AI for Networking with Josh Mayfield at Techstrong Con 2024

Welcome, welcome, welcome. Hey, everyone who's, uh, curious enough to stumble in here. Uh, today we're gonna be talking a lot about what is real and what is just hype, uh, when it comes to using AI in networking use cases in INO overall.

And so, uh, I'm Josh Mayfield. I'm the Senior Director of product marketing at kinic. Glad to be with you here today.

Uh, we do have some opinions when it comes to the AI world, the hype of most of it, um, and the realistic ways, uh, that it can actually be leveraged in a lot of networks that we see. Uh, today. It's really fascinating when you see the kinds of ways that people are using gen ai, um, to really open up the kimono and demonstrate a lot more versatility when it comes to extracting answers out of their networks.

So, um, with that, we'll just get into those, uh, just three points today. So there's been a tsunami and like tsunamis, uh, you don't know necessarily where it all began. You just see the tides coming in, um, and then they just keep coming.

And that's what we're noticing with ai. It just keeps coming. Um, it, it's broken shore and, uh, the flood continues to, to come through.

Over $20 billion last year in startup funding went to AI-based startups. Uh, really colossal, uh, ground shift that we're seeing, uh, really changed the landscape because of this invasion, uh, as some people call it, the, the invasion of the copilots, the invasion of the ais. Um, what implications did I have on, uh, on INO and networking?

What, what can we draw from that? What, what good parts can we take and what things can we discard? Um, and then assessing all of that and saying, okay, is it real or is it hype?

Um, are there aspects that are, are worthy to pull upon worthy to use? Uh, what are some of those use cases? And, and in this age of, of ai, uh, data and telemetry are the new gold.

Uh, so what does, what does all that entail? Um, and so with that in mind, we will get into some of these things. And for starters, how did the tsunami begin?

Uh, how did we get here? And, uh, it wasn't all that long ago really, when you think about it. Uh, think about the, the internet change that happened over the nineties, right?

You had the, the DARPA net sort of becomes publicly available in the late eighties, really catches on in the early nineties. You have Netscape, and its IPO that kind of set off the internet age in 95. But it wasn't until, you know, the mid 20 teens that we really started figuring out what to do with all this cloud thing and this internet thing.

It took us a good 20 years of basically a generation to get there. Um, a human generation. Meanwhile, AI in, in less than a decade has made leaps forward.

Um, and so how do we get here? It may seem like it was just a, a a day ago or, or just a year ago, but, but really a lot has been compressed into a short amount of time. So let's go back to 2015.

Uh, TensorFlow and PyTorch really started laying the groundwork to be able to fulfill tasks based on a neural net computation. And it really started edging things forward to where there could be a generation of something, a task generation, uh, and models of that sort. Transformer changed everything.

Um, when the transformer model gets introduced in 2017, it just unlocks all the, uh, research and development that just exploded after that fact. When we started to understand that, uh, attention seeking was, was greater than anything else, um, and, and getting the, the models to predict, uh, the next move. And, uh, and then when we get the, the gans out there, this really starts to get into multimedia type, right?

We're starting to get photorealistic imagery, and we're starting to get the kinds of computation that are possible in parallel processing with GPUs. We start to really see the, uh, the capability of the compute. Um, when we're getting into, uh, into gans.

Um, open AI comes out and hits with GT three, and everything changes. Um, when it comes to dozens of language translation, uh, the ability to write, you know, the last chapter of that book that would have been, um, we start to see long form content creation. We start to see the ability to have nuance in language.

We start to see traces of memory and, and the ability to have back and forth. Uh, and, and this is just really just a, a few years ago, um, if this was a toddler, we'd be talking about it in months old that it is, that's how young it is. We, oh, it's, it's 30 months old.

And that's how we would, we would be talking about it. Um, we wouldn't even be talking about it in terms of years at this point. That's how quickly all of this is, has transpired.

Uh, Bert GPT-3 start to become foundations themselves. What has been really, uh, remarkable, uh, when it comes to the evaluators and the assessors of these different models is just how good the models themselves become. And now you get a, a simulation simulacra where, you know, the model replaces the thing that it was modeling and, and becomes a standard in and of itself.

With BUR and GPT-3, we start seeing that emerge, uh, where these foundations, uh, start to take shape. Governments start paying attention in 2022, what the heck's going on here. And things are starting to, to grow out of hand and, and out of control.

And, and in particular, the, the kind of control that that regulations can, can write policies around the effective use and the proper use of responsible use of this new technology. Um, we see it already with the eus AI Act, um, and increasing awareness, uh, by a lot of legislative bodies, governments, uh, around the world starting to, uh, to form opinions and considering new policy. Uh, and then 2023 comes, and I had to go with smaller font here, because that's just when everything started blowing up.

Uh, and so just a few highlights here. Um, meta's advances in the open source area and with its LAMA models has really advanced, uh, being able to crowdsource a lot of capability when it comes to AI modeling and, and the consumption of, uh, generative ai. Uh, not as popular as open AI's chat, GBT, um, but it has such a baked in, um, user base and an ability to, to learn quickly from such a diversity of, of modeling because of that, uh, that user base that meta has, um, and Thro getting integrated into Amazon time will tell, um, how all of that transpires, that if it looks anything like what Microsoft has done with OpenAI, um, there, there could be a possibility there.

Uh, when it comes to what could transpire. And speaking of Microsoft, a co-pilot, uh, went to a, a tech preview and then a, a limited release general availability, and now it's made its way into some of their security products, um, and all kinds of things. Uh, we know about the, the fun stuff with Barden and Jim and I, some things good, some things, uh, not so good, but making news nonetheless.

And like I said earlier, over, over $20 billion in one year alone spent on AI startups. Uh, it's a, it's a growing area, and it's growing rapidly. We went from some instrumentation in 2015 to $20 billion in annual startup investment in eight years.

Um, being able to have that kind of meteoric rise and to have participation by some of the industry's largest giants tells us that we're onto something here. And I'm not, again, I'm not, I'm not not quite ready to bind to the hype. I'm one of the most cautious AI skeptics you'll ever meet, but also one of its biggest proponents of its potential all at the same time, wrapped up in one in one head.

But, um, but there are some things that we can draw upon. There are some things that we can look at, um, when it comes to using AI in the INO in the networking space. What, what are some of these things we can pull out that might be real for us and we can leave aside the hype?

Um, one is, again, when it comes to the world of ai, data is the new gold. And when it comes to the INO and the networking world, we have to consider a lot of factors and all of that factoring is precisely what an AI is good at. Being able to understand all of the principle relationships, the key value pairs, the meaning and implication of this, and that the relations of them are models.

And in so doing, when you can embed an AI and, and train it on these types of models, you can get it to start looking at networks the way an INO manager does. The way an INO architect does, the way that a troubleshooting engineer is going to look at the appropriate paths for a particular flow. Um, the way in which you can determine yes or no, is it the network when it comes to app latency?

You know, latency is the new downtime, right? So when it comes to understanding what's going on, we're talking about a model and we're talking about data that's far flung, but is related and relatable. When we can apply ais to that model and teach it in such a way, it doesn't really even need to train on our data itself.

That's the great thing about it. We already have an existing model since we have all of these data points and we have all of this telemetry, when you get all of this piled together into one place, it then just comes down to learning based on repetition. And we're talking about network flow fundamentally, right?

So we're talking about snapshot observations in the trillions per day, uh, being able to take reference points continuously within that model, and then just being able to train it on what this is, and that is, and it being able to discriminate all those repetitive learning instances. And whenever you embed that and look at it across the whole domain starts to learn, starts to learn what's preferable, it starts to learn what a configuration is more optimal. It starts to learn the things that you're doing when you're troubleshooting and seeing how to be an apprentice and bring an insight to you and so forth, because it's mapped that model and then understands how you're using it and wants to help out.

Um, it's kind of like clippy, I can see you writing a form letter, can I help? Um, but being able to say, I can see you're investigating a latent app, here's an insight that you may not have considered when it comes to a particular disruption in US data center for being able to draw upon those kinds of insights based on the behavior of the systems and the behavior of the analyst on those systems is the bread and butter of what AI is intended to do. Um, and so when you think about the, the use cases that could explode from there, when we look at all that different telemetry and we put it all in one place, we, we start to get a lot of versatility and we start to enable the organization make a lot more, uh, meaning out of the data that's in that, in that network telemetry, understanding what's inside all of this and what does it mean to me being able to understand the digital forensics, being able to understand my cost and potential consolidation or containment, perhaps rerouting triggers that reduce my overall cloud consumption, helping me to understand my internet exchanges and peering my costs related to that capacity planning and so forth.

Being able to detect DDoS and particular threats related to that anomaly detection, monitoring gear, being able to do things in this supplemental way and being able to do so versatile by just asking a question by just writing something in a prompt of, you know, show me the traffic patterns of the last six hours that are going between my various clouds, but highlight the ones that are most significant to GCP and AWS, but bypass app, A, B, and C. But make sure you highlight app D when you do that. The AI is, that's learned on the patterns.

It doesn't need to know the precise data to learn on the pattern and then provide you back with a relevant answer. And to do so in a common taxonomy, being able to say from all those networks and all that telemetry across all of these different things you're trying to do, I'm going to, to have this intelligence filtered through all of that and come up with the appropriate response and the right connection where I'm I going to take action and remediate. And really, when it comes down to, I'll leave you with this, when it comes down to, to AI in networking, is it real or is it hype?

I think it really comes down to a, to a vision test. And there's really just four components. So I, I would submit to you that there, these four things are, are relevant aspects to the world of networking and using gen ai.

First is having some sort of query assistant, nobody except the fewest of the most senior expertise level, people can effectively plow through all of that telemetry and all of that data writing queries that are based on some proprietary language. But we can all speak our own language. We can all write our own language when we're interacting with these systems.

And so that just greatly broadens the possibilities of what you could answer, because now it, you're not limited by what you can ask and how you can ask it. The second is being able to build upon that in a journey to pursue, whether it be an investigation or troubleshooting an incident, or determining whether or not there is a, an, an appropriate path for this particular, uh, direction or another through traffic engineering. When you go down these journeys, you're prompting upon prompting upon prompting and being able to then apply that in a saved way to another question.

And now the model builds models upon models. Next becomes, again, through that process of companionship and being with the network engineer and the architects, um, and with our I and o leaders, as we troubleshoot and solve problems, it's learning from you. So in that AI vision, can it be learning from me and be assisting with me and, and calling out possibilities of root cause analysis, et cetera.

And then finally, can it, can it take action with me? Can it assist me in the action that I wanna take to remediate the issues that we've discovered together? If it can pass this vision test, then it's not hype anymore.

It's rooted in reality of the things that I would be doing anyway, that it's going to accelerate supplement or augment in some way. So I can go faster, I can go broader, I can go deeper without going longer in time and being able to do that, to answer any question about my network by this source of truth that is speedily going down the paths that I would go down anyway, and doing so in a way that's insightful because it's learning from me as I go. Um, that's really the, the eye test, you know, the vision test when it comes to, um, are these the areas of focus or is it, hey, plug in this AI and look, all your problems are solved.

That's not true. Um, but instead, are there concrete examples of where this AI is deployed, how it's used, and is it in these areas that broadens my versatility, that allows me to go in depth, but not add any more time. In fact, helps me reduce time.

So those are just some thoughts about, you know, is, is AI and networking real or hype? Um, hope you, uh, got something out of today's, uh, conversation and want to thank you. Um, we always have the, the final thing here to, to say thanks.

You can follow, likes, subscribe, and do all of those things. Um, but I think one thing I'd leave you with is that, um, it's both, it's not real or hype. It's a little bit of both.

And we need to be discriminating about what we call out, what we can use and toss aside what, what needs further development. Thanks a lot.