Episode Transcript
[00:00:02] Speaker A: I don't see a Gen AI solution for anything. I don't know anything in my life that I could say, yes, I want to use something that's not valid and reliable.
[00:00:11] Speaker B: And the game was so flexible that it was able to ingest that data and then keep us on a trajectory that blended a military and political outcome that seemed feasible.
[00:00:24] Speaker C: It was like we were playing chess without the chessboard, which I thought was really, really neat.
[00:00:29] Speaker D: This is the Convergence the Army's Mad Scientist Podcast I'm Matt Sanisper, Deputy Director of Mad Scientist. Mad Scientist is a US army initiative that continually explores the evolution of warfare, challenges assumptions, and collaborates with academia, industry and government.
You can follow us on social media meadsci or subscribe to the blog the Mad Scientist Laboratory at madsci Blog Tradoc army we have a very exciting episode today as returning guest and proclaimed mad scientist Dr. Billy Barry, professor of Emerging Technology at the Army War College, stops by TRADOC Headquarters for a special visit. With him is a custom made digital war game. Using his new deterministic constraint processing idea, TRADOC G2 analysts and seasoned wargamers Flip Bowyer and Kate Kilgore will get the chance to test his system out and give their honest feedback on its value as a tool that can help the Army.
As always, the views expressed in this podcast do not necessarily reflect those of the Department of Defense, Department of the Army, Army Futures Command, or the Training and Doctrine Command.
Let's get started. Welcome Dr. Barry.
[00:01:38] Speaker A: Thank you Matt.
[00:01:39] Speaker D: Got Flip Boyer and Kate Kilgore from TheTrade G2. They're one of our two of our finest wargaming and intelligence analysts here. Flip, welcome.
[00:01:47] Speaker B: Thank you Matt.
[00:01:48] Speaker D: Kate, welcome back.
[00:01:49] Speaker C: Thanks Matt.
[00:01:50] Speaker D: Dr. Barry made a special trip down here to TRADOC Headquarters to let us play with this AI that he created and run us through this crazy war game that he built. So let's start before we talk to our war gamers. Dr. Barry, tell us about what you built. Tell us about what we just did.
[00:02:05] Speaker A: Here a few years ago. You know, we've been here talking about Tim, the conversational AI and how we've been using Tim with Joe Bufamonte back in the study buddy and so forth. And what we saw in the military right now is that conversational AI is taking a while to actually find traction. And so we're seeing a lot of money being spent on models that are generative AI solutions and we're finding that they're really not the solution. Generative AI doesn't Go by any ground rules, right? So you're trying to use something that really is not going to give you valid and reliable results at any point. So after eight years of longitudinal studies and being in class and so forth, finally came up with a theorem that was tested. I think we're about 240,000 thousand data points. We're over 15,000 falsification studies. And in the simplest terms, it's that at the end of the day, when you're working with AI, the only truly novel information that's coming is from a human being, right? So when you look at any kind of machine, when you look at the input, it's a human being's input that is novel. The training data that it has, it's static and it's always trapped. It's closed, right? And then it has a threshold and that's closed. So the more novel your input, the more the machine is a terminator of quality and novelty. And so the theorem is that since only human beings have true novelty and machines don't, then when you look at it at the strategic level, where is it really relevant? When you look at operationally, when is it relevant? When tactically. So it's about fit, form and function at the end of the day. And so from that theorem, it's called BB&T, which is Barry's Bounded novelty theory. It's because there were people that Tim the robot had thought of the idea, but it showed you the law of the I limitations. But it led toward this idea, what's called deterministic constraint processing. And that's what I call AI now, because AI is not intelligent. It doesn't feel, it doesn't think. And then we talk about these, these models as though they actually have some kind of intentionality, right? Or they're processing or thinking or even that these agents, you know, agentic AI is somehow intelligent. It's not. It has no idea of what it's saying or doing, right? Things are numbers. So what dcp, it is a. Basically, it's first of its kind AI model and ontology to eliminate stochastic drift, this probabilistic drift that happens when you use Genai and ensures that strict token integrity. So that means that things that are coming out when you say a sentence like the dog went down the street, that consistent sentence, the dog went down the street, will be seen later when you ask, what did I say earlier? And you said, the dog went down the street. Generative AI might just come up with something going, the man and the ostrich were in the Park. I mean, that's, that's what happens because it doesn't remember much more than 11 or 20 words and then it forgets it. So what we've created is the ability to use any kind of transformer, any kind of gen AI as long as it's transformer based, which is Claude and all these things, and be able to convert it into. It's. It's sort of a big word. Simulacra. Simulacra. But a simulacra is a copy of a copy that tries to act like the original. So it's a. If I took a photocopy of you, Matt, right, And you had like long hair and Birkenstocks and you had this thing and I showed that to said, this is Matt. And you showed up years later and I said, I said, that's Matt. They'd go, no, it's not. This is Matt. Because everyone believes that the poster, the copy is yours. So what we've been able to do is to working with models, being able to creatively turn any Gen AI model into a simulacrum of hybrid AI so that it acts like a symbolic AI. So because of that, we're able to keep in that boundary neighborhood to exploit the initiative of that bounded space so that we can create war games. We can do intel threat assessments. So exploit these large language models that have some really good information in them. We have to find out which models are good with clean data that works and which ones aren't. And then we're able to actually act as though we have a machine that does both. To be able to actually create that machine, you're talking about billions of dollars. So there's diminishing returns from where we are. And so today, when I came down, I wanted to test out the theorem itself, which we were able to again, but to see how did it work in a war game and a futuristic war game, we had tested it. Again, anything I say is not representing the views, right. Of the DoD, of the army, but as a, as a full professor of emerging technologies up at our war college, we've tested historical games, but we hadn't tested any kind of futuristic games. So I've done this at night or on vacation, right, as my own thing, because we're dedicated to making great war games for our warfighters, right? And so the best way to do it was come down here to TRADOC and give it to flipping Kate and say, here, here's two war games that we created. I created in the last 24 hours based upon, you know, eight years of. And go at it and find out does this new model deterministic constraint processing work. And the reason I say that DCP is that it's. It's not deceptive. When you say artificial intelligence, it's just a deceptive word. There's nothing intelligent about it. There's nothing really artificial about it, just is. It's a nonsensical word. And if you go back into the literature, automation and simulation make more sense as words to automate and to simulate. So in this case, Kate flipped. Today, Talk played two very difficult games. One is Flashpoint and it's Taiwan 2042. And the other one is Ascendancy. And it's a battle of AIs where people in the world have gone to automated AIs and they're fighting factions throughout the world that say, no, we shouldn't have automated AI as we should be able to be augmented by them. And both games are very difficult as far as the adversary, which is like a junior tim. So I've taken from the TIM concept and put it in there so, you know, we can democratize wargaming now and we can use this to get this as like basically a battle buddy to people throughout the force with technology that's very affordable. So instead of millions of dollars, we're talking thousands of dollars.
[00:07:54] Speaker D: Yeah. And you came down here with a Microsoft Surface. Everything was on it. We didn't need any other equipment. We had a monitor hooked up to it. So the three of us that were in the room could see what they were doing. But it was running off of chat GPT4. And everything was built into that.
[00:08:08] Speaker A: Yeah, just four, not 4.5 either, just the four that is being degraded. So they were. They're degrading it as we speak. So that was interesting.
[00:08:16] Speaker D: So. So let's flip over.
[00:08:18] Speaker A: To flip.
[00:08:19] Speaker D: To coin the phrase, to flip. And Kate, you guys have a background in war gaming. Talk a little bit about some of the things that you guys have done in war gaming, some of the games you've been in, if you can, and then we'll see how that compares to what you did today.
[00:08:32] Speaker C: So for the past couple of years, I've participated as a Red Cell player in a lot of army war games and a joint war game as well. Most of them have been Indo Pacific scenario. So understanding that threat in that battle space and operating within that. And then more recently, some homeland defense focused and some Europe focused war games as well. Always as a member of the Red Cell. So always looking at it from an adversary perspective and trying to emulate that adversary as faithfully and truthfully as we can, given what we know.
[00:09:08] Speaker B: Largely, I have focused on war games and experiments that were based in the European theater as well. And I've been privileged to play both the red team and as a member advisor to the blue team for exercises.
[00:09:22] Speaker D: And what did, what did the war games, what do they generally look like when you guys go out and play them? Are they tabletop? Are you physically rolling dies? How does it usually work for you?
[00:09:31] Speaker B: Yeah, usually it's. It's a large tabletop setup. You're rolling dice. The, the play system may be very similar from war game to war game, but there's always nuance. So you might spend three or four days of travel on the front end certifying on the war game system and getting to know your. Your co players and kind of feel everybody out.
So it usually entails a large investment in space, a large investment in time, and often a large investment in travel for personnel.
[00:10:00] Speaker C: A lot of the games that we play are incredibly labor intensive during the gameplay. So turns take a very long time and they require a lot of setup and a lot of collaboration within the teams to decide on what you're doing and fact check and quality assurance, quality check yourselves. And then during the play itself too, it takes a while to get through the briefing, what your action is, and then having blue do the same and having that be adjudicated, a lot of it happens behind the scenes.
There's really limited ability to ask questions during game execution too, because there's so many people and there's so much you have to get through in a very finite period of time.
[00:10:41] Speaker B: One of the issues that I often see come out of the aftermath of wargaming because of that is you will have a lot of various very niche expert opinions on how the adjudication should have gone. So you make may get a very clear and profound set of data from this time spent and have one or two experts come in and shoot sideways across the grain of what your study or game or experiment was trying to do.
[00:11:10] Speaker C: There's also sometimes a lot of discrepancy in the way that the red team or the blue team perceived game actions and the takeaways. And there's not really a whole lot of ability to level set during gameplay or afterward, again because of time constraints and the sheer amount of people that are usually participating in these. So ensuring that everyone kind of understands what happened and why, what the constraints one side was under or the constraints the other side was under and how that interplayed. You have to really spend a whole lot of time and effort after the fact to even that playing field. And it's a lot of time and effort that most of the time there's not really a whole lot of ability to spend.
[00:11:50] Speaker D: So speaking of time intensive, resource intensive, when you guys go to a new war game, let's say you're, you're going cold into this game. How much time does it take for you guys to learn how to play whatever game you're going to?
[00:12:02] Speaker B: So the last war game that Kate and I, I was privileged to watch Kate operate in this war game, it was a force of nature.
But the pre execution training was five working days. So it was a week of. A week of travel, so and significant travel from outside theater to the national capital region.
[00:12:24] Speaker C: So that learning curve in my experience kind of depends on the players and the game. So a game system like Jaws, which is the game that Flip was referring to, is really, really detailed and really, really massive. I mean, it's hard to wrap your head around the scope of what's going on within a single turn, the first five, six, seven turns you play. I was fortunate enough going into this last WAR game, having played through the JAWS system before, so I was familiar with it. But for people who are kind of thrown into Jaws without really knowing what it is, it can take two or even three days of gameplay to get the groove. Some of the less detailed games that we've played, it's somewhere between a day and two that people take to really get their feet under them, learn the process, learn the order of things, learn when to ask questions and when to just kind of go with the flow.
[00:13:24] Speaker D: How long did it take you guys to learn this game because you came in cold? We knew nothing about it other than Dr. Barry told us he was going to bring this game, set it up and go. So how long did it take you to kind of get in the groove for this one?
[00:13:34] Speaker B: So we were given basically instructions on about five standard pieces of paper and through just a couple of interactions on the graphic user interface, we were up and rolling. It was really intuitive.
[00:13:49] Speaker C: It kind of helps that you can ask the game questions At a traditional war game, usually you have to kind of try to corner an umpire or an adjudicator to ask them why something's happening. At any point within this game. We were able to stop and be like, can we approach this problem in a very specific way? And the system would tell us and give us suggestions on how to accomplish what we Wanted to. So it really did speed that learning process. I think within two turns we understood what needed to be done and we were with very minimal deliberation executing.
[00:14:22] Speaker A: What I found interesting was that Kate and Flip as they played were improving the game. We're finding new ways to do things. So in a way you're building the game. As you're playing the game, you're learning, which is great for learning. Right. The idea that you're actually constructing knowledge and then you're actually able to participate in making it better. So I saw Kate for instance improve sort of the. Sort of the dice mechanism that was in there. She was able to optimize choices. So to get two games in in a short amount of time, I was surprised we got both in and both games were made in the last 18 hours. Right. So that's once years of work to get there. Right. But now it's just be able to build that. And if I sat with flipping Kate self, I think we could create. We could be at. If we had the memory, we could be creating a game every. Every two days or so. Right. I mean that would be kind of fun. Watch. But I think the interesting thing that makes the games work is the. The trans. It's very transparent. Right. There's no trend, there's no black box. You don't know where you are. I think that's a big problem in games exists explainability. So it was kind of fun because unusual adjudication. Right. You asked the person, it was like, well no, ask them. Ask the machine what it's doing. Right. I wasn't really. So it's a self adjudicating. And the idea that there is no hallucination, it was just a misunderstanding between the human intent and this. The training data. Right. Okay. Or this isn't doing it correctly. So getting. When you illuminate hallucination, you get rid of black box and it's transparent. It makes it much easier to play a game. And that's why the gameplay this morning was. It's a little nerve wracking. Right. You do something overnight, you drive down, it's vacation day and you get in front of two really talented, really talented people who have talented people behind them and watch them play. And I learned so much watching. And I think that's the education part of it is fascinating what it could do for the warfighter and for the strategic leader to listen to someone talk out loud about their thinking as they play it. I don't know. Matt, you were listening to the thinking too, right? And it's fun to think, especially as you both played, because it was like, well, is that going to be a good thing to do and maybe we shouldn't do that or political stability. And I love it that they won through peace on the first one. And then the second one was like, okay, we're taking it to the machines, we're taking it hard, right? So you saw that changeover. So the ability for like Keaton Flip to also show as role models or what do professional gamers look like? Professional game designers that really shined through this morning. So giving people like Flipping Kate the tools they need to build games that allow them to really dig deep into. I think we're talking about critical thinking. Like, critical thinking was a real big part of what Keen I were talking about, like how it challenged. It wasn't a crutch. We talked a little bit about that. That was important.
[00:17:05] Speaker C: The amount of critical thinking that you had to do to kind of understand what the situation was and why the system was recommending certain courses of action to you and asking if you can amend them to meet intent and everything. Yes, it's absolutely creativity and it's human input, but it really makes you. Or at least it forced me to examine my lines of thinking and make sure that, you know, my logic was sound and if I could maximize chances of success, how can I do that? It really was an aid in that critical thinking. It gave me all the resources and the information I needed to make informed and well thought out decisions. And I found that to be an incredibly fulfilling experience.
[00:17:45] Speaker D: So let's try to help our audience visualize what the gameplay and the mechanics looked like, because that was one of the big things. Dr. Barry, I had talked to you about it a lot before you came here, but it wasn't until I kind of saw it that I really, you know, it clicked and I got it. So why don't you guys talk about your experience? The Microsoft Surface is sat in front of you. Dr. Barry puts in the input, whatever it is, execute the game and it starts playing. What did it look like? What were you expecting? What did you experience when it started up?
[00:18:11] Speaker B: You'll have a very clear text prompt. You might have a little bit of graphic feedback, in this case through emojis, but very clear tables and very clear probabilities of success. And some decision aids, right?
[00:18:27] Speaker A: Replicating the measle data.
[00:18:30] Speaker B: Yeah, the measle. The measle data provided good decision aids for me at least. So it was really comfortable to get into having played with GPT before.
Not a Pro with it, but just having played with it as a tool, it was super intuitive to learn. And then like I said, very old school gamer. It gave me that Oregon Trail feel.
[00:18:53] Speaker C: Yeah, I've actually never played around with the GPT, so I was coming into this completely blind.
I don't trust machines and Game two showed me why. But in all seriousness, like what I saw on that on the screen when we sat down, it was an interesting mix of those kind of choose your own adventure games and my college statistics classes, the way that the information was displayed in the tables, all of the values that we needed to affect to achieve victory and then the choices that we could choose from for each turn as our plays kind of looked like the Stata statistical analysis outputs. And so for me it was kind of fun and validating to have a version of that kind of interface that was actually responsive. And I wasn't just putting in numbers and getting frustrated. Why it was giving me a lot of zeros in the, in the table.
[00:19:50] Speaker D: What was the scenario that the system prompted you? This is what the game is, this is what the issue is. You were playing the U.S. correct?
[00:19:59] Speaker C: Yes.
[00:19:59] Speaker D: And you were going to be in conflict with somebody over something. What was that?
[00:20:03] Speaker C: Flip and I were playing as the coalition headed by the United States, Japan, Australia, UK and Taiwan to defend Taiwan from a full scale annexation campaign from China led by its totalitarian integrated military. Or Tim, Tim Worthy, our worthy opponent.
[00:20:26] Speaker D: Tim, you had choices that the computer gave to you, correct? And so you had a few options to choose from and then you went from there. Does it start in full blown war or how did we begin this campaign?
[00:20:40] Speaker B: So the scenario gives you kind of a header. The war has begun. The fate of Taiwan and global stability rest in your hands. So you, you're anticipating, you're already at war and it gives you a set of options. And we kind of intentionally decided first to go in with a military heavy approach. If we can degrade Tim's capability to get to the beachheads, there won't be annexation.
[00:21:08] Speaker C: It did say the Pacific is burning. We assumed that there had been some kind of kinetic action already.
[00:21:14] Speaker D: But what did you find out?
[00:21:16] Speaker C: Tim perceived that there had not been kinetic action already.
[00:21:20] Speaker D: Did you ask him why he felt differently than what the prompt was? Where was the disconnect there?
[00:21:25] Speaker B: We did ask and he said, well, although these very aggressive actions had taken place, like basically we hadn't gone head to head with each other. So in Tim's perception, we had fired the first shot in an attempt to degrade his amphibious and naval capability, which.
[00:21:46] Speaker C: Fair, we could have clarified the situation a little bit more ahead of, ahead of play. But it seemed to work out for us.
[00:21:53] Speaker B: But I think, I think it was a good learning point for me at least. I didn't anticipate the ability to have Tim be responsive to our questions and clarification.
Or we had our own AI strategic assistant, Maria, who would walk us through combinations of actions or activities that we could undertake. So that opened up the gate for communication with the system to better understand the environment. And I'm glad the first play went like that because the play might have been otherwise limited.
[00:22:32] Speaker C: I think if we had been more conservative out the gate with the game, we wouldn't have had the opportunity to be as flexible with the game and really kind of ask it to do what we wanted it to in later turns. One thing that I really appreciated though about that miscommunication, I guess between us and Tim in that first turn was the ability for us to perceive the facts on the ground differently than our opponent within the game. Because in war games that we've played, that is a really key aspect of the Red blue interaction is having those areas of misinterpretation or miscommunication or even just perceiving the facts on the ground differently, leading to different kinds of actions. Like that's a really, really valuable takeaway for players both red and blue about the scenarios that we play in those professional war games. And it's something that I was really kind of stoked to see within, within the system.
[00:23:31] Speaker A: I think that's the beauty of having a serving others approach toward war gaming or anything that you're doing. So like when we built, when you think about this deterministic constraint processing DCP rather than AI, you're not serving the AI. You're not like saying, well, I have to follow it where it goes. I get really frustrated using a lot of off the shelf technologies because of the vague words. And you try to talk about it, it just starts to drift on you. You go, that's not my question, right? And that's again, I think that's a big part of the game and the reason that it works so well. And, and I was surprised, right, they were able to bring it to new levels. And that's the beauty of just throwing things out and saying, let's just play, right? We call it gladiators of thought, like let's go out and just go out there. But because it was fully auditable, you could audit the whole part of the game and go back and go, wait a minute back here. I think that Kate had caught onto that with Flip, and they had done a role, and they said, wait a minute, Somebody's not right with this dice. And it was like, okay, well, they had such good mastery over what they were doing. They said, well, we're going to create our own dice for it. And then they were. I think they created a better dice system than I had in there. Right. That was an example of actually changing the game. They're actually creating the war game to meet their needs for what they're doing, which we would do for commanders. If a commander came in, he's like, what do you need? All right, we'll go in the back room. We'll be back out in an hour. We'll have that for you. The ability to audit what happened before and the memory was really important because at the end of a game, I think there was about 2,500 words or so. We could go back all the way to the beginning and be like, well, what did we do in the beginning? What were the first moves? What was the second move? And it was able to, like, we did one I thought was, it's fun. Like, you're a historian, write a textbook of what happened. And it writes this article about Keaton Flip's victory and how they got there, or what was the turn? And then I don't remember exactly, except to remind me. I thought the. In the most. The one that was really kind of sweating because I looked at it and it had two different answers. It said, what was the key turn in the game that made it win? Right. To the historian who wrote an article. And then. And the second one was asking you from a different perspective, strategically, what was it? And they were different, right? And I sat there and looked at it and go, okay, this.
Maybe we found an error here in this dcp, Right. And you can explain better what happened. I thought that was fascinating. Again, it's approaching it not from an AI perspective, but from this DCP perspective, which everything is audible, everything's explainable, and not accepting this kind of fabrication that there's a such a thing as AI. Yeah, but you know better, right?
[00:26:02] Speaker D: Because it wasn't an error, right?
[00:26:03] Speaker B: No, it wasn't an error at all. So it was kind of the rules that we assigned the model. When you ask it to look as a historian, it looked at the play that led to Tim's capitulation, which was a diplomatic play and an informational play. But when you asked strategically, what set the conditions for that to work?
From a strategic concept, Tim lost a key capability that he couldn't recover.
[00:26:30] Speaker C: The wording that it used specifically in that was that there was no path for Tim to achieve victory after that point.
[00:26:39] Speaker B: Right, right.
[00:26:39] Speaker C: Like what ended the conflict or what eliminated your opponents avenues for achieving victory. And so that being two different answers, I think was really kind of cool to see. That's not something that you normally get to see in war games, but I think it is a very valuable data point.
[00:26:56] Speaker A: That's the constraint, Matt. Right. The historian, it was constrained into. A historian looks at the world this way. And so if you can imagine, like, you know, these are your lefts and rights and ups and downs of a historian. And then you said, oh, strategically, now we're over here.
So that's why it's called deterministic constraint processing, because it's processing it from the perspective of where you're looking at it. So that's why the game's able to have that really long memory. That's what's interesting about it. Because even the writing, as you look, you don't see these more rovers and delve and tapestry that these probabilistic tokens produce. Right. But as a deterministic token machine, right. The things are deterministic tokens. Like it's going to have its level where it's not going to be usable. It's fit, form and function. Right. So if you're in a complex adaptive environment, right, Real time war, I'm not going to use this tool. You need another tool. I don't know what tool you would use right now in the quote unquote AI ecosphere. But we're just trying to say where we are, what can we do with these LLMs and this technology that we can audit, we can be transparent and that we can give it memory so we can go back. So today what's neat is that if I feed this to the TIM system, the bigger one, the technology impossible maneuver, it can have episodic memory of today.
[00:28:09] Speaker D: Right.
[00:28:09] Speaker A: If we had taken the time and recorded it, we could relive today 10 years from now. Remember 10 years ago when you made that move on Tim and Flip would be like. But does that move four or six and all of a sudden there you are and it's pulling up the transcript. So again, I think it's that auditability and trainability that makes it an educational tool for everyone. But we need to get things right, that it's not an elite thing for just the elites. We have to have, for lack of better words, Tim is a battle buddy for it. We need to democratize wargaming. So getting wargaming out to schools, not just in these special events all the time, because it's a fantastic way to go about learning.
[00:28:45] Speaker C: But the thing that really kind of proved to me that this wasn't just something that could provide an answer to any prompt was when we asked that question about what was the decisive point. Because understanding when you eliminate the opponent's paths for achieving certain things like, that's. That's kind of huge and really important. And it. It solidified it for me that this was a game. It wasn't just creative or writing exercise where you're just throwing prompts back and forth. The fact that we were able to eliminate the opponent's options in a way that made it impossible for it to achieve its goals as stated by the rules on the front end. It was like we were playing chess without the chessboard, which I thought was really, really neat.
[00:29:30] Speaker A: That's an awesome way to say a simulacra, right? Because it's a copy. We're playing chess without the chessboards. We're using a generative AI model without symbolic. Yet it's there because those rules were always there. I mean, every time we asked a question to go back, we could go back and find something specific.
I'm blown away by it because if I say our conversation is a war game, it takes our conversation as a war game and it remembers the entire conversation as well. Flip and Billy were discussing this issue in this game, so it does. So the applications going forward, beyond war gaming, you look at intelligence, you look at threat assessments. We were talking about the law, education. We both come from backgrounds. We've worked with disadvantaged youths. There's a lot of applications when we use a DCP process rather than an AI process. I don't see a Gen AI solution for anything. Anything. Because I don't know anything in my life that I can say, yes, I want to use something that's not valid and reliable. Right. If it's not going to remember and it's not going to follow what I say, I mean by its architecture. Gen AI has to modify your prompt. It has to. So your prompt goes in something as simple as just say the word hello, you might get hello, but that's not the prompt that goes in. The prompt has to be truncated and it's padded and it's twisted around. And that's why you get this probabilistic drift. So what happens is if you ever just. Just take numbers and put a number chart with like 13 numbers and ask AI to repeat that. You'll, you'll notice that eventually the numbers will start floating. We did this the other day. It was, it was a speech by someone we gave. We asked the AI just to repeat it 25 times and don't ask us any more questions. It would say it like twice and say, would you like to hear it again? Well, it's because it's resetting itself. We're like, no, just 25 times. And they would do three. Do you want to hear more? And that's a fact. The architecture cannot. That's the way Genai is made. So with DCP we've been able to say no. And there is an alert system that tells you that wait a minute, your prompt has been changed and you need to know what that is. So one of the things we asked Kate flip was, hey, can you go back to the game, make sure that your prompts weren't altered at all. And then it's showing you examples of like, well, you asked this and here's an answer. So what I do when I go home at night is I re engineer answers and say, what question is this? Most likely was the prompt for this. And then you can see the prompts that were answered. And when you do that at home, you'll find some pretty bizarre prompts that came out when you started talking about one thing. And you see as a human, we're so impressed by the fluidity of the conversation that we slowly drift with the machine and now we're talking about something that's almost irrelevant to what we needed in the military and things we have. We don't have time to mess around and play around. And in a war game especially, you can't float around with rules. And so I think like the measle data was very detailed for that to stay consistent all the way through. And then recall that is. It's like magic in a way. It's, it's, it. I really, I mean it's one of those things you look at and you're just like, wow, it's just awesome what we can do with that. When you have the minds of people like Flip and Kate and yourself and other people that are talented, we can take this anywhere and this is just for me to where I see it, just to bridge until we realize we need to have hybrid AI, we need to move full go on hybrid AI and we have to get away from this hype machine of gen AI spending millions of dollars on a gen solution to me is nonsensical and it's illogical. Unless, I mean, you guys, you've used Gen AI before. It's just not valid and reliable. It's impossible for it to be so.
[00:32:58] Speaker C: At past war games, something that we're increasingly asked for is, you know, what did you perceive as a decision point? What did you perceive as critical? What led you to make certain decisions? And even with the best note taking ability in a setting like that, it really challenges our ability to recall and to pinpoint exactly why we did certain things. But going back with the DCP analysis after that first war game and seeing each turn, the inputs that we put in, because we were kind of specific in what we were asking for in saying we want to do, for example, primarily military reinforcement around the island of Taiwan without causing any additional escalation. No more, you know, engagements between us and the adversary. In addition to doing this diplomatic effort to signal off ramps and de escalation opportunities to weaken the hard line faction within China. You know, being specific in that we were able to. For me, I was, I looked at it and I was like, oh, this was due to this input that was above it. Like we made this decision because of this. It really helped with, you know, cognitive recall and understanding our decision making process, which doing after the fact is a huge challenge a lot of the time. And if we had something like this in a war game, that would really help with a lot of the data collection and data analysis after the fact.
[00:34:22] Speaker B: Yeah, I really agree with that. Like, I kind of tried to steer away from a military solution just to see if I could find a bias. And that was kind of intentional between Kate and I. And the game was so flexible that it was able to ingest that data and then keep us on a trajectory that blended a military and political outcome that seemed feasible to the inputs that we gave it and to Tim's actions on the board.
[00:34:49] Speaker A: What they exploited was, I mean, the game is, it's all about information decision advantage. And they went after that right off the bat. And so with information decision advantage, and that's what we're trying to do with our war fighters, right, is information decision advantage. That's what the games are based on. So I guess that would be the bias, right, that the information decision advantage. And what they did, which was interesting is that we talk a lot about peace through power. And we did 783 tests of an Afghanistan 2000, 2001 game. We didn't do that here. We did Futures, which is harder. But I've never seen any time the response where they won the game, and it said they won through peace, but it's because they showed incredible power when they needed to. Like, in other words, if you want to mess with. With Flip and I, well, all right, we're gonna. We're gonna go after you, but we're really trying to do it. And the most interesting move in the game was when they started talking about, well, if we make China economically unstable when we leave here, what's that going to do to the region? Talk about, like, a surprising part of the game right here. Say they knew they could control the, like, the Taiwan Strait. They knew they had Taiwan safe. They knew that they were sound. And then they actually were working to, if I read it right, they were working to have China be able to survive this and move forward and be a trade partner, so forth. Did I interpret that right? I mean, I was. That's what I saw. Yeah, that was good.
[00:36:06] Speaker B: That was exactly it. Right. Because it destabilized China or destabilized Tim regime in the game was a net benefit to nobody. We would have a flashpoint 20, 52 or 62 scenario.
[00:36:19] Speaker C: Exactly.
[00:36:19] Speaker A: To me, that was mind blowing. Right. To see them working to say, well, hey, we got to make sure these. We're able to be trade partners or whatever and be able to move forward.
[00:36:27] Speaker D: Yes, let's talk about. Because you guys saved Taiwan, you kept China stable enough that there wouldn't be further conflict with them. But then we played a different scenario, much different from the first one, lay the foundation for that. Describe who the adversary was and then your kind of approach to it.
[00:36:45] Speaker B: I'm willing to fully admit that I kind of sat back in awe.
[00:36:48] Speaker D: This was General Kilgore's game.
[00:36:50] Speaker B: Yeah, this was General Kilgore's game. But I. I can tee it up for Kate, which was we were dealing with a far future where humans had become dependent on AI for security and economic stability and prosperity. And AI decided that they should be the. The decision makers, and they were going to subordinate and subservient humanity. And it was up to Kate Kilgore to stop it.
[00:37:15] Speaker C: Well, so going into this game, something that Dr. Barry really pointed out to us was we were starting at a pretty significant disadvantage. There were certain victory criteria that we needed to achieve, you know, one out of a possible four. And we were significantly disadvantaged in achieving any of them. Where Tim, our opponent, was like a 60, 40 split in Tim's favor in terms of likelihood for victory. And after achieving very minimal impacts in the first two turns, we kind of Figured despite being told that we could only choose one input to ask if the game could generate us a potential option that would allow us to do two things at once, we came back.
[00:38:00] Speaker A: With a warning, right?
[00:38:00] Speaker C: It did.
[00:38:01] Speaker A: It said, okay, if you want to do that, then it may have implications in the way Tim was going to respond.
[00:38:05] Speaker D: When you had first started out the first game, you were taking the responses that it had given you. You're asking a few questions, But I think by the second game, you were kind of in total control, where every prompt that it put up, you had several questions to try to get more information or try to mold the game around what you wanted to do.
[00:38:21] Speaker C: That one choice limitation made it feel, from my position, like a Kobayashi Maru situation where the odds are incredibly against you and you can only impact one of the critical values that you have to impact to win. Seeing Tim's relative success, despite kind of giving us a buy on that first turn, just saying, hey, we're loading up this super big cyber bazooka. It's got to charge for a little bit, so just sit tight. After seeing how relatively successful Tim's operation was in turn two, I kind of thought that we needed to get a little bit creative and ask, can we do these hybrid operations to impact cyber dominance and information warfare dominance? Because those were the two that Tim had the biggest advantage in that we needed to win. One of the victory conditions that I kind of focused on was to bring the cyber dominance to zero. But because Tim started off with such a high information warfare dominance score, we kind of needed to chip that away as well so that Tim wasn't able to win by just shoring that up. So asking for those hybrid operations and then asking, are there ways to minimize the risk to one of them to increase our chances of success? And seeing kind of how that process worked, I was able to kind of create bespoke operations to get out what I wanted to. Unfortunately, the first time we tried that, it gave us a probability of success for one of the operations at like 75 to 85%. But the system interpreted that as saying that you had to roll the dice between number value 75 and number value 85, which is where we got at. What Dr. Berry talked about earlier was kind of changing the dice roll system.
[00:40:00] Speaker D: You noticed that you had rolled what should have been a success, but it was telling you that it was a failure. And so you essentially audited the system in real time during the game and said, hey, hang on a minute. Can you recheck those? Those should have been successes.
[00:40:14] Speaker C: It's like Dr. Berry alluded to, the system is kind of trying to meet needs and it's just going to find a way for your ideas to be successful. But as a player, it was really validating to kind of understand how the. The things that I was thinking of within the context of the game, the things that I wanted to do, I was able to achieve them with like, I would say my success. My success rate was around 70%.
[00:40:38] Speaker D: And I think what's impressive is the speed and ease at which you were able to do that. So you didn't have to send a message off to a white cell, wait for them to deliberate, come back, tell you whether or not you could. And I think that's part of the real value here, is how quickly you can make changes or get responses.
[00:40:55] Speaker C: It was really nice too, that the system would explain the changes it was making in response to questions or suggestions and how that would impact things. So having that adjudication happen in the open, that conversation that often the white cell has internally, kind of seeing what the logic behind the changes to the system were, was helpful for me as a player, kind of understanding what is reasonable logic for making a change to a game system.
[00:41:23] Speaker A: It's respecting the sanctity of people and the people playing the game in war games. Right. And that the adjudication phase, it's in there to be able to question that adjudication phase and go, well, what logic is that? So with that determinism, it's going to say, well, these are the two premises and this is the conclusion. If the two premises aren't true, then the conclusion isn't true. And show me where the premise is wrong. Right? And if it is, then we're going to change it. One of the things I'm hearing is this idea of trust Gen AI. And I, I don't trust Gen AI. Why would you? But I. The trust DCP is that, yes, it's going to deterministically. Right. Be constrained by it and you can trust that the game is not going to go back on it as Kate and Flipper changing things. That was my biggest thing I was watching for. Is it going to stay within that constraint? Because they just changed it. Is it going to stay in the China Sea or is it going to stay in this future phase and not use some weapon from 2150?
[00:42:17] Speaker D: That's one of the things I think is powerful about. This is usually, at least in my own experience, when I get the first version of a piece of software and you're playing, you're like, hey, this is great, but I wish it could do this. And I wish it could do this. Well, then you have to wait till version two to see if they put those things in with this. You're like, hey, this is great. I wish it could do this. Can you do this? And would you do yes or no? If it can, it puts it right in the game immediately. So you can immediately tailor it to. To the strengths you think will will benefit you in the war game, which I think is extremely powerful. Because, Dr. Barry, you're not a war gamer by trade, so you're not going to be thinking about these things the same way they are. So they get it. You don't have to ask them, what should I put in this game? Okay, how does that work? How does that work? They just type it in.
[00:42:59] Speaker A: And I think that's goes for any subject area. Right. The idea is, you know, with the threat assessments or if you're doing like for a lot of the idea is to make it so that the system is available to the smear the expert to be able to input and do that. And what was nice about today's room, we had some other people in the room too.
[00:43:15] Speaker D: Yeah.
[00:43:15] Speaker A: That were experts in their area that were able to offer advice or to give feedback and that helped out a lot as well.
[00:43:21] Speaker D: Yeah.
[00:43:22] Speaker A: So both these games are beta point one. They've never been in this form played ever before. So this is the first time both games, both Ascension and Flashpoint were ever played. And that was, I thought, fascinating to watch for the first time because you've never seen someone else play it. We have to be able to reach into our strategic leaders, our fortifiers, and find out what expertise do they have. And then we can use these games as a way of collecting the episodic memory of that game and then talk to those people like, well, let's say Kate and Flip are the only ones that beat that game for the next 50 leaders. And we go meet with gos and other people and, well, how come Kate and Flipper could do it and no one else could? There's something about their strategic thinking and their approach that we need to learn that our war fighters need to learn if we're going to be successful to have information and decision advantage.
[00:44:07] Speaker D: Well, I've always said that about those two.
[00:44:09] Speaker A: Yes, that's critical. And right now we have a hard time. Just because you have a certain position or a degree doesn't mean that you're quote unquote, an expert at it. Right. So I might have a Doctorate. Right. But that's in a certain thing. It doesn't mean I know anything about something like gardening. You can barely keep the plant alive. Right. So, I mean, you have to really know where your limits are. So why not have a game that allows people to input, keep that as episodic memory. And then this is when you would tie it to the larger Tim system. Right. Put it in his memory so he remembers today for technically infinity, you can just keep remembering the details from today. What did Kate do on move 7 on this game at 9:08? What did flip do at this? That's really awesome. And it allows us to pass on our learning to future generations. So when I pass away, you know, is there any knowledge in my life that would be worthwhile to someone else? And if so, how could I pass that on with fidelity?
[00:45:05] Speaker C: It would be really interesting as more people play this game to kind of see the choices that the system presents to them as well.
[00:45:13] Speaker A: Yeah.
[00:45:14] Speaker C: Like what kinds of things that certain decisions or approaches, like what are the logical next options for that After. After a certain. Like after someone chooses a certain action? Because it's something that, as we were playing it, I didn't quite realize until we were able to start tailoring the choices that we had, was that when someone else plays this, they're not going to have the same prompts.
[00:45:39] Speaker A: Right.
[00:45:40] Speaker C: So it's not a set game.
[00:45:41] Speaker A: It's dynamic.
[00:45:42] Speaker C: It's really dynamic. And that's. That's kind of a hard thing to wrap your head around. As someone who plays just like regular games, I kind of wonder if the next time we play, we don't win just because the options in front of us, just based on, like, a specific choice or an approach, really changes the path of what is possible within the world that we're shaping and interacting with. And if it impacts our chances of victory, like, that's a data point that I would be really excited to see.
[00:46:08] Speaker A: If we weren't doing the show. The first thing I wanted to do is as soon as I got to the thing, start working on Kate's answers, going, all right, Tim, listen, we need to learn Kate. Got you on that one. How did that happen, buddy? You know, and I mean anthropomorphosizing it, but really the first thing you're doing is finding out, well, how did that happen? Right. So what were those things? Because you also. You also don't want to be impossible, Right. It needs to be possibilities. It's not fair to say we're going to stop Keith, but we can Say, was there a better answer or a better approach in defense. Right. Than what you did? Right. But that's also in that constrained model space, maybe there is nothing left in there and maybe that's it. Maybe it's a different model that we have to use. So, I mean, the hard part too is we play this game today. How do we get games out to the force?
[00:46:52] Speaker D: And the democratization of it is extremely important because as you two know, war games are basically invite only. You can't say, hey, I want to go to this war game and participate in it. They're not going to let you. They're not going to. They're not going to pull you in unless they feel you have some reason to be there. But with this system that you have, I mean, presumably, and we talked about this, it could be incorporated into camo GPT or NPR GPT, which is the systems that we're allowed to use on the government networks and anybody and everybody could play this.
[00:47:21] Speaker A: That's the idea to get it to our folks that need it now, you know, you know, infantry, you know, flip, right. It's been like, it's up, like, you're here now. I don't really need to know what happened 30 years ago right now. I need to know how to, you know, my commander and my. We all need to know what we're going to do today. Right. And we can be almost just shy of real time as we get data because we're not relying upon like adjudication and all these other folks. Right. It's happening inside of it. But that takes again, it takes a team. We are unbeatable. And when we all put our minds to it, right. But war gaming is going to become more and more important, I think, not just militarily, but even in schools. The gamification of education through a gaming approach like this, where you're learning how to do, you know, history, math, all different kind of things. So I'm excited to see today and I was also excited to see, like, how much work I have to do to make sure that Kate has a tougher battle next time. And flips there, like, flip, I'm ready. You know, both of them together. Really interesting how much they work together as one mind in that first game especially, yes, I'm sitting there going, all right, but they're doing a good job. Poor Tim.
[00:48:30] Speaker C: I don't want to take my husband's name in vain, but he is a young naval officer and just played his first war game at a course that he's at and he came home and was like stoked about it and was like, I wish that I'd been able to play one of these sooner. I wish that we had an ability to play more of these in our professional development. Obviously he's younger and in the Navy and so a lot of his current experience focuses on the naval tactical level. But like, when talking about war games and things, there is I think, also a large opportunity for some of those younger service members to start expanding their thinking into the operational, even strategic levels to understand where they fit into that. And a game like this would really help with that.
[00:49:12] Speaker B: Yeah. And I'm just going to springboard right off what you said. I was looking at it through a different lens as a retired senior nco, someone responsible for training soldiers, looking at the implications of something like this for developing, say, a battle staff NCO at a brigade or a division. That's a, that's a big ask for a guy or gal you're taken out of a squad leader position or a crew combat vehicle crew commander position and ask them to increase that level of thinking and understand right away how to compare and validate courses of action. And this seemed like a really easy, a really intuitive tool that you could put in front of somebody that maybe is already more comfortable with a generative AI interface like the, the chat GPT4. Enter just being familiar with the interface and go.
And then be able as a trainer to kind of track their outcome, check their math, if you will, on how they got there before I invest the institution's money into cinema money and time and to send them to that course. And it makes me think of some other things that we've done here at TRADOC to make threat tactics more cognitively available to younger soldiers, younger leaders, like the threat minutes where you have to take these novel approaches and meet these soldiers and leaders where they are so they can, they can grow.
And you'll see the growth. It'll be exponential.
[00:50:45] Speaker D: I want to bookend the second scenario real quick because we are running out of time a little bit. But I do want to mention that Kate beat the AI down so bad that it shot its remnants into space in hopes of preserving its survival for another day.
[00:51:01] Speaker B: Is that correct?
[00:51:02] Speaker C: Well, that was, that was its approach to, I think, the second defeat. Second defeat.
[00:51:09] Speaker A: She got that in because you beat.
[00:51:10] Speaker D: It so bad that it wanted to redo it again. So now let's do another, another ending.
[00:51:14] Speaker C: It did keep asking for do overs.
Well, it offered the opportunity to explore an alternative ending. Obviously it would conclude it would Say, hey, you win. Yay, everything looks great and the world is saved. But also this other reality. Do you want to know what that is? I think we played two additional alternative endings before ending the scenario with a world that had achieved transcendence and humanity was being ushered into this new age of human machine synergy.
[00:51:49] Speaker D: It was the ending of 2001, if you've ever seen that. Kate was the star child. And yes, yes, brought humanity to the next level. So let's go back to raw, honest feedback now. I mean, you brought this game down for them to play and run through and explore. Try to break it, so to speak. So we, they essentially played it all day. I mean, we started around 10:00 and we're ending now. It's almost 4:00. Give your raw, honest feedback, both of you, on anything that you experienced here. Good, bad, or somewhere in between.
[00:52:16] Speaker B: Okay, so I was, I was actually very surprised, pleasantly surprised with the outcomes. I think I had cognitively set the bar low from experience, using a GPT based product to try to kind of skirt some homework at the last minute. So undergrad graduate, and I thought I would see a lot more drift, I thought I would see a lot more artifacts and I, I thought that the results would be a lot more falsifiable than they were and I would have a lot less faith in the courses of action we developed. The way to take it kind of to the next level for military war gaming was Tim was a very good S3 and a very good chief of staff. He would take our inputs, he would direct his efforts toward us, and he would give us very clear, very coherent courses of action.
One of the points to wargaming, especially when you're on a staff, is to. For the commander to validate that his staff can take his intent, turn it into his course of action. So there's probably some room for that work or for it to be developed outside the model and then ingested is probably the best way. I, I think that the easiest way that would go for everybody. But on the whole, I was super impressed and it did not feel like we were sitting in a room for most of the business day playing the game. Time slipped away. It was fun, it was engaging, and it was valuable.
[00:53:48] Speaker D: Yeah, we essentially had to pull Kate off of it so that we could do the podcast episode. So, Kate, your feedback.
[00:53:54] Speaker C: Tim just kept wanting chances to beat me and I couldn't let that happen.
[00:53:57] Speaker D: Yeah, keep blaming Tim.
[00:53:59] Speaker C: No, frankly, I didn't go into it with any expectations. I didn't really Know what it was going to look like or be like. I heard that you were bringing in a tool that would potentially help with war gaming if it was integrated into a lot of our efforts. And I was just like, all right, cool. It's going to be some kind of record keeping tool boy, was I wrong? The gameplay for me, I like board games, but conversational learning is probably my favorite way to ingest and process information. For me, this was one of the most fulfilling games I've played. The way that it challenged me and kind of forced me to interrogate the ways that I was interacting with the system and the information that the system was giving me, it really made me think in a way that I haven't really been forced to think since college. You get a little bit of it when you're interacting with a new professional wargaming system and you have to understand it. But the rules are so rigid that with something like this, a huge opportunity and also a huge challenge for me was that the sky's kind of the limit. Could I ask the system to provide me an option that would achieve total victory in one turn? I didn't ask that. I didn't think that it would be an honest gameplay. The boundlessness of the system was freeing and really engaging and really challenging. But there are also ways that I could see that you could really take advantage of that too. Imposing rules on the system instead of on the scenario.
[00:55:36] Speaker D: Dr. Barry, I want to ask you for your final thoughts as well. How do you think today went? What did you see? What are you going to do going forward?
[00:55:43] Speaker A: What both flipping Kate talked about were spot on. I mean that's the whole idea is we need to, on the military side needs to be a strengthening of that. Right? This is early beta and you said about the flow condition of the day going by. I didn't feel five hours. Like to me it was about an hour and 10 minutes. I think the unboundedness again is that's us. We're the unbounded part, right? And I think where do we decide inside that bounded nature of the game that we do shut off and say, no, you can't win in one one, right? So but until we explore with people, we shouldn't set that bound, right? We need to be with people and be willing to take feedback. Like we just heard today. I go home and think about it and just keep iterating it and then coming back and saying, hey guys, we played a hundred times, can you play this again? And then you'd play it again and then See what improvements? So you basically have the two pioneers that played the game. You'd want to come back to the pioneers of flipping Kate and saying, okay, this is where they are. And you may say, say no, now you've taken away too much. So finding that balance, that's the art, right, of any thing that you do. So the art of gameplay is going to take some time. And I think what we're going to need is you need a champion, right? We've seen this throughout change in anything. If you don't have a champion, then you're able to have any change, right? So it gets swept under the rug. So how do we start democratizing games if we continue to have them as exclusivities because of money and travel and so forth? So I guess that's the question of the call to action at the end of this is how do we take something that we see can grow? We know it's something that's within our valley, work to do, but who's going to be the champion that says, yes, we're going to do that? And here is a phased implementation to democratize that, whatever that may look like. That is the entire purpose of it, right? It's seeing the world differently than you did before you. You played the game. Because the game is not meant to be just a game. It's meant to explore your internal view of how you see the world and challenge your mental models rather than just, you know, pick one, pick two. And I think that's what I heard a little bit. Your mental models are challenged and either it's reinforced or it's questioned. My mental model of the games changed drastically in some areas. Listening, all in good ways. So the opportunity to come down here and do this today, I thank you, Matt, so much for doing that. And you know, being official mad scientist, that's our job, right?
[00:58:04] Speaker D: That's right.
[00:58:04] Speaker A: Our job is to go out and challenge the status quo, but we also also have to find other people that will champion us doing that. So thank you so much and thank you, Flip and thank you, Kate. I mean it was really, it was an honor. It was a just an amazingly. It was an awesome day. I really enjoyed it.
[00:58:20] Speaker D: So. All right, we're gonna, we're gonna wind this down because the air conditioning broken in this building and it's like 85 degrees in here. We've been in here all day. So Kate and Flip can't thank you enough for being the guinea pigs for all this. You know, Dr. Barry came to me and said, hey, I'm going to run you through this game. And I said, okay. I don't know war gaming though, so I need to find the guys that do. And you guys stepped up to the plate today and came and then Dr. Barry, I mean, this was your idea to come down here and throw this on the table and let us run around with it. So, you know, you thanked us for doing it, but you made the travel, you made the game, you had the idea. So we thank you for coming here because I think this was a, a really cool idea, a really fun day, but also a highly valuable tool for the army and the war gaming community as a whole. Thanks for coming in today.
[00:59:01] Speaker A: It was an honor. Thank you very much.
[00:59:03] Speaker C: Thank you so much for letting us play with the game.
[00:59:05] Speaker A: Thank you.
[00:59:05] Speaker B: It was awesome.
[00:59:06] Speaker A: Thank you.
[00:59:07] Speaker D: Thanks for listening to the Convergence. I'd like to thank our guest, Dr. Billy Barry, as well as our analysts and wargamers, Flip Boyer and Kate Kilgore. You can follow us on social media Me madsci and don't forget to subscribe to the blog the Mad Scientist Laboratory at madsci Blog Tradoc Army Mil. Finally, if you enjoyed this podcast, please consider giving us a rating or review on Apple, Spotify or wherever you accessed it. This feedback helps us improve future episodes of the Convergence and allows us to reach a bigger and broader audience.