Techstrong Gang – April 24, 2025

Hey, everyone. Are you ready for McGruff AI to take a bite outta crime? You're watching Text On Gang?

Hey everyone, happy Thursday. Welcome back here to Textron Gang. It's Alan Shimel and we've got a great Textron gang today.

Told you we're gonna have McGruff the AI crime dog. If you know who that is, you're of a certain age. Uh, we've got more, a lot of AI stuff as usual car man.

It just sucks the oxygen outta everything for good reason. Perhaps though it is pretty, uh, pretty big. Let me introduce you to our gang for today, though.

We got a great gang of people to talk about stuff. Uh, first we'll start off out, out west. He looks like he's up on top of the Golden Gate Bridge again, watching over his vast domain, our Silicon Valley man in the know John Schwartz.

Hey, John, how are you, man? I notice anything different About this. Yeah, I could tell you look like you either got a facelift or a new camera.

I'm not quite sure. I got both, but don't tell anybody. Uh, yeah, no, thank you.

You send it to me and I we set it up. So it's, uh, I It's a huge difference, Sean. I know, I know.

Sorry about being in the cage. No, no. Hey, don't worry about it.

But you look, you look fantastic. Thank you. It's great to Be here.

Thank you. All right. Moving over from Silicon Valley to Denver.

Well, not well, somewhere near Denver, right? What is that? Northwest?

Northeast Denver. But you take the Northwest Parkway to go from to Boulder. Actually, no, I took, yeah, that's, that's on the other side of I 25.

But you're close. You're in the neighborhood. It's been a while since I've been in Boulder.

Mitch's been a while. Anyway, he's back from, uh, be asking me for directions, by the way. Yeah, but that's a whole nother issue.

No, let's not go there. But, um, anyway. Weren't you just out in Silicon Valley, though, at AI Infrastructure Day?

Yeah, I was out in Santa Clara at Google talking about AI infrastructure, some of the great things that they're doing. We talked about some of that in reports and great Kubernetes work that they're doing too. So some things they launched at, uh, Google Cloud next.

So Kubernetes man, love that stuff. Very cool. Just a quick, uh, blur.

Um, you know, on Textron TV today, we will be live streaming, uh, I guess it's day two of maybe it's day three, no day two of, uh, of AI Infrastructure day. So stay tuned for that. Speaking of ai, she is our editor for Techstrong AI coming at us from Ohio, Sona Saha.

Hi Sona, how are you? Hi, Alan. I'm good, thanks.

Thank you for having me. It's great to have you on here, and thank you very much. Um, moving on from sag, we're gonna go there.

He looks like he's in his natural element, the data center there, our friend JP Morgenthal. Hey, jp, I don't know if we ever really told people kind of your background, give them, why don't you give them a little bit of a, a JP history, if you wouldn't mind? Sure, sure.

Happy to. Thank you. Um, so, uh, it, it's a long sorted history, uh, that has taken me, you know, from software engineering, uh, through Wall Street Consulting to, uh, dot com startups and huge managed service providers as a, as a lead around services.

And, uh, back through, uh, into, uh, now a, uh, entrepreneurial effort again, uh, providing AI consulting. So it's a, uh, and, and along the way, a, a a lesson learned from, uh, a friend of our all, uh, Dave Comb never stopped being an analyst. So, uh, you know, I actually was a legitimate analyst, industry analyst.

I got paid to be an industry analyst, but even when I wasn't specifically an industry analyst, I acted in that role, uh, because I think it's important to put some context around this technology stuff for, for people pragmatism. Uh, I think we'll hit a little bit of that in today's discussion with the ai. Uh, and speaking of the Kubernetes stuff, Mitch, go out there and look at my hot take from yesterday on Kubernetes, in which I, uh, took on the, uh, you know, Kubernetes is ridiculous.

PAs is satisfactory, and, uh, uh, issue again, dating my, or including my 2014 post about why PAs is the way you should be going. So, uh, uh, yeah, you may love Kubernetes. I'm not, I I think Kubernetes is a great tool for PAs providers, so, oh, you and I should have a good point, counterpoint argument.

I love that. Let's stay tuned. We're bringing it to you, the thriller, the thriller in vanilla.

Alright. Um, We will, we will put that on a future show for sure. Absolutely.

Mike Mike's head's spinning. Um, speaking of Mike, no joy in Mudville, Mike, they lost again. No, but you know, as I was sitting here listening to jp, I realized that I've known him forever.

So I'm suddenly feeling a little older than I was a minute ago. Well, you're both in the primary of your life. What are you worried about?

Anyway, Mike Ard, our chief content officer, he, He needs to spend more time in Florida, Alec. Yeah. Yeah.

Well, no, he, he did, he had enough of it, and, and I had packed the wagon and moved back up. Um, I'm On. Anyway, let's jump into today's first topic, which is, uh, as I said, AI takes a bite outta crime.

Mike, what is this one? Well, it turns out the police department around the country and a few other places have figured out that they can start using AI to engage people who are probably thinking about doing something, shall we say, uh, illegal online, including people who might be trying to cultivate children or various illicit activities. John has a story on all of this, but John, like most things with ai, there's a good and a bad here.

Walk us through it. Yeah, you know, it's interesting. So the headline that you, you circulated was AI takes Bite Outta Crime, but I think this is maybe a bigger bite than most of us would like.

Um, there was a, uh, investigative report by an operation called 4 0 4 Media. They did this for Wired Magazine, so I'm gonna give them all credit. They, uh, looked into a company called Massive Blue.

So they're a New York based company that has a product called Overwatch. It's being pitched to police departments. And basically what it is, is they're deploying these virtual agents that infiltrate and engage criminal networks across channels.

So maybe chats, email, what have you. And what they've done is they've, they've created this, as you said, Mike, to prevent things like human trafficking or even to protect schools. But what's happening also is that some of the police departments, according to this reports, which cited a couple of examples, how are looking beyond the suspects.

The usual suspects now are looking at quote unquote, college protestors, radicalized political activists. And, um, they're looking into it for, um, the case of immigration related activity and some of the, some of the fictional characters that were created. These personas, AI personas included this 36-year-old child free divorced woman who's outspoken, lonely, and body positive, who's also involved in baking and activism.

There is a 14-year-old boy whose child trafficking AI person, persona, and AI pimp persona, what have you, escorts external recruiter for protestors. College. Protestor, in a sense, is kind of a big brother.

It's, it's, it's morphing beyond the original intent. And along these lines, we've come across some other examples Beyond this one company, Palantir is assumed an increased role with ICE to track down individuals, provide logistical supports in mass deportation efforts. This is, of course, the big data analysis company co-founded by Peter Thiel, another company called Geo Group, has assisted the Trump administration in tracking immigrants according to New York Times story.

So, and again, I don't object to the use of AI agents to find legitimate bad guys for the original use, but it's, this seems to be, uh, expanding or morphing into things like infiltrating protest movements. Um, and it, it was a fascinating story, and it just kind of, again, shows you what AI can do and how far it can go. Good and bad.

Mike, I, I, I just, Mr. Mr. Legal, legal, take it.

No, No. I'm not gonna give you legal opinion here. I'm, I'm just going to give you some plain good old Bill Clinton style, common sense.

Don't blame the tool. Blame the fascist. That's, that's an excellent point.

Where I was gonna take this is I really want to divide the story up into two parts. The first is the technology. We are all struggling to understand this AI stuff, but I, I, I, I think this is a great, um, archetype for what this technology does well, right?

It represents personas based on an amalgam of voices that it has seen. Okay. Um, so this is what generative AI technology really does well, I know we hear a lot about using it to vibe code, but at the end of the day, this is the sweet spot for generative AI is creating and emulating that vo different creative voices from, uh, the history of training.

Okay? Now, the second part of this is how is it being used? And I think Alan just splendidly ex, you know, gave a representation of, you know, what we should be thinking when we see these technologies in play.

I have a question, though on the legal side. So I will toss this to a, where is that fine line between enforcement and entrapment if I'm creating an AI agent that Engage someone? So again, is it a fine line?

Yes. But is it a, uh, a well litigated line? I would say yes, ai, whether, whether it's an AI created persona or an undercover police or, or law enforcement officer, uh, taking on that persona, you know, the, the, the line between entrapment and, and actually being, you know, criminal activity.

Is it, it's not a clear red line, certainly, right? It's, it's a shade. The shades of lines there.

You Know, maybe it, maybe it kind of collapses the lines a little bit or maybe come together, Alan. 'cause what I'm thinking about is, I don't know that we're approaching minority report, but you can use AI to take information about publicly available information about people, their, you know, so, so their social media, et cetera, and then build the persona specific to that one person you could really personalize. You think they weren't doing that already?

You think they weren't collecting public data and then, oh, I Think They're doing it building a money platform, But then you're dependent upon whoever the undercover person is, their ability to Emulate that. Certainly AI has made it easier to do. I think it could more efficiently, effectively Emulate that.

Yeah, it's made it easier and it's probably more effective. Mm-hmm. But, but here's the thing.

It's the old story why we can't have nice things, right? This is, this is a great technology to catch child traffickers, child pornographers, uh, criminals, right? And who doesn't want to do that?

Do we wanna clean our streets and I guess our internet of, of criminals and these kind of scum who are doing these kinds of things? Absolutely. And if AI and AI is a great way we could use to do it, but if you don't have safeguards, if you don't have individual rights, if you don't have things like due process, you, and, you know, you don't hold up undue search and seizures and, and you know, you don't have that backbone that made America what America is, the, the potential for abuse here is so great, is so great that you almost, it disqualifies all the good it does if you can't keep it in the guardrails that our constitution and our democracy is supposed to guarantee us.

And that really is the issue. You know, with so many protests going on at weekly or even twice a week in almost every city, I think it's gonna be really tempting for certain individuals in law enforcement to use this tool to, to infiltrate. And I'm not saying it's, it's happening now, but I think there's a great temptation for where it will happen and it will escalate.

So yes, there are great uses, but again, is you can never leave well enough alone. This seems to happen with technology, not just with ai, but previously they're just, they're just, the boundaries are, are are broken. One of our founding fathers, uh, I forget who said the quote, said, better that a hundred men go free than 100 No better, that a hundred guilty men go free than one innocent man goes to jail.

Yeah. That, that's a principle of our system. com era where the killer app wasn't porn.

Obviously you haven't been playing with AI very much. It's, it's coding, it's not Porn, it's coating porn. You know, there, there're on goat Porn, I, I I, not that I ever look mind you, but there are plenty of AI porn applications.

There are. It's just when you, when you think about it, it's just another persona, right? Yeah.

It's exactly. And that's, that's one of the personas in this example here. Yeah.

I mean there absolutely porn is a driver of it, but Gambling and porn, that's what built the internet. And kittens And kittens and cats. Yes.

But you know, again, seriously on this, we, the, the, the, the abuse here is, is it, it's a shame because, you know, can AI help us identify serial killers, kidnappers, you know, all of these kind drug dealers. But Profiling is, is huge opportunity for profiling. Um, is, is, but again, even, you know, if you, you know, understand what, you know, shows like Manhunter, the history of the start of the FBI's, um, you know, a, you know, part of the agency that that does this social profiling and behavioral profiling, what you have to understand is that it, it's just a baseline.

It doesn't mean that somebody can't operate. And all that AI's gonna do is, you know, speed that time to that profile up may recognize if you give it enough data, things that a human might have missed, uh, about the case and, uh, provide additional, but it doesn't necessarily mean that the criminal themselves isn't operating or staging that persona, staging that behavior to throw police off. Well, I think we're missing to that point, uh, jp there's the other side of this is you, you add a, we're talking about it as exists today.

Add agents to this. And you could very easily use AI to go out and create a persona or an addition to an existing persona. I could go out and pick one person and say, no, they don't have a blue sky account.

Alright, lemme go set one up there. The agent goes out, starts posting things we want, we could actually start to build something we could extort them with, or that, you know, we talk about planting evidence on someone, a nefarious person in law enforcement could do something like that to, to, to add evidence, quote unquote, against someone. So, uh, we're, we're in a precarious position of this whole identity, and what really is you and what did you really do can easily start to blur, uh, with the use of ai because it can not only create personas, it can simulate activities and actions of those personas.

And I think that's where it's gonna be really difficult to un untie that knot. Yeah, no, that's, that's a great point, Mitchell, but again, let me be clear. It's not AI that's doing it.

It's the person using ai. Le let's look at another angle on this as well. I find interesting as hopefully our audience here will, um, and that's the cost and the overhead to doing this.

We could talk about a group like Massive Blue saying they can do this for the New York Police Department, but think about the scale of that. Think about, you know, the requirements for the underlying inference engines to be able to do that in a timely fashion, right? To be able to go out and scrape all this data, and each one of those going through the LLM doing the, uh, you know, uh, the, the, the inference do, generating the next level of content, taking that netting it to the memory, going back around, going out to the next, and iteratively going through somebody's Facebook, and then iteratively go it.

And that's one person. You multiply that times what the scale that they're promising. N-N-Y-P-D-I don't think we have the infrastructure to deliver this, uh, in today's market at an affordable price.

So it's a great concept, but who's paying for it? Well, first of all, that price is going down. Like, you know, JP every Day, well, 10 years, it'll be 3 cents.

That's not my point, Mitch. The point is, you, what? It's, we're again, looking at futures and everybody's talking about them now, like they're here and it's doable.

That's not a scalable design in today's market without at least $500 million of overhead, you know, annually going to being thrown at this at the scale of NYPD. Well, if I'm on the criminal site, then I get bots to go compromise accounts, get free accounts, and all the LLM services and do all the things I want it to do. Right?

I guess law enforcement could do that as they wanna really act criminally as well, so, well, Yeah. So I, I think the money from the, from the Leo side, you know, from the law enforcement side of the house will be allocated, right? It will be allocated.

It may, it may take a little while, but that money's gonna go into it. But Mitch, your your point is, is a point I think well taken, which is don't think the bad guys aren't using this too, right? So, I have a hypothetical though, Just a hypothetical that'll that's admissible.

Go ahead, counselor. There You, there you go. So don't I have to have a victim to have a crime, but if the victim is an AI agent, there is no victim because the ai, you've Never heard of Victimless Crimes.

I'm just putting it out there. I've never heard of aliases. Hmm.

So what, what you're really asking is how long will it be until an AI has rights Or I can't indict an AI agent, so, oh, there's, in, in Inevitably there's gonna be a lawsuit, right? Somebody's gonna say, I was, I was entrapped by by a virtual agent. You know that that's gonna come And Yeah, But a but a lawyer's gonna stand in court and say, well, where's the victim?

I don't see any victims just saying, NVIDIA's the victim. Wasn't There a movie like this already? I'm sure there was.

Well, how many crime shows or movies are like some, some something, uh, thief or victim or whatever, that isn't really a person. It's So a child trafficker goes online, finds this Agent Bo, that's ai, and as a whole conversation where it gets caught goes to court. And defense argues he knew it was an AI all along, Right?

So he didn't really do it. I mean, star Trek next generation data doesn't remember there was that whole trial of is data a being or can they just move him along and stuff. They question every, the entire series is questions on alien, sentient.

Mm-hmm. No, but data's not an alien. Data's an ai.

Right. Great. Well, I look forward to the show when we're discussing that particular topic.

Next. Well, everything I need to learn about everything I needed to know about ai, I learned watching Star Trek. So right After we debate passing Kubernetes.

Yeah. All. Alright.

This is descended into, whoa, What state is Context Window. That's what I wanna know. Yeah.

Alright, We're gonna take a break here on Text Drunk Gang. Let's talk. Ooh, when we come back, we'll talk a little bit about AI coding battles.

You're watching Text Drunk Gang. Hey folks, we're back and talking about one of our favorite subjects, again, AI coding tools. There's been a raft of them coming out lately.

Everything from Elon Musk and Grok tools to open AI is now trying to theoretically at least buy a couple of companies in this space. I guess whatever they're currently doing, coding isn't quite living up to their expectations. At the same time, companies like Docker that people already use to build applications, they're now injecting AI coding capabilities into their tools.

Mitch, my question to you is, will people go get new tools or will they just wait for the tools that they currently know and are familiar with to get AI capabilities? And maybe I'm not gonna need all these fancy new startups. Uh, grasshopper, you must snatch the pebble from my hand to get that answer.

Now, to, to J's point about coding being the killer app with ai, there's a couple of tra trajectories that, you know, are they on a collision course? Probably. But the paths that are happening, and this is an area I spend a lot of time in in the analyst business, is enhancing the current way we develop code.

In other words, enhancing ides, either with a command line component to talk to the LLM and do vibe coding or, um, doing agent development or directing agents through it, whether it's, um, GitHub agent, agent mode, things like that. And then there's also the new IDE now windsurfer is an IDE, it looks like, uh, uh, vs studio or vs code, excuse me. Um, but effectively, it, it, it functions much like you would use Gemini or, or kind of any LLM type tool to generate code for you and taking you more out of the normal kind of technical environment of an IDE.

That's the whole vibe coding environment, which is just move into a natural language type of interface. So that we're, we're moving from generative AI was, you know, fancy as a chat bot, then as fancy as a copilot. Now it's fancy as a studio product.

Everybody's coming out with their studio products, whether it's Google or, or anyone else for that matter. I, I think the real, the real challenge here is the AI companies, the, um, open ais of the world, the, um, anthropics are moving into being development tool companies, if you want to call 'em, they may not be for just for developers, but that's essentially what they're becoming. And that's why Open AI is, I think, trying to leapfrog and catch up.

Uh, but talking to Windsurf, potentially buying them, they'd looked at, um, I think it was, could, could city another company. I don't remember which one they looked at before that, I think. Oh, that's right.

Right. Yes. Um, and so who knows what was, well, where they'll buy, I think its directionality is they're becoming the next generation of software development companies in a new form, in a new format, in a new way of doing it.

But that's the trajectory. Now, you talked about Docker. Docker is very much on that first path of let's take the tools Docker desktop that we, uh, developers use all over the world.

Very, very popular type technology. And let's add not just containerization, but let's add ways of managing, um, MCP, um, MCP instances of different implementations of, of model context protocol. Because it's, it's a nice protocol for accessing certain things, but you don't, it isn't one kind of universal server that runs for everything.

So that complexity is increasing by introducing MCP across untold number of environments. So Docker stepping and say, there's an opportunity here. Let me create a catalog for you.

Let me do some AI agent integration to manage all this. So we're kinda racing down all lanes all at once. And I think there will always be sort of an IDE based always is relative, um, around for things that are very much today's code centric until people start to migrate to more of the natural language interface.

And we'll see if that takes over. But that's certainly what's competing for developers, eyeballs and end users for that matter. Yeah, you know, I'm sorry, go ahead.

So, Lana, were you gonna say something? Um, uh, yeah, sure. Um, so yeah, definitely all these companies are launching their own, um, set of products for development.

And, uh, but I still feel like AI assisted coding is sort of a marque area. Um, so it seems like they're, uh, the developers are using, um, AI for coding in, uh, two key ways. Like first for, um, prototyping.

They start with a rough concept and get AI to generate initial code base and get a working prototype ready in, not production ready, but decent for, uh, getting early feedbacks. And then there is another set of people that are using tools like, um, a Windsor for co-pilot and, um, in their everyday, uh, development workflows. So these people use AI to, for code completion, code generation, refactoring, and, uh, generating tests, documents, what have you, uh, basically as, um, a co programmer, uh, both ways are good, uh, for boosting uh, the speed of delivery as long as if there's a human developer in the loop using their years of experience and knowledge, um, to manage the outcome.

And, um, that's the only way to really keep the code maintainable. Uh, 'cause if they're not, uh, which sometimes entry level developers fail to do, uh, in that case, debugging AI generated code is a massive headache. And not to mention it also has some major, uh, security implications.

So I think, uh, it all comes down to, um, how these companies that they're, that are building these solutions are able to address these issues. 'cause for, uh, non-technical people who are still using co uh, using, uh, AI to code, uh, for them, it's like majority of the way it, it kind of, AI definitely helps, uh, getting to the, to the closer to the finishing line. But that 30% of the way it's still, it's where all the pain points lie.

Fair. You know, I, I look at this, right? And if you look at the whole world of things AI can do, right, of what the potential for how it's gonna change civilization, our last segment about crime and abusing it or using it, right, for, for good there, we spend an inordinate amount of time about how AI works in coding.

And maybe that's just because of who we are, right? com cloud native now. And we, we, we live in the tech world.

And, and to us, the idea of AI writing code, of, of machine, of code writing code is I guess, a paramount consideration. But is it, are we putting too much emphasis on that being what AI is for? Well, I, I, I absolutely totally agree with this because when you look at the work of what developers do, such little time is sent in front of a keyboard actually working on code.

There are so many other tasks and, and that's, we're just talking about the individual productivity of one individual developer of what AI does is a co-pilot that's working with them. 7 sonnet and the, um, the cloud code, uh, preview and also some of the releases from GitHub is, let me take the steps for you to do these things. Not just generate code, but let me write the test now, let me run those tests.

Lemme set up the environment to create all this. It's those tasks that also take time. And that's just scratching the service code reviews, all of those things.

A lot of, a lot of spent time is spent on triage and understanding how these things connect together and where the issue is and where which code bases it is. And it's stuff that I don't maintain. It might be elsewhere, it might be an open source, whatever it is, is the kind of full conte context of the code that you're working with or the system that you're working.

It could be in a config file, right? It doesn't have to be code. So it's, that's the real work.

And I think that's, that's where the productivity is. But I think even beyond that is if you can make a software team more productive now you've amplified the impact of AI beyond single individuals, individuals helping their own work. It's like everybody's got a be better spell checker.

Well, let's, let's, let's talk about grammar for a moment. And that's what we need to really produce software. And that's, I think where the real benefits of AI can make a massive difference.

I think you're spot on, but I take it one step further, I'm just dubious that all these little startups and or open AI and philanthropic for all the money that they have are gonna have become the dominant AI coding tool vendors. Because frankly, there's already people well established in that space, and none of these guys have a sustainable competitive advantage over Microsoft or Docker or any of these people. So I'm kind of like scratching my head going, I'm looking at all this money being port in this space, and I'm kind of saying, all these guys are gonna get rolled up before you know it.

And I feel like the process is already starting. But jp you may think I'm, I, uh, I, I I like that there's these tools available. I mean, early on, I and I speak for I think a bunch of us, you know, it's, uh, it, it's faster than going to stack overflow and looking up the answer and cut and paste.

Um, I don't always like that it how it modifies the code, but I, you know, uh, but I'm an advanced developer who's been programming for, you know, 20 plus years, right? So I have, you know, I have expectations of how the code should look and the, um, and basically, you know, also you didn't write the code. So now what you find yourself doing is taking time to understand.

So if I go out and I vibe code, right, and it generates a chunk of code, I don't just accept it. I have to go analyze it. I have to look through it.

I have to decipher what it's, what it produced, and decide if that's good or not. So, um, you're shifting, I think, some of where your time is spent, right? You're becoming more of the analyst over the output of what these things are creating, and you're spending less time fingers on the keyboard.

I, I think that's okay. Uh, uh, and in fact, there's aspects of it that's really good. Because if you can actually, you know, in your head breathe through code and, and, and analyze whether it's gonna produce the output you want, then you have a pretty good understanding of the underlying fundamentals of the thing you're building.

I think that the issue is that we're seeing, again, hyped promises of people who don't know what the heck they're doing. Look, I pressed the button, it generated this for me. Okay, great, there's a bug in it.

Go fix it. I'm just gonna paste the bug into the LLM and ask it to find it for me. Well, that's where hallucinations start to take over.

And the results that you get are, I would say 50 50 as to whether you're getting a valid output or not. So, you know, pros, cons, interesting development. I use it, it helped mostly the LMS I use still can't write any WordPress, so it's kind of fundamentally useless for me.

Um, and you know, because it's like, all right, build my website. All right, here's some HTML. No, I don't want HTML, I want WordPress doesn't understand WordPress.

So, um, you know, with regard to what you get, uh, is still, you know, within a, a certain realm of, you know, what's popular, JavaScript, you know, Python, some c plus plus C sharp, if you're lucky, uh, a little go some rust, you know? But you know, the more esoteric you get in the environment, the worse your, your results get, Right? So John, you want Coke, but they're only serving Pepsi.

Is that? Well, but, but I think, here's the thing. You gotta remember, guys, today's April 24th, 2025, April 25th, 2028 or 29 or 2030, you're gonna look back to this day and this age and say, wow, it was really infantile then, wasn't it?

It was really immature. We're not even in the gly teenager adolescent phase yet. We're still in the baby crawling steps phase.

I don't know. We've learned to walk even, right? And look, All story you can say that, you can say that about AI in general, right, Ellen?

I mean, every, every, no matter what the topic is around ai, it's like, it's just, just starting, it's just picking up. I mean, we're not exactly a couple of be quantum leaps, Right? The hype machine always takes over.

But when you have CEOs saying, when you have CEOs saying in six months, all, you know, eight 90% of the codes can be generated by ai. No, it's not. Yeah.

Yeah. For one developer, maybe, Wait, hey Mitch, what do you think the actual percentage is gonna be on if, if you had to make a guess? So if you, if you conclude code completion, I think you're in the probably 30 to 40% now people use out, maybe it's higher than that now, already when you talk, if you're saying they're doing vibe coding and it's building all the code for you to, to J'S point about, yeah.

Now when there's an issue, 'cause I've used some of these tools and, and when there's a bug, it says, okay, I'm gonna fix it. And it sits there in an infinite loop trying to fix the code that it can't figure out how to fix, right? It's, it's still early.

We we're, I think we're years away from it's doing everything. It's doing all the code and we're just doing analysis of that. I, I also, what what's also interesting about this industry is my view, and I'm not discounting the money that's invested in it 'cause it is literally billions of dollars.

The LLMs themselves, what's the differentiator between all of the people creating LLMs today? You know, may and, and how many developers have the time to go figure out, well, should I use Lama this for that? Should I go use Gemini this for that task, this kind of code?

You know? And, and tomorrow it's gonna be different anyway because there's gonna be a new version of the next thing. So to to say that the, those, those AI companies are gonna become development tool companies, I kind of don't get, don't get the connection there.

If, if all you're gonna do is a vibe coding and, and entering natural anguish text and you give it a kind of a nice gooey to to it, I don't think that's where it's headed. And I think as we get some true leaders in secure code, quality code, things like that, that start to catch on, that people actually use, then you'll see some real trends and some sifting out in the market. But we're not there yet.

Not even close. So, yeah, so this is an interesting, I had my my analyst hat popped on last night. 'cause I saw yet another agent development platform, uh, AI agent development platform.

It must have been now the 25th or 26th that I've seen. I mean, even when we had app servers emerge, we had seven, this is 25 of the, of the, of a tool. Now as an analyst, the one thing I try to do immediately, and I've been able to do this in the past with different technologies, is identify the differentiated and the, you know, in, in the past most often you could find a different angle, a different hook.

Oh, I see you're doing it a little differently than this one because you're doing X, Y, and Z. You look at all these agent development tools and the level of differentiation is almost nil. So they all, I mean they're all approaching the same, you know, problem domain from the same perspective with the same tooling.

And they, and they, they're just, you know, emerging on the market like, um, or, or multiplying like bunnies at Easter time. So JP let's, let's assume that trajectory stays the way it is. 'cause I think it is, is for a while.

Then the, then the, the differentiator, uh, shifts to, well who has the developers already? Now you're talking about the GitHubs and people like that, or Microsoft that have Well, biz Microsoft, right? Well, yeah, exactly.

Wow. It, It's who can build community. Sometimes they don't think they're, but, And interestingly, N eight N has one of the largest the, uh, agent development communities, uh, right now in, in all of the AI development.

And that's because they emerged as a o open source tool and they built the open source project and managed it and built a community of people who were using this tooling. Uh, and, and it's an active community. It's really hard and everybody wants to do it, to build an active community.

Um, but it's, it, it, it really is just like finding gold. It's like you hit upon it. You, you got in there at the right time.

It's a matter of timing. It's pure luck to some degree. Your thing caught fire, right?

Why did your thing catch fire? I, I, I deployed N eight N It's okay. I I, I, it's not the best.

Uh, I, I, you know, I think this is emerging techs and maybe those other companies will start to, uh, be rec recognized for what they do. But it doesn't matter right now, the, the volume. Look, if the best, if the best software or best company one, we'd be using OS two today.

So let, let, let's move along. We're gonna take a break here on text. We to be using Vista Warp.

Remember the guns on warp. Anyway, we're gonna take a break. We're gonna come back.

Let's talk about digital AI twins. It's not Arnold Schwarzenegger and Danny DeVito, but we'll be right back. You're watching Textron Gang.

Discover Textron Group, the epicenter of tech innovation. We are your go-to for reaching IT leaders and practitioners worldwide. Our secret impactful content that sparks awareness, engagement, and top quality leads with us.

You'll access editorial websites, streaming videos, virtual events, custom content analyst research, and more. Join our satisfied clients. Let's revolutionize your tech journey.

Contact us today and tell your story to the world in the most powerful way with Textron Group. Hey folks, we're back and we're talking about, well, digital twins as advertised. And jp, I'd love to get your take on this.

We have a story over on digital CXO talking about how AI might make this, uh, methodology or technology or whatever you want to call it, more accessible to more folks. 'cause I feel like digital twins had a huge promise early on. And it works great for companies like Boeing, where I'm building some massive engine, but getting it out of the edge where the average maintenance worker is, has been a little more challenging.

So can we push digital twins out to the edge now with a little help from ai? Uh, first of all, digital twin conceptually is still is a, is a great rep, you know, way to think about how to design and build and test, uh, in an engineering and non-engineering context, right? You could create a digital twin of a city.

You create, so civil planners could be using them that the number one issue. And, and, you know, GE Digital when they were around were, you know, driving hard at, uh, at this technology, you know, enabling companies to build and deploy and manage. And, and it's a, it takes a lot of philosophies and it's bringing them together.

It's, uh, it's about componentization and more than anything, it's about creating the, uh, physical representation in a digital form and then having it respond. Now create that is where I think, uh, the industry had difficulties implementing this technology. It's data intensive.

I mean, and in many regards, you were building a model and it's a, you were building the NLLM right? Now, you here, you, or we're learning now from the masters of creating LLMs, right? As they start to grow and get more, uh, uh, larger and put more context through what it really cost and how difficult it is to prepare the data and train it and test it.

And what that all comes down to right now, you take that and you shrink it down on a smaller scale, and you had to do that for your digital twin. But where was your data coming from? If you were planning out the, you know, how a a your engine was going to operate the digital twin, great.

Where is all that data coming from? Preferably, you're loading up some sensors, you're tracking data over a period of time, you're loading that in great, on a small scale, really difficult, on a larger scale. So what AI can do now to help digital twins is that it can go and extract that data structure, that data enable that data for, uh, to help define the twins.

It's, it's really a model, creating a model. So we're using LLMs to define digital models for digital twins. And I think it makes the, uh, the technology more, you know, more widely available to companies to now use because that overhead, uh, that data ops that was required initially is now the overhead to get there is significantly reduced through the use ai.

So yeah, I think they are a good pairing. Mitch, I thought you put up something on LinkedIn I saw about making it easier for ll for companies to create their own LLMs. Now was was, was that you, I don't think that was, I don't believe so.

Oh, I was another niche I'll take credit for, but I'll take credit for it. Not, But, but, but it is becoming simpler to take the foundational LLMs and customize those with my own data to create what feels like a custom LLLM for that specific use case. And I think, you know, I'm seeing that in spades everywhere.

Especially funny enough, as we go into RSA, there must be at least a dozen companies now that are kind of doing something equivalent, whether creating an LLM for cybersecurity tasks based on an open source LLMI think digital twins will be the same. And it might be that a vertical industry, like I don't know, retail or the airline industry, might get together and create some foundational LLMs that they're gonna share that are trained in their particular industry. Jp, what do you think?

Uh, I've done it. 2, uh, Quanta quantitize to, you know, four bits, which means that I, I have the memory to be able to do that. As you start to, uh, as these LLMs get larger, even the quanti quantized ones, you need more memory, right?

I mean, the next level up from that's, uh, seven gigabits. The next one up is 13 gigabits, then 48 gigabits, right? You, you, you need hardware to back this up.

Not necessarily the GPU. You can do inference on CPU, it's the vector space, uh, uh, that you're, you're searching that needs to live in memory. So you need more available RAM to be able to really run these things at, at higher efficiencies.

But I think you can find a nice middle ground, 13 gigabytes. You can do some PEFT on it, which means that you're actually modifying the weightings based upon your particular use case. You could do some prompt training.

Um, and between those two, even with on a, on a low end, I think you ha you can get really good results. Um, cybersecurity is an interesting one for me. Uh, I actually built a sim years ago when I was with Kinetic.

We were gonna, we were building our o we were partnered, but with, you know, building a sim for the DOD. Um, and uh, you know, the, you know, looking at things like how do, how do you efficiently rep represent and understand a low and slow attack, uh, is really difficult. And I'm, I, you know, I'm trying to think through, is the LLM able to recognize a pattern like that because it's not really well known now.

It's getting better with agents, right? And you have memory, but you are watching over a long period of time. And then you're identifying that pattern.

Is this something that AI is really gonna be tuned to do? Or is it better, you know, something quick like, Hey, I just saw three events, LA failed login events. Is this an attack or is this just somebody who has fat fingers?

'cause me, I'll log, I'll mis log in at least 10 times on my phone, my fat fingers. So I think the inflection point is if we begin to see LLMs that can adjust the, the weights and measures in, in, in, in the models themselves. 'cause that's the part that the humans do today.

That's the, that's the training, if you will, that we actually do. And LLMs can play a role in generating prompts, creating data, all kinds of things to help train models. Uh, but models don't create lms don't create other elements by themselves.

'cause they can't do that part of the work yet. Not saying it, they won't at some point. But heuristically, how is that emulated or, or done?

Well, the interesting thing there, mish is, uh, and I learned as recently by talking to the, uh, to Claude, uh, the RAG can do this because RAG actually creates, uh, a mini vector space on the data that you just provided through the rag. So it, it searches that space as if it were part of the LLM. It, you know, you don't need to go and, you know, and modify the weights.

If you add a document, it will actually be able to index and search that document relative to the what you're trying to achieve. So if your data is something that is r able, is that a term? Now we're ragg able, if your data is rackable, then you have the opportunity to say, you know, go out, get me the last 5,000 events from Datadog and I want you to analyze it and tell me if you see any of this.

Unfortunately, everyone else can go see those 5,000 events too. 'cause you've now uploaded it into, you know, No, no, the Alan we're talking about doing this, this on your local, you're talking about doing this in, in your data center, right? I have a single rack with, uh, you know, a co maybe an H 100 in there, just speed things up and I have a ton of memory and a bunch of SSD or you know, or memory based storage, right?

And it's fast as anything, right? And I'm using a Reddit for my, for my memory for my agent. I can do pretty well, pretty locally just for that scenario.

I'll tell you what I am looking forward to because most of the twins, the digital ones are reflection of some static environment and that's helpful. But soon I expect that I'll be able to poke at something in the digital twin and it will change the real thing for me automatically as it connects to a bunch of AI agents and various automation frameworks. And I'm not gonna have to physically go out there and do much Anything.

So we came back to AI porn. I Thought you were gonna go to baseball Or something. Yeah, that's I do.

We're go to baseball guys, I gotta pull the plug. We've got what a great show today. John Sagid, JP Mike, thanks for joining.

I hope you guys have enjoyed it. Remember, we've got Text Drunk TV coming up right behind this with a full lineup, including AI Infrastructure Day Live continuation. Um, but until tomorrow for Text Drunk Gang, this is Alan Shimel.

Have a great day everyone. We're out.