Unlocking Networking Potential With Generative AI

Speaker A: Welcome to the AI Business Podcast where we feature conversations with interesting guests at the intersection of artificial intelligence and business. Hi, I'm Tom Tolley for AI Business. I'm excited to welcome to the podcast Sharon Mandel. She's the CIO of uh, Juniper Networks. Welcome Sharon.

Speaker B: Hi Tom. How are you?

Speaker A: I'm doing great. First of all, maybe you can give us some background on Juniper Networks and your role at the company. Sure.

Speaker B: Uh, Juniper Networks is uh, an AI native networking company delivering solutions for the enterprise, um, cloud data, uh, centers and for um, our wide area networking. Uh, we also have security products M and I am the SVP and Chief Information Officer of the company. So I lead our IT organization but also work very closely with our product organization. I like to think of myself as the first and best customer of Juniper's products, particularly the enterprise ones. Uh, and uh, working also very closely with, with our CTO organization on um, Genai, uh, where uh, I focus very much on how we're going to apply those technologies inside our company.

Speaker A: Maybe you can explain, uh, go deeper on how Juniper Networks is currently using AI for networking solutions, uh, particularly for optimizing network performance, traffic routing and bandwidth allocation.

Speaker B: Sure. So um, Juniper Networks, uh, I think it's about five years ago now, acquired a company called uh, uh, MIST Systems. And Mist is where the AI native part of our networking story comes from. Um, they had uh, left to found their own company, both believing that uh, you needed a cloud native solution, um, in order to help better manage networks. Um, but that cloud native was um, as much as it was about the architecture, they used uh, to build the software to help manage it. It was about collecting the telemetry information from all of the devices so you could understand in real time, um, where problems were. And then from that data they started applying machine learning algorithms in order to solve problems, um, hopefully automatically and worst case, letting the network operator know, um, perhaps before even the end user knew they were there. So an example of that is, um, they can understand if a particular access point is getting congested because there's too many people around it and perhaps reroute the traffic to another network access point, um, in order to not have some number of users having their connections interrupted. Um, they're also able now It's a very API, um, surrounded system. And so through those APIs they can connect with other applications and actually collect what our uh, uh, Chief AI Officer likes to describe as labeled data from applications that they can connect with that network to telemetry. So we can see when somebody uh, might Be having a problem, let's say, with a zoom call. Uh, is that something in the network? Is that something in the application? And perhaps resolve it quicker, either with the vendor or, you know, your own organization who might have built that application. So very exciting stuff. Uh, you know, there are reports of reducing trouble tickets, you know, from our customers on the network by 90%. Um, and it has dramatically reduced tickets in our own organization as well. So, um, really changing the experience, uh, and where people spend time, uh, thinking about the network, uh, much more about creating the right architecture and a lot less about running around, solving problems and deploying people remotely to install networks, making that experience, um, the time to value from when you get your network gear to, to going live much faster.

Speaker A: Yeah, I actually, uh, talked, uh, to one of the founders of MIST years ago, and it was kind of like looking into the future. What he was explaining is what you were just explaining, but this was years ago and they were really ahead of the curve in terms of networking and applying machine learning and AI. So it was really fascinating.

Speaker B: And now we're starting to take that back further through the network. So, um, another acquisition we made a few years ago was a company called Abstra that, um, does a lot of work in how you, um, manage and deploy your data center. So you describe how you want your data center to behave. It auto configures, but it also collects a lot of information. So my team now can bring. That information is now coming into the MIST cloud and also being correlated with the rest of the network telemetry. So, so now we can better understand issues in our data centers, um, related to the network. And it's eliminating, you know, the traditional additional tools we go out and buy where you might be aggregating logs and trying to find correlations and having to do a lot of software development of sorts to do that. Um, MIST is doing all that and revealing those insights to you today.

Speaker A: Right, Right. Great. Uh, obviously the demand for AI is off the charts, off the scales. I was just at a conference this week and of course that's the only thing people talked about. And we're now seeing this huge boom in infrastructure. It kind of reminds me a little bit about the Internet boom when the infrastructure component took off with Cisco and Juniper Networks back in the day. Um, so what are you seeing in terms of networking solutions and demand, and how has that changed in the past few years?

Speaker B: Well, I think, um, there's a couple different angles to come at that from. Um, first is just as a CIO and I share common concerns with everyone else where the number of solutions we have to support, um, on top of the network, um, it's growing. Our staffs and our budgets aren't um, and so you know we have to constantly while innovating and bringing new technology to bear on the business, we're constantly having to make room in our budgets and you know you have to do both the part of the do more with less which is maybe not the traditionally hasn't been the exciting part of the job and innovate and I think AI is really kind of coming to bear and a solution like mist on how you're both bringing innovation and exciting new technology to bear and actually making room in your budget for other things at the same time. So um, I think that's a new opportunity and a new frontier for all CIOs. I think if you look at um, uh, the infrastructure provider part of the house, um, the other place that Juniper is making a lot of investment this year is in order to do gen AI which is you know, different than the kind of AI that was used. You know the machine learning, the traditional things we have been doing the last you know, five to 10 years using M those models, building, training, using and then getting inferences from those models requires some very vast complex uh, infrastructure. These uh, what they call AI clusters, multiple tiers of compute, um, that are all networked together. And we see a huge opportunity uh there where uh, you know currently there's a single player really reaping all the gains from that in the market today, uh, using a networking technology that they bought. But, but we uh, believe that Ethernet and a standard technology is really going to be the ultimate best place for that. Um, we're proving that out in our labs and with some of our customers now and we really hope for that to be a big growth segment for Juniper going forward.

Speaker A: Great. Um, in terms of uh, Juniper again you mentioned we got these heterogeneous environments that ah, evolved over the years, uh, multi cloud hybrid network environments and so on. How do you see Juniper and AI kind of improving the visibility and insights across these very complex uh, environments.

Speaker B: Well so you know, as I was describing, the kind of combined mist and abstra use case and the things that people who work for me have to do to manage, you know, the volumes of data are just astronomical, um, coming from these solutions. Uh, and humans just are not good at processing that kind of data. So it really becomes about where do you bring that data together to do that analysis, um, in a timeframe that's relevant to the challenges that we're facing with, with our applications or our end users, depending on where in the world they're sitting. Um, you know, nobody is sitting all together in one building anymore where you have kind of a controlled surface area to manage. Um, and the kinds of things we're um, doing in the network today, you know, understanding people's locations in order to give them more customized experiences or personal experiences or timely experiences, um, that you know, all requires the collection and aggregation of this data and this visibility and tools on top of it that are, you know, doing those real time analytics and making those real time decisions. So um, that's, that's where I think, you know, it goes. Now. Um, the other, the other important thing is, you know, this data will never live all in one place. You know, no one in my position is going to pick it up and move it all into one application provider to provide a solution. So the network is absolutely essential for moving that information around in order to have that visibility for the AI, wherever it lives, to make, to make decisions.

Speaker A: Perhaps you could share, uh, a case study of how one of your AI networking solutions has made a really big impact for a customer.

Speaker B: Uh, well, yeah, I mean like I said, I'll just go back to the um, you know, there's a variety of different stories. So first is um, and this may not be so much uh, the AI, but just the cloud deployment of networking solutions. Um, it really changes uh, where you have to position people. And um, so to the degree that your solution can understand both who the customers of that solution are, what they've owned and are entitled to, and then when a particular piece of gear shows up, you don't require expertise out at that edge where the equipment has to go, but that, you know, my mother could pick it up and put it in front of a camera and scan a QR code and she could have her, her network device up and running immediately. That has transformed the speed and the pace and the reliability of how you deploy networks, which goes to something we're all trying to achieve, which is reducing time to value from between when we figured out what the solution to a problem is and delivering it. Um, but then I think what you really hear about is um, the, the, the uh, the time away from the grunt work. Right? Like when I say reducing 90% of trouble tickets, the kind of problems that people in companies who use our AI solutions spend on these average problems. You know, so the, the, the methodology of building mist was kind of what are the trouble tickets. You know, let's, let's just look at volumes, you know, which, which problem has the highest iteration, how do we take the data and apply a machine learning algorithm to that and take that not off the plate of just that one customer, but the entire, anybody who is going to experience that problem in the future. So now where we're spending our time uh, uh, as network operators is really on the kind of long tail of problems that are truly unique to something that's in my environment and really solving. Can I get a way around that uniqueness. Right. And so what my people spend time on is hard, interesting, challenging, meaty problems, not kind of redundancy which I think makes. When you hear our customers talk about this, it transforms both the employee experience who are users. Cause they're experiencing fewer problems, there's more reliability. This foundation that we rely upon is really a solid foundation but also to the people who have to run the network, they're happier network employees. So that's where I really see the kind of fundamental um, differences that people know they have a network that they can rely on and that um, the people who are doing the job of maintaining that network are really working on interesting things and not mundane stuff. And I think that's what makes work exciting for people.

Speaker A: Yeah, employee experience ex is always very critical if you want to get technology to work. Right. Um, so in terms of uh, maybe take out the crystal ball, I know when it comes to predictions uh that's a hazardous activity when it comes to technology. But maybe um, you know, when it comes to AI and networking and networking industry and so forth and what you're seeing at jupe Jumper networks, you see, you know, maybe what, what can we sort of, what might we see in the next few years? You know, keep it, you know. You know uh, you always want to go where the puck is going. You don't want it to be where it is today. So maybe get some ideas on that.

Speaker B: Yeah, well, you know I can write my own I have a dream speech. Right?

Speaker A: There you go.

Speaker B: So you, you hear a lot with Gen AI about um, the most popular coding language is going to be the one that you speak. Um, not the esoteric ones we've created to sort of be a translation barrier between us and the machines. I think you're going to see the opportunity where even today with uh, the products that help with this self configure and intent based things that requires you to understand and fill out templates. You know, I think I'm going to talk to my machine and say I need you to build a server for me in the data center that has this much capacity, um, you know, and it's just going to materialize, Um, I think you're going to see a lot more freedom of choice. I think that the network is going to become much more about this software layer and I think that the providers who own that software and that intelligence, um, perhaps not only run the network with the hardware that they build, but they may really be controlling and driving um, networks of other companies and some that may have even been their competitors in the past. Um, so I think it becomes much more of a software driven ah, world. Um, I believe, like I said, I think you speak English to your network and we've already started doing that in mist. Um, and now we're looking to kind of tune um, LLMs to be, to even have more capability and to be able to ask more questions of your network. Um, you know it's going to be a lot less how do I move data analytics somewhere or create a lot of reports and you're just asking questions about um, okay, if I'm, you know, I can imagine a scenario where I have a branch that has a certain number of employees in it and a certain type of configuration and now I know I'm going to go set up an equivalent thing in Costa Rica, you know, it can tell me, you know, what, what is the network circuit that I might need to buy? Um, you know, how, how many aps am I going to need for that location? Um, and it might even create the order for you to send back to the company. Um, you know, based on you asking about different things. There's just the possibilities of what we're going to be able to do, um, I think are going to be endless. Um, and it's going to be much more like talking to your, your colleague than the way we've had to interact with machines, um, with specialized, you know, commands and languages than we've ever had before.

Speaker A: You know there's uh, you know these co pilots are just proliferating, sorry to keep track of all the different copilots that exist uh, today. And this has been a very short period of time since these LLMs have emerged for Juniper or for your department from your standpoint, uh, what are you using internally not just for networking, but just everyday processes to improve productivity and efficiency using AI. And what you found works and maybe what you found just, it's not cutting it right now.

Speaker B: Yeah, um, well, so I think we're in a huge learning curve. Um, at Juniper we have a mandate from our CEO, um, that every department must have a gen AI, specifically genAI use case in production this year. Um, and that's proving both fun but also challenging and interesting. Um, I think, uh, where these tools are really useful is where, in the areas where you have your data estate in order. Um, with mist, I think you have this advantage of the data is almost clean by definition because you've designed the devices to send the data that you need and to use the core standard networking metadata that that exists. I think it's a bigger challenge as you start to take AI and spread it across a broad set of domains where that data may not have that same level of integrity. Um, I'm finding we're going to have to do a little bit of Marie Kondo to our unstructured data in the company, um, to really get some of these use cases in production that deliver accurate results because the output of any AI is only as good as the data that you put into it. Um, so I think you're going to see us spending a lot of time really finding that data that's pristine and getting it into the right places. Um, we're playing with, um, and working with a number of our vendor partners. Um, Microsoft, uh, is a big player, um, with ServiceNow. Um, we're very interested in how that interacts. I, um, think you're also going to find this. Um, I'm waiting for the war of the chatbots, because again, depending on the system that the chatbot might be associated with, um, that chatbot may be looking at a different set of data. And just like two people looking at a different set of data, they're going to have to agree. But I'm seeing like exciting things where the chatbot might become a full citizen of the chat room that um, you know, the war room in a, in a problem solving situation, um, people are enabling these technologies and everything. It has so many APIs. Um, it'll be interesting to watch where it's confusing, but I, uh, was just on a call before this where we were talking about the recent, uh, Air Canada bereavement story, uh, where the chatbot gave a different answer than the human that you got on the phone. But I would argue that if you called maybe five different humans at Air Canada, you also might have gotten five different answers. And we have to figure out what our definition of good and reliable is for this technology because I think we're looking for perfection from the technology where we don't expect it from each other. So, you know, figuring out what that balance is in your corporation is very interesting. And so Um, I think it's a lot of experimentation. Um, I think while the technology on the right data may work really, really well, I think as these vendors go into companies, you're going to find out how it works on your data. Um, and now applications where you know, basically if they worked for one person they would work for another. We're really going to have to do more hands on work to validate that these things are doing what we hope and expect they are doing. Um, and maybe doing some tweaking to how we manage some of our information ourselves to be able to have them do a good job.

Speaker A: Yeah, I think that's a really good point and I think uh, you know, as we get farther along on this journey, we're realizing that this is not the silver bullet I think we were, we were promised, as it always is, that seems to be the case. And I, you know, I saw one, one uh, area where uh, you mentioned the Air Canada. And I think in that, in that situation it was probably something where they probably should not have sued, uh, you know, let this go to court. But uh, but regardless, uh, I do think, like I said, it does point out something really core here. And one thing too is, uh, Microsoft um, has run into these problems as well with its GitHub copilot where it'll create code that may not be secure, uh, that could have vulnerabilities to an enterprise. Microsoft um, actually had a good response to it and they said, you know, AI is like a human. It's based on data that it gets from humans and so its responses are not always going to be correct. It's going to be like a human. So you need to have a human, you know, code review. You know, just because it's AI doesn't mean you avoid all those other things

Speaker B: he was doing before at a dinner I was at the other night, you know, they were, you know they were commenting on the term co pilot. Right. Um, there is a co pilot in the airplane you fly. And, and the pilot may give the co pilot the reins for a while, but the pilot is still the one accountable for delivering that, that flight safely to. And so we have to remember as we're using these tools that ultimately we are the ones responsible for and we have to be careful to not assume perfection and that we are. You know, my mother used to tell me, don't believe everything you read in the paper. Right? Go do some other research if it seems really out of the realm of normality. And I think, you know, we, not just with AI, but you Know, our social media, you know, we have to be reminded that at the end of the day we, we're still accountable for how we use the information that's, and the answers that are fed to us. And um, sometimes they require double checking. And you're gonna find that certain tools in certain domains are, you know, almost perfect and you're gonna find other ones that rely on other types of information. And I think that puts the onus on people like myself who are deploying these tools to really also deliver education to our users. Um, and not just say, hey, here's a cool new toy. Um, but here's your responsibility using that toy. And um, already um, you're seeing um, like risk measures being associated with models so that maybe you can determine how much you need to pay attention versus um, not. So, um, yes, a lot to come. Um, and for all the things that this technology is going to do for us, it's also going to create challenges that we have to govern and manage.

Speaker A: Yeah, great. And I like that idea of risk measures because, uh, that's a good way of looking at it. You got to assume that some of this information is not going to be what you wanted. So you got to do the monitoring observability and uh, you know, make sure, you know, things are on the, going on the right track where you take, uh, you know, make course corrections on it.

Speaker B: It's almost like credit ratings for the, for the models and, and you, you, you have to decide whether or not you're going to, you're going to take that risk with that model, right?

Speaker A: Yeah. It's not foolproof. Absolutely not yet anyway. It has AGI. They keep talking about it, but I haven't seen it yet.

Speaker B: Uh, yeah, we're a long way from that.

Speaker A: That's right. Um, but yeah, maybe just the final question is just, you know, some thoughts about kind of the what comes next with AI and Juniper. A lot of exciting things, a lot of acquisitions and they're all coming together. Maybe just a couple comments about, about that.

Speaker B: Well, I'm, I'm actually not the product manager, so.

Speaker A: Okay. All right.

Speaker B: To answer that. But, um, I can tell you that uh, some of the exciting things we've just put into production that have just come, come out. Um, uh, you know, bringing some of the security decision making closer to the edge. So our network access control that's built into Mist has been really exciting and has allowed us to really again improve the user, the end user experience of the network, um, by eliminating a handful of tools that tended to create friction between the user and getting access, quick access to the network and then the resources that they um, move to on the network. So that's, that's been pretty transformational. A lot of interesting developments in location services again to um, help provide context for users and applications as they interact with people moving through physical um, space, um, bringing 5G together with Wi fi for transportation, you know, completely transparent movement as you move maybe out of your physical location, um, into a more wide open space and having um, those networks interact seamlessly with one another. I think you're going to see a lot of things there.

Speaker A: Well Sharon, it's been a great discussion. Really appreciate having you on this uh, podcast and uh, you know, a lot of exciting things. Definitely on, you know, networking and all this infrastructure is this, you know, an area that's just exploding right now. And so it's been great to get your insights on this and uh, for everyone else out there, have a great day.

Speaker B: Thanks for having me.

Speaker A: To get more AI news and insights, visit our website@aibusiness.com until next time, thank you for listening.