How AI can change frontline military employment

Terry Garton We've talked a lot about AI on this show, and it's often on an imaginary level, right, theoretical level. What we're talking about deploying AI on the battlefield is actually providing services to service members at the work they're doing, perhaps during the battle. What do you think is the most pressing operational or cognitive challenges service members face today?

Tyler Saltman As you know, I think the current AI problem is becoming too generalized. And these Mega Frontier models, Openai, Anthropic, and X models are too generalised, too big to use and expensive. More importantly, they are not a domain specific to fighter jets. I saw Grok hallucinating something that happened on the news just a few days ago. And of course, I've seen ChatGpt do some crazy things. And the reason is that they were trained on all YouTubes that transcribe into 10 trillion tokens, and of course, much of this internet data. So they took the internet and became a huge corpus of data as a data set, and trained on it. And of course, these datasets have a lot of harmful biases, and the models do some crazy things. What we do is build these models from scratch and create datasets, which makes them culturally recognized by American fighters. We are one step further with the Army, the Navy, the Marines, of course the Air Force, the Air Force. They can all be different cultures, different acronyms, different ways of doing business, different ways of dealing with the fight. And finally, you take it a step further based on your profession, MO, your military profession specialization. I, I was a logistic person in the Alpha Army of 90. I needed AI that was very different from combat medics, fighter pilots, or combat engineers. And that's really what we're solving. Domain-specific personalization for fighter jets.

Terry Garton I was also an Army logistics guy, maintenance guy and I remember having an entire van full of publications and maintenance manuals to film when we went to the field. The other day I spoke with people from DHA who were talking about AI-enabled medical assistance. So, how do you think about really deploying this in the hands of fighter jets and service members? And how will it help them make their work better on the ground?

Tyler Saltman As you know, we've experienced it [Basic Officer Leader Course (BOLC)] And through trade-offs, learn how to logistics the military. There are lots of exams, there are lots of tests we had to take, there are lots of doctrines that we have to memorize, there lots of [field manuals (FMs)]training manual, you will name it. But now these [Large Language Models (LLMs)]these agents are the ultimate compression function of knowledge. They can train all of this doctrine. Then, through natural language processing, we can now ask questions similar to human conversations like this. And AIS can actually help you. For example, just create a road plan to do X, Y, Z. [Light Medium Tactical Vehicle (LMTVs)] Do I need to move X amounts of equipment to this location? How much fuel do you need? And soon, will it help you come up with the correct answer? This will allow you to create a typical operation image. And of course, we can do the same in medical care. You can train all this medical data and make people understand how to make what they need more efficient.

Terry Garton Well, you mentioned that all the different services are very different cultures. Do you think this individual level AI will help bridge some of the differences across the various service branches?

Tyler Saltman Absolutely not only a service branch, but also an ally of our NATO. For example, if you get one of the models translated into Hebrew, that's not a big deal because you don't think of it like an IDF soldier. Now we can work with the Israelis to train us to get the data and create their culture. Also, the South Korean military is the same as Korean translations, but does not let the context get lost.

Terry Garton What is the challenge of bringing these different cultures, languages, and even equipment sets all together?

Tyler Saltman The challenge depends on becoming personalized. And how do we do that? Well, what we do is take these [Large Language Models (LLMs)]we distill them and they become much smaller. Next, reinforce them with Loris. This is an adapter and accesses fine tuning. And we deploy this to your device, your laptop or your phone, so you don't need the internet and your data never leaves. And it's a challenge in itself, and although this runs like a ChatGpt experience, it makes the internet incorrect on your device. And the difficult part is reducing this so you don't lose the magic of these models.

Terry Garton I'm talking to Tyler Saltman. He is the founder and CEO of Edgerunner AI. In combat scenarios, individuals make big decisions under stress. What risks are there inherent risks to deploying this type of AI at this level?

Tyler Saltman You know, I don't think the big risks are unfolding that. And here's the reason. It is better to make the wrong decision right away rather than making the right decision too late. For example, let's say you're running a convoy and the truck goes down and moves an explosive ordinance. So, for example, can we repackage this explosion ordinance on another truck with a mortar round? Or if that truck is attacked, will it make the explosion worse and take out my entire fleet? As a young Li, you need to call the shot. You ask three different [Non-Commissioned Officers] (NCO)]you get three different answers: yes, no, maybe. Well, now I have to make a decision. Now, the AIS was right on my device, so in reality, yes, you're fine, is it okay to go, I won't do it. Now it's not a silver bullet, but now you can make a much better decision right away. Of course, I can pivot later. As you know, it's the difference between life and death.

Terry Garton Well, you talked about having AI on your device, and you also talked about it not being connected to the internet. How does it work in a rejected environment?

Tyler Saltman That's its beauty. It works in a rejected environment, as if using Wi-Fi. So, this is live on the device, like Microsoft Word and Microsoft Excel. You can open the interface with the laptop, like chatgpt. And of course, the internet is not necessary. Therefore, when you are in a disconnected environment, you are OK to go.

Terry Garton As the batteries in these devices can be exhausted, what concerns do you have to deploy AI at that level replaces personal knowledge and initiatives?

Tyler Saltman The good things about batteries – we've built different runners that optimize to use different parts of the chip, improving battery life and preventing the device from overheating. Therefore, fever can be a problem. But you develop the good points. I think I'm not so aware of this that I'm smarter, smarter and less informed. I think it's important to keep people in a loop. Think of this as your expert AI assistant. But it's not a silver bullet. You still need to check. Even as I am a young eu, I asked something to the Supreme Warrant Officer or something in my NCO. Similarly, like these AI agents, I think these AI will get even better over time. And of course, you can refer to where we found the information via search extension generation, what we call RAG. So, combining these agents with RAG can also help you see where the source is.

Terry Garton You are clearly a veteran military training model. It probably hasn't changed much since I went through it. But that conversation about keeping humans in a loop – what changes do you think are needed in the military training process to make sure they keep humans in a loop?

Tyler Saltman So I was in charge, I create a SOP. You will put your system on your laptop. And when you use it, what we've built into our system is to give a thumbs up. Give a thumbs up every time you get a good answer. If that's wrong, thumb down. This is called the reinforcement of learning human feedback. It will then be used to suppress and retrain AIS in the future. Therefore, humans will help you enter the loop to keep AIS up-to-date and future grounds. Because by simply removing humans and relying on AI to retrain itself, AI will actually become a fool, reaching convergence with other AIs, they will all be the same. So it's essential that humans stay in the loop. Yes, it's a great plan, a good OP order, it's a good risk assessment, or it's not good. To prevent the AIS from falling off the rail again.

Terry Garton So do you think this is actually a reality?

Tyler Saltman Therefore, our system works today. We are actually deployed in live environments. [Special Operations Command (SOCOM)] Overseas, and we're excited about it. I think within 2-3 years, that will become the norm. I think AIS will become part of all Warfighter toolkits. Also, joining the military will be included in the course. And I hope so, as we fell behind them in drone racing, we cannot fall behind China and this AI race.