Challenges of Building and Deploying AI Agents with Fortinet’s Jim Richberg

Hello and welcome to the latest edition of the Techstrong AI video series. We're here with Jim Richburg, who is Global Field CISO for Fortinet and also head of Cyber policy. Jim, welcome to show.

Well, thanks a lot, Mike. Glad to be with you. We went from being overall by everything chat, GPT to now.

We wake up in the morning and there are more LLMs than you can shake a stick at, and we're trying to figure out which ones to use. And not only are there LMS in the zoo, but there are other predictive models that we've had for a while and causal and all kinds of interesting stuff that's out there. So how should we be thinking about tapping into all this, uh, stuff?

We seem to have an embarrassment of AI riches all of a sudden. Well, we do. And and some of them are, are the new kids on the block and some of them have been our friends and have been around for a while.

And let's start really with the two main types of ai. You've got the, uh, you've got the predictive or discriminative AI where you try to, and I use air quotes here, learn a model. You try to learn enough about a data set to be able to predict or discriminate what category a piece of data would fit in.

So I have data about animals. I want to be able to say this is a dog and not a cat, even though they both have four legs, fur and a tail. I can tell based on other things I can discriminate between them.

Um, and we've called that one just big data analytics and we've been doing it for a very long time, decades. Uh, the new kid on the block is this generative AI where I want to understand the data well enough to be able to generate a piece of content that would fit in it. So I can say from that data about animals generate me a cat.

And that's why we get the hallucinations where we may get five legs, two tails, whatever, because it's missing some characteristics. But one is about saying, I can look at a piece of data you present me with and tell you where it fits. And the other one says, make me a new sample that would fit in a category of my choosing.

So they, they're different, but, but similar. They're both, they're both ai that you'd use 'em for different things. I think people have a reasonable understanding of when to use the predictive models.

But generative is a little trickier because a lot of the output is probabilistic and we seem to be experimenting with where to insert that. And a lot of times, maybe we don't have the right use case because we have a deterministic use case there where the output needs to be the same a hundred percent of the time. And that's not something an LLM does at this point.

No, I mean then that's just in the nature of how you initialize the starting positions for them. But I like to say, look, the generative is about generating content. It's about knowledge, it's about creating it.

It's the people like you and I who are knowledge workers who are going to find this, the tool that's really going to further what we can do. Um, it's it, it can help on automation, but a lot of what we want to do with automation is really that predictive or discriminative. I don't want myself driving car pausing even momentarily at the intersection to generate a new kind of response for how it should deal with traffic and what different road signs mean that should be predictive based on looking at stimuli.

But the generative one is going to be good for saying do the kind of mental work that human beings do. And, and, and I mean that's a good way to do it. One is about knowledge and the other, they're both about, they're both about information, but one is about content creation And how smart are these things getting, because we hear a lot about the reasoning engines are getting better and better, but that also seems to be dependent upon the number of parameters and the amount of memory that's made available.

So, um, that seems to all come at a, at a different cost scale as we get smarter and smarter. So how should we be thinking about these reasoning engines? So I, I think you really need to make sure that the AI you're using is fit for purpose.

You know, and let's even set aside, let's figure you, you made the initial leap of saying this is a use case where generative AI is my friend. This is a use case where my old standby of predictive discriminative exhibitor tool, let's say I've decided this is about novelty, this is about new content. Even then you could say there are at least currently two types of, you've got the chat GPT that we're all familiar with, where query engineering has become a real science.

It's pretty interactive. And then you've got the chain of logic type ones probably best embodied by something like deep seek where it breaks a problem into chunks. It eats the elephant one bite at a time.

It it, it tries to make it into pieces where you can logically parse it and, and, and approach them that way. So for problems that blend themselves to logical approach, it does really well. Programming, helping to create new pharmaceuticals because we know chemically how different molecules and atoms can be combined.

It's just a matter of there are myriad ways to do it. So you gotta figure out what kind of problem am I trying to solve. Those chain of logic type things are really good at dealing with those structured problems.

But if you ask it to write a novel, write me a song about a purple cat on the moon drinking margaritas, it's not going to give me a better outcome than the chat GPT style. So make sure you're using the right, even within one type of ai, the right class of of tool for the right problem set. How Will I kinda use these LLMs and all these capabilities?

'cause we hear about AI agents all the time. The, and I can't help but wonder just how dynamic is it all gonna get because will different agents call different LLMs based on different use cases? And how might I switch different LLMs as the LLM itself gets smarter?

I think that, I think that's where we're going. The ho you know, AI is getting to the point where you may want the human to do it, but you probably want the AI or something, um, at the model level to be able to say, I want to pull from this LLM pool for this and that, and maybe I want this to be the private pool and I want to use the, you know, the back end for that. So I think we're going to, we're to your point, we're going to want a lot of that to happen seamlessly so that we can tailor the, uh, you know, the uh, the query against the, the right set of data, uh, in the right model eights.

But I don't think we want to have to do that manually. So how would that be accomplished? 'cause I, I can see in my mind at least that there's gonna be a need for some orchestration of multiple agents and somehow or other that orchestration is gonna automate some process on a mostly end-to-end basis.

Yeah, and I think, and, and I think there's a, there actually is a, a fair infrastructure that's in place to glue this kind of stuff together. I mean we, we've seen this even with the fact that the LLMs have gotten so large that in some cases you can't run a model in a single data center. So there's that very fundamental kind of glue software that could, that synchronizes execution across multiple physical locations, even though it's simultaneously training a single model over and over again.

So I think we're going to find that same kind of specialization in the software infrastructure that allows all of that to be done. I mean, to me, the more interesting thing is you want that to be done seamlessly because security is gonna be a problem if, you know, one type of data, one type of model, it can, can leverage one type of security controls. And yet I want to pivot to something else.

And they've got different parameters, different policies that inconsistency can make a gap. And the gap is where it can get exploited. So I worry more about not the efficiency of execution, I worry more about problems, right, uh, cropping up that are gonna make holes for, for malicious exploitation, And how will we combat those threats?

'cause we've already seen things like jailbreaking and all kinds of fun attacks aimed at these things. But, um, it seems to me maybe that we're gonna need AI agents to secure the AI agents. I mean, how is that gonna work?

I think that's part of it. The the biggest challenge, Mike, I see on security is that it it's a good news, bad news story. I think the, the good news is we have imperfect but improving controls that you can bring to bear on everything from securing the infrastructure to, you know, se query securing yourself against data loss during query.

The problem is we don't yet have the equivalent of the shared security responsibility model that evolved for cloud, where now it's pretty clear, depending on what I'm doing, ranging from hosting all the way up to, you know, platform as a service, software as a service, what's mine as the tenant to do versus what's yours as the host to do. Now the early adopters in, in cloud migration got the arrows in their back when they said, well, I've gone to the cloud, what do you mean you're not responsible for my security? We don't have that for ai.

And that's, I think, the problem you have to order a la carte. It would be like going into a restaurant and saying, the good news is this restaurant has the food I want to order. The bad news is they handed me a menu and it's an alphabetical order, buy food dish.

That's the problem right now. It puts too much onus on the user, whether it's an individual, whether it's an enterprise to have to specify security. And at one point you go, I don't care who's providing it, I just want somebody to be responsible for providing it at every layer of the stack.

This may seem like a silly question, but I'll ask it anyway. So in your model with the shared responsibility, I kind of have to opt into security and I wonder if maybe part of our problem with shared responsibility is we should be asked to opt out of security rather than in Everybody cares about the security of their training data. Theys, they, they care about not having whatever they're doing compromised the people who own the models and are spending all the time training IT care.

Everyone cares about, I don't think anybody's gonna wanna opt out and say, I don't the model weights don't matter. It's like, yeah, that's your secret sauce. So I think even if you just say, I'm a consumer, I don't want someone stealing my account and running queries on a model that where I'm paying for it.

So I'd be hard pressed to find a good circumstance for anyway. Uh, that's s that's separate from anonymization. You might wanna say, I wanna run this anonymously, but I don't think anybody's gonna wanna say, I don't want security.

I mean, I I I'm hard pressed to think of a good use case for that. So in my mind, that means that I should get the security by default. Correct.

And rather than have to say, oh, who's securing the infrastructure? Who's secur, who's securing the model, the weights, the DLLM, you know, my queries itself, it's just, it's you, you don't want this to be onerous, Nor do I want it to be a repeat of that cloud security example that you had where basically, you know, we're spending a lot of time trying to figure out who's responsible for what and inevitably it just falls in the middle and nobody covers it. Right?

But I think we, Mike, that that characterized the early days and now I think it's pretty clear if I just sign up for hosting, you know, they go, okay, you own the machine, we're gonna make sure nobody physically gets to it. We're gonna keep the power running. You own everything else, all the way up to, hey, we're providing, you know, the infrastructure, the software, et cetera.

All you gotta do is, you know, it's, at this point I think it's pretty clear who owns what, depending on what level of service. And, but I don't think we want to take the number of years it took to evolve that for cloud, given how much emphasis we seem to be placing collectively on using AI everywhere. And that's setting aside, are we gonna be able to power it?

Right? That's the separate part of this conversation. Yeah, That's true.

Power will always be an issue somewhere. So as you kinda look at this whole pantheon of LMS and AI models, what's your best advice to folks about how to go about kinda operationalizing all this stuff? 'cause on a certain level it's kind overwhelming.

I mean, you know, start by saying, and I feel like Mike, as organizations a number of years ago, started to look at Gen ai, they were giving it a free pass on showing ROI, they just felt like the fear of being left behind meant that companies were going to invest in using Gen ai, even if they weren't sure that they were using, that they were using it appropriately, that they had the right use cases. So start by saying it's clear in my mind, is this something where it is about generating new content? Is it about, you know, automating?

Is it about that, the ability to say, I can see something, a new PI can see a piece of data and figure out where it fits predictive discriminative versus I need knowledge generated. And then even within that to say, okay, is this problem one that lends itself to that logical, highly structured world? Or is this something where it's more freeform and a chat GPT so ensure that the tool you're using is fit for purpose.

I mean, I think that's really, you, you, you don't wanna put this against the wrong use case and find out that you, you're, you're wasting resources. You may get an answer, but you may be overpaying. We may collectively be burning more electricity to get that answer.

And if you literally pick the right tool for the job, Folks, even in the age of ai, that old piece of wisdom that says, look before you leave, still a plus. Hey Jim, thanks for being on the show. My pleasure, Mike.

All right. Thank you for all watching the latest episode of the Text Strong AI video series. You can catch this episode and others on our website.

We invite you to check them all out. Until then, we'll see you next time.