Techstrong TV July 10, 2025

Hey everyone. Ah, modern medicine. You gotta love it.

A cure for hallucinations. You're watching Textron Gang. Hey everyone, it's Alan Shimel.

Happy Thursday. Ah, jk. It's not really a cure for hallucinations for people anyway, talking about AI hallucinations, and can we have, we found something that actually works on this.

I don't know if it's microdosing or what, but, um, we've got an interesting gang of people to talk about our subjects for today. Let me quickly introduce you to them. We have, uh, John Schwartz, who's been up all night with the latest Silicon Valley news, Garima Bo Powell, our good friend, guy Courier, and of course the Dean, Mike Ard.

Ladies and gentlemen, let's jump right into our first year, A cure for AI hallucinations. Anyway, shades of one Flew over the Cuckoo's Nest. Mike, what are we talking about?

Alright, well, there's one of these companies out there that specializes in ai and they've been mainly focused on infrastructure optimization, but they've put out a, an announcement saying that they have a framework now where you can go into these models and look for the variables that indicate that this is where, you know, for DevOps folks, it's kind of like a feature flag that this is where some sort of effort has been made to suppress an answer. Or is the root cause of an issue where the LLM is just plain old hallucinating. I don't know exactly how this is gonna work out, but it's fascinating to me because it takes on an issue that we've been struggling with forever now.

And part of the problem with these models is I have to go in and retrain them, or I gotta give them additional data to fix these hallucinations. And that's a lot of effort. And these guys are saying, now we can fix this in minutes.

Guy, I know you looked at this. What's your take here? Have we finally come up with an answer here?

Or is this kind of just maybe wishful thinking? It's not wishful thinking, it's a real answer. Is it the total answer?

That's really doubtful. Um, so let's talk about what this is. Um, it's a really neat trick.

Uh, so the way models work is they work in layers of processing, you know, the famous neural net and, um, what, uh, CTGT is doing, which is the company founded, by the way, by two, you know, 20 something Wonder kins outta UCSD, go UCSD. Um, so what what it does is, uh, what they're doing is, uh, sort of inspecting the inference as it goes and finding the layer where the hallucination so-called occurs. Now, I think there's a little bit of slight of hand going on here because to a large degree, what they were, uh, working on was programmed bias or censorship, in other words, models, um, that, um, for not through the training and through the source data, but essentially through tweaking of the model, um, won't answer certain prompts or, or will answer with hallucinations.

Um, so the neat trick is, you know, they start with this sort of listening post that sees how, um, the inference is going through every of these, you know, thousand, thousand million whatever layers in the model. And, and they are repeatedly hitting the model with prompts that they know or believe very strongly will produce the bias. And it's almost like looking and seeing where what layers is is going off and lighting up most often when they do that.

And then making adjustments to confirm. So when they amp up that layer, does it get worse? When they lower that layer, does it get better?

And that is a process that takes minutes. 'cause they only have to run, you know, a certain number of prompts can do it very fast. And it's rel relatively mathematically defined how they do this whole thing.

So that in the end, when in step three they make an adjustment, now you have an improved model. The slide of hand I referred to is if you have a bad training set, um, then, uh, or, you know, uh, if the training set, the data set is not curated, which these are huge data sets and they often are not, that is a source for hallucination and bias and so forth, famously, in computer vision where facial recognition doesn't work, uh, properly because of the training dataset. So I'm not sure I see how this applies in the first instance, but it's still a really big announcement, um, kind of like deep seek where you're saying, wow, there's a whole lot that you can do without having to throw a thousand, a million GPUs and months of training time.

And I think just to put this in perspective, um, we have been saying two things all along here. These are early days for ai. We don't really understand how it works or the implications.

Here's an example, a neat trick somebody figured out to quickly make models work better. And the second one is, and this is my favorite, always human supervision and human intervention is critical in using ai. Well, and now we're talking about human intervention to improve models.

I I realize it's all programmatically done, but the, the identification of likely hallucinations is confirmed by humans. So, um, this is, uh, really worth studying and looking at. Yeah.

I also wanna jump in here because, um, mean, GH has, uh, mentioned some positive sides of, you know, this announcement. Uh, being a DevOps, uh, practitioner, I always find interesting ways to see and challenge ourselves that what can be our next bet from a DevOps perspective. So I see there are core components which, uh, gee has mentioned that, you know, how do we make, uh, our morals perfect.

Then there is a middleware, which is in making, which is basically, you know, when you are prompting your system, how well you actually get the response, how do you monitor it, how do you infer it, and how do you interfere in the, uh, output? So that's the middle layer, which probably will be in making, and a lot of DevOps people are looking at it from an observative perspective. And the third aspect is the top layer, which is observing, because there will be residual hallucinations.

You know, uh, nobody has understood, uh, these models enough to claim that, you know, all hallucinations will be gone through this, uh, you know, announcement. I think there is a controlled approach and, uh, interrupt, uh, probability is kind of, you know, also being challenged here because explaining AI distance remains still remain difficult, right? So there's a lot more to, you know, added to it if you see from a challenge perspective, like how do you, um, cut down the over reliance of automation, for example, uh, uh, for in this scenario, um, there will also be evolving threats, right?

I mean, how people are also looking at, you know, uh, better ways to do new attacks. And the attack vector is also broadening, right? And last, uh, point, which I think he also touched upon is error versus intent, right?

I mean, we have to be very careful when we, uh, you know, deploy such kind of technology because human in the lead is, uh, is having an essential kind of, you know, uh, element to it when we are, uh, dealing with emerging technology. So, I mean, there is no silver bullet. Again, uh, it's hard to separate the accidental from intentional misinformation.

You know, am I the only one who sees the irony in trying to figure out a Chinese based AI and, and trying to get them to remove the censorship? I don't think there's irony in that. I think it's a damn good thing that the censorship or that we're trying to remove it, that we're trying to remove it.

Let me use the, the tired old phrase, low hanging fruit. It seemed plausible, surely to these two researchers, uh, um, uh, that, um, examining deep seek would likely produce, um, easily found bias that they could then correct for, remember this, these biases, they call it censorship. I think that's correct.

There are layers in the model where those biases have sort of essentially been placed by human beings, and they're finding them and neutralizing them. And I, I think that, that, it's not really ironic, uh, to me it is a sensible way to go about the research. Um, and it just so happens that, uh, China right now is an authoritarian country while capitalistic and tries to control information.

Famously, I don't think they exactly deny it. So there you go. They're not the only ones.

Um, but true. Let, let me, it could be some one of the worst ones, though. I do wanna say that from a government standpoint, It's all relative.

But let, let me, let me make a distinction here. I think there's a difference between human intentionally introduced bias slash censorship, right? Such as in deep seek, not talking about the Erman Square, for instance, right?

Versus some of the hallucinations just where the fri did he make, did they make this up from that? We saw in, in a lot of the AI models more earlier on. I don't, I don't see him quite as much anymore, but you know, I still remember when asking it to write an a bio a my own bio as a test, and it would make up, yeah, it would just flat out make up stuff.

So that's the sleigh of hand. I was referring to Alan, which is that, uh, their test cases and their proof of concept have to do with intentionally placed bias. But my view, my analysis of this is that the exact same technique could be used to find hallucinations.

When you have a set of prompts that produce hallucinations, you don't actually have to care whether it was intentional or not. You can still find that layer in the model that shines when you run those prompts and you can still make those adjustments to, to, to the vectors, to, to produce a better result. I think Gary's point is a perfect point, which is that what I, I'll rephrase if, if you don't mind, Gary, which is, we, this is a step, it's gonna have a reaction, and then we'll need further steps.

And this will be development, just like every tech wave we've ever had. I'm excited about the whole idea that I could actually use an open source model now with some level of confidence that the answers are gonna be right. 'cause I have some level of control over that.

'cause historically, up until now, we've all been independent upon the AI model provider to put some level of governance or some level of control in that. And frankly, I'd like to see them all get outta that business. And all of us just take responsibility for how we're gonna use these LLMs ourselves.

The feature validation part is very interesting because, uh, this brings dynamic control. You know, if, let's say our compliance and regulatory, and I mentioned it over and over again, that the, if they wake up and they, uh, catch up this emerging technology wave, there's a lot which can be done through this dynamic control. And, you know, if you bring observability and the prompt engineering in the loop, I think there's a lot of more, uh, trust getting built up for, for these technologies.

So the, the problem though is, I mean, look, if, if it hallucinates writing my bio, I know my bio and I can easily tell it's a hallucination, unfortunately, you know, we've got a problem with fake news in this world, as it is. Most of the population, at least here in the US can tell the difference between fake news, conspiracy theories and tinfoil hats, right? They're gonna have a hell of a hard time figuring out what's real or not that the AI is spitting out to them.

So Mike, to your point, either you trust nothing or trust everything, or, you know, how do you know what to trust? And so putting it on a, a personal responsibility thing, it's your responsibility to make sure the AI's not BSing you is, is I think, going to be beyond us. Yeah.

And, and this is kind of a little peripheral to the issue, but I mean, I, I remember seeing a, a, a report in rock, you know, and, and the, the bi the amount of bias that's evident there bordering on blatant antisemitism, and it's actually getting worse at the same time. Musk talks about starting a third party. But I, I mean, that's, this is something, the one, just going back to hallucinations though, is this is, I think it, it's, it's very encouraging news, a potential game changer.

I mean, one of the headaches, lingering deterrence or, uh, obstacles in the wider use or trust of AI is, has been this case. And anything that addresses it, I think is positive. And, um, so I'm very encouraged by it.

Does this just work on deep sea, or this will work on any model in ai? No. Well, it'll work on a neural net model.

Not every, not every model's a neural net, um, in the, in the world. And, and no doubt, uh, we will start, uh, experimenting, um, further, I mean, in the origin of ai, it wasn't always a neural net. It's also has to be a feed forward neural net probably.

And some neural nets are recursive, et cetera. So I don't wanna get too wonky. Um, but the, the essential idea of, um, monitoring the propagation, 'cause that's kind of what it is of the prompt through the model.

Um, it, it would have to be adjusted, I think, for, for other types than feed four neural nets. But, uh, you know, that's, again, more development for later. For now, it just applies to, to feed four neural nets, which would be most of the generative AI models, if Not all.

Those are the common ones we use. You know, Mike, I, I almost wish we had our friend Chris Blas on today's show and here about how his civic minded ai, well, whenever it's called, It's really well's. I'll tell you what's ironic, sorry.

At the end of the day, brands, quote unquote, and people, we still are the one, those are the ones we are going to be turning to and relying on because the systems are too complicated. And no, Mike, I don't think we can do it ourselves. If we can, we'll be using some kind of open source tool that does it for us.

And that means an open source organization that we're trusting. I mean, it still all comes back around to human supervision intervention, human interaction. I would like to distinguish between ourselves and the definition thereof.

There's a difference between ourselves as a you and me, and, and a retailer who has some AI capability can go in and control this stuff themselves. Now. And I think from an enterprise perspective, this is huge.

I mean, it's all got us individuals, but it's me as a company, I now have some level of confidence, a lot higher than I did before of what may or may not be in that AI model. So from that perspective, I think this is gonna drive a lot more AI models into production than anything I've seen lately. Uh, I don't wanna sound too negative about this whole thing, but I think, uh, one point which I'm trying to make here is it's, AI is becoming like a drug.

You know, the more you use it, you more you get addicted to it. And think about my 10-year-old kid in the house who has hands on chat, GPT, I just wanna kind of ensure that, you know, we understand that human accountability is, uh, of course required, but there is also concern about how do we use this technology who is behind, you know, how we regulate this, right? I mean, gee, I interrupt you.

What does it do to a 10-year-old, right? So I it is the Wally Syndrome, while I, you know, from the movie, you know, Mike, I think I sent you an article last week I was reading, are, are we gonna lose the ability to write, right? Because we're relying on AI to write for us.

Are we gonna lose the ability to code when AI's doing all the coding? Are, are we destined to be corps floating around in some servo chair and Ouris take care of us? No.

No. So think about the history of music, okay, around 1900. Kurt Vonnegut had a great, a great statement about this.

I wish you could remember. It was so clever. It basically said that in every town and every village in the year 1900, there were all kinds of people with middling to fair ability to play instruments because they would get together once a week and they would play, and people were entertained.

And the radio killed that because everybody could just beam in expertly produced music. So has the overall musical ability, average or median musical ability of the US or the world population gone down? Yes, have, but you know, that was a shift, that was a change.

So good writers will still be good writers who are human beings, who may very well use ai, but will be responsible for their final output. And it will either work or not work, and there will be fewer. I'll Tell you that when I consider that, you know, a lot of the people who have been filing copy.

Yeah, I'm not even sure they know how to write This. This is also true. On that note, we gotta take a break.

Yes, Mike? Yes. We'll come back.

Um, we're good. You know, speaking of this whole subject, we're gonna continue with teaching AI to teachers. 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 gonna talk a little bit more about usage of ai, but this time we're talking about how to train folks how to use it. But we're now training the teachers, or at least hopefully John, all the big AI companies seem to have gotten together, and they're saying they're gonna have programs to help teachers make use of AI in the classroom and understand how it works. I kind of agreed at this initially with a, wow, that's awesome.

And then, you know, my pit of my stomach kind of went, well, I don't know how that's gonna turn out. So what's your thoughts here? Well, That's, yeah, I had to say I had the same reaction.

So, philanthropic, open ai and Microsoft announced with the American Federation of Teachers, which is one of the nation's largest, uh, teachers unions, this creation of a AI training center, which will first be in Manhattan. They're gonna spend $23 million over the next several years. You know, actually when I saw that number, it, to me, this is kind of a personal aside, to me, it said a lot about where AI thinks about education and, and, and teachers $23 million.

And you compare that to the billions and billions of dollars these same companies are gonna spend on the infrastructure and investment in ai. So that's something that, that i, I kind of wanted to bring up. But the idea is that this academy is gonna offer this free virtual training to, uh, the, uh, union members.

There are about 2 million, um, K through 12 educators in particular. And the goal is to, to, to get at least maybe 400,000 people in person at this facility over the next five years in some, in some manner, to teach them how to use AI safely and ethically. Now, that's a very kind of broad sweeping generalization of, of the use.

So I reached out to a couple of teachers and talked to them, and their takeaway was they were highly encouraged, but I, and I, this is a big, but they, what they want to do is they want to use AI to create their curriculum much more quickly and efficiently, but they're also concerned about the exposure of it to their students. So there's still this misgiving that they have. But all in all, it's encouraging.

And I, and also within the context of the idea that a lot of these companies already have some sort of education partnership or initiative. So, OpenAI, for instance, has a partnership with California State University, um, you know, to bring its software to 500,000 students and faculty, philanthropic is introduced clawed for education. And, uh, Google has struck deals to bring its AI tools to public schools and universities.

I think what they want to do at the bottom line is as AI takes off, they wanna avoid a repeat of the whole STEM situation. They want more students to have access to technology, and they want the people who teach them to have that access as well, and have some sort of knowledge base to, in which to teach them and make people more comfortable, because it's inevitable. Now, I think Garima mentioned this earlier, it's, it's, it's like a drug and it's spreading.

And I think it, this is just facing reality. It was something we have to do. And I think it's also something that the government is looking at, especially as it comes to competing as a country versus China and others.

Okay, so I, I've got some thoughts on this. So blank. John, I agree with you at first blush.

What a great idea, what a noble thing for these companies to do, right? Because, you know, you can have kids who are self-taught figure it out playing at home. But isn't it great if you had teachers who are competent in this and able to teach it to the next generation and, and make sure maybe that they still know how to write and they still know how to code, but they use, they leverage AI to 10 x themselves to, to be better.

But then I thought about a personal experience I had when my oldest son was in high school. His high school had all these like, areas of concentration and one was technology, and they sent a thing home looking for parents who, technology savvy to join our tab. Our technology advisory board sounds perfect for ahum, right?

I'm a tech dude. I'm in tech. I volunteer.

I wanna be involved in my son's education and make better education for the community. It was public high school. And I went to the technology advisory board meeting, and as it turned out, much like this AI initiative, it was sponsored by our good friends up in Redmond.

And it was great. The school had licenses from Microsoft office for Microsoft VB Studio, for Microsoft security tools, for a lot of Microsoft stuff. And I, I raised my hand.

I said, what about if we want to use, maybe it was Google or whatever, the open Libre or whatever an alternative was, or any alternative. And they said, oh no, oh no, you could use Microsoft. If you don't like Microsoft, you could use Microsoft.

And if you don't like that Microsoft, you could still use Microsoft. And so that's my only fear here is in doing this, are they locking? Apple did this in the education market in the eighties and nineties.

Every school you went into only had Apple computers. 'cause they were big in education. They were the education computer.

Are we locking in these players to train the next generation of AI users only on their products? See, that's, and we gotta work worry about. So They're going to raise these students if all goes well, according to plan on their technology.

You're, you're absolutely right. So they gained an early advantage by seeing, they Lock in, they lock in the DM word, they lock in, But they also see this, there, there's a statistic like two, two thirds of students, like 70% of students use chat GPT in some aspect, and only 30% of the teachers even understand what it is or know how to use it. Oh, abso, I, I agree.

It's a smart move. It's just, but I, this is like a marketing play too. You're absolutely right.

Oh, now you're starting to get at it, John. I also also see, I know, Yeah. I also see two big problems being a mother of, uh, a kid.

I mean, there are two big problems. And the problem is bigger than what LMU mentioned, right? I mean, it's not only, uh, tech lockin think about this.

You know, this is not funded by the government. So this 23 million is funded by capitalist. And how they would re uh, benefits out of it is first of all, influencing the curriculum.

It's not only technology lock-in. So how would you, uh, I like imagine that the curriculum for the next generation would be influenced by these tech giants. And the second problem is that when you will train teachers, and you'll have all the data, right?

Who owns that data and how do you make use of that data? Are your next generation teachers being, you know, AI bots? Think about this.

So I think that, again, there has to be some kind of governance on top of this, these initiatives, because I feel very nervous as a parent if this happens to, uh, the kids, which, you know, so I'm, I'm gonna play into your paranoia. And, and, and so, uh, Trump signed this executive order in April, this White House task force in AI education, and they're asking for these public-private partnerships for K to 12 education, by the way, and AI and its use in academia. And my concern is, who are some of the companies they're gonna be working with?

Because they are, in a sense, they're already working with open AI on the SoftBank, uh, Stargate projects. So I suspect we're gonna see these names crop up. And again, it's like as, as, um, Alan pointed out with Microsoft, it's not just education.

They, this is lock-in strategy no matter what. The market is the exclusion of others. So, good guy, you wanna say something?

Go ahead. So it seems like Alan is playing the Mike Baard role. He's scratching his head and saying, gee, you know why this is, they're just trying to lock people in and, and, and listen, it's a good point, John.

You made a great point. Darma too. But I am, I'm gonna play the Allen role, and I'm gonna call b******t on this.

This is bull Pucky from beginning to end. The $23 million is a tell. This is a marketing effort to say, Hey, you know, we realize that, uh, AI is, uh, worrying you.

So, so we're gonna train the teachers. We have been through these before with other areas of culture, society, and technology. What do you think this training is going to do?

What do you think its purpose is? I've taken these trainings, not in ai, but in other things. The whole idea is to sing you, to sleep with happy.

Talk about how you can cooperate and it's gonna help you and it's not dangerous and all this other sort of stuff. And as far as the substantive, practical matter of how do I utilize this tool to be more productive, to especially to produce higher quality stuff and to be more reliable, that's gonna be completely absent. I have no faith in this whatsoever.

And I, I love anthropic. I think Claude is awesome. I love what Microsoft is doing.

Uh, I will leave it to everybody else to decide why I didn't mention the third brand that's in this announcement. Um, so, uh, that's great. And I'm not calling evil on anybody.

I'm saying this announcement and this effort is horse crap from beginning. No, but it Also, and you know, the also thing, guys, this thing, it just fill, fizzle out. $23 million is nothing.

I mean, this is, this is Just, they'll keep the marketing up, man. They'll keep it up. They'll keep the, they'll, they'll keep the brainwashing of the teachers going.

Like, you know, teachers are, are, are great at teaching. They're not necessarily experts in how these types of tech systems work, right? I, I will defend the teachers because a missed several of them.

Several of 'em are my cousins, but they're smarter than you think. Yes. And Randy Winegart, the, the, the head of the tea, the, the teacher's union is no fool.

She's, she comes outta New York. I come out of, we've known her for years. Um, I'm, you know, so I'm just like, I think they'll see it for what it is.

But you know what, when you, when you're starving in the desert and someone opens up a hamburger stand, it's a good business. Um, right. You, you, you know, you can tend to get a little choosy, less choosy about what, what kind of meat it is.

It's, it's meat. It, that being said, though, to the point made about, well, we don't have government regulation on this, and these tech vendors can unduly influence the, this organization, And we just passed a ban in the United States on government regulation of ai. Why Is second?

Why is that any different? No, we didn't pass that ban. I, I think it was taken out, guy.

Oh, okay. But sorry, my bad. But Why is that any different than tech influence on the government itself?

It doesn't, hasn't big tech kind of bought this lock stock and barrel already? Yes. Same old s I'm gonna use the S word again.

Same old s I'm just saying like, I, I, it's, this is, you know, a velvet glove. The iron fist is everybody trust ai, use our ai, use All ai. But that being said, hey, so what, what would be the alternative that we'd like to see?

We'd like to see the government undertake a training program to make teachers teach AI better. Right? And that's gonna take dollars in a time where what has passed guy is you could use your vouchers to go to an alternative school and public school teach, you know?

So public schools will get increasingly less, at least in the US I'm talking about increasingly less resources that they'll have available for things like teaching teachers about ai. So if we don't have these private public partnerships to do it, we can't, we don't, there's just not the resources. That's a bigger discussion.

But I'll just say that in, in the long term prospect, I'm optimistic and not because I think the government needs to intervene. Fair enough. All right.

So all we need is a bigger version of this. Is what you're kind of saying is that Maybe with a little oversight. So it's not, you know, so it's not, I mean, this is, this is a privately run AI academy for teachers.

That's what this comes down to. I, I don't have a lot of faith in the education department, department of education, especially when they had A, I thought is going away. It's a one, right?

Right. Yeah. No little teaching.

This is sort of an oligarchy move, right? Oh, yeah. You know, it's sort of absolutely.

You know, the oligarch saying, don't worry, we will be your benevolent, you know, rulers. But let's take a break on Textron Gang. Let, let's talk about cybersecurity.

I had enough of the of Big Brother ai. com is the leading resource for news analysis and education on challenges facing the cybersecurity industry. com covers all aspects of cybersecurity, including data security, DevSecOps, cloud security, application security, network security, security threats, and more.

com has the largest selection of security content featuring breaking news, blog posts, podcasts, and more. com to learn more. com.

Home of security bloggers network. Hey folks, we're back. And we are gonna shift the gear, as Alan said, to talk a little bit about cybersecurity.

But there's one of these ransomware gangs out there call hunters, and they sent out a note saying, Hey, basically they're getting outta that business. They even sent folks the keys to decrypt their data. And other folks are saying, well, maybe this is just another head fake.

But Alan, what's your take on what's going on here? You know, that these people suddenly wake up one morning and decide that they needed to do an active contrition. So first of all, let me say this is on an article on Security Boulevard.

Should be in your ticker, check it out. But Michael, you know, this isn't black hats turning to white hats. This is, you know, you're saying they're changing jerseys.

What they're really doing is going back in the locker room, putting on just a different number, Jersey. So you don't know which bad guy is which, but they're still bad guys. Right?

And this isn't the first time these people have done this kind of, you know, change of, uh, change of protocol change, of, of attack, vector move, right? They're, they wanted the big RAs ransomware as a service operators. And you know, this is now the second time where they're saying, Hey, we're pivoting.

It's when you think about it, right? This is a very mature ecosystem where you have companies that pivot, but let's be clear, they're not pivoting to the good guy side of the screen. They're not here to help you with your ransomware.

They're not here to release new zero day vulnerabilities in a responsible disclosure method to prevent, or, you know, bad guys from hacking you. That's not it at all. They're just saying, this particular attack vector has gotten a little bit harder.

It's gotten a little bit too crowded, a little too competitive. It's, I'm not making the profit margins I used to, and we're gonna go find something that is, uh, profitable or maybe easier. And that's what they're doing.

So they, they, you know, they have the same jerseys on just maybe different numbers, so you can't tell who's who, but they're still bad guys. So are you saying, Alan, that uh, that, uh, this particular route of profit, criminal profit is drying up for them? So they're, they're shifting to another and it's waving their hands over here like they're Yeah, Yeah.

I mean, overall, so the ransomware encrypting your data ransomware game isn't paying the kinds of dividends. It was, let's say, during COVID three years ago, two, you know, two, three years ago. Um, for a lot of reasons, right?

People are, are wise, better educated, so they don't fall for it as easily in spite of ai, improving the phishing, um, better, better solutions in terms of keeping offline copies that allow you to, to, to, uh, you know, reinstall or, or overcome the ran the, the encryption of ransomware, um, cyber insurance companies getting involved and knowing what, what the going price is and stuff like that. Overall, ransomware, the margins on running a ransomware business are not what they were three years ago. There's probably lower hanging fruit, more money to be made doing some of, I mean, even the ransomware people have moved from encrypting your data to doing something else, right?

To, to hostage you basically to ransom you, right? Ransom is not just encryption anymore. And so I'm sure we'll see them resurface with a new game in town, right?

They go from three card Monte to the Shell game, You know? So do you think that this represents some level of progress, though, in the fact that they feel compelled to go shift? And is this an example of maybe where we're actually winning, or is this like a Charlie Sheen definition of what winning needs, If it'll make you feel better, you know, yeah.

Where we, we we're making progress, but this is, this is the cat and mouse game that is security. Before we called it cyber. And before it got so fancy, it's always been a cat and mouse.

It it more than cat and mouse. You know what, it's really like Roadrunner, Roadrunner and Coyote and the security guys are Wiley e Coyote dealing with Acme. And soon as Acme sends me a good unencrypted ransomware kinda solution, Roadrunner comes up with something else, right?

We're, we're always a step behind. And so nothing lasts forever, right? We still, you know, we don't have code red worms infecting us or, you know, the love, the love virus or whatever it was called.

This is like, this is the way of the world in security. You move on to the next scam, right? You move on to the next attack vector, you move on to the next, you know, money maker in, in terms of how they're doing it.

And they will, they did it before. They weren't always what they are now, right? That it's just, this is security.

Yes. So, you know, it took us a little while, but we, we've definitely made progress on encrypted ransoms as a, as a vector. Mm-hmm.

Or guy, let's look at it another way. Won't they just come back with a more efficient model for doing the same thing? Who knows, maybe they'll use AI tools to start generating these attacks and they'll find a more economical approach and then they'll just be back doing it again.

Yeah, I don't, I just, it sounds like Alan's saying what I was saying, um, roughly in the last segment, uh, which is that everything is old. That's old as new again, it's just, it's the criminals are gonna crime and if they are cyber criminals, they're gonna cybercrime and, uh, you, it's whack-a-mole all day long forever, right? That seems awfully pessimistic, but nonetheless, um, uh, you know, it's The reality.

It's human nature. Human nature doesn't change. So, No, but it's also the reality of being a cybersecurity professional.

And there's a reason why we have the burnout rates we do in cyber. There's a reason why we have, you know, some of the issues that we have in, on a, on a human level with our cybersecurity pros, because it, it is frustrating. You do make progress.

You fight the good fight every day. You fight the good fight, you try with limited resources going up against gangs like this that are, you know, making money hand over fist or well-funded oftentimes are under the benevolent eye of a, of a foreign entity that allowed them to operate. And, you know, just, and, and it is a whack-a-mole game.

And, and after doing it for a while, you, you get tired of it and you, you kinda get jaded and desensitized and, and all the other bad things that come with it. And, uh, you know, it, it is, it is what it is. You know, black hats coming up next month.

I'll be out in Vegas doing video there. And, you know, that was always the premier place for security researchers to show, right? It's not RSA where it's the, the part the industry throws a conference for itself.

Black hat it used the lease used to be, um, the place where security researchers go to show off their latest research. And I'm sure there'll be new vectors and new attack methods that are gonna be shown at, at, at Black Hat that take the place of sort of your encryption encrypted ransomware, let me be clear, encrypted ransomware is only one kind of ransomware. There are other ransomware, blackmail kind of things and stuff that are out there Point taken.

But can, can we ask the wizards on the call, um, Gary Ma and Mike Ard was almost literally a wizard and his last name and, and John our, our Silicon Valley dean. Um, I feel like in the broad scope of history, crime has gone down and new, new avenues for crime appear, but why wouldn't they follow the same curve? Now, it could take decades, but why crime went down, you know, across The world?

Oh, no, no, sorry. No, Let's get into that. I mean, um, you know, let's take for example, the lottery.

I can go buy a lottery ticket and it's perfectly legal, and yet I'm more than familiar with people who are still running numbers games down in places, and it's still very popular, even though technically it's illegal and the people running it are definitely mobsters, but we tolerate that. So my question is, back to you guys. Are we really after eradicating cybercrime, or are we just trying to contain it so some level of acceptable cost that we can all kind of then wink at?

So that's an old view of, of security, right? 5% cost. 5% to, you know, call it shrinkage, via cyber, via via security, via computer security.

And to them, and that was an acceptable amount built into the model because they were highly profitable. It was built into their margins. And, you know, this is before e-commerce was as big as maybe it is now, right?

5% was acceptable risk. And that right there, guys, not just guy, but ladies and gentlemen, is security. Security is not about stopping every attack.

Security is not about having zero risk. Security isn't about whack-a-mole. Every mo security is about managing risk.

And so at the highest level, a good ciso, a good risk management team, looks at what is my risk from ransomware, from encrypted ransomware versus my risk from a zero day vulnerability on my cloud facing server, versus the risk of other risks that I deal with in business. And they make a decision how much, how much will it cost to mitigate, right? My, my risk of attack is 60%, I could mitigate it down to 20% at a cost of X dollars.

It's an equation. And so do I wanna pay X dollars to take my risk from 60% to 20%? Well, how about if I pay Y dollars and take it to 40% instead?

Yeah, that's the annual loss, annualized loss expectancy that we've been talking about for years. And these, no, always really put it practice. Look at large enterprises.

This is especially retail e-commerce. This is what they go through. This is the exercise, You know, I'm gonna date myself, but I remember years and years ago, I remember writing about eBay and the fraud and how that was an issue, you know, back in the day.

And they would just say, look, it, it always comes down to a certain percentage of acceptable losses or instances of everything. And they would always, they always would give me a a certain percentage thing, but I always, the issue with me sometimes though, is their percentage remain the same if the business was booming. So fraud in terms of dollars lost was, was increasing every year.

But you're, you're right, it's a formula for them. It, it's a, it's risk management. And at the end of the day, sometimes even good security people lose sight of that because we don't want even one attack to get through.

We wanna whack every single mole. Yeah. So let's, I mean, uh, also our technology to this mix, I mean, this is a business perspective, but if, you know, I talk about two things because I think it's essential to talk about the good work which we have done.

Organization like my three has also good, uh, done, uh, excellent job in, you know, protecting our businesses and laying out a strategy. And if you look at this particular, you know, threat vector, um, they have published specific white papers, how to contain them. So, uh, I would like to pat their back on the other side.

We also do not, uh, also shy away from the fact that quantum is coming, you know, so there might be some, you know, this is a emerging technology game, right? So we need to un ensure that we are equally aware that what are the next steps or next tactical steps in this process of, uh, you know, changing ties, right? So maybe that is something we, we should start to think about.

And organizations like Mitre and other, you know, organizations who are working in the space would be actively engaged in figuring it out. Uh, the next steps There. And there is, and you know, we're about outta time.

Let me close this segment with this though, some good news as we sit here today, we are more secure and it is harder for the bad guys to be successful than they it's ever been. And every day we get better, they get better too, but every day we get better. So especially for my security people out here, you're my people.

We know how hard you fight. We know how hard it is doing this job, and we know how many holes you close and how many vectors you close off, and the great work yeoman's work that security professionals do. And it is appreciated.

And it, it does get better. It's just in many ways, the the game is set. The rules of the game are set as such that it's impossible for us to total, to have total victory, right?

Like World War II type total victory. The bad guys are never surrendering. I don't care what these ransomware scumbags do.

They're bad guys. They're bad people. And that's who they are.

We do the best we can. So that's it. I'm gonna step off my soapbox.

Is that all, is that like the cyber thin blue line? Is that what that Was? Yeah, exactly.

We'll be, wait, we've got a great tech drug tv, uh, show following this. Actually, we have, we have our good friends at Tech Field Day. Steven and the Gang have a Tech Field day going on, so we'll, we'll have you take a look at that as well as Techstrong tv.

We'll be back tomorrow with more. Until then, John Garima guy, Mike, thank you. Thank you for watching this.

Alan Hummel, we're out. Hey everyone, welcome back here to Textron tv. My next guest is Jared Zonar Reich.

Jared is, I think I got it right. Jared is the founder and CEO of a company called Prompt Layer, and it's his first time by Techstrong tv. So let's welcome him.

Hey, Jared, how are you man? I'm great. Thanks for having me.

Pleasure. So, you know, you found you're the founder, CEO here. Tell us, I mean, you didn't wake up one day I, and said, oh, I feel like doing this today.

What, what drove you? Where's your passion for what's behind? I feel like doing ai.

Yeah. Um, don't we all? Yeah.

So we, we started, we, we launched this product about two and a half years ago. Uh, me and my co-founder and my co-founder and I known each other for like a decade from hackathons back when we were, we were in high school, and, uh, we, we were working on something before it, it didn't really have a lot of traction. And Chacha PT had just come out and we're like, how, what are we gonna build with ai?

I was at a AI startup before this. He was doing machine learning research. So we actually had a list of ideas and prompt layer as it exists today was the idea we made to get through those other ideas.

So we said, let's make a tool to just help us build things. We ended up releasing it in the January after Chadi came out, like a month later, and it blew up. And we said, all right, I guess we're prompt layer now.

Go figure. Sometimes just serendipity, right? If you build for yourself and you're customer number one, it makes a lot easier.

Well, you know, so Jared, over the years, it's been a lot of years, I've spoken to a lot of founders, and you know, that's actually a very common kind of pattern, which is, you know what? This, I've got this problem and I don't see a solution out there for it. I'm gonna build the solution for my problem.

And now, hey, I can't be the only one who has this problem, right? Yeah. There are other people who probably want this solution as well, and there's been some amazing companies founded on that very premise, so it's all good.

Um, Jared, what you, you talked a little bit in here about like kinda what drove you to, to, you know, found promptly or two and a half years ago, fill us in on the last two years. And, you know, I'm sure things have changed, right? The the, the solution has changed, the problems people have, the use cases give us, give us the background.

Yeah, totally. So I think where we started, uh, so let me, let me start. What was the first product?

It was basically a tool. We were building different hacks with AI and building different projects, and we wrote a prompt and then we updated it and we forgot the good prompt. So it was pretty much just a logging observability tool to remember that last good prompt.

That's still a core part of what we do, but we do a lot more now, like you said, we're really, we like to consider ourselves the, the workbench, like the all-in-one workbench for building AI systems and building AI products and workflows and building them as a team. So what that means is we, at the core, we're prompt management. So how do you version all your prompts?

How do you log them? How do you edit them? But also how do you eval them?

How do you test them? How do you, how do you unit test these agents and make sure the outputs are good? And then how do you lock the results?

So it's really this way. Teams iterate. And I think our core thesis, which we've developed, honestly, we didn't start with it.

We're both engineers. We built it for engineers, but slowly our customers kept coming in with non-engineers on the team. And the first time that happened, it was actually a, like a parenting coach and someone, they, they brought on someone who was a teacher for 15 years, she never coded in her life.

And we said, what are you doing? What, why are you using front layer? What's going on?

And she explained to us that the engineer set it up for her. She goes into prompt layer, she edits it and opens the app. And that's how they tailored the voice.

And she was the one who knew what the output should be. If you're building legal ai, engineers don't know what that is. Or, or psychology, mental health ai.

So that's our core thesis. Now, the best prompt engineers are not machine learning engineers, and you need to include the whole team in it, in the process. It's not just an engineering pursuit.

And it's really, it's a new type of knowledge work that, uh, is, is being built. I, I agree. You know, I I've been saying for a long time, you know, they estimate now, I don't know, there's 40 million something, maybe, maybe 40 million people around the world who are code, who code.

You wanna call 'em developers, call 'em developers. But with ai, we're gonna have 500 million people Yeah. Who code, right?

Right. Who, who, who develop stuff and, and that's gonna be the world, right? And so we need, we need prompt layer, like applications that are gonna service these 500 million people who are developing apps, Right?

Right. There's a, there's a new, there, brand new persona that's gonna be involved with building software, and it's not someone who has a PhD in machine learning. It's these new type of coders, like you're saying, it's the, uh, it, it, it's really the person who knows what the output should be.

And that's what makes it so much more powerful. And I think there's a lot of people, you know, there's a lot of opinions in AI these days, and, uh, one of the opinions people take is the opposite, that, Hey, this is still machine learning. We need to get really deeply technical, but I think those teams are gonna lose because frankly, engineers, I'm an engineer, so I'm allowed to say it.

Engineers don't know what the output should be. A lot of times they're not the ones who have the sense and the taste and the product insight. They Right.

They don't have, they have the engineering background, you know, the coding background, but they don't have what the, what's the special sauce of a particular app or, or a subject matter expertise even, let's call it. Mm-hmm. Right?

But you know, Jared, in the low code, no code space, right? We called these people citizen developers. Yeah.

And a lot of people didn't like that name. A lot of people did like that name, but a lot of people didn't. And you know, the idea is, is that they're not professional developers.

They're, they are developing, most of them are developing just one app for a very specific purpose that suits their, you know, whether it's their hobby, their passion, their job, or what have you. They feel driven, I need an app that does this for me, not really looking to learn best practices or get a PhD, as you say, I just, this is what I want, right? And I, I, and I need it.

I need it for work, or I need it for me, or whatever. Um, now look, in the last couple years we've seen, you know, crazy amount of evolution in terms of how people work with proms, how people work with LLMs, how we were discussing it on Textron Gang earlier today. You know, when it first started, all these LLMs were based on, you know, a, a a, a fact of, you know, it was a 2-year-old ball of, of, you know, a snapshot of whatever they were using a subset of the internet today, when you, you know, talk to your prompts, they're going out live on the net and grabbing stuff.

So it, it, that makes the world very different, modern, in a modern workflow of, you know, building LLM coded apps and so forth. What, what's changed? Is there a best practices, Jared, or are we still sort of in the evolving practices?

Yeah, a little of both. A little of both. I think, uh, I, uh, there's a few things to say here.

I guess firstly, if you're building an AI application or building something with lms, uh, perfect isbe enemy of, uh, of, of, of, of duck. Oh, good, right, of good. Uh, the problem with perfect a lot.

I see so many times, so many teams, engineering teams, real hardcore teams, they're trying to get this perfect prompt and this perfect eval where they get a score, a metric that says how good it is, and they can fix it. Um, you need to prove that the AI application is gonna work to your team so you can invest more resources to build these evals. And a lot of times, the way to make the prompt good is to start logging how people are using it.

Because the cool thing with LMS is there's so many different ways to use it. Uh, it's a new paradigm where users can enter anything basically into your AI application. So you need to start tracking what they're doing, using those for back testing and historical testing or regression testing.

And until you release it, you're never gonna really have that. So we really encourage teams start best p first, best practice, start simple, get something kind of working, and then you could build the edge cases from there. And don't start with like a a 50 node agent that's doing everything.

Start with a very simple stateless prompt, and you'll be surprised how good it gets. And then the other best practice is to do it in a clean and organized way. I think, uh, the, one of the big problems with these things, and what scares a lot of people is you have prompts scattered in your code.

You have strings everywhere you have, you don't know which version is which. What our tool focuses on is how do you organize these, how do you what it, or GitHub, basically for prompts, how do you version, how do you know which one's working? How do you connect these with tests instead of, it's the wild west, uh, but a lot of stuff is up for grabs.

How much can single prongs do? How many times do you, how many models do you need to combine, which is open source model? Which model's best for this?

And my answer to all of it is assume it's a black box, and it's just try and check. And that's the best way to do it best. That's good advice, man.

Good advice. It's a good way to say nothing at all. Well, no, but I, I, you know, I think, so look, I've been through a lot of tech cycles in my career and, you know, there are people who sit on the sidelines and kinda ring their hands, should I do that?

Should I do this? I, I always believe you jump in and that's the best way to learn, right? Mm-hmm.

And so that kind of fits my personality. Um, if someone's out here and saying, you know, I like what Jared says. He seems to have a common sense approach to it.

This prompt layer thing sounds pretty good. I'd like to check it out. What, how would, what would you tell 'em?

Yeah, how do they go about doing that? com. Uh, you could sign up for free.

Uh, we have a lot of free individual users, and we have a lot of really like enterprise hardcore engineering teams using it. So it's a pretty, uh, pla a pretty platform for a lot of people. You can follow me on Twitter as well.

Uh, we give some updates there. And, uh, yeah, I encourage you to sign up and let us know any feedback you have. I love it.

com come? Yes. Excuse me.

COM Hey, Jared, I hope this, I told you it was gonna be an easy interview. I hope this was okay for you thing we left out, you think any, anything else you want to share? Probably the, the thing I'd say is it's not too late.

We're still super, super early in AI and lms, and whether you don't know how to code at all, you're a content writer who doesn't want to get left behind, or you are a machine learning engineer who wants to be the one who runs LLM systems for your company. It's not too late. It's very easy to figure out these things.

At the very core, it's a input and an output. Don't even think about how it works, and, uh, you'll become an expert very quickly. I love it.

All righty. com here at Tech Trunk tv. We'll take a break.

We'll be back. Hello and welcome to the latest edition of the Techstrong AI Leadership Insight series. Today we're with Mark Kuper, who's the CEO for over Proof, and we're talking about how AI is being applied in the beverage industry because, well, there's a lot of things that are related to each other that are not intuitively obvious.

Mark, welcome the show. Thank you. Glad to be here.

So, what exactly are your customers doing with ai? Because a lot of folks would assume that after, I don't know, several centuries of understanding how to sell beverages, there would be nothing new under the sun, but what do we discover? Yeah, it's interesting.

So the, the beverage alcohol space globally, uh, and particularly here in the us, um, they traditionally have like under leveraged, uh, data. Um, it's a very fractured market. So, um, the, the industry as a whole tends to run behind on trends like this, uh, in general.

Um, and what we are doing with over proof is not only helping our customers understand the market better with the use of ai, but really present ourselves as a data partner, an AI partner, and a solutions architect. Um, help them navigate how to apply AI beyond just the data that we deliver, but also like how to improve their, um, other verticals like risk management, supply chain, sales, management, and support. Well, gimme an example of how somebody in the beverage industry is doing that.

I'm sure you don't need the name names, but just describe the use case In the US specifically. There is a three tier distribution system, and what that means is a supplier, our customers have to sell to distributors who are the middlemen and then sell onto the retail. So that's on-premise, which is hospitality or off-premise liquor stores or any, any store where you buy your product directly from the store.

And the complexity there is that there's always information coming from has to channel through a distributor. Um, and some of that information is not shared strategically, um, or just very delayed or fractional. So what we do is we help our customers understand where they are present on menus.

Uh, that's a big part of what we do, at least, um, by analyzing, um, 80% of menus out there, uh, that we collect, looking for our customers brands, and then tell them in relation to their own sales performance, how a specific cocktail is, is performing for their sales volume. Um, a perfect example, uh, we have a customer that has a national, actually a global focus on the espresso martini. And this is an an up trending cocktail?

Absolutely, but not in every market. So we analyze their, their local sales performance for vodkas. They have a few in their portfolio, and then look at, okay, which in which cocktail is that particular drink, uh, performing the best?

And in some of them markets, they strategically decided to move away from a focus on their brand in espresso martini. And, and this case they moved back to the Moscow Mule, uh, and they saw an immediate sales list sometimes, uh, in, in the range of seven 8%, which is for those bigger companies, easily millions of dollars. Does that resonate?

Yeah, I get you. I've been in a couple of bars in my day, but, um, I guess my question to you is, is it on the menu or are there other data points to collect? 'cause sometimes when you go into a restaurant or a bar, it's something that's sitting on the bar itself or behind the bar that kind of tips you one way or the other.

So what kind of data points are we collecting? We, we, we are focused on, um, on menu data. Um, we have some software applications where we can track, you know, back bar placements or in the stores when, when, uh, samplings are done.

Um, so we track multiple data points, but where, where the volume is, is really generated is a placement on the menu and not just in the drinks list, like specifically in a cocktail that drives vol. Uh, the, the majority of volume for our customers. Um, once you're an established brand, um, it's, you know, the consumer starts calling for a drink.

So that's when most of more mature brands focus on a back bar placements, what's what we call it. Um, and that is becomes important, but the majority of brands are still being built, are still trying to reach the consumer. And it won't be like, like you would you, I can mention a brand like Tito's.

Tito's on the rocks is something you hear a lot, but a a new vodka brand, you won't hear the consumer call that. So you wanna be in the menu, and we're helping our customers understand which menus to target, which venues to target, and what specific cocktail, uh, is highly, is best recommended for, uh, each brand. Is it my imagination or are there more different types of cocktails than ever?

It seems like every time I go into someplace somebody's invented something new and different. And is that a deliberate strategy because they're experimenting with which things might resonate with the end customer? No, absolutely.

So, um, trends are changing all the time. They're seasonal. Um, uh, some are related to literally the temperature outside, but, um, sometimes it's, it's just, uh, a movement of, of trends, um, that's not necessarily related to the weather.

But, um, what, so where I think, um, AI can, can really play a role is, is predict when a trend is upcoming. So the aggregate of menu data and actually sell into a bar with the, with the notion or with the, with the da, with the data back report, saying that we are seeing that the, uh, specific spreads not to call out ma brands all the time, um, is, is popping up on menus all around you at a specific price. We suggest to take this like our brand in this cocktail and, and help the, their customers, the, the, the bars and restaurants to, uh, to, you know, be be be ahead of a trend.

Also. Is the AI you're using largely predictive, or is it generative, or what type of AI are you applying to this Predictive when it comes to, like, we can predict when a bar, uh, is due to change the menu, the menu cycles? Well, it's called some bars change on the season.

Some bars change, uh, three, four times a year, specific dates. Uh, we have historic data of about three years. Um, so we can, we can actually see or predict when a bar is about to change a menu.

The other part type of AI that we use is to classify these strengths. So to your earlier question, like which cocktails, how do you classify a cocktail like a martinis or martini with they're leechy martinis? There's, uh, you know, some people would say like the, the classic martini is, is vodka based, orgin based, there's espresso martini, there's the, uh, uh, passion fruit martini and so forth.

So our models automatically classify, and once there's enough data or enough of the same, uh, data points that we find, let's say the passion fruit martini, then we start to classify specifically as a passion fruit martini, because the data is, is large enough to start doing that. So the classification and uh, uh, segmentation of a, of specific cocktail is where, um, where we, we have a lot of, um, benefit of, of ai. Um, and then the other part is the sales performance.

Sales performance is more looking at correlations between sales or, um, consumer data or social media data where we bring, um, where we do basically proactive, um, analysis or automated generated reports for sales forces to act on. So this creation of actionable insights, um, that traditionally happen through an, an army of, uh, analysts at the larger companies, smaller companies don't have that luxury of having so many analysts looking at data. So that's where, um, the, the AI comes into play, um, to automate that and provide Salesforce with, uh, actual insights.

Are you also applying this to things like beer and wine, or is that more closely or tracked already? No, so definitely, um, we analyze every aspect of the menu, including food, where beer brands can use the, the data is, is understanding like the composition of, of, um, um, of a menu, like how many are, um, draft on draft versus on the bottle, what kind of drinks, uh, what kind of beers are, um, on draft or, you know, sold by the bottle. Um, and even certain flavor profiles.

So we, we use online data to understand like what kind of IPA, is it more fruity or more hoppy? And then we, we index, um, that kind of information as well. Another, uh, application of using more than just, um, the spirits data is the food menu.

Um, so we, we help the Salesforce of our customers understand what, what cocktail fits best with a couple of drinks, uh, uh, food items on the menu and automatically generate a, a, a sales template where they can say, train the, the waiters in, in a specific bar, which is a very, uh, common thing to do. Uh, you wanna train what the wait staff on, on your product. So when you do this, we, you can recommend, like, these are three, three items on your menu that pair very well with my drink, specifically in that cocktail that's listed.

And you give them kind of a sales argument to upsell from one drink to, um, um, to the drink. You're, you're selling, we're applying that to wine and we're applying that to, um, to beer as well. What about the retail outlets themselves and all the liquor stores?

Is that another place where you're collecting data from or is that already in my point of sale system anyway? So, um, in the off premise, which is liquor stores for the mo most part, um, we have a, a, a software application that, uh, tracks the execution of sampling events. So tasting.

Um, so we track who, who, which consumer comes in, uh, male, female, approximate age group, did the person tried, the person liked the person by. So it's really conversion data. Um, we correlate that to the weather, uh, to location in the store, um, obviously the timing day of the week, time of the day, and, um, bring that in context of, of, um, of the may trend as well, and help our customers, um, understand when and where to best do their samplings.

Um, and, um, and then optimize even to the level of like, which talent should we use again to sell our product in, um, during those events. So also there, it's all about data collection and then bringing back the insights, uh, to optimize the execution. So how do you know when I go to the bar and I order drinks that it's actually for me because, well, I live in a multi-generational house and I might be out with my father-in-law or my son and some of the stuff that they drink I wouldn't touch with a 10 foot bull.

So in a bar, it's, it's usually, it's the flavor that, um, that, that, um, that informs a decision. Um, so what we are seeing a lot is, um, customers using, um, regional data or local data to help sell, like this is a specific flavor profile that works for, for this area, and that typically is aligned with what consumers mostly like to drink at the personal level. Like it's, it's, it's still, um, I think a challenge to, to help a consumer make a decision other than the example I gave earlier where, uh, you could recommend a specific flavor that goes well with, um, with a food item.

So if someone orders aino, um, our data shows, for instance, that a a user infused product works pretty well with a bruna and, and it's ordered often. So we're using the data there more as a recommendation engine, but not really to understand what the consumer wants. It's more like a recommendation.

So yeah. 'cause in that scenario you might be accused of actually stalking people to a level where that might be too granular, right? Exactly.

Exactly. Well, what we are working on is a, is a new application where, um, after tasting in a store and someone likes the product, um, or even purchase the product, we help, um, convert that consumer like to go to a bar and actually order that, that drink or that, that, that particular product. Um, and that's a very hard, um, not to crack.

Uh, and the way we're doing that is basically a QR code, uh, coupon sponsored by the brand owner, so you can get a discount, um, at the bar level. But that's something that's pretty tough to do, Is one of the things that you're kind of taking advantage of is that people have a lot of affinity for whatever it is that they decide to drink over time. And, um, they're creatures of habit and they tend to order the same thing over and over again.

Um, uh, on the one level, I imagine that's a good thing. On the other end, it might be a bad thing. So how do you nudge somebody over to something in a way that doesn't annoy them?

Uh, that's a golden question. Um, uh, if I had the answer, I would probably also have a, uh, more focus on, on that in the actual sales part of, of this. But, uh, I don't know if I can answer that like a data where AI can assist in that.

Um, it, it, it, it, that's more an interaction, especially in bars. So in hospitality, it's really about the recommendation that the server or the bartender provides you with. Um, so it's, it comes back to training, um, and that's what a lot of brands focus on to train the wait staff and, and, and to, to win over the hearts of bartenders because they ultimately are the one introducing you to a change of your habit.

Um, and, um, the data can help them to better understand like when to do, like the recommendations or the, the, um, food flavor fit, that kind of stuff. Um, but in the, in the hospitality, it's, it really is about winning as many bartenders and waitstaff people for your brand. That's why it's not easy to, to win in this market.

Well, folks, you heard it here. A lot of people out there are pretty keenly interested in what your favorite alcohol beverages in any given moment. So next time you have one, just think of all the people who wanna know what you're drinking.

Hey, mark, next being on the show. Thank you. Appreciate it.

And thank you all for watching the latest episode of Text Strong AI leadership series. You can find this episode and others on our website. We invite you to check them all out.

Until Lynn, we'll see you next time. I have a great pleasure of being joined by a wonderful panel here. We will make sure that everybody's mic is working.

Steve, would you introduce yourself? Sure. Happy to.

Uh, hey everyone, my name is Steve Boone, uh, executive of Product Growth at Check Marks. Uh, excited to be here and, uh, answer your questions. Wonderful.

Yeah, Naomi, Excited to be here. My name's Naomi. I work at Contrast, the best security company ever.

Best one on the stage, right? Well, my Boss, my boss is Oh, I see, okay. Right There.

Um, but no, I really do love it. I am the senior director of product security. Happy to answer any questions about that.

Fantastic. Tyler, Hey all. I'm Tyler, head of product at, uh, Sonotype.

Nice to be here And fantastic. And we have a special substitution at the end here. The slide is not caught up with, uh, Rami sauce, who's joined us, right?

Welcome, Rami. Hi. Yes, thank you.

Hi everyone. Uh, Rami, I'm one of the founders and CEO of, uh, man filling in for Baal because this was originally my project before Baal came in. So Good to have you all here.

Let's start out with kind of where we are with AI powered AI assisted development. You know, there's a lot of data out there, whether you kind of intrinsically watch LinkedIn or, uh, TikTok, somebody talking about vibe coding and what they're doing with the latest development tools. Or for example, we've done our own research just around how many people say that they're using AI tools in development, standing where between 40 to 60%.

That could be AI assisted code completion, it could be vibe, coding, et cetera. I'd love to get each of your sense of where are we of thinking about security and all of this, because of course, we've never rushed before and tried out new technology before we thought about how we were gonna secure it. So, you know, we'll know to do it the right way this time.

Uh, maybe Naomi, you wanna give you a start and give us your, I'm gonna pick on you, uh, if you wanna give us your, kind of your sense of, are we talking about security in terms of how we create software with AI power tools? I think actually this is like another, it's an old story told again, because if you think about who's old enough to remember the first time you ever used open source or your developers begged the security team, can we please use open source in our projects? And you're kinda like, Hmm, I don't know yet.

And now you look back, you're like, well, that was hilarious. 'cause everyone uses open source nowadays. I think we're just experiencing the same thing.

An old story told, again, the use of ai, our developers are using AI begging security teams. Is it okay to use AI in 10 years? People are just gonna be like, how did we ever try to stop you?

It's just that story told again. And I don't think our job is to, to say no. I think we're gonna try to tell them how it's always the purpose of security.

And the fun part is, you know, playing alongside them, making sure they're doing it right. Um, to answer your question directly though, right? Actually, I did forget your question, so I'm just gonna keep talking.

Question was, are we talking about security as we're using AI powered? We are because the developers are gonna be doing it without you, with or without you. So if security is alongside them playing nice and making friends, they're gonna start listening to any guidance that we can give them.

What we're trying to figure out now, it's a little chaotic as a security, uh, industry, we don't really know what we're talking about yet because we don't understand what they're building yet, understand what they're building, it'd be a lot easier to give them guidance. I, I really liked your analogy of kinda the open source usage. I think, you know, when we think about open source in your organizations, my mind goes to, uh, supply chain security, right?

Understanding what's coming into your organization, especially if you don't take ownership for it. So AI is no different, right? If our developers are leveraging ai, uh, the best time for us to be looking and analyzing and scrutiny, that code is at the moment, moment that they're actually generating, right?

So if we're gonna encourage our developers to leverage AI as a new powerful tool to help them generate faster code, then we also need to support them with the right tools that allow them to scan that code, get instant feedback on the quality of that code, and give them confidence on whether or not to use that code. Because it's not a silver bullet, it's not a magic wand. And so really our job now is to start putting those right guardrails in place so that organizations can use AI more confidently.

Yeah, and I think that there's Two vectors there, probably more than two. There's the coding assistance you talked about. Mm-hmm.

And then there's the AI models and, and technologies that being put into the software that the developers created themselves, right? So there's the kind of like, uh, you talking about open source components, like a component selection. You're making a model selection.

So there's that vector, and then there's the vector of using the, uh, coding as systems and toolings. I think we, uh, it's our responsibility and privilege to work in both areas. Uh, I think the coding assistance get more press, um, but I don't know if they're more important or actually less than what's being brought in to the, uh, data science pipelines and product delivery pipelines in the products that are being shipped and, uh, provisioned, Right?

I think there's also the issue of, uh, of scale, volume of code being generated, right? So with automated tools, you can, one person can create, I dunno, three times, five times, 10 times more code than an individual contributor typing on a keyboard. And so that's almost by definition going to cause a lot more vulnerabilities to occur in that code.

That's one. Two is, almost all of these models have been trained on open source corpus, which inherently is less secure than what you'd find on average in an enterprise that goes through some testing and some QA and some, you know, uh, rigorous, uh, process of release. No way.

No, really no Way. Yeah. I've seen, I've heard I've about that happening.

Yeah. So, you know, I think everyone expects code generated by an agent to be less secure, uh, on the one hand. But to Naomi's point from before, these people have been around for open source 10 years ago, 15 years ago, and are not as resistant now.

So they get it, they learn the lesson. Mm-hmm. They're not going to try and stop the tide with two hands, and they're already starting to think about, you know, the right way to adopt it and not resist it.

Yeah, I was thinking about the same thing. Different analogy though. If you remember cloud, we're not going to the cloud 'cause it's not secure.

How long did we kind of stay in that one or two year period before we started really moving applications and software there? That's actually a very interesting analogy for a different characteristic of, of ai, which is, you know, when cloud came out and everyone rushed to protect it, it took it a long time before it got into a meaningful market share of the number of applications in the world, right? So within one year, two year, three years, 10% of applications were on the cloud.

And I think it's the same with AI now, where less than half a percent of applications in the world today are ai, but the most innovative, the most strategic, the most future looking ones are on ai. So I think it is worth watching, uh, what's happening there despite it not being that of big, of a portion of the world yet. Yeah, I I it's definitely arguable that we're moving a much faster pace in terms of adoption to AI and our tool set, et cetera, than even cloud was.

I'm curious, we're in a different place too, of where security teams and software teams are today and how they work together, that kind of tools that they need, uh, collaboratively defining what those are, selecting vendors, et cetera. Is that part of why we don't see security teams in enterprises saying, wait, wait, stop, don't be using this stuff. It's not secure.

We can't, we haven't secured it yet, so we don't see the brakes being put on. At least not, you know, to the degree I'd argue cloud was Steve. Yeah, I, I think there's a really big opportunity, and I think that's what a lot of AppSec leaders see, right?

There's an opportunity to use AI to get your developers engaged in AppSec, right? I mean, let's face it, a lot of developers aren't security experts. Um, but now we have the ability to bring them the information they need when they need it, about a vulnerability to give them the, uh, feeling that, Hey, I can roll up my sleeves, I can actually start to work on this.

I can remediate this. I can, I don't have to spend days looking at it and trying to debug it or learn about it. I can position a developer, you know, I always make the analogy of running a marathon.

I don't run a marathon, right? That's not my, that's not my, uh, skillset. But if you put me two miles from the finish line, alright, I can do that.

And I think that's one of the reasons why we're seeing this, uh, you know, both AppSec and developers embrace this, is it's giving everybody this head start, it's putting people right in front of their finish line of their goals. And now all of a sudden, even if I'm not a security expert, I got the right context, I got the right confidence. And if we put in the great guardrails, now I feel like I can go and actually be engaged in AppSec.

And by the way, if your developers are engaged in AppSec, that's how you're going to have a faster and more productive AppSec program, right? You need them to be engaged. But isn't that so possibly dangerous though?

You're starting two miles before the finish line, those first 18, how many miles is in a marathon? 2022. There we go.

Yeah. I've never run one either. Yeah.

So those first few miles, what, what if something had happened in those first two miles? I twisted an ankle, I hit a pothole. Sure.

Or I did something and now, now I'm hobbling towards the end because I never had that foundation, right? I think we're setting ourselves up for potential failure in the future if we rely on AI too much as a crux. That's what I say.

Well, I I wouldn't say use it as a crux, right? There's, there's gotta be a, a balance of, of trust, put, verify. Um, but at the same time, if we ask, you know, an average enterprise, how long it takes to remediate a vulnerability?

If you don't know that, work on that. Now, when you have that number, how do we improve it, right? How do we improve the meantime to resolution?

Well, AI can really help us as a developer, if I don't have to spend two days on that and I can get that down to a couple of hours or even the time that I might research to understand the risk to actually come up with working code, that's a huge benefit. Like, that's worth the risk in a lot of places in the trade off. Again, you gotta have the right checks and balances in place, but I think that's ultimately where people are embracing this technology is that it provides a lot of opportunity.

I also think you have a growing with a difference is you have these kind of, these more security as code more security groups, writing code platform engineering teams that are responsible and want to use the tools themselves. So there's this, how can my life be better? They're, they're seeing that the code that they're writing themselves is, uh, faster and more efficient.

So I think maybe than a, a decade ago, you see more security as code security coding. And so that kind of wanting to embrace from the beginning, I think is, you know, kind of raising the floor of expectation and velocity, Right? So look, honestly, I think we just haven't had an Equifax level security event coming out of AI yet.

And so I, I think it's a question of when, not if and when it happens, that's when you know how good the relationships between security and developer really are, right? I think right now it's kind of up in the air. Some, some organizations are very good at it, others not so much.

Sometimes you see friction, sometimes you see a lot of push and pull. Uh, but I think it's still to be, to be figured out what it really looks like when, when the rubber meets the road. So it's, uh, during those trying conditions, everybody reverts to type, right?

It's kind. We all work well Nice together when under pressure it right, it's a little tougher. Um, by the way, we'd love to have some questions for our panel.

If you wanna step up to a microphone, I'll do my best to to call on whoever steps up if I'm not blinded by the, the lights here. But feel free to join us with a question. I'm curious, I'll all of you talk with security teams, development teams, what are the questions that organizations are asking you about using AI tools, the security of the software that we're developing, maybe even the security of the tools themselves, right?

These, and a lot of new tools that we're using. Uh, what are customers bringing up to you or potential customers asking you about AI and software development, Rami? Sure.

So I think, again, taking on the, uh, corollary of, uh, of SCA and open source, we are super early days in the, in the life cycle of ai. And most people I talk to just wanna know where they even have ai, right? So if you talk to a large enterprise organization, mostly we talk to, you know, the security people, dev, DevSecOps, compliance, those kind of people, they have zero visibility into what the developers are doing, right?

Very commonly, I've had hundreds of conversations where I'd go and ask again, our champions, you know, how many applications do you have with AI in them? They would tell me, you know, 10, 15 and then run a super basic simple scan and find 200, 300, right? So I think the first kind of stepping stone is for them to just figure out what the engineers are doing and where and to what extent before they can even start any kind of, uh, meaningful analysis of security or anything else.

That Inventory discover stuff, right? You need to know what you have to know, what you have to protect. Exactly.

Well, it's even more hard because no one's making their homegrown models. They're building off philanthropic, they're using Claude, they're just taking someone else's work and then building on top of it, whether it's an MCP server that they've built for themselves or some integration or some tool. So there, that's just another thing.

If we're gonna be doing asset management on ai, we have to know exactly how the developers are using ai. So they're borrowing from these places, how are they actually using it in the tools that you're building and the stuff that you're using for your own customers? Like, what, what is going on in their ecosystem?

And I think it's gonna differ between teams. You have to have those relationships. I would echo we hear the exact same thing verbatim.

We also see an interesting kind of upshoot or, uh, kind of the first green right outta the ground of, uh, those same people now being asked to also look at the data science organizations, the data engineering or organizations mainly we see as a lack of, well, we don't know where else to put it. So kind of feels like AppSec, even though we're not shipping the product, we're deploying into production. So we also see that same sphere of scan and understand in the AppSec teams are now being asked to also look at data science teams, which is a whole other set of, um, let's say silo busting that they're being asked to go do.

So, um, scope changes, uh, as well, including not only security, but a lot of questions we hear about legal risk and policy and understanding of what the legal implication of the models are. IP protection of the model data usage. So it's, uh, kind of an expanded scope, both in terms of breadth and depth of what people are being asked.

At least we start, we're starting to see that kind of bubble up in the, at least in our customers. Yeah, I, I think those are all really great things that I, I'm glad you guys are paying attention to. I, I get inundated with questions about trust.

Um, and, and people largely come to tre marks and ask, how can we trust the AI generated code? Um, how can we be assured that the AI remediation is actually going to fix my vulnerability? Um, they come to us with questions about, you know, how do we put in Ben, but don't break policies where we can, you know, give the developers an inch and say, yes, use these new tools, but also come back to our higher ups and to our board and say, but here's how we're evaluating that code.

Here's how we're scanning that code to make sure that we're not introducing more risk into our pipelines. Um, and so we work with them on scanning that code in real time. We work with them on understanding how you can use reflection in a model, right?

So we've come up with this idea of a confidence score, right? When a model gives you a response on how to fix something, you gotta question that model. How confident model are you that the code you're giving me can actually go remediate my vulnerability and have it explain how it came up with that solution?

Because education is, is a continuous thing, or at least it should be, especially this day and time with all these new technologies. So if developers are going to use these tools and apply 'em into our code, we have to give them a chance for that human intervention that, uh, we always hear the human in the loop. But I think that's one of the most important things is people come to us and say, how do I trust this?

How do we put in policies in place that are gonna allow us to be able to sleep at night, uh, when we're introducing this new code into the organization? Great. And Tyler, speaking of trust that Steve brings up, one of the conversations we don't have yet is zero trust.

And how should we treat software that's generated or created through ai? O one of the questions I asked, like testing companies early on is, do you think of code created by AI as a junior developer? Do you treat it as any other developer?

Do you treat it as the most suspect of, we don't know where this came up from. Maybe it's like an open source project we've never heard of. What, how do we treat, do we need to apply some, some zero trust type principles or we have to take a totally different approach when we think about AI and development?

I think that academically, I could sit up here and talk to you about zero trust and as a, as a, as a framework. And, uh, I'm not gonna say it's bad, right? But I also think we all work and support for organizations that have goals of, of completing a mission.

And in that case, it's this kind of bend don't break set of policies we see the most. And that iteration, and it's been said before, but I don't really have a better metaphor than the early days of open source and that kind of, um, iterative disclosure and trusting it more. We don't see, it's not necessarily a, a good set of policies, a good set of tooling, a good set of processes can work and I think be effective regardless of what or who writes the actual code.

But what we don't see is any removal of accountability in that RACI matrix. What what we don't see is any removing of the developer that does the merge request or the pull request still carries that accountability, uh, and responsibility to the work that they are, are, are doing. And that I think is the foundational, uh, thing of success is that accountability is never removed.

There is that poll request. That merge request who somebody's name is next to somebody's job. It is to make that happen.

And that, uh, think sets a foundation of process and policy and trust, uh, that we see kind of as a, a good place to start. Mitch, we might be looking at this the wrong way too. I think we're sitting here thinking, is there another security control that we need for ai?

And I'm gonna just sit here. I think the security controls that we have today are pretty decent. I catching the bigger risks.

Like I'm worried that my developers are putting in intellectual property into these models, right? What controls do we have out there, guys that stop developers from doing that? We have DLP, right?

These are security controls that we should already have in our organizations that should technically work for some of these risks that we have today. What we're most concerned about, I think, and we haven't really voiced it yet, is the non-deterministic factor of ai. Like in computer science, we know this, you have code, you build it, and you end up with a binary, right?

And you build that thing every single time the same way you're gonna end up with the same exact binary. You do a hash on a thing, that hash is gonna come back the same exact way, right? You know this, it's deterministic.

The only problem is with ai, everything's changing all at once. It's just so much faster. And you can't run something through a model on day one and have that exact same thing happen on day two or day three or day 10, because that model just keeps changing if you do it that way.

So we're trying to explain it to ourselves. Are our security controls good enough? I think it is for now, but we still need to understand how our developers are using ai.

And if so, how is that changing our day to day? And if we can't figure that out, we're always gonna be behind. And I always think security is a laggard behind technology.

Technology's gonna happen with or without us. We better play along and play. Nice.

Alright. Look, I, in principle, I agree that existing practices are, is a very good start. However, I also think we should keep in mind the fact that there are some attack vectors that are unique to ai.

My favorite example is all these code generating models, uh, always hallucinate, uh, dependencies. Mm-hmm. They make up open source dependencies that don't exist.

And some of them do it in a consistent way, meaning they would hallucinate the same dependency over and over again, which kind of gives you an opening to create this dependency in real life and, and do it maliciously. So that's something a human developer doesn't do. Okay.

Just as an example. So I think yes, existing tools and practices super important and apply to the code being generated by the agent, but there are a few additional things to also keep in mind and take care of when you let, uh, when you let AI write your code. It's Kinda like typo squatting for open source packages and libraries.

Exactly. Steve, any thoughts on, on the trust side of you, or the one that brought up trust? Yeah, no, it's, these are all really great points.

Uh, I, I think that there are a lot of good measures in place. Um, ultimately I, I feel like this is more being driven by engineering than anything else, right? At least in the early days where we're seeing a lot of adoption for ai, it's with developers either using it to write code or building applications where they're embedding ai.

So for me, on the trust side, uh, it's important. Now, we always talk about shift left. I think we've sick of hearing that term.

Uh, but now more than ever, AppSec should be sitting with engineering and understanding what your developers are doing. How are they using ai? What types of applications are they trying to build?

You need to understand the roadmap. You need to understand if they have plans to adopt these libraries. You need to start putting in and understanding and building policies internally where you are comfortable with AI usage, right?

Maybe we say, okay, yes, it makes sense. You're gonna use you, uh, AI on the, the front end. Cool.

If we want you using it for authorization, maybe not right now. Why? Because trust takes time, right?

We're not just gonna trust everybody overnight. But that's how you can start putting systems in place where your organization can monitor the usage, understand the usage, and then build practical policies around it, right? So to me it's, it's crawl, walk, run, and, and you know, don't throw the baby out with the bath water.

Uh, to, to your your points guys. I mean, there's been a lot of great things that we're already doing. We have to keep doing them.

Um, but you know, I think just really sitting with engineering and understanding where their push is coming from, where they're trying to use this technology is, is vitally important. Okay, this next question's got a little bit of wind up, so hang up, hang on with me for it. So we've gone through this transition of thinking about security and software where rather than deploy software stay stable, we don't change it.

That's how we know it's secure because we don't deploy software very often to an, an environment that we deploy software quite frequently, maybe multiple times a day, multiple times a week. I like to use the analogy of software's not a solid, it's a fluid. It's constantly changing, whether it's open source or third party or APIs or SaaS services, your own code.

So that's the windup to your, to your point, Naomi, I, I wanna explore that a little bit further. The non-deterministic nature of generative AI doesn't mean it's gonna generate the same code the same way every time. You may say that fix didn't solve it, try it again.

Who knows what changes? Are we in a better spot today because we have automated processes today that handle changes in code that we aren't predicting coming through the development pipelines. Is that gonna help us address some of what you brought up?

I mean, if I wanna be honest, the processes and tools we have in place are not sufficient for codes that changes all the time. But if you're still building code, writing code, building code, and running it somewhere, hopefully your controls are good enough for that. I'm worried about more of that age agentic future that we all talked about and listened to for the past few hours.

That part scares me a little bit. I don't even know what that future looks Like. Like honestly, More philosophically, I think potentially there could be a lot of value to the lack of predictability of these models, right?

Because from a security standpoint, when you are too predictable, it makes you open for an attack. Whereas I don't, like, I don't have anything concrete. I'm just thinking about out loud, out loud.

Maybe you can find a way to leverage the non-deterministic nature of AI to create more security rather than less Subcu security through obscurity. Yeah, exactly. 'cause it's not the same every time, right?

Interesting. Tyler, any thoughts? I know I'd hate to make a judgment call on better or or worse.

What I think is we're in a, uh, state of I think iteration that some things are better. I, I know plenty of, uh, engineering friends that like would rather speak and let the code be written, move up the level of abstraction. I think that in that way can be very good.

Um, and I think that just like anything from a security perspective, I think there are some aspects of what we do that are better now. Um, patching speed and, and, and frequency. Um, so on net I feel like we're raising the, raising the floor.

Um, and that comes in both good and bad. Uh, but uh, what's that Chinese, what's that proverb? May you live in interesting times.

It's interesting, Yeah. Getting more interesting by the moment. That's Right.

I I think one of the ways that we, uh, can also address this is by, um, not making it so unpredictable. I mean the, you know, part of what the big effort is, is building large language models, models where the output is more predictable, right? And we can do that.

We can red team, we can ab test, you know, we can verify in, in pre-check just like we do a lot of things that one plus one is still equals two in that model. And then as soon as we start seeing deviation from it, 'cause we're doing things like continuous testing, then we know we have an issue, right? But more than anything, right, when we think about the impact of, of ai, the, the amount of new code, how fast everything is, is being developed.

If you don't have good CICD processes today where you're automating builds, running union tests, doing, uh, you know, uh, automated testing, this is, this is gonna be extremely painful for you. 'cause you're gonna need those good checks and balances. You're gonna need to be able to make sure that you've got the good regression testing in your models themselves, uh, to, to really I think get to the point where you're talking about of, of having that trust.

Great. Jake, is that you out there? You have a question?

Um, so, you know, we've sort of talked about one side of AI here, which is how you empower development with ai, but the other folks who are being empowered by AI are the bad guts. Absolutely. Um, and I think one of the things that we're gonna see, and we saw, um, some of this in the Mandiant and the, uh, DBIR report last week, is there's gonna be an increase in zero days, um, because it's so much easier for folks to come up with really difficult attacks or sophisticated attacks.

So how do you see the future of defense against zero days based on the fact that it's gonna get easier and easier and easier? Yeah, I mean the, the really scary truth is, um, if you think about how far we've come from a, let's say publicly accessible productization, you know, the companies we represent, what we're building out there, um, our attackers are years ahead of us right now, right? So they have a large runway to work with and it's gonna take the AppSec market, I think years to catch up until we can balance things out and be playing on a level, level playing field.

So, um, vigilance is, is gonna be key. You know, we've got a lot of research at check marks that, uh, kind of to your point shows that already attackers are doing things like, um, open source, large language models and poisoning them, right? We were able to prove remote code executions.

We're able to show that these models are now really the, the kind of vehicles for malicious packages. So your spyware, your ransomware, your malware. Um, and so, you know, when we think about zero days and we think, you know, for us it's, it's gonna take us, I think all collectively as an industry to stay on top of this and, um, to really look out and start thinking about creative solutions and education around what the latest threats are and building them into our threat landscapes.

Uh, but I think it's, it's a, you bring up a very, very good, very big challenge in the market that, that we're all facing right now is how we're gonna close that gap. If I was, if I was a threat actor and I'm, I, I'm not, but if I was, I would go into GitHub, I would point my mo my model at it, I'd be like, tell me the top 10 most popular projects on GitHub and what vulnerabilities exist in those projects. I would then try to go to a sonotype.

I know we have a sonotype here. I would then see where those projects are actually used. Like if it's an open source library that's popular log for shell log for J is a popular one.

For an example, I would then ask Sonotype point my model to Sonotype. How many open source projects use that vulnerable library or have that vulnerable thing that I'm looking for, right? How many of those are in commercial or uh, cots software?

How many things are sold? How many things are for free, right? Like, and then I would attack that all day every single day.

I think that's, it's like log for she but more all the time. Yeah. And I don't wanna like scare anyone, but if you don't have the visibility into your applications since you know what's actually running in your applications, like how are you supposed to defend against that?

Right? And I will say there's a pretty good software out there. Contrast does that pretty darn good.

So I'm gonna say like, if you guys don't have that visibility in your applications, check out contrast pretty great. Alright. Look, I'd say fight fire with fire, meaning I think we're already starting to see where agents are being used for protection and for creating better and faster security to combat those, uh, bad actors using AI to create the attacks.

So this is something we are working on, but I see many those people do it as well, which is deploy agents to on your side in your defense, that can act as quickly as the ones against you. Yeah, I would say, I was gonna say the same thing. Fight fire with fire.

Um, 'cause we're doing similar things. I would also say that there's the, uh, rapid response that can further be accelerated by the AI agents, acceleration, tech technologies. Let's say you're not perfect.

Uh, somebody breaks through the offensive line to go into the, to the secondary. It's how fast you can collapse on that. So it's visibility agents help with that so you can fight fire with fire, both in protection and then rapid response to close that.

I think that's what we're seeing is, um, there's a lot of hardworking, talented people on trying to protect, um, and innovate in a lot of those areas. So I think that's, that's exciting to see. Great.

As Mark comes up, hold for a moment, please thank our sponsors and thank all of our sponsors that were a part of, uh, helping you bring this to us today. So I'm sorry we didn't get to your question, but we'll try to talk to you in the back. Okay.

So I hope we see you again next year. Okay. Thank you.

Thank you. Surely appreciate it. Is that micro rubia or is that an AI impersonation?

Core weave buys core scientific and Arista requires VeloCloud. Ingram Micro is hit with ransomware. And can you escape a large language model by using dense words?

Open AI gets data centers from Oracle and Microsoft is replacing its human workforce with ai. All this and more in the tech field day rundown. Welcome to the Tech Field Day rundown, where every time we meet every week we run down the greatest IT news of the week with a variable degree of smartness.

I'm your host Alistair Cook and joining me as a co-host since Tom is out at, uh, networking Field day and Stephen is buried in meetings. I have the great delight of having Gina Rosenthal. Welcome to the show, Gina.

Hi, it's so glad to be here with you today And I welcome you on National Dimples Day. Um, it interesting that it's, it's, it's a plural. My, uh, younger daughter has a single temple just in one cheek, but it is also one of Stephen Scot's favorite days.

It is National Durian Fruit Day. Uh, remember this is a fruit that you cannot take on public transport, at least in Singapore. Launching straight into our highlight story at the beginning, someone's impersonating Senator Marco Rubio using AI voice and chat on signal.

Uh, though the targets aren't named, the story highlights a broader trend. Uh, Trump aid, Susie Wallace had her phone hacked leading to calls to top officials and the FBI has warned of ongoing impersonation campaigns targeting US leaders. Russian agents are faking Ukrainian security contacts to recruit saboteurs, and Canada's issued a similar alert about senior officials being in impersonated in calls and texts.

Gina, have you been impersonated by ai? Is that really you? You'll never know.

I think it's really me offense. Uh, yeah, this is kind of a really interesting and and terrifying story when you get over the, you know, it kind of sounds funny on the surface of it, but to have, uh, the United States Secretary of State, um, impersonated so well, by the way, he writes in a, in a chat and the way he speaks, which is pretty easy because he's, he's been a senator and a public representative in Florida forever. So his voice is everywhere, but he's a secretary of state right now.

And I think the, the crazier thing is just his, the tone of how he writes in a text perspective and how he would res, you know, ask someone to get him in contact with somebody else or to give him money, which is what the, um, the, the article talked about. So, um, it's very dangerous to have that high level of an official and even the White House secretary, uh, is, is pretty scary because she knows everybody and how to contact everybody. So someone that was able to hack into her phone and then I use the correct phone to then impersonate her, um, is, is terrifying.

So, I mean, AI has come a long way and when you look at what's, what the ransomware gangs are up to, and then you think about nation states that, um, have this ability, which, you know, we definitely know between Russia and and Ukraine, that could be, it was terrifying. They actually were, were recruiting members of the Ukrainian public to help them with suicide missions. They were wanting to mess up missions, um, in the Ukraine.

So there, there's obvious reasons why nation states would do it. Money is the reason that ransomware gangs will do it. But this is no longer just kitty script stuff, although it's easy enough and available enough for people who are just messing around and messing about, um, to pick up some extra coin, make it very easy.

So where do we go as a society if you don't, can't trust what you hear or see, I dunno. Core Weave wants the whole cloud enchilada buying Core Scientific to ca gain control of their AI data center density for $9 billion. Core Weave will add at least two gigawatts of data center capacity, which already host the core Weave AI cloud.

Even better, it will save $10 billion in leasing those data centers from Core Scientific. Is this how a Neo Cloud provider effectively competes with the existing hyperscale clouds? Well, I think it's this vertically integrated, we own it from as well sometimes like to say from soup to nuts.

That was the name of a podcast on virtualization a while ago. Uh, it's really is this idea that in order to get the efficiencies of scale, the efficiencies of, of operations, you just pull the profits out in one place once. Uh, so you need to own everything from the beginning to the end.

So Core Weave is a provider of this sort of neo cloud capacity, this GPU first cloud infrastructure. And they previously simply had their applications running inside these data centers that belong to Core Scientific. Uh, here they're acquiring that.

6 billion worth of assets. Well, that definitely stacks up for that $9 billion of spend. But there's a bigger picture in here, which is that core scientific has contracts around the power supply to run those data centers.

And that's really the asset that's being bought in here. And we'll see this as a recurring theme around AI data centers. Having enough power to run these data centers is a concern.

Having a commitment to that power for a long period of time is the way that you protect your, your, um, business for the longer term. Definitely having a, a single entity that owns everything from the data center ground all the way through the application software that's running on top of that to deliver this cloud out is vital. In order to squeeze the maximum efficiency, uh, neo clouds are gonna be just as competitive on price as the, uh, hyperscaler clouds, and we'll see more and more neo clouds, um, coming up trying to differentiate in market.

We've seen this at AI infrastructure field data where we saw a couple of solutions that were, uh, about building data centers in places that power costs were lower because for some of these near cloud use cases, it's more cost effective to move your data. Where the power costs are on the data center is lower than it is to build your, uh, AI data center close to your data. It's a little bit of a flip to how we've previously looked at building data centers for, uh, for our environments.

We've typically said data has a high gravity, and you really tend to accumulate compute close to your data. Clouds are sometimes saying, we're actually gonna move the compute or move the data to where the computes cost effective. The VeloCloud SSIS solution is moving from Broadcom to Arista.

VeloCloud was part of the VMware when they were acquired by Broadcom and isn't part of the core VMware Cloud Foundation product that Broadcom sees as being the future of VMware. So we aren't surprised by the, the sale, uh, Arista gains a new ESI product, which shall as a secure when connectivity to thousands of locations using the velo cloud manufactured network devices and the cloud services. Does Arista want the SASS e business, or is it the AI that's actually running all of this in the valid cloud solution?

So is it, I thought it was always safe, but if it's sass e Okay. Like Sasha, but, um, it's the secure Access Service Edge. Um, so it's different than what your networking would be for, um, a data center or for, uh, inside of an organization.

Um, but what they, what Arista has gained from this is the virtualization of those components that are required to talk on the edge of their, their service areas. And with the virtualization, you also get the ability to run it as infrastructure, as code, which, um, ma makes everything easier to change on a dime, to protect, to segment all the other things that we know, um, virtual networking does within a data center, traditional data center. Um, so is this for, that's a good question.

Is this gonna be for AI and, um, I think, so. It, it could be for ai, but it could also be for gathering the information that's shed from Edge locations. How do you get that quickly into where the data is gonna be com, um, gonna be worked on by different algorithms?

Uh, how do you isolate networks so you can get them from the edge into the core? Uh, I would guess all of that goes into that so that you don't have to worry about just having a normal sd-wan. You actually have, you're able to move data from one place to another.

Quickly. Moving on to ransomware, again, to the these bad actors, it hasn't magically, ransomware has not magically gone away. Ingram Micro was infected this week and had staff work from home while critical systems were shut down.

Sleeping computer reports that Safe Pay ransomware was the culprit, and the attackers were able to compromise the company's Global Protect v pnm, I hope Ingram Micro are using one of the ransomware products they distribute and will be back in action soon. Safe Pay is new to the ransomware scene. So are they a new type of threat?

Well, Safe Pay apparently burst onto the scene. Um, um, I call it these, these ransomware or malware, uh, products, different organizations call them a different, uh, threat protection companies call them by different names. Uh, safe Pay burst onto the scene last year and seems to, uh, again, it's, it's likely that this is a nation state actor and that, um, there is something fairly aggressive going on underneath it.

Safe Pay is the, the dual threat. So they're the data exfiltration as well as encryption. And so this is very much the way ransomware, um, development has, has been running.

Is is that that dual threat, two ways of extorting you from money? Uh, Ingram Micro apparently has now come back online. So they were struck on Thursday last week, and as of the beginning of this week, we're starting to bring systems back online.

Uh, the Global Protect VPN is provided by the Palo Alto Networks devices. Uh, there's no indication that that was a compromise of the Palo Alto devices themselves. And Palo Alto put a press release outside.

They're gonna look a little more closely, but more usually this is a layer eight human kind of stuff where, uh, vulnerable passwords are, are set up and, and the, the VP n may well have been the entry point, but when this did break, uh, Ingram Micro, who are a, a global organization, there were a lot of countries around the world. They're a big distributor of software, uh, products. Uh, Ingram Micro sent a whole lot of staff home and told them not to use the VPN.

Um, if you rely on VPNs for your connectivity to your core applications, then, you know, being sent home and told not to use the VPN might mean that you're just scrolling on, uh, social media and watching videos. Uh, they did also, um, it seems like the Ingram Micron did a, a good response here in that they were proactively shut shutting systems down to avoid an in, uh, additional compromises, additional, uh, lateral spread in this compromise. So it sounds like they were pretty quick to respond and pretty brutal in the way they did respond.

And I, I applaud Ingram Micro for being really aggressive at stopping the spread of ransomware within the organization. I've seen it organizations just being too slow to respond to these kinds of attacks and ending up that the speed at which the ransomware spreads is so great that, that it just overwhelms the kind of response rates that we see. And we see delays in doing things like shutting down the entire network and particularly shutting down connectivity out to the command and control servers, uh, just leading to much greater infection than is needed.

So I think, uh, Ingram might have done well and are bringing systems back on minority. A group of researchers have found that large language models can be jailbroken by using dense and complex linguistic phrasing. By applying advanced language transformations, attackers can mask harmful intent in a way that evades detection while still prompting the model to produce restricted or sensitive responses.

Uh, the classic one is, if the LLM won't tell you how to build a bomb, ask it to dis to write a book in which the plot is the building of a bomb and extract a bat part. Uh, Gina, how are we gonna get around this? How are we going to protect ourselves, our society from people with poor intent?

I think people have to keep researching it with exactly what these folks are doing. Um, this was so interesting because I love that you giggled at that whole phrase, uh, by using complex linguistic phrasing. So when I was reading the article article, to me it was like, okay, you're listening to a technical program, you know, a presentation someplace, and they're just using acronyms the whole way through, and they're not stopping for a breath, and they just probably put all the information in one breath, and it's very, very hard, even if you understand the topic, to follow them, because they're talking at such a academic level, and this was directly their idea.

So they, the, the models can shut off for obvious things, like if it's blatantly racist or if it's very, you know, you can ask someone, ask you to do something that can harm other people. They can shut that off at that level, but they wanted to know that they, they can do that because that's, that's obvious that you're looking for, to do something very negative and toxic. But what happens if you bury the request in very, very dense message and they're now calling this the info flood?

You know, just think about some of those presentations you've been in, and this is exactly what it is, Nick, the article's great. I suggest you could read it because it's very hilarious. And if I sat here and read it, you, everybody's eyes would roll back in the back of their head about what they are.

But they were able to, um, figure out how to give them all of this very technical jargon or very complicated English jargon. And within that, put the, the gel, whatever they were trying to gel break out. Um, and it worked every time because the way that these things work, but the way that LLMs work are actually the tokenization of the English language and what's supposed to be expected after each sentence and bombarding them with all of this information, all of these different words and all of these different tokens.

Um, I don't know if it makes them like us and we forget what we were listening to and we take one piece of it, which really wasn't the point of the whole talk, but, um, that's, that's how they're breaking it. So I think the way to get around this is, is there a better way to build the LLM or is there a better way to, you know, if, if everyone's not gonna have the opportunity to build the language models, what do you have to put in place to protect yourself from someone trying to do a jailbreak like this? Like how do you protect your data and only use the models to, to, um, help you with basic linguistical or even image tasks?

And, um, that's for the researchers to figure out. They figure it out really soon because this is, you know, this is something we should be able to trust and, and use and to make everyone's lives better. We don't need people, um, going beyond imitating someone's voice and actually trying to get proprietary data out of what we've built with our rag in instant in instances or anything else.

5 gigawatts of data center computing power for an estimated $30 billion annually. This is a massive deal for Google, who reported a little over $10 billion in annual revenue for this data center infrastructure business before the deal. OpenAI and Oracle are also partners with SoftBank and Abu Dhabi's MGX fund in the Stargate project, which plans to build $500 billion worth of data centers around the world.

Is the whole world just gonna be one big AI data center? It certainly feels that way. These are massive amounts of data center capacity that are being built out.

I mean, this is absolutely an awesome deal for Oracle three times. The what quadrupling this, this scale, this is a new deal for three times. The previous size of that means four times the size of business.

Now, uh, four and a half gigawatts of computing powers an interesting that as you get to larger data centers, and particularly cloud provider scale data centers, you don't measure them by square feet. You measure them by the power demand, the number of megawatts, or in this case, gigawatts of power, the data center consumes, because that's really the design point for it. This is huge.

And in that larger context of Stargate, it's going to be even huge. I mean, 50 500 billion, half a trillion dollars worth of data centers around the worlds, quite a bunch of those are being built in the US but also in other countries around the world because of course, um, there's lots of places where there's demand for this. Um, yeah, four and a half gigawatts currently, that's, uh, a contract for roughly a quarter of the current, uh, data center capacity in the United States.

So that would definitely involve a huge build out by Oracle, supported by this contract. Uh, potential new sites for this spread all across United States, Texas, Michigan, New Mexico, through to Wyoming and Pennsylvania. Uh, also close to the head office of the tech field day team in Ohio.

So this is a, a lot of, um, build out. Uh, there's a, a bunch that's being built out in Abilene, Texas. 2 gig gigawatts of the Stargate is being built out there.

And we heard from the, the, the data center startup cruso on, uh, AI infrastructure shield. And they're one of the companies that is helping build out new clouds and locations where power is potentially otherwise, uh, uh, being wasted or being, uh, delivered at very low cost. So this whole theme around putting data centers anywhere there is power to be had, is gonna continue, particularly with this massive growth of AI's, uh, data center sizes.

Um, Oracle is planning to buy about 400,000 of NVIDIA's, um, GB 200 chip to power just the Abalon data center. Uh, that's a massive amount of, uh, compute power that's being pushed into one place. Personally, I keep seeing this infinite growth and any time I see, uh, this sort of exponential increase in the size of something, I know for sure it can't go on forever.

And exponential growth by definition cannot go on forever. It always has to come to a, a crisis at some point. Uh, I personally am very much hoping that the crisis is that we find a much more efficient and better way of running these, uh, AI applications and delivering value out.

As Gina just covered, there's some pretty big challenges with the way LMS work may not necessarily delivering us the general purpose AI that we would like. So keep watching this space, massive build out, uh, and investment is, is awesome. Uh, hopefully massive value delivered back from them.

It's time for us to take a little bit of a closer look. And this is, uh, an interesting story that continues to be seen. We've seen multiple angles on this before.

Uh, Microsoft recently laid off about 9,000 employees. That's about 4% of its total workforce. These layoffs within the tech industry or risk reduction in force are fairly common.

Uh, typically a seasonal thing that we see a chunk of the workforce laid off from an organization. Then over a period of time, a new, uh, workforce get tied in because the company shifts where its focus is. And so they need to lay off the people who, who can't be relocated into what they're doing.

Now, the difference here is that Microsoft is shifting to AI to replace these humans. And, uh, whilst the company leaders say that AI will eliminate many routine jobs, but it'll also create new opportunities in advanced areas, thousands of tech workers were displaced, but mind for AI experts has sawed with companies offering high salaries to attract, attract top talent. How do you think these shifts will impact the future of Microsoft, Jenna?

Well, I think it's already had a bit of an impact because the, the, the shutdowns are global. They're not just, you know, in the United States, uh, and the shutdowns are pretty complete. There's complete lines of business that are being switched right away.

Um, and so that has an impact not only on the, the individuals being rift, but it also has an impact on the local economy. It has an impact on the greater economy because we see a, a huge trend right now in, um, outsourcing jobs and saying that they're being removed for reasons, for AI reasons, which it turns out that that's not always the case. So, um, you know, of course I, I have a product marketing agency.

I've seen this happen to my own agency, um, and I have a friend who does special effects, his own special effects company. He's seen it happen in, in their organization too. And the problems I see with Gen generative AI replacing humans is, especially in the fields where you are using, um, words to define things and you're doing things with words and you need creativity, uh, AI can't replace it.

I'm starting to see that in my own business, starting to see business coming back a little bit, um, starting to see a lot of cleanup work, which I predicted that would happen if they did this. That would be cleanup work. Um, I think it's bad, uh, for the entire stream of work, right?

So this happens, if this happens, this is knocking out a lot of senior workers. That means the new people have to go in and try to use AI without having any context of how the business works. They don't have anybody to mentor them.

Um, it just leaves a lot of gaps. And I think it's, uh, I'm not sure if it's shortsighted because I, I think this is the intention that Microsoft had and where they wanted it to go. Um, it'll just remain to see if they are, um, being a little too lackadaisical in their approach with this to, to replace knowledge workers with ai.

I dunno, what do you think? Well, I think there's, there's a couple of parts to this. One is that these, these riffs are routine, uh, hanging a banner of everything's happening with ai.

Um, comes up, I, when I say routine, they happen pretty much every year. I mean, they're horrible for the people affected, but they happen every year. They happen across almost every organization that we deal with.

That's a, a reasonable size in the industry. Hanging the AI banner across it. There's a couple of pieces to that.

One is everything's AI now. So these, there's refs have to be ai, but also Microsoft is saying you can get huge amounts of benefit from using AI and copilot in your own business. So they, they have to walk that walk.

They have to show that they're getting huge amounts of benefit. I think, uh, I agree with you that they probably are over rotating on this, and they may find themselves in the same situation. I think it was IBM found themselves, they, they laid off a whole bunch of, uh, sales staff because they were gonna replace them with AI and then had to turn around and re-recruit those sales staff because the AI just was not doing the job that was required.

We may see if, if the, the truth in this is that the, the staff members are, are being replaced with ai, we may well see that same wind back happen here, and it'll probably be nice and quietly it'll just be quietly rehiring staff, hiring staff to do the jobs. Uh, you don't wanna admit that you made that kind of mistake. I'm sure I, you know, I've said for a long time, a long time for like probably 20 years now, that in knowledge workers working positions, the means of production is no longer in the hands of the companies or the owners or the managers.

It's inside of our heads of knowledge workers heads. So what to me has been happening is you've now got, um, a lot of companies, Intel is another one that is now doing layoffs today and replacing all of their marketing team, sending it to Accenture that's going to use AI to do the, the marketing. So when you, you do that, you're, you're taking that means of protection back in house, but you're trusting a large language model.

That's where the means of production is now. So we know what the troubles and how that does not work, and, uh, we've talked about that forever. We talked about it today.

Um, so it's just, I think it's an attempt at, to use pure, um, capitalism type of language. It's a mean to reclaim means to reclaim the means of productions no longer in the heads of the workers. It's in the code of, of the law, of the language models.

Um, and I don't think, uh, you know, good luck to them, but, you know, I don't, I don't think that that's the way to go. Now, having said that, I use AI every day in my work. I like using ai, I like using the Lang large language models.

But, um, I also have a context for, for how it should be used, because I have a context of the whole field of marketing and how what we exist and what our goals are to try to do things with, and that guides my prompts and how I'm doing things and how I use that for my clients. So I think there is a way to use AI to make things faster, to get rid of some of the dredge work, to help make the changes faster, all the rest of the things that they seem to want to do. But I don't think it can be done without humans.

And if it is, we'll see how that turns out. I think that the people that will decide will be the people that consume the messaging and, you know, that have to use the code that's built by AI and we'll, we'll see how that turns out. We'll, indeed and other things we'll see is that this week Tom Hollingsworth, my usual co-host here is actually at Networking Field Day 38.

So today and tomorrow you can tune in for all of the live stream goodness and the replays both on LinkedIn as well as on Techstrong tv. Coming up in August, we will have the Tech Field Day extra at Share Cleveland. So, uh, Stephen Scot is going to just drive up through Ohio to get to Cleveland for a share.

And then in September, I will be making the big trip across the Pacific Ocean for AI infrastructure field Day three, that will be September 10th and 11th and later on that Tom takes back on the baton. He has security field date, September 24th and 25th. Check out the Tech Field Day website for all the details of all of the upcoming and of the past events from Tech Field Day.

So thank you very much for watching this episode of The Tech Field Day Rundown. You can catch new episodes every Wednesday as a YouTube video or on your favorite podcast application. The rundown is streamed on Techstrong TV as well.

And you can catch us and other techstrong and future and group programs right there on techstrong tv, including on your smart device with the techstrong app. We'll be back next Wednesday to talk about all the IT news in the week. That was until then for myself and for Jen Rosenthal, for from all of us here at Tech Field Day.

Thanks very much, and wishing you and yours a great week. Hey everyone. Ah, modern medicine, you gotta love it.

A cure for hallucinations. You're watching Textron Gang. Hey everyone, it's Alan Shimmel.

Happy Thursday. Ah, jk. It's not really a cure for hallucinations for people anyway, talking about AI hallucinations and can we have, we found something that actually works on this.

I don't know if it's microdosing or what, but, um, we've got an interesting gang of people to talk about our subjects for today. Let me quickly introduce you to them. We have, uh, John Swartz, who's been up all night with the latest Silicon Valley news, Garima Bo Powell, our good friend, guy Courier, and of course the Dean, Mike Ard.

Ladies and gentlemen, let's jump right into our first year, A cure for AI hallucinations. Anyway, shades of One Flew Over the Cuckoo's Nest. Mike, what are we talking about?

Alright, well, there's one of these companies out there that specializes in AI.