user. Don’t expect to be part of generating or developing AI solutions. While an innovative Marine in your company could create custom code that solves your problems in theory, meaningful implementation would require an unreasonable amount of time and effort to overcome administrative red tape.

However, there are some tools you can use immediately, and this article will outline a few. Some things we recommend are not technically AI, and would be more appropriately categorized as really good software. If you’re a computer-savvy reader or stickler for technical definitions, please just let it slide.

If you do have a great idea, there are people you can contact. Specifically, you can reach out to the Marine Corps Software Factory. They are in a much better place than a rifle company to get a software solution implemented. This article will provide some details as to who they are, what they do, and how to talk to them.

In general, future AI tools could be used to speed planning, automate menial work, or aid in processing large volumes of data. There are some thought-provoking examples of current use or visions for future use that this article will describe. The goal is to arm you with some answers to the question “What do you want from AI as an end-user?” Hopefully a good understanding of what you could have can help you articulate what you want.

What even is AI?

AI is best conceptualized as a general-purpose technology that has applications for a broad range of human activities rather than conceptualized as one particular thing. Much like the printing press was a game-changer for the information domain or the combustion engine was a game-changer for applying force to objects, the lines of code that make AI have the potential to impact all sorts of human endeavors. The greatest potential is in the cognitive and information domains.

As a starting point, Encyclopedia Britannica defines AI as “the ability of a digital computer or computer-controlled robot to perform tasks commonly associated with intelligent beings.” These include things like visual perception, written or verbal speech recognition and responses, decision-making, and realizing an insight based of some perception. If you want a more technical description, here is a good video with a professor talking about AI terms.

Breaking it down further, here are three ways to bin your understanding of AI:

1. Dealing with text and pictures.

2. Dealing with big data.

3. Novel software.

Bin 1: Dealing with text and pictures

Much of the latest buzz around AI revolves around Neural Networks. Neural Networks are known for their ability to learn from large amounts of data, and are used heavily in processing text/pictures, generating text/pictures, or gaining insights from massive data sets. They include a lot of hype generating things like large language models (such as ChatGPT (openai.com)) or generative images (such as AI Image Generator (deepai.org)). The name “neural network” comes from the way scientists in the 1940s thought about how the algorithm works. Wikipedia has a great deep dive into that history. The name also implies biological sentience like that in an animal brain. This is unfortunate and incorrect. Neural networks are just math.

They work by taking a large amount of training data, using math to figure out what makes items in the training data coherent, and then making a new algorithm that can recognize or generate something similarly coherent when prompted. For example, I could use a neural network to train a new algorithm that could write emails. Say I used all the emails in my outlook sent before 1200 as training data. The process of training the new algorithm would find if an email starts with the word “Good” the second word is: “Morning” with high probability, “Job” with medium probably, and “Waffles” with near 0 probability. When I then prompt the trained algorithm to write a new email for me, the generated email will start with “Good Morning” since that has a high probability of being coherent.

The idea that we could tell the computer to draft up a document, review a document, recognize something in a picture, or help us visualize something by generating a picture has huge potential. More to follow when we talk about ways to use AI now.

Training an algorithm with a neural network is very resource intensive and can take days or weeks. The resource intensity during training is due to the need for significant computational power to deal with massive datasets and complex models However, once algorithm training is complete, it is a set of steps and some math (although in ChatGPT’s case, a math formula with billions of parameters).

Using a trained algorithm takes few resources, since it is basically just a big formula. Similarly, modifying an existing trained algorithm takes fewer resources than making one from scratch. This means that organizations or individuals can tweak existing trained algorithms to something more useful (or more nefarious). A recent example of this is someone who took Chat GPT’s algorithm and removed all the guardrails preventing the algorithm from producing racist or harmful content (like how-to steps for cooking meth) More to follow on this when we talk about future uses.

To sum up, a first good bin for thinking of AI is technology that can facilitate processing or generating text and pictures.

Bin 2: Dealing with big data

The term “AI” also appears in the context of quantitative analysis. Computing power has gotten so fast and efficient that analyzing data sets using statistics has become very easy. Statistical techniques that used to take a lot of time now can be done almost instantaneously. This allows for people to apply all sorts of different analytical approaches and tools that were impossible a few years ago due to how long it would take to do the computations.

In other words, the term “AI” sometimes just refers to doing statistics quickly and with enormous datasets. Why would you use just a linear regression model when for almost no cost and no additional time you can apply a logarithmic, a polynomial, a random forest, and a nonlinear model? No need to worry too much about what those words mean… the point is why wouldn’t statisticians look at a problem in a bunch of different ways and see which makes the most sense?

A satirical meme about how AI could be viewed as just rapid statistical analysis. Image Source

The implication of new and faster ways to apply old statistical tools is that we can quickly make sense of large datasets at little cost. Statistical analysis that previously would have taken multiple weeks and several post-graduate level degrees to understand can now be available to the beginner. This makes it easier to gain insights from data.

It also means we can also go back and look at old data we have been collecting for years and gain new insights. Think being able to easily do something like what they did in the movie “Moneyball” in areas other than baseball. Maybe we have had all the information we need to solve a problem but just haven’t applied the right model. A lot of the work being done in the Marine Corps with data right now is just that… applying analysis techniques to data we already had.

To sum up, a second good bin for thinking about AI is finding insights from quantitative data.

Bin 3: Novel Software

Finally, many people use the term “AI” to describe good software that does a new thing. Especially since the term “AI” entered pop culture, it seems to be applied to everything recently developed. This probably has to do with that part of the Encyclopedia Britannica definition saying AI does any “task requiring human intelligence.” You can make an argument that almost any task not currently automated fits that description.

Any software that automates a process seems magical at first. But as the magic becomes mundane, what we first thought of as AI becomes just software. We don’t give a second thought to things that only existed in science fiction 30 years ago. Most new cars automatically break if it senses you are about to hit an object. A Roomba can navigate and clean a room. Alexa can search the internet and answer your questions. As these things have become more commonplace, we stop thinking of them as artificially intelligent, and think of them as just a good tool. A great read filled with examples of this is the book “Army of None.”

However, newly automated processes are hugely important, even if we eventually just think of them as good software. Automating things makes us faster and enables decision making. We will talk about this when we discuss future visions. Expect AI tools to help us sort through options and enable us by taking care of simple tasks, like the way that JARVIS helps Tony Stark in Marvel movies.

The idea of automation leads to another important concept when thinking about AI, the difference between “autonomous” and “automated.” Truly autonomous machines would be self-sufficient, could learn, and could make decisions in many different environments (Terminator, C-3PO). Automated machines can do a narrowly focused task if well-known criteria are met (Roomba).

Autonomous systems that approach sentience just haven't been figured out yet. Machines can be very impressive at specific tasks, but they do not have the flexibility to do other things outside of their lane. To use some examples from “Army of None,” a chess supercomputer cannot make a cup of coffee. A self-driving car cannot decorate your house. A human might get beaten by a machine at each individual task of playing chess, driving a car, making a cup of coffee, or decorating a house… but a human can do all those things.

The line between “autonomous” and “automated” isn’t always easy to define, however. Afraid of machines that make decisions to kill without final approval from human beings? We’ve had that since 1983, it’s called the Tomahawk missile. Once you launch it and send it to a certain box to loiter, the missile senses targets in the box, compares the targets to preset parameters/guidance, and then decides to strike the best target. Is that a highly automated process or an autonomous killing machine?

To sum up, a third bin for thinking about AI is good software that does whiz-bang stuff.

Clear as mud? Ok let’s dive into some applicable stuff.

Being prepared for the impacts of AI

AI has the potential to revolutionize the cognitive aspect of war, kind of like the industrial revolution impacted the physical aspect of war. Much of this will happen at various headquarter levels, most of them high level (like COCOM high). AI promises to speed up command posts and staff functions in garrison or combat by making it easier to absorb, process, and analyze information. Automating the menial work associated with information allows humans to focus on making decisions.

An easy example is in fires. Pretty much every aspect of the kill chain besides physically firing the weapon (finding a target, queuing surveillance assets, getting location data, transmitting location data, etc.) could be made faster through automation. A faster tempo across kill chains results in more effects on target. Here is a link to a great podcast about just that: Ep 112: Paul Scharre on AI 101 - School of War Podcast.

Be ready for new training programs and methods informed by big data. If done correctly, well-collected data can definitively provide an answer to what training is most effective. However, that might not remove the emotion that some people feel about a change. The update to the Marine Corps Rifle Qualification is a good example of using data to justify a change that people could feel strongly about. More on this when we talk specific, but here is a link about it: Brutecast S5 E15 - Advanced Marksmanship Training Program

As software applications at the tactical level become more common, the Marine Corps will have to develop better ways to maintain them. Our communicators already do this to some extent with their gear, but it would be more widespread. Be prepared to have digital property consisting of software applications like the way you have physical property consisting of a program of record equipment. You likely will have to spend some time maintaining your digital property just like physical property. You also may be able to tinker with it provided you don’t make huge modifications, much like you can tinker with but can’t weld things onto an MRZR.

Overall, be ready to be flexible. An AI solution may look good in development, but not work in the field. Even if the technology is great, the rollout of any initiative or program can be friction filled. People might not immediately be enthusiastic about something different, even if they eventually find it to be objectively better.

Help the Marine Corps with good ideas

If you do have a great idea or made some good code, there is a place you can contact for further actions. It is called the Marine Corps Software Factory.

The Marine Corps Software Factory is quite new, but the vision is for it to work directly with industry leaders and create software solutions for the Marine Corps. As much as possible, they take on projects that can be completed quickly (weeks or months) and can be generalized for use for a wide audience. They were established in Austin, TX just last July. Ideally, when they make a good software solution, they find a program of record to take it on for longer term sustainment.

If you want to contact them, they encourage direct contact by their organizational mailbox (mcswf@usmc.mil). There is basically no bureaucracy or restrictions on reaching out. If you prefer to have something more formal, here is the Software Factory's product intake form. You can contact them to help them identify and define a problem. If a Marine has made something useful, they can talk about getting it to production.

If you want to check out what they have created already, the best way is to reach out to them directly.

They don’t have a streamlined marketplace yet, although one is in the works. A lot of their work so far has focused on plugins for the Tactical Assault Kit (TAK), which they turn over to Marine Corps Tactical Systems Support Activity (MCTSSA) for long term maintenance. If you’ve heard of TAK, that probably is exciting. If you haven’t heard of TAK, more to follow in the section about specifics.

Here is an overview of what they have done so far accessible by your .usmc email:

MCSWF Overview 16_9 May4 2024.pptx

One caveat about the Software Factory is that they are not the best point of contact regarding SharePoint/ Power Apps. For those, the best place to go is Enterprise Information Services.

How might a rifle company use AI generally?

In general, prepare to be an end-user of AI that you didn’t produce. You also should prepare to feel the effects of AI solutions implemented at higher headquarters. This might mean new training programs optimized using big data, preventative maintenance customized to each vehicle you own, and a whole lot more we will discuss in the section “How might a rifle company use AI specifically?” Finally, be ready to help identify good ideas to the people who have the job of developing software applications for the Marine Corps.

Being an end-user

The normal rifle company simply isn't in the right place organizationally to do much besides being an end user. Unless you have a bunch of Rifleman who can code, know how to influence the DOD acquisitions process, can integrate things across the DOTMLPF, and have access to the USMC’s databases, most AI or software you use will come from a source external to your company. If you check out the DOD AI Adoption Strategy you’ll see it generally is aimed at engineers and acquisitions people. This is ok… it probably is unfair to put the burden of something like “developing algorithms that allow individual commercially available drones to exhibit semi-autonomous swarm behavior” on your LCpls.

Being just an end user is fine, and there is power in exceptional application of existing tools. You probably weren't personally part of the design process for a smartphone, but that doesn’t mean you can’t figure out some cool ways to use it. You probably can teach Marines how to shoot even if you didn’t take part in creating engineering specifications for a rifle. Innovation through novel application of an existing tool can be as valuable as making something new from scratch. Additionally, you don’t have to worry about all the technical and organizational friction required to develop things.

Overall, try focusing on implementing and integrating existing software or AI solutions in garrison or in the field. Also, as cliché as it may seem, good fundamentals on basic infantry tasks will make your unit fertile ground for integration of technology. If you had to choose, being good at the basics and achieving limited AI integration will probably be more empowering to your company than trying to be an Innovation Hero who invests the unit's time into trying to transform how the service uses AI.

Maj Ryan Shannon is an Infantry Officer currently serving as 3d Reconnaissance Battalion Operations Officer. He is a graduate of Naval Postgraduate School where he earned a Master of Science in Operations Research. He can be reached at ryan.a.shannon.mil@usmc.mil.

---

Have ideas or experiences that could benefit other infantry and reconnaissance leaders? We need to hear your voice! The Connecting File depends on the willingness of readers to share lessons learned. Read more here about how you can contribute to our community.

---

The opinions expressed above are those of the author and The Connecting File. They do not reflect the views of the Department of Defense, the United States Marine Corps, or any other government entity.