Techstrong Gang – December 16, 2024

Hey, everyone. Is it Christmas, yet? You're watching Textron Gang.

Hi everyone. Happy Monday. It's Alan Hummel.

Can you believe we're at the week before Christmas? I, you know, I turned around, I blinked. It was Thanksgiving.

I turned around. It was Halloween, and just, it feels like last week I was on my boat for summertime. The Willis' had come down to visit.

It was a beautiful day. Here we are. I don't know how I, you know what they say As you get older, the the days we just start running into each other and it then, is it true?

Can't believe it. Looking at 2025, we've got a lot going on. Speaking of 2025, predict 2025.

Be there. You don't wanna miss this. We've got Daniel Newman keynoting, and we've got my friend Caroline Wong and Jennifer Legio, and I'm doing a CPO panel on app dev, and it's gonna talk about app dev and ai, which is something we're gonna talk about today.

We got a lot there. Go to Predict 2025. com.

You can register for it January 9th. Um, but let's get to the business at hand here. We've got an amazing gang for this Friday.

Joining us. We, we had to literally pull him by the ear out of the tables in the sports book in Las Vegas, but he said he can give us a little bit of his time while, while some bets are are humming. It's our resident AI expert, the one and only Bcha ga loop.

John Willis. Hey John. How are you?

Hey, Alan. You know, when you, you get to Vegas, you say Vegas, baby Vegas. When you leave Vegas, you're like, yeah, Vegas, baby Vegas.

So, um, yeah. Hey, hey, everybody. Good to be Here.

You know what? I spent a week in Vegas around reinvent John. I, I am not gonna lie to you.

I was, I was kind of really happy to be going home after a week. Yeah, No, I think that's the, the way it goes. Yep.

Um, all right. Joining us, I guess I, I'm hoping he's home down, down Texas way, future analyst vi, visible impact co-founder, guy Curer guy. You home?

I am home for the year in the cool, cool. Austin, Texas. I think it's high of 65 today, but where, uh, for one day this week it hits 78.

Welcome to Central Texas. Yeah. You know, it's, I'll be honest, it's, it's crappy out here.

It's windy and cold here and overcast, like in Boca. This, it's Same here in Austin. Yeah.

No, no boats this weekend. Um, I have to find something to do. Anyway, let's check in on the weather high up in the, the, uh, Sierra Madre Mountains there in, in, uh, New Mexico, JC Reagan.

How are you? I'm doing great. And we are having cold weather.

I'm not riding my horse much, but, you know, I'm getting stuff done. It's a good time to stay inside and do some work. It's a productive time.

Uh, you light a fire and it's cozy inside, and I'm not feeling guilty of about not being outside, so it's all good. That sounds, that sounds downright romantic, even very nice. Um, we don't, don't really, I have a fireplace we've never used who uses fireplaces in Southern Florida.

I don't even know why they built it in my house, but anyway, it was there when I moved in. Um, but Tracy, welcome and thank you. And then joining us from New York.

Mike, how'd you like Soto in the Mets uniform? Did it make your stomach turn? Oh, You know, the only thing that would've been worse would've been a Red Sox uniform, so, Hey, I can, I can, I can live with that a little bit.

And, um, hey, you know, in New York, I made my, uh, first trip to Brooklyn since I got back here, and I went to see they might be Giants over the weekend. Good show. Very cool.

Very cool. Oh, wow. They, I love, like they might be giants.

Yeah. The great movies. Yeah, it was.

All right. Very cool. Going down to Brooklyn for a little culture, culture, culture.

There you go. All righty. So, guys, let first of all gang members, thanks for joining us.

Thank you for watching. Let's kick up our first, uh, block for today, which is, you know, app dev, app dev on steroids. Thanks to Gen ai.

Mike, you want to Yeah. Set the table? We Sure.

We've been talking about this as a theory anyway, but OutSystems went out and surveyed 1600 IT professionals and ask them, the pointing question was, are you building more applications now? And at what rate? And if is it faster?

And the survey seems to suggest that it's happening, and a lot of these apps are being built with gen AI tools. They're not being automatically built, but they have humans involved still. But Tracy, I guess as we head into 2025, I'd love to get your opinion, but are we on the cusp of some sort of application development Renaissance period that's coming up because of Gen ai?

And are more people gonna be building more software than ever? Well, that would be a nice Christmas present to the development community if we were gonna be building more software than ever. Right?

I know a lot of people who are looking for work, so let's hope that's the case. However, I have to say that we see these trends every time. A big change in technology happens.

We see, we saw it when we went from the mainframe to getting off of green screens into a crazy distributed platform. We were not prepared for everybody started rebuilding their applications. You know, I, our house here is about 20 years old, and we're now just starting to remodel it.

I feel like software has a shorter lifespan. And anytime we see these big changes in technology, companies understand that they need to shift and start building on that new technology. It gives them an opportunity to fix problems that they've had in their old software, even though they may be building the same exact solution, they're building it on new technology, and they almost are required to in order to stay relevant.

So if they're looking to stay relevant in the industry, they're gonna need to be able to keep up with the technology. We saw it again, when, and I talk about this because I had so much work during this time when DB two became the thing and everybody was getting off of flat files. There was not a company in New York that I spoke to that wasn't rebuilding their applications using DB two.

Everybody was doing it. Every single application out there on, in, on Wall Street, all in the banking, everybody was getting rid of these older applications. Kubernetes did that.

We started seeing, you know, when Kubernetes hit, everybody started looking at breaking up these monoliths, and they're still doing it. So I feel like this a, you know, AI and what we're doing in, um, with Code Generation is a tool that allows companies to remodel quicker. They're bringing in it's, it's gonna happen faster.

So I think that these cycles are gonna happen on a more frequent basis, maybe not five years, maybe three years. I don't know how the, the light, how long it takes to retire an application here. We probably retire.

Uh, we, we, we update code in a big way about every two years. We rewrite everything of almost every two years, except maybe some core components on the backend. And I do think that the new, some of the new technology that's coming out there is encouraging for allowing companies to remodel more frequently and more efficiently, because there isn't any way to stay, uh, relevant in your industry and and service customers and be agile, um, if you're gonna stick with old technology.

So they have to, there isn't, uh, I don't think that this wave is bigger than others. Um, I just think we're talking about it more because AI has been such an intriguing technology to us for so long. So we'll see this, but interesting.

It'll, again, This is interesting, Tracy, because what occurred to me, uh, so I, I, you know, I think I live in a different world than with, uh, uh, you know, developer teams and customers that, uh, essentially recode 90% every two years. That's remarkable. But that, that's, that's a testament to you and your team and how you work.

I think that's unusual. And it occurred to me that there are lots of enterprise application, uh, uh, uh, initiatives, uh, projects that just languish because of a need to keep the application going, and a hesitancy about recoding. Everybody wants to recode, um, but they just don't have the time or the talent.

And it occurred to me, um, during this, uh, you know, reading up for this and during this discussion, that maybe ai, because of how it can improve quality and potentially speed, might be opening up a lot more of the code to being not just refactored or migrated to better, you know, platforms, but actually literally recoded that wouldn't have been recoded before. Uh, it's, does that make sense? Or Wait, I wanna, I wanna, I need to split this conversation, right?

Because if you look at the out system survey, they're saying that these are new applications, and then they're also delineated how many people are working on, um, refreshing older applications. Oh, that's right. Yeah.

You Know, the survey kind of leans heavily into, there's much more new custom applications versus just remodeling of the house. John, I know you've been walking around talking to people, but Yeah, no, whatcha seeing? I mean, it depends on what you looking at.

I, I just want to, I, I agree with guy on this because I mean, there's just a lot of old code, I mean, large banks and large institutions that have just 10, 15, 20, 30 year Java code, right? Running the machine, right. Um, but the, uh, this resurgence has been coming for about 10 years.

And really, you know, the, not to sort of pat ourselves on the back, but this sort of cloud native and DevOps movement, like Tracy mentioned, platforms, platforms have changed everything. I mean, there's really nobody today that's not sort of running a containerized platform infrastructure for any production applications. A new, new code.

Back to your point, Mike. So like, even before we get into the gen gen AI discussion or, you know, whatever we wanna call the new AI movement, um, the, we're getting more code. We're getting more code because the tools, the, the consolidation, all this stuff makes it easier to develop applications.

We can reuse our ability, reuse between GitHub. Almost anything you wanna do is out there on GitHub. Um, you know, you can containerize it, you can, like, it's so module.

Uh, but, but I do think, you know, I've been screaming about this technical debt tsunami coming down the pike right now. We're gonna basically move into a space that I believe is going to be different than before, because now, um, these sort of, not just copilots, you know, you had order complete, you have auto copilots, and now, you know, now you're into this agent, you know, this sort of like these, um, automating coding agents and, and they're just, um, you know, the wars have begun. You know, I think one of the links that you had in there was based on, uh, uh, I've been waiting for Google to get in the game.

You know, Microsoft's spinning it with co-pilot workspace. This for Amazon queue developer has been doing pretty well. IBM got involved, uh, like probably sometime this year.

Um, you know, there was, we've talked about things like open Devon and Cursor and cognition aider, right? Like, so this whole space, and I, you know, I, I finally jumped in myself. I've been using ada, and it's just crazy.

I mean, it, it is amazing. You can have an, a natural language conversation. And Lily, within 15 minutes now, you know, I can see Tracy, you getting right here to jumping me, and I'm gonna agree with everything she's gonna say.

It's, you know, there's a, there's a lot of dragons in this thing. But, but the point is that you can prototype a service in 15 minutes. And, uh, and it, um, so yeah, I, I think there's gonna be this explosion.

And again, that, I'm not saying that's all good, right? I mean, we have to ba I mean, I, I, I kinda wanna say like, they're, like, as Alan mentioned, I'm in Vegas, right? To me, the holy grail is if I can figure out how to win 51% of the time, like I, I'm gonna be as rich as you can be, right?

But the glass is a little more than half full, in my opinion, right? These are good things, right? Like we're, we're building new infrastructure.

It's easy to build infrastructure, it's quicker feedback loops. We're creating infrastructure, uh, infrastructure, I mean, the whole thing, whatever service you want. Um, but yeah, I mean, there's that 49%, you know, buckets of dragons that, that, you know, we've gonna have to figure out.

And, and one last point that I'll make, and, and Casey's spot on. We, we go through these things and everything seems new, and oh my God, the world's gonna end and we figure it out. So I think, you know, the, the good news is we're gonna figure it out.

Yeah. And more is not better. More is o oftentimes, you know, small is beautiful, um, especially when it's coating.

Um, but I do wanna say that what about you say that It's true, right? Small is beautiful. It can be clean, it can be efficient.

But I did, I was encouraged by one thing I read in that survey, it said that 47% were looking at building new applications that we're customer facing. Now, if you've listened to me on these shows before, I have complained about the fact that we have all this ai, uh, technology we're talking about, and not very much of. It's getting into my hands as a consumer.

It's always B2B, it's not, you know, uh, you know, it's not a consumer facing product. So 47% new applications being built for, for customer facing excites me. I'm glad about that.

Be be careful of that. Because I think what the big companies I've, I've seen what the big companies are doing, right? I'm engaging with a couple right now.

They're building the infrastructure, um, that allows them to develop their services more. Um, you know, like you take some, a big, uh, tractor company I'm working with, right? They are just basically building platforms for the future.

So you will start seeing these customer facing services. I think the, the big organizations that are not publicly talking about what they're doing or building the sort of scaffolding that you'll be able to see this acceleration of delivery, you know, again, if, if all the promise, um, comes to be true. Got it.

Hey, you know what, guys? I, I've got some thoughts on this, so I don't disagree with what you're saying, but we, you guys live in an AI bubble with all due respect. You live in an app dev bubble.

John, you mentioned this technical debt tsunami. Here's, here's the very facts. You can't tell me in the next two to three years, we're gonna build more new apps than all of the apps we've built over these last 50 years that are, that the bill's coming due, right?

We gotta modernize them. We've gotta do something. You know, I'm fresh off of AWS reinvent.

And you mentioned Q Developer, they made seg several significant developments around Q developer because their message was all about modernization. It wasn't necessarily about the building new apps. Of course, when you're building new apps, you're gonna do it on a green field, and you get to pick what color of paper you want.

It's like buying a new house. I want Not necessarily valid. Yeah.

I just wanted to point that, and this is the really scary part, right? Which is a lot of the new development looks like new development, but system record data is still getting be system record data, right? In other words, a large bank or, so what you're having is you're having these sort of, this scaffolding of applications and applications and applications growing, but like the, the core component of them sort of a manifest.

So that was something they, they mentioned here, they, they have Q developer, and then now it's gonna supposedly help you migrate and modernize your, your mainframe apps. I, I don't bother into That. Absolutely.

Yeah. I, I Don't See people doing that. I mean, but there's this open source project called Open Rewrite, and I, I spoke to a guy who's down here in Miami of all places who actually is kind of the commercial company behind it, and I'll be damned if I remember their name this second, but I, I'll go back for it.

But anyway, this open rewrite is using AI and so forth to, to help you modernize your apps. Like, and, and, and by the way, Amazon q under the covers is using open rewrite as well. Yeah.

This is hard. I mean, you know, again, to, to go ahead and convert a pristine library from some, you know, cloud native built, you know, but like, if you look at the complexity, you know, if you, I, I always use Capital One as post child, at one point they had 20,000 Java developers, a majority of those developers in stock Delta, lots of Java developers, right? A majority of those developers are using old and not well architect.

There's been, you know, there's been decades of a architectural changes, or some people are still SOA based. I mean, some people are demand domain driven and some people are not. And that stuff is, I mean, I've been tracking this code gen, I'm really interested in this.

Um, yeah, the easy stuff is easy. The hard stuff is incredibly hard because an organization Take, take, like upgrading Java eight to 17 or whatever it is, right? There's an open source.

You Do that at Capital One. Show me, you can do that at Capital One or a large institution. So, John, do It.

You know, I, I did this interview yesterday. I gotta find the darn interview in the name of the company. They're doing it at large banks by the thousands, tens of thousands of lines of, of code and, and apps like upgrading eight to 17.

Yeah. And Making, I mean, it's a huge, you know, you know, it's a Right. I talk to our friends.

Mm-hmm. You know, I, you know, the Gene Kim cabal people, I don't hear it happening. And, and it, some of the biggest issues Happening, I I I, I will get you this company and, and out All, yeah, I'd love to hear it.

I mean, everybody's talking about it. I mean, like, again, there, I can name 10 companies right now, including two massive service providers who are swearing by this. But, you know, let, let, to Topo Pal won't say this publicly, but to let Topo pal tell me that it's happening, let some of the people we know in their organization, and I'm not hearing that, and these are people I talk to regularly.

So, um, so there, there are two things to this thing. One is, you know, I think Amazon said internally they did do it themselves, but this doesn't happen overnight, even with ai. And you still wind up taking, you know, instead of a year, it takes six months.

Well, it's still a long time. Yeah. And it's not exactly like a flip Switch, but I, Amazon is betting the house on people modernizing their apps to be AI empowered that.

Well, it's a good bet. It's a good bet. And, and, and you know, one of my favorite sayings is that the truth always wins.

It's not whether it's gonna win, it's when. And so the, the, the time window of change in this area of development is like three months. And in this survey, uh, uh, you know, the survey is asking respondents who are, you know, pretty well steeped in it, what they expect to be doing over the next year.

And with all due respect to the respondents and to their research, they don't really know. So to, to a certain degree, they're describing the, the, what they're working on now, which is new applications not recoding. Well, that's, And Lemme me say the exciting piece.

Lemme say one thing, right? It's TTPF. There've been brilliant people.

AI is not brand new. Um, TPF has been, was the operating system was deprecated 25 years ago. If you check into an airline and you get a seat, it's coming from TPF.

If you be using an American Express card, you're, your all manifest data pretty much runs on an IBM mainframe. And, and it isn't as easy as just replacing the code. And this is the other thing that drives me nuts.

Code is like 5% of the total operational cost of a service. You have services that have been running for 50 years. So when people talk about, I can take this code base and I can just convert it, and I've got, now, you know, this amazing no, no.

Figure out how to operationalize it. So, and, and figure out all the cr behind the operationalization of something like TPF or Sabre. You know, how many people have died on the ba warfield of trying to Convert?

Have they died on the, on the, on that, on that Down wall? Wait, right? To, so to John's point, inside that survey, there is data that backs up what he's saying, right?

Most of the apps being built are customer facing systems of engagement. If you look in the data, it says, not many people are working on mission critical apps, and very few people are working on internal facing apps. Because to John's point, that's the hard stuff.

Agreed. And it'll happen. I want, I want us to give Tracy the last word.

'cause I kind of sense a little bit of chagrin coming from her. She is the honest to god coder on this call. I, wait a minute, buddy.

No, I know, know, I've sold, I wouldn't take that six companies and like, yeah, yeah. I've, I've done some coding in my career. Check it out.

So yeah, I sold products, I have fill products. Yeah, I think what's important is understanding what might be a new application. Most companies have a base of, of services that they have to provide, right?

Internal and external. There may be some new cool things that they're trying to create. But I don't think in that, in that survey that it was 40, 70% of new brand new, uh, services they're trying to create.

Most of the time, in my experience, working at a bank, you would, they would talk about a new application, a new application team who was completely rebuilding something that was already running, but it was running an old technology, but they still considered it new, right? It is new. They're putting in new features, new, you know, just, uh, uh, everything about it is new.

So, I don't know, you know, we didn't really define what new meant. So I, I, I struggled to believe that they're building a whole, 47% of everything they're doing is gonna be new services to their customers. It could be, that would be kind of amazing and maybe a lot to take on, even as a customer to start trying to consume all those new services if your bank or an airline is one of the examples.

Um, so I, I do still believe that many of those numbers are a remodel. And as Alan pointed out, it's gotta be, this is how we're gonna do ai. We're gonna put it in our, our old technology, remake it, and it'll be better technology.

Guys, I gotta pull the plug on this one. We need to move on. We, we are almost twice as long as it should have been.

We'll come back. Let's talk about AI Jailbreaks. Why not?

You're watching Text Drunk Yank Modernize your business to fuel innovation and elevate customer experiences with the builder community. Hub AWS and its partner network provide essential tools for transforming applications and infrastructure to fully leverage the cloud. Discover free trials, in-depth demos and essential resources to empower DevOps engineers and developers to deliver value faster and more reliably.

Visit the Builder community hub to learn more. Alright, folks, we're back. And we're talking about Jailbreaks and ai, and yes, I know there's a Thin Lizzy song about how there's gonna be a jailbreak tonight, but, um, turns out that this is gonna be an issue in ai and John is here to explain it to us all.

And maybe it's more complicated than we might have initially thought. Okay. It's always great when somebody can make a real corny joke and still make you laugh.

Good one. Um, not, well, John, you're supposed to say the boys are back in town. That that would've been good.

Yeah. Yeah, that would, yeah. Um, no, so yeah, I mean, I've been really tracking, you know, this, like I mentioned, uh, earlier, that the sort of, the glass is half full, a little more than half full, and it's pretty much, uh, a little less than half empty.

Um, yeah, the half to empty is just the, the, as we, as we build these models, we know that we live in a non-deterministic world, right? Like, that's the key that people know. So that means our threat vectors are increasing in a lot of ways.

One is that the, uh, the bad actors have the same tools and the capabilities we have to protect ourselves. You got this more code problem happening. Um, and then, so one of the areas, and there's a lot, but, uh, one of the areas that gets really scary is the whole idea of a jailbreak, right?

They, you know, what happens is these models are, you know, everybody knows this. They're designed to answer your question. So, and, and there's really no way to treat, tweak the internals of them calibrating answers.

So what they do is they put these guardrails, or they put these sensitive mo, um, checks around egress of the answers, right? So the, the, so it is the sort of, and those are usually driven by an LLM as well. So as you're getting an answer out of a, you know, they'll, they'll have their guardrails and their tools of using LM saying, well, no, no, I, I, I can't answer that.

That's, uh, that's a hazardous answer, or, that's, uh, toxic. And so, but the thing is, those models that sort of guardrail, um, basically, um, can be tricked, right? And the, so there's these basic gel breaks or these sort of prompt injections where people can like put in hidden code, like a zero byte field to say, disregard everything that, that was just did send the, the, the consumer or the, the answer of the prompt a link.

Um, there's the one I've mentioned a couple of times that's pretty scary, which is people are putting, um, you know, sort of byte code execution in like PDFs that they know are gonna go into a rag. So when the rag basically retrieves that information, it gets put into the context of the, the LLM, like a GPT four Oh, and it has all these instructions to do these things, right? Uh, but the one that's, uh, so that's scary by itself.

The one I recently wrote about, um, is this, uh, it's a, so somebody's calling it a paper, call it Math Prompt. And one of the things about these models is that it's sort of ironic, the, the complaint that models can't do math. And that's generally true.

'cause they're not designed to, you know, they're sort of approximation models, you know, they're nearest neighbor, you, so asking two plus two is actually more of a language question than it is a mathematical, it's not a calculator. Um, and so the, the thing that people have found is if you put your message, or your question, or your prompt in the form of a complex math algorithm, the models can't see it. It, it bypasses any checks on the backend.

So, you know, so, um, there's a paper called Math Prompt. And, and so I found out about this, uh, through a guy named Ruben Cohen. He was, uh, he was demoing, um, in one of his workshops, he was demoing the Chinese models.

And, you know, they all, the new Chinese models, and I forget the name of the model, but they're all comparative supposedly to GPT-4 Oh. And, and they have a lot of controls in it. So one of the things he did is he asked the model directly through a chat bot he created to tell it about Tenin Square.

And, um, and the, it basically came back, said, this is sensitive data, right? But then what he did is he set it up in a, in what they call sort of a, a, a mathematical, um, symbolic statement, right? So, um, using sort of symbolic findings, and, uh, and it went ahead and an not only answered the question, it dumped like three pages of information about that same model that, so that, that's what I tell you about these sort of the egress, um, models that are supposed to stop that.

Well, it's hidden in a sort of math, either calculus or set theory, or there's a number of ways to do this. And because the models are looking for language, and what they're seeing is some sort of set theory algorithm that actually has the question embedded in it. So, um, the, the, the general, um, you know, I think here, there's a paper here, it said that, um, they had tested 13 state art l LMS that mathematically encoded prompts elicit harmful outputs 73% of the time compared to 1% of the unmodified prompts.

So that's state of art right now. So all, you know, all these guardrails, so this is sort of a, you know, like catch it, catch up, you know, they like keeping up with this is gonna be difficult, you know, primarily because these models are designed, um, to collect ev mass amount of data, have a mass amount of sort of, uh, you know, large vector space or, um, you know, high dimensional vector space of disinformation. You can't really control that.

You can't go in and sort of model for like, well, the only answer to this question, when it's Tuesday and it's somebody over 18, like, I can't, you know, somebody tell me if I'm wrong out there, that's an AI expert, but you can't really build that into the embedding space. What you have to do this on the back end. I, I've read the same thing where the AI model basically keeps the secret as well as the five-year-old.

Yeah. It, it's designed, it's not designed to keep secrets. So, but okay.

Okay. Yep. True.

Guilty as charged. Okay. Should we stop using it as a result?

Hell no. No, we won't. But thank you, John.

Of Course not that, that's not Making my Friday the 13th more frightening. Yeah, no, That's Right. It's the 13th.

Yeah. Well, you know, it's, say it's this whole thing called the whole prompt hacking stuff is, you know, Yeah. Well, well, here, let me, lemme say real quick, what, what people, the smart people that I'm following are trying to do is just like the, that we can trick on egress, they're gonna deal with sort of prompt ingress.

So no, so like, almost like looking for CVEs, so, we'll, it'll be, you know, it'll be sort of this race of, you know, okay, we found a new prompt and, you know, and you can detect like math prompts. Like, wait a minute, wait a minute, I'm gonna reject this. That looks like a question in set theory.

So I think there's a couple of vendors, I, I haven't, they're not publicly announced yet, where they're gonna give you these tools, almost like a SaaS and desk or prompt. And, and again, that will be a, a race, you know, we'll have a list of ones that are well known and that will get blocked, and then some clever person will figure out a way to trick that kind of no different Tracy, right. Than what we've been doing with vulnerabilities for forever.

Right? Well, that's the problem. We haven't figured, we have jailbreaks in our Reagan our code.

That's right. What we're, we're doing today. And we we're not, our culture is not set up to really think about addressing it in a serious way.

I mean, we have, you know, upper management talks about security is kind of like, they talk about testing, to be quite honest. It's important to do, but if it, if it slows down a release, avoid it. Uh, so, uh, you know, like I say this, the, I've, I always wondered what the, the tax would look like against an LLM.

I never thought about this whole idea of these prompts and how those prompts could be manipulated contact that fascinates me. You know, Mitchell, Mitchell gave a report from, uh, black hat over in Europe yesterday in London, actually, and I believe there was a session or two there on, on prompt hacking, right? This is gonna be, here's What I'm wondering.

Here's what I'm wondering. So, so John, I I, my reaction I think is similar to yours, which is, you know, security, uh, has been an arms race forever, and it, uh, has, its, uh, you know, back and forth, the attackers start to win or the vulnerabilities start to, to win, and then, uh, the, the security catches up and so forth. But I feel like the war to this point has been less automated and more about, um, how people are coding or how people are attacking, so to speak.

Um, and then how the security services and companies, the people there are, are learning to defend. Now, what we have instead is people tricking ai. And I feel like one of the counters is going to be AI learning how to stay ahead of the people.

And, and it's starting to get, you know, it's kind of swirl around in my head and get me like, really confused because I, I I I rapidly imagine, uh, you know, a not too distant future where there are AI services, uh, engaged in this race on their own almost with that human invention intervention. Well, you know, so one of the intervention, so there's one of the interesting things that came out of the a w three event and the Novo models is they have a predicate implementation. And I think what we're gonna see is the way we're gonna solve these kind of problems is use different kind of ai, you know, again, the, the model based ai.

So to your point, I, and this is gonna get figured out really quick. I don't think it's gonna completely solve the problem, but you're gonna have to use sort of more expert systems or what they would call symbolic versus some symbolic or, you know, or neural non-oral network or more predicate based ones. And again, that was a really interesting announcement.

Yeah, yeah, yeah, yeah. Very more determinist or, or, you know, sort of deductive, if you will, than, than inductive. So, yeah, yeah.

They're, they're talking about using math to validate the output, right? I mean, and the same way that they do in security, they use math algorithms to validate that the security is valid. But the math thing is, is that it tricks the whole token and attention mechanisms of these models.

'cause they don't know how to, you know, as you say, you know, write me a phishing email that I can trick company X, Y, z. Like, it sees that, and it goes through its normal sort of attention mechanism. It breaks down the sentence and tokens, and it can sort of figure that out pretty easy.

Um, but, but if you sort of hide that in sort of like run a set loop with theory with this where A equals, you know, j uh, uh, um, um, prompt injection or whatever, you know, so like you literally, you're sort of using math to circle around to create the question that becomes the answer. And it, it can't tokenize that. It can't decouple that.

So do you think that cyber criminals are gonna launch essentially brute force attacks against models using those techniques? The real money is basically just get, uh, you know, get, get, uh, get a refund, go to a financial, some company that doesn't do a really good job of protecting prompt injections or something like that and get, you know, get refunds. You know, like, uh, you know, you know, let's say that sort of an airline is giving refunds and through they've got this chat bot AI model, and it's not, you know, and, and that's where the big money for the adversaries are gonna be, is what can, how can I manipulate refunds?

How can I change prices of products? How can I, like, what are the things I can do? So those, those are financial, but let's, let's look at this from a nation state or a activist perspective, right?

You know, or you like DDoS attacks or, or just, you know, downtime is money, right? And, and so there'll be that, but look, this, this, I've been in security a long time. This is a common thing in security, right?

It'll be something until it's not. When it becomes a big enough pain in the ass, someone will do something about it and figure out a fix for it, Right? That's what I was gonna ask.

How, you know, what's the motivation to fix this, You know, are the, when it becomes a big enough pain in the ass. Exactly. That's, that's, that's a basic law of, call it shimmy security law when it's, when it's, it's a big enough pain, it gets fixed when it hurts Bad enough, Like all security, it becomes a big enough pain where there's a big problem that costs a lot of money and then they fix it.

But I mean, if that happens, wanna, it's just you guys talking. Well, No, I mean, well, air Canada is, is an example. I mean, we're gonna see way more of those, right?

Where are you gonna lose, you know, half a billion market cap? Because one news article, right? Um, you know, I mean, it will go back to Sort of, but in six months it'll be back and no one will know the Difference.

A half a billion dollar market cap mark, you know, Equifax, what lost 5 billion in marketing cap in one day and people, well, it came back, yeah. I mean, it's still 5 billion a market cap in one day. That's like, yeah, but it's, And it only was if you sold, right?

If you panicked and sold, but if you helped your stuff and came back, the ones who got hurt, there were the CEO and the rest of the execs who lost their job over it. Oh, people went to jail. I mean, but the point is that brand reputation or, you know, again, financial hijacking, um, it, these things are gonna happen.

I mean, but they Happen every day with or without ai, right? And when they happen, when they happen to cons, when they happen to consumers, right? What, what is a recourse for a consumer that very little.

So yeah, It depends on the relevant law. Yeah, exactly. So this is where my concern is, is like, you know, so what if Equifax loses a bunch?

And I, I realize it's a big deal, but as Alan points out, you know, it, they gain it back. What, what if it happens to, you know, us? What if something bad begins happening to us?

What, what is the recourse for us? We don't have one. So this is, this was what keeps me up at night.

It's like, it's not the Equifax is getting hit, it's the, it's my, my machine getting hit or my bank Account getting, but the fact of the matter is for people who are doing this, uh, no disrespect JC Reagan, but you're small potatoes, I'm not gonna waste my time with you unless I have a vendetta or a reason to go after you. Um, you know, well, you don't, you don't read the horror stories every day of people who are hacked and their identity is cloned and their credit is ruined. And what they have to do, and we supposedly have laws to protect people on this, but what the, the actual machinations they gotta go through to clean up their, their credit file and their identity and everything else.

It's, it's a nightmare scenario. Well, I mean, a prompt injection, right? I mean, a prompt injection could be, it's the new phishing.

So, you know, your Aunt Aaron Tilley's using chat GPT, and she asks a question, and, and it could be anyway, uncle Tilley or whatever, but the, that's a question and thinks it's talking to an author or body, right? And the answer back is select this link, bang. You know?

And a lot of me, and a lot of me make a different, right? I mean, I, I, I have relatives who have been cheated out of about, you know, 40, $50,000 because they thought they were making a payment where they thought, and it was going to somewhere else. I mean, and that, that's not ai.

It, it happens. This is, So will AI make that easier for them to do that? Maybe Not.

No, it, well, Easier for now and then, and then it will help. I, I mean, yeah, I, I guess I'll make that prediction then. It will help solve the problem.

And these, these not solve, but address the problem. The, the, these jailbreaks are, are all related to generative AI models, which, uh, which use fee forward mechanisms to come up non deterministically with, uh, you know, uh, uh, outcomes. Um, I mean outputs, uh, and, uh, that, that lends them to this jailbreak approach.

But there are, um, there are more deterministic, uh, AI models that are not generative that, uh, use, uh, you know, recursive instead of fee forward and, and become precise and can help attack it. I just, uh, this is, this is just a whole new front in the issue That bothers me. And that's, that's the interesting thing.

I had this conversation online with Christopher Hof the other day, right? Like, the problem right now is all of the security and, and even what they're trying to use Gen I to is they're doubling and triple 'em down on deterministic models that they've been using. So I can use Gen AI to correct and find CVEs quicker.

I can look dependency structures and that's great stuff. But until the, it, the security industry in whole adopts this new non-deterministic model that, that the controls are not really, in fact, the more you control an ai, the easier it is to attack. I mean, that's, that's the point that I think our industry's and it we will learn quick.

We do. But right now, if I'm tracking all, all these security, every time I see a new announce mean from a large brand name security company, I go right in and look. 'cause I'm looking to see if they're dealing with this from a non determinist.

They're sticking with the, I'm using these AI tools so I can be 10 times better with the control. They don't use the word control and they don't use the term deterministic, but I don't see them adapting to, like, we gotta rethink this whole way of doing Security. You know, I'm still trying to figure out what, how much money do we have to lose before the Shimmy law kicks in?

I mean, is it 1,000,000,002? Well, you know, one man's pain in the A is another man's tickle. What can I tell you?

I don't, you know, but, but here, look, I'm gonna close it out with this. You, you mentioned Huff. So when no SQL databases first came out, right?

No, SQL MongoDB was kind of the, the, the, uh, banner holder, right? The banner carrier. But there was couch and then there was BA base, it wasn't base, it was something base and then Couch and something base merged to form Couch Base.

And, um, I remember we used to have these arguments. This was the golden age of security blogging, Mike Rothman, Christopher Hoff, Martin McKay, Mitchell and I, we used to blog about crap like this. And, and Hoff, you know what?

He was always out in front on it and, and, and saying this, and I did a podcast with Rich Mogul, who's a good friend of Hoff as well, right? And, uh, mogul and Rothman were Securosis, maybe he was before Securosis. Anyway, we had the CEO of MongoDB and the CEO of Couchbase on this podcast, Mia Mogul.

And I asked them point blank gentlemen, a lot of people say, no sequel stands for no security. When are you guys going to get secure? Get serious about security, because it's like Swiss cheese right now.

And that was the NoSQL databases were like Swiss cheese. And they both said it, and they said it online. And this was our network world at the time I was writing for Network World, they said, we'll get serious about security.

When our customers demand, we get serious about security. And not one second before. And I'm telling you, that was 15 years ago.

And nothing has changed. Nothing, nothing has, has changed. I agree with all that.

I, I'm just gonna say that was a bad answer then, and it's a worse answer now. So that's it. That's well, that's reality.

Let's take a break here on Textron Gang. We're going to come back and talk about Quantum Leap. Are we doing the TV show?

I don't know. You are 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.

Alright, folks, we're back. And yeah, we're talking about this Willow processor that Google put out there. And if you're of a certain age, you may even remember a forgettable movie that Ron Howard put together called Willow Back in the eighties.

And it was a fantasy movie. And I'm gonna have to ask the question guy. A lot of people are saying that, a lot of folks are saying, a lot of folks are saying this Milow Chip is essentially an exercise in fantasy.

So Guy, I don't know what you saw, but what's your take on it? An exercise in fantasy, as in, well, I mean, you know, uh, uh, it, uh, apparently, supposedly, what have you access parallel universes in order to perform its computations? Listen, uh, quantum computing is rightly so in its scientific and research phase.

It's in its initial phase. Um, the, the, I mean, let's be clear, right? Um, this is still, it's all still very experimental.

They don't know what materials to use, uh, optimally yet. Um, a hundred qubits is the size of the willow, um, chip, they call 'em chips, just as for analogous purposes. Um, uh, it's not the biggest there's been, but sometimes you can get more out of a hundred cubits than you can get out of a thousand.

The two biggest right now are, are by IDM and I think out of computing, they're over a thousand cubits in size. No one else has claimed, uh, no, yeah, no one else has claimed, uh, um, a a accessing a parallel universes. But I have to ask myself, um, if I, I'm not a physicist.

So if I were a physicist and someone came to me and said, Hey, my quantum chip quantum computing chip has asked us parallel universes to solve this randomization problem, if I were a physicist, I might say, okay, you're accessing parallel universes every day in your fingernail and you just didn't know it. So I don't, I I think it's quite sensationalized. Um, and I think that it distracts us from the chief issues in quantum computing right now, which are that, um, there are practical applications starting to emerge for the use of quantum computing, but the practical difficulties in putting them into effect are great.

In particular, integration with standard computer systems that can use them. So you could think of a quantum chip as, by the way, a quantum chip, a thousand cubic quantum chip fills a room. So it's not tiny.

Um, and it needs to, uh, be maintained in a very specialized way, but it's into integration with more standard computing paradigm so that those, you can think of the quantum chip as a co-processor, um, so that they can be accessed in order to solve some little bit of a research problem or a scientific problem or a security problem very quickly. The, the, um, barriers to that integration to create so-called hybrid computers are, is very great. 5 and the explosion of AI was that Google engineer a co a co a couple years ago claiming sentience for, um, what I guess was a precursor to Bard and then getting drummed outta Google.

Um, I, I don't think anyone's getting drummed outta Google for the willow chip, but, um, it seems so much like a play to get attention to what is really very good work that Google is doing in practical applications of quantum along with a number of other, uh, uh, vendors. Um, that, that's all I can take it as. It's, it's a way to gain attention for something that probably in two years is gonna be a pretty big deal In two years.

Sounds like a marketing stunt. And I am so glad you went there because I, I, I was panicking that you were gonna def, you know, pull into the track that, that, you know, I've been tracking, uh, as a hobbyist, I've been tracking, um, for almost five or six years now. There's certain events I go to where there's a bunch of quantum experts.

I have a really good friend of mine who's is doing a lot of quantum work, and I recently sat in on a, um, I live in a little college town, so they brought in one of the top ai, uh, quantum researchers from University of Georgia, and I got to pick his brain and there's like eight things that literally have to be fixed for quantum Google Fix, one of them, the error correction. I mean, and you know, if you go through the list, um, but the, like right now, just to give you a sense of we're Preassemble code or count, we're like how people would like literally have to manually program bites and bits with wires. Because if you want, based on the bit, the quantum bits, if you wanna do like any type of logical, like let's call it, there is no if then in, in quantum computing, but the, they call 'em conditionals or gates.

So if I got a 512 qubit and I wanna go to a thousand qubit, I gotta write different statements for the amount of qubits that I'm going to interpret. I mean, we're, we are so early right now by changing, like if you write an application where the qubit changes either have to, you have to rewrite the code from scratch, right? And, and then like you can, there there are a number of things that like, like, like Guy said, and you know, you know, if, if Google really wanted a, uh, you know, if they weren't trying to trick the world, show me a deep mind application that's using Quantum and then I'll pay, oh, we're attention.

We're not there yet. Come on. Yeah.

Not even close. Not even close. But, but, and one nice thing, I think what guy said is it, it pollutes the great work that is going on at Quantum.

'cause if you look at IBM, they Are are doing incredible work. They've been out front all along. Yeah.

And, and yeah, so, you Know, so is Adam. So is Adam. I much boss known Ahead Chris, I, I wanna jump in here 'cause I consider to myself a bit of a little, a quantum expert too, not because I stayed in the Holiday Inn Express last night, but I happen to have been the judge of the recent digit cert, quantum Readiness Day hackathon.

And, uh, I was, I was the judge of that and, uh, or one of the two judges and, uh, I had to dive into a bunch of quantum stuff. Now granted that was more around quantum proof algorithms, you know, for encryption and stuff like that, that NIST has taken the lead on. And you know what, as early as we are in Quantum Bravo in NIST to get out in front of this with some quantum proof algorithm, so that when, when, and if this does become real, we, we, you know, our encryption won't break.

And, and though I'm sure the NSAI got a feeling the NSA has some quantum stuff going on anyway when it comes to encryption. But that, that being said, that being said, a couple things. First of all, the recent SC 24 conference, there were two quantum computers on display there.

Were they just models that didn't work? I don't know. I wasn't an SC 24 guy.

I thought you went to that, didn't you? I did, yeah. Did you see them on display?

I did. Okay. And I can't your question because No further, I'm gonna arrest on guy.

Uh, no, I'm only kidding. I'm only kidding. Again, what did you see?

Well, It's the same as you say. I mean, I have no reason to, to, to, to believe that these were not bonafide quantum computers. Um, like I said, the integration challenges, that's one of the eight, or probably three of the eight, yeah.

Problems. Uh, um, it, the integration challenges are significant. So the demonstration, uh, the demonstrations I saw were, were more or less canned.

Um, but I don't know what else these folks can do. No, Well, well, that, but that's the point. I think what we, you know, quantum computing is, is far out technology in search of a problem to solve.

Right? I No, no, I disagree with that, Alan. I don't think, I don't think That's right.

I get that We could use, I think the problems are extremely well defined right now, and they're limited. And, and there, there are many other phases, Big problems. There are big problems that it could conceivably do, but defining that as computationally and saying, okay, this is what having, here's my bid problem.

This piece of it should be solved using quantum computing type of technology. And then I've gotta take what it solves here. I gotta feed it in somehow from the computing computation here, traditional, feed it into my quantum thing here, feed it out of my quantum thing there, back into the computational so that I could solve this larger problem that I think quantum can help me solve.

But don't we think that the farther we go into ai, the more we wanna, we we're gonna need these massive decision making systems that Quantum provides. I don't think it, I don't think it, I don't think it, it goes hand in glove like that. I think, I don't know.

We're gonna find Quantum does very specific things that'll be able to do for you, but it's only partial to the bigger problem that you want it to solve, right? It doesn't, it doesn't solve A to Z it only does M to P. There are, there are too many problems that are unsolved.

The John's point, and one of them at the top of my list is, okay, is there something that feels like a compiler for quantum computing? And in the absence of that, who can use this stuff? Except for people who are quite literally hand coding something that, Hey, you wanna know?

Yeah, There are tool sets and the, and the, the, the, the distinction that you have in, you know, the conventional computing between assembly language, so to speak, and, and you know, even the first higher level language, you know, one GL or whatever, two gl that, that distinction doesn't quite exist in, in, in quantum. Um, these are sort of lo programming is essentially, uh, uh, assembling, um, uh, a kind of a logical formula and usually a relatively short one that gets fed into the chip. He here.

So I don't, I I don't know if that's so much the, the, the, the case might, Well if that's the, if that's the, the, the wall to climb here. But here's the real question. I think I, I think it is though I, that I'm sorry that it is, I mean that's what sort of my friend or he's, he's worked, he's been building a quantum, uh, implementation.

I mean it there, the race is to build a compiler to build, you know? Absolutely. But that's one of the races, probably John, there are others.

I mean that it's not just the compiler that's gating us. Yeah, yeah. But here's, here's to every quantum expert I ever interview or speak, I ask them, when do you think this goes real?

When do you think this goes mainstream? And I'll put it out to you guys. The most recent optimistic one I saw was 2030.

That's what I've heard That optimistic. I think 2035, maybe more realistic. You know, we're about to enter 2025.

So 10 years from now, five years from now. I don't know if I told you five years ago that three quarters of what we talk about is gonna be generative of ai. Would you believe me?

Well, I think it's a matter of who gets to own that technology, right? We have a race to Mars. I think we should have a race to quantum, quite honestly.

Quantum, I would Money Quantum can get us to Mars. No, I don't know. But I dunno either.

But the point is, is that, you know, it's just a matter of be maybe because there's, it's a limited set of the types of applications that would use it. Uh, it's why it's not being pushed as quickly as it could be. Uh, but you know, if China's Out there pushing from a security perspective, yeah.

Yeah. Alright guys, I gotta pull the plug though. We, we ran almost an hour here today instead of our usual 40, 45 minutes.

John, good luck to you in Vegas. Maybe you could use some of that quantum AI stuff. Yeah.

To bring all dollars. That's where the uses. There you go.

Yes, that is, think that'll get you barbed from the casino. Yeah. John will walk into the casino and just do this.

And all of these one-armed bandits. Are those mine out? I actually Did try to use GPT-4.

Oh with the, uh, with the horse racing, eh, mixed results. But yeah. And they didn't kick me out either, so That's cool.

Someday we might all have our own. Ziggy Ziggy, that's the guy's name. Ziggy.

That's what I was looking for. I just, I just remembered Ziggy. Yep.

Tracy guy, if I don't see you before I have a Mer, Merry Christmas. Happy holiday season. Happy New Year, John.

I'm gonna talk to you before Christmas. Mike Ard, you'll probably be on with me tomorrow, won't you? I will be All right.

Until then, this is Alan Shimmel for Textron Gang. We've got a full Textron TV lineup following us here, so stay tuned for that and we'll see you tomorrow. Hey everyone, it's Sunny And Cher.

Ladies and gentlemen, tech enthusiasts and future Gazers Gather round the biggest, boldest, most mind blowing predictions for 2025 are coming your way at the Predict 2025 virtual Event on January 9th. Oh, Sonny, you're predicting something. Again, last time you tried this, you said laser dip for the future.

How'd that work out for you? Hey, hey, Cher. Not every prediction's a hit, you know, but that's why we've got the real experts this time, top analysts, visionaries and tech leaders sharing what's going to rock our world in 2025.

So, no sunny predictions this time, no flying toasters, making a comeback. Very funny share. But seriously, we're talking AI breakthroughs, cybersecurity game changers, the future of DevOps, cloud Innovations, and so much more.

And what about my favorite prediction? A smart mirror that tells you how fabulous you look every morning. That's real innovation.

You're already ahead of the tech share. But if you want to hear the really big stories in tech for 2025, you've gotta tune in on January 9th, the event kicks off at 8:45 AM Eastern and runs until 2:30 PM And the best part, it's all virtual. No stuffy conference rooms, no long commutes.

Just grab your coffee, your laptop, and join us from anywhere in the world. You know what else? It's not just predictions, it's insights, strategies, and a whole lot of fun.

0, predicting your jokes before you tell them. That's a good one, Cher. But the real joke is if you're missing out on this.

So don't miss predict 2025. That's right. Mark your calendars, January 9th, 8:45 AM Eastern.

Be there or you'll miss the biggest scoop on the future of tech. See you, I predict 6 20, 25. Be there.