Episode 997 ·
Self-Driving Networks and Trusting AI with Bob Friday of HPE Networking
Can a network learn to fix itself well enough that the people running it are willing to take their hands off the wheel? Today, we're talking to Bob Friday, Chief AI Officer at HPE Networking and CTO of Campus & Branch. We explore how IT teams can move from approving every fix to simply watching over the work, why agentic programming calls for a different kind of hiring and a new way to test what you've built, and what it takes to turn a good product into a real business.
All of this right here, right now, on the Modern CTO Podcast!
If you want to know more about HPE Networking, check their website here
About Bob Friday
Bob Friday is Chief AI Officer at HPE Networking, a role he took on after HPE’s acquisition of Juniper Networks. He co-founded Mist Systems and became Chief AI Officer at Juniper and CTO of its enterprise business after Juniper acquired Mist. Juniper acquired Mist, an AI-powered WLAN company, in 2019. As Group VP and Chief AI Officer at Juniper since February 2022, he focused on improving client-to-cloud experiences through AIOps.
At Mist he led the push to apply AI to networking. The work centered on self-learning wireless networks. That work includes the Marvis virtual network assistant. He has argued that AI assistants like Marvis can take over repetitive tasks, freeing IT teams to focus on innovation and revenue. He hosts the video series Bob Friday Talks and the podcast The Q&AI, and Juniper describes him as a part-time vintner and one of the industry’s most prominent voices on applying AI to IT.
Transcript
(Bob Friday at 00:00:00) So then we started basically keeping track of every user on the network every minute. And then it became, you know, can we use simple and fancy math to figure out what network feature is causing a bad user minute? Right? So I'm a big fan of efficacy. And so right now I've got domain experts that actually score because, let's get your thought. My opinion right now is you can benchmark all the stuff you want, and usually most of these solutions will benchmark into the 90% range.
(Joel Beasley at 00:00:38) I am curious person, and I was looking at your background. You have a fascinating history. You're like part entrepreneur, but you've done so much. And I'd hope that we could just start with you giving me a brief overview of your journey and how you got here today.
(Bob Friday at 00:00:52) Okay. I mean, so I always tell people, like, I send a Christmas card to the FCC every year because I built a whole career out of what I call unlicensed spectrum. Right? Because the journey started back in the eighties when the FCC released this unlicensed spectrum rules, you know, and that was the beginnings of the Metricom Ricochet adventure, you know, where we had visions of hooking meters up to the reading meters remotely. Then that quickly pivoted to, like, let's go hook up all these laptops. I don't know if you remember back in the nineties, Mosaic browsers and that stuff was coming out. I would tell you, you know, wireless internet, mobile internet, it was a good thought. You know, it was a better thought when Apple came out. So that was kind of the adventure of market timing is critical. So that was kind of the Metricom Ricochet adventure, which led to the Airspace startup adventure of, you know, let's go help enterprises connect all these Wi-Fi users up to the internet. That's when... Because back then, it was like these... You know, if you look back then, it was like all these Wi-Fi, all these companies, Microsoft, they were deploying these very large networks. It looked a little bit like service providers. Right?
(Bob Friday at 00:02:10) So similar to the Metricom Ricochet adventure, which was like a wireless ISP adventure, enterprise start. So that was kind of the Metricom adventure. We sold that to Cisco. That was kind of my first big company adventure. Right? That was okay. Yeah. You see, I'm going from a startup into a big company. I would tell you that that adventure there, what I learned there is when I first got there, HR told me I overreacted. Some product manager told me they were going to delay my project for two quarters. And I don't know. Have you ever worked in the startup?
(Joel Beasley at 00:02:50) I've only worked in start... I haven't had the big enterprise experience.
(Bob Friday at 00:02:55) Okay. No. So you get... You know, when you're in a startup, right, you know, if you have more than a year's worth of money in the bank, you're doing pretty good. I mean, so if someone tells you they're going to delay something for six months, that's like, okay. That's half my money. We don't delay anything for six months. So that was my beginnings of, you know, the difference between startup adventure versus big company adventure. So there I was the CTO of Cisco. So I learned that in big companies, it was more about industry innovation. It's not product innovation because I was CTO there. It's more about standards, industry, and how you can really help product innovation. So that was kind of the Cisco adventure. And then after Cisco, I started a company called Mist. So we went through the whole... You know, we acquired Meraki. I saw the beginnings of this cloud. Okay. We know cloud is going to be important. That began the, you know, the paradigm shift from, you know, hey. We just don't want to need to manage network elements. We've got to manage this cloud user experience. So that was the beginnings of the Mist adventure. You know? Can we go out? You know? And that was really... I would tell you, that was really a bigger different thing about, you know, if you look at the Airspace adventure, right, when I started Airspace, that's kind of what I call the new techno... Emerging technology adventure. Right? You know, Wi-Fi was emerging. There was a whole bunch of us trying to figure out, you know, figure out that's... That new emerging technology. The Mist adventure, when I started Mist after Cisco, that was an architectural adventure. I mean, there, you're competing against big Cisco HPE. You know? You know, the real bet there is, okay, guys. If you're going to do that, you've got to make sure that it's an arc... It's not a feature. You're not going to beat Cisco or you're not going to beat Cisco or HPE with a new feature. It's got to be a fundamental architectural change. So Mist was more about emerging technology, which I call market timing. Right? When you're doing an emerging technology, are you at the right place at the right time? You know? That's kind of like Metricom Ricochet. Right? Wireless internet, it was a good thought but it wasn't the right time. Right? You know? Cellular, you know, it was a better thought when, you know, Apple came out with the iPhone. So if you kind of look at my adventure, you know, it started with Metricom Ricochet. Airspace was kind of the emerging technology adventure, Wi-Fi coming out, Cisco, and then got back into the Mist was more of the architectural. Okay. Are we going to disrupt the big guys with an architectural change?
(Joel Beasley at 00:05:42) Were you able to do that?
(Bob Friday at 00:05:44) I'm now the chief AI officer at HPE. And so that's what we were talking about before. Right? You know? So I've been competing with Aruba for about twenty-five years, and now I'm here at Aruba after twenty-five years. And so, yes, I think we... I think that bet paid off. Right? You know? Yeah. It's... You know? We sold Mist to Juniper then sold it to HPE. And so, yes, I think we can safely say self-driving cloud AI ops is an architectural change that's, you know, disrupting networking.
(Joel Beasley at 00:06:19) Is that the current one that's happening right now, the self-driving networks?
(Bob Friday at 00:06:23) That's why I have not retired yet.
(Joel Beasley at 00:06:25) Okay. I don't... You're not... You don't strike me as the type of guy who will retire.
(Bob Friday at 00:06:30) No. I think you eventually age out at some point.
(Joel Beasley at 00:06:36) We'll find out. Alright. So self-driving networks, help me understand what these are and how they're different.
(Bob Friday at 00:06:42) This is my Waymo Uber analogy. Right? I mean, have you done the Waymo Uber thing yet?
(Joel Beasley at 00:06:51) I have. Yeah.
(Bob Friday at 00:06:51) Oh, yeah. Right. So, you know, can we bring that same experience into networking, you know, where IT... You know, the enterprise IT guy starts to trust Marvis AI assistance to actually start to drive their networks, you know, fix it with... It's what I call kind of, you know, going from human in the loop to human on the loop. Because if you look at Waymo Uber, there's still a human somewhere in the back keeping an eye on things. Right? They're not totally... They're not in the driver's seat anymore driving it, but there's still someone there keeping track of that Waymo Uber. So I think we're in the same phase with networking. Right? You know? IT guys, you know, five years ago, you probably would not have jumped in that self-driving Waymo Uber. Right? They were still driving around trying to convince themselves it was safe.
(Bob Friday at 00:07:43) When you dropped... When you jumped into that Waymo, what made you decide it was an okay thought to jump in a car without a driver? Because you did some research.
(Joel Beasley at 00:07:53) I'll be honest with you. I haven't been in the Waymo, but I have a Tesla with full self-driving, and about 70% of my driving is FSD. And so when you're talking about the trust thing, you're exactly right. The way that it started was, oh, yeah. I just do it here. I just went from one parking lot to the next, and it was very... And then it just grew. It took me two years before I was really, like, put full self-driving. Now I'm like, we're just begging Musk. Please just remove these stupid controls that I have to, like, look forward. I just want to do my thing. I just let it drive me because it's been in every situation, and it handles some situations better than I would have. And it's great.
(Bob Friday at 00:08:37) Yep. I've got it too, and I'm in the same boat. It's like, you know, I was going down the street, and it stopped. And what it stopped for is because it saw some kids in the bushes that I didn't see. You know? And so that's, okay, you guys. These sensors are a little bit better than what I am doing. So... Yeah. No. I think the full self... So that is kind of the self-driving networking vision of, okay. Can we actually bring that Waymo Uber experience into networking, you know, where humans actually start to trust Marvis to actually take care of their network. You know? Yes. Will you trust me to fix a VLAN? You know? Will you trust me to send a ticket for the bad cable? So I think we're on a similar journey. I'm a big believer. I don't know if you've done any of this agentic MCP stuff yet.
(Joel Beasley at 00:09:28) I'm a... So my history, twenty years software engineer, and I have been building more in the past year than I've built in my entire career because of all of these amazing systems.
(Bob Friday at 00:09:38) Yeah. I mean, if you... You know, I talked to my software engineers here that the ones who are really into leveraging these tools, they're probably two, three, four times more productive in terms of generating code and speed now. I mean, so that... That's probably back to the... You know, AI is not going to replace you, but you better learn how to use it because, you know, if you don't keep up with these new tools, you're going to be... You're not going to be competitive. You know, you keep up with the next generation of kids who are learning how to vibe code. And I would say MCP, you know, when you look at this stuff, you know, for me, it's more of a, you know, where it's going from deterministic programming to nondeterministic. Have you written a skill file yet?
(Joel Beasley at 00:10:25) Oh, yeah. I'm far beyond that. I've got teams of agents working together. Yeah. But it... The real art seems to be when to know how to blend your deterministic versus nondeterministic systems. It's like when... They're different. It's not that one's better than the other. They're different tools in your shed, and the job that you have is dependent on the... You know, the tool you use is depending on the job. Yeah.
(Bob Friday at 00:10:46) Yeah. So they'll know... You know, to put... You know, what I tell people now is, like, hey. You know, going forward, I'm not sure I'm hiring computer scientists or hiring English majors. Right. I mean, because now it's like, guys, I need people who actually can write good English and understand the frontier models and make sure they can create prompts and skill files. You know? And my other analogy is winemaking. Right? I mean, it's like, you know, if I give two winemakers the same grapes, I will get slightly different results. If I give two agentic programmers the same tools, I'll get slightly different results. Right? Okay. Here's my MCP server. You guys write whatever skills you want, and you'll get slightly different results. Right? So you guys... You know? So that's... You know? I think we're getting from a very deterministic, you know, Python was very deterministic. This is very nondeterministic. It's a little artsy. You have to have a feel for it. You have to understand the models. You know, and you have to understand how to write good skill files.
(Joel Beasley at 00:11:55) Absolutely. And then it will give people... You've got engineers who aren't creative, and you have engineers who are creative. And I think the engineers who are creative are just going... This has just amplified me. I mean, the best explanation I have come to after building and selling three products over the course of my... Not as big as the products you've done, but small 20, 30 person teams. The best example I have is I can now do on a Sunday afternoon what would take me six months, handful of engineers, and maybe $500,000. So that's the compression. So when we talk about, like... And I know it's different for everybody. Different engineers, different people... But for me, that's what the current compression is that I can deal with.
(Bob Friday at 00:12:40) Yeah. And what I'm finding right now is my domain experts. Right? You know, the people, they're like, I'm not waiting for your product. I'm going to do it myself, and they will invent code. Now it doesn't scale. It's... There's one thing to do an agentic POC. It's another thing to try to turn that into something you could scale and sell at scale.
(Joel Beasley at 00:13:03) Well, 100%. I got so frustrated with a CRM. I won't name the name of the CRM. It's a major CRM. But I was trying to integrate it into an application that I built, and I was getting so frustrated being limited by what they have exposed to allow me so they could charge me based off of usage. So I just got frustrated. I put it down, and in an afternoon, I just rebuilt the only... They have a thousand features. I don't need a thousand features. I use six features, and I was able to rebuild the six features of that CRM into my own product. Now I have no constraints. It wasn't even about money. It was just about freedom.
(Bob Friday at 00:13:39) Well, I mean... And you're seeing a whole bunch of startups, right, who basically got the business model. You know, I will replace your big SaaS provider with, you know, whatever function they're doing that can be reengineered very quickly and replaced.
(Joel Beasley at 00:13:54) So I want to talk a little bit about this Marvis you've mentioned a couple times. Can you explain what that is?
(Bob Friday at 00:14:00) Yeah. So Marvis is our brand name for our self-driving solution here. You know, and this has been kind of the, I would say, the vision when we started Mist back in 2014, was, you know, like I said, when I left Cisco, you know, the bet we made was that, you know, we had acquired this little company, did very good cloud management, but not really good day two operations. Question was, you know, can we really do a better job at user experience on day two? You know? So when we started it, it started off about, you know, can we actually start measuring user experience? So then we started basically keeping track of every user on the network every minute. And then it became, you know, can we use simple and fancy math to figure out what network feature is causing a bad user minute. Right? So think this Zoom Teams. Right? You know, something goes wrong, we're trying to figure out, okay, what's causing this problem here? You know? And that's where Marvis came in. It started there, and then Marvis became that fancy math. You know? Can we use, you know, some... You know, one thing we have is large experience model. You know, we take Zoom and Teams data, join it with network feature data. We train a model that can actually predict video collaboration user experience. You know? So think large language model. Right? You know? So you're trying to actually just train a model to do that. Now that is now becoming a tool in the agentic framework. Right? You know? You think about this large experience model, it's becoming a tool in our agentic toolbox to start solving root cause to find out what is causing a problem and can we fix it. You know, first of all, we've got to figure out what the problem is, and then can we fix it? And that is where this agentic framework, you know, these agentic harnesses and skills and tools are leading us to now solve problems that we couldn't before. Right? You know? We got to a certain level of all the... You know, all the deterministic Python stuff. I would say we got probably 70%. I mean, so since 2018, I've had the support team using Marvis every... For every support ticket. And so we got close to about 70 some percent, close to 80%, but then it got kind of stuck. Right?
(Bob Friday at 00:16:26) We just solved all the easy problems. And then there's this long tail of more difficult problems. And that's where I think this agentic stuff is gonna really help us start to deal with where you have to look across three or four databases.
(Bob Friday at 00:16:43) And these agents are good. Right?
(Bob Friday at 00:16:46) I don't know if you're an IT person or a networking person, but easy problems you can figure out. Right? Hey, network's down. I can...
(Bob Friday at 00:16:54) It's the harder problems that you have to go look at three or four different APIs or something to sort out. Why are you having a bad Zoom call? Is it a client, a wireless, a LAN, or a WAN? I gotta go check all four things to start to sort out what's going on. That is where it's starting to help.
(Bob Friday at 00:17:11) What I call the needle in the haystack problems.
(Joel Beasley at 00:17:15) And this Marvis, that's one of the things. So it's got all the basic stuff already solved, and now it's working to advance to get the needle in the haystack problem solved.
(Bob Friday at 00:17:23) Yes. And that's where... Okay, now we have an agentic, you know, think Claude Desktop. We got skill files.
(Bob Friday at 00:17:31) These skill files are becoming from our domain experts. Right? This is our support team. I think that's the other problem is, eventually, you need domain experts. It's one thing to know how to write a skill file, but you actually have to understand the problem you're trying to...
(Bob Friday at 00:17:46) Typically, these files are basically describing the steps to solve a problem and the tools to use to solve it. Now, obviously, an interesting thing I've seen with engineers now, ideally, you wanna let the frontier model do the reasoning and the troubleshooting plan. A lot of engineers are still trying to force the model to be deterministic. Right?
(Bob Friday at 00:18:10) Step A, step B, step C. Ideally, you wanna let the frontier model come up with a plan to solve the problem. They don't trust... They're not quite ready to trust the frontier model to come up with the troubleshooting plan.
(Joel Beasley at 00:18:23) And the thing that makes this so difficult, Bob, is it's advancing so rapidly. A conversation I had over the summer with my good friend Derek, who's an engineer, we were talking about... And then by the time we talked last week, it's like things have changed. The models are just getting so good and they're narrowly good at different things, the different models. All the frontier models.
(Joel Beasley at 00:18:49) And so just keeping up with it and what you can do and how you should do something, it's every quarter you could reinvent the workflow. And I just watched yesterday an engineer at SpaceX explain how they're organizing the AI agents together to achieve outcomes. And I had been doing that, but the way that they did it, I was like, oh, it's so much better.
(Joel Beasley at 00:19:13) It's just... So it's advancing so fast. How do you keep on top of that? You just surround yourself with smart people?
(Bob Friday at 00:19:18) Right now, I'm starting to surround myself with the domain experts.
(Joel Beasley at 00:19:23) Domain experts.
(Bob Friday at 00:19:24) I mean, yeah, but it's not really the engineering team. What you find out is your domain experts around you are playing with the latest and greatest stuff almost before the engineering team. I don't know if you've heard of Jev yet.
(Bob Friday at 00:19:40) I mean, this is like another... Yet another tool, right, that someone in the team found that does a better job with tool and skill selection. Right? I mean, should you let the frontier model do it, or are there other cheaper models to do it better? And that's back to your point.
(Bob Friday at 00:19:56) Right? People are building models optimized for some function that may do a better job than a large generalized model.
(Joel Beasley at 00:20:07) Oh, absolutely. I built an app to scan my email and to surface things to me, and I found out really quickly that you need a lightweight, like, I call them a classifier model, to classify which email should even be directed to the stronger intelligence because the cost can rack up really quick.
(Bob Friday at 00:20:26) Yeah. I mean... So I'm a big fan of efficacy. And so right now, I've got domain experts that actually score because, I just get to your thought.
(Bob Friday at 00:20:38) My opinion right now is you can benchmark all the stuff you want. And usually, most of these solutions will benchmark into the 90% range. But it's really when you get down to humans scoring things that tell you the difference between the solutions. I mean, most good solutions will benchmark pretty good. And that was a year or two ago.
(Bob Friday at 00:21:00) I was trying to do some doc search thing, and we had about six different solutions. We had some startups and some... Startup, two homemade solutions, and some big companies all trying to solve this problem. They all benchmarked in the 90% range.
(Bob Friday at 00:21:16) And I think that's the other challenge. How do you QA this stuff? In the old days, I say, here's my Python code. Here's my unit test. Please verify that my code does what I claim it does.
(Bob Friday at 00:21:28) In this agent world, how do you QA it? I mean, the whole software QA paradigm is gonna change here. Right? How do you make sure that your agentic solution is actually working?
(Joel Beasley at 00:21:42) I mean, there's tools for that.
(Bob Friday at 00:21:43) But I mean, as I say, there's benchmark tools, you have LLM judges. What I'm telling you is until you have real human expert score it... I mean, you could benchmark LLM judges, and I'm telling you, though, until you actually get real domain experts to score it. I mean, because I assume you're doing some sort of LLM judging. Right?
(Bob Friday at 00:22:01) You've got some tools to judge whether or not...
(Joel Beasley at 00:22:02) Yeah. Well... And the best judge is me. So a lot of the stuff I mentioned is personal technology that I'm developing. Less than five users, often me and my wife or me and my kids.
(Joel Beasley at 00:22:14) And we're just talking. We'll meet. We'll discuss it. I'll show the kids how you talk to the AI to get it to do what you wanna do. And that's a lot of what I'm building are small things that I could never build before because it would just take too much time and money versus the value, but now we can spin it up and do it real quick.
(Bob Friday at 00:22:33) Yeah. I mean, in that case, you're the domain expert. Right?
(Joel Beasley at 00:22:36) Right.
(Bob Friday at 00:22:36) Right. You're... Yeah. In a bigger effort, you gotta have more domain experts. Networking, not everyone knows about networking.
(Bob Friday at 00:22:44) Right? You have to have real domain experts, especially around wireless, switch and route. Right? You have to have route experts to let you know the routing answers are correct or wireless experts for the wireless stuff. And I will tell you a funny story on, personal thing.
(Bob Friday at 00:22:59) My sister and I were on a six hour drive talking to Grok, and I tried to convince her. Yeah, guys, you don't have to have... Just make sure your questions are concise because halfway through the ride, all of a sudden, it says you're out of some batch of tokens. It's like, you guys...
(Bob Friday at 00:23:16) Try to... You don't have to personally get to know your AI here. But you do, Bob.
(Joel Beasley at 00:23:24) The voice is so good. You just feel like it. Yeah. Gotta be nice to the AI.
(Bob Friday at 00:23:28) They're all burning tokens. They're just burning tokens.
(Joel Beasley at 00:23:31) They're just... The Grok drive is awesome. That's something I recently have gotten into, I'd say, in the past two weeks is hooking up Grok and talking to it while I'm driving.
(Bob Friday at 00:23:42) Yeah. It's interesting because you have something in your mind. You wanna know the answer to it, and they'll grab the answer quickly for you.
(Joel Beasley at 00:23:49) The delay... It's honestly an annoying delay. We just need to put the microchips right into our brain. I don't know. The world is so strange. What other products are you working on? You got Marvis. I love that. Is that out in the wild, or is that just for you guys internally?
(Bob Friday at 00:24:06) The Marvis is out in the wild. Marvis is an actual... We are working with big enterprise companies right now. I think, as I said, I think we're in the phase right now of the Waymo Uber.
(Bob Friday at 00:24:19) We're in that driving phase where they had drivers going around millions of miles trying to make sure it all worked. I think we're in the phase right now of networking. Humans are gonna wanna be in... Yeah.
(Bob Friday at 00:24:30) They're gonna want someone in the car with Marvis. They're not quite ready to hand the keys over to Marvis yet.
(Joel Beasley at 00:24:37) And you guys are designing the features of Marvis in a way where people can have this same experience I had with Tesla. We can use it a little bit. You called it human on the loop versus human in the loop. I thought that was fascinating. I've never heard that until today, and I'm writing it down. This is my favorite new phrase.
(Bob Friday at 00:24:54) Yeah. I mean, human in the loop is where, you know, it's like, hey, we have the answer, and with your permission... Or, here's what you need to fix, and they're in the loop. Right? They wanna look at the evidence. They wanna see everything.
(Bob Friday at 00:25:11) And they may even wanna make the change themselves. That's what I call human in the loop. Right? They're still a driver in the seat making decisions on what to do and whether or not they give Marvis permission. Now...
(Bob Friday at 00:25:24) Once they trust Marvis, they can put it on what I call the trust list. And once it's on the trust list, then they're on the loop. So they're in the loop where they wanna see every evidence... Before anything happens, they want to prove it. Once they're on the loop, Marvis is making the changes, and they're just keeping an eye on the changes.
(Bob Friday at 00:25:45) And this is like the Waymo Uber thing. Right? There's someone in the background keeping an eye on that car in case it gets stuck. Right? If it gets stuck at an intersection or something, there's actually a human keeping an eye on those who will step in and help.
(Joel Beasley at 00:25:59) Yeah. It's scream activated. Yes. Okay.
(Bob Friday at 00:26:04) No. It was almost screaming in the car. Time to sneak in.
(Joel Beasley at 00:26:09) Yeah. Oh, that's good. Why do AI operations fail? Like, and how do you build that trust? Different projects going wrong inside large organizations.
(Joel Beasley at 00:26:18) How do you manage that?
(Bob Friday at 00:26:20) This is what I call efficacy. You gotta have... I mean, I've been a big believer of you got to use your own AI. In networking, my customer support team has been a proxy for my customers.
(Bob Friday at 00:26:42) Because my position has been, hey, if my support team is not using Marvis, why would I expect my customers to use it? So, ultimately, that's where I think they get in trouble.
(Bob Friday at 00:26:58) And this is what I'm saying. There's a paradigm shift. Right? In the old days, you would try to have a QA team check everything out to make sure before you shipped it, it worked. In this new world, you've gotta have your domain experts using it.
(Bob Friday at 00:27:12) So we're here at, my support team, my SEs. Guys, if my support team SEs are not using Marvis, why would my customers trust to use it?
(Joel Beasley at 00:27:25) Okay. So you guys are using your own product. Is that making it better too? Are you training it better with your own team?
(Bob Friday at 00:27:34) That was critical. I mean, that was the... As I said, when we left Cisco, that was the... I mean, there was two bets we made. I mean, architecturally, we made the bet that, yes, you guys, you need a blank sheet of paper to design a real time cloud native architecture.
(Bob Friday at 00:27:52) Right? That's hard for big companies to replicate. Right? If you have to change something with a blank sheet of paper, it's really hard for a big company to go back to a blank sheet of paper. So that was the one bet. So, okay, guys, we're gonna bet that we need a new cloud architecture.
(Bob Friday at 00:28:11) That's gonna be hard for the big companies to start from scratch. The other bet was organizationally. In this case, you're... This is basically tying your support team to your data scientist team.
(Bob Friday at 00:28:25) Right? Because, organizationally, these are usually two separate teams. Product team, support team live in different buildings. Here, my support team and my data science team sat right next to each other. And it was really the support team that made sure we got the right data in the cloud.
(Bob Friday at 00:28:48) So I kept track of... Okay, you guys. I can't have my support team logging into APs and switches. Because if the data is in the AP and switch, then I can't solve it. Right? I gotta get that data back to the cloud. So that was kinda the first major step.
(Bob Friday at 00:29:02) Okay, guys. I need... And believe me, it was hard. If you ever worked with IT teams, taking it... We're basically taking away their SSH keys so they can't log in. It's like taking candy from a baby. You think I was stealing their lunch.
(Bob Friday at 00:29:19) You're basically saying, I don't want you guys logging in to the devices anymore. I want you to tell me what data you need back in the cloud. So that was probably the other key bet. Right? There's kind of the blank sheet of paper for the cloud, and there was also this organizationally...
(Bob Friday at 00:29:35) That's also hard for big companies, right, to get two of their big organizations to play nice with each other.
(Joel Beasley at 00:29:42) Do you think that there are any tasks that should stay in human hands, or are they all going to end up automated?
(Bob Friday at 00:29:48) I suspect in the networking, as you move towards the core router, things have more impact. Right? The core router, you probably will be a human there for a while because touching the core router can impact the whole organization. Touching an AP at the edge may affect one or a couple users. You can kinda see as you get closer and closer to the Internet, that core router becomes a thing.
(Bob Friday at 00:30:16) So I think there'll be a human in the loop there. But I think ultimately that we're gonna slowly start automating almost all these tasks. I think we're gonna get to the same point that Waymo Uber got to where they finally proved that, hey, that Waymo Uber has a better safety record than a human.
(Bob Friday at 00:30:41) I think we're getting the same thing as, like, guys, most networking problems are some human mistake. You'll get to the point where we find the same things. Like, hey, guys. Who makes more mistakes? The human or Marvis?
(Joel Beasley at 00:30:59) Oh, that's interesting. That's a very... Like, that's very clear mathematical way to look at it. Because right now, the self driving cars are doing significantly better than humans.
(Bob Friday at 00:31:11) Yeah. I mean, I tell my buddies, keep that driver's license. It's gonna become an antique because pretty soon it's gonna be illegal to drive. Right? And be like, okay, you guys.
(Bob Friday at 00:31:19) It's gonna have such a better safety record. It's gonna be hard to convince them to hand out driver's licenses.
(Joel Beasley at 00:31:28) Do you know how I think it's gonna roll out? I think it's gonna roll out by lane. I think you're gonna start with a lane, like an autonomous only lane, and then it'll... Slowly, there will be... Then two of the four lanes are autonomous, then three of the four.
(Joel Beasley at 00:31:40) And then you're gonna get all sorts of interest. I don't know. That's one of the ways I think it could roll out.
(Bob Friday at 00:31:46) Yeah. I mean, I think in some places, they're building autonomous highways. Right? Especially for trucking. You saw Arizona.
(Bob Friday at 00:31:53) Yeah, where they're starting to build, especially for corridors where they have a lot of trucking going on, right? Starting to build autonomous lanes. I think I saw that.
(Bob Friday at 00:32:03) I mean, you can see that happening. It'd probably be cheaper to have autonomous highways where they have big trucking, where they're going between A and B.
(Joel Beasley at 00:32:15) That is true. That is so cool. And then it's going to become so attractive to use those, you will get the autonomous car vehicles, right?
(Bob Friday at 00:32:25) I've never gained yet. I think I'm at 90% full stop driving. I'm not sure. Are you?
(Joel Beasley at 00:32:33) So you've done so much. You've started your own businesses, you've worked in large corporations, and you've done that, I think, two times each almost. What's the biggest lesson that you've learned coming out of all that?
(Bob Friday at 00:32:45) I would say what I—and this is what I tell my startup friends—is there's a difference between building a product and a business. You have to have... I mean, ultimately, if you're building a business, you have to have a product.
(Bob Friday at 00:33:01) But all my friends... I mean, building a business is different. If you have to build a business...
(Bob Friday at 00:33:08) And another interesting story is I had a friend who was trying to do an RFID startup, and he was trying to raise money, couldn't raise money. And eventually I told him, I said, it's not you. They just don't trust you to build a business. They trust you to build a product.
(Bob Friday at 00:33:29) So he went and hired a CEO, gave him the same deck, and he went out, raised money. And so that's my... You have to realize, building a business is slightly different than a product.
(Joel Beasley at 00:33:43) And you also said something earlier about building a product versus an industry. Could you elaborate on that?
(Bob Friday at 00:33:51) Well, I think maybe I said, aerospace was called emerging techno... When you're building... If you're doing emerging technology, Wi-Fi was an emerging technology. There's a whole bunch of us trying to start companies around Wi-Fi.
(Bob Friday at 00:34:10) Back in the 2000s, it's all about Wi-Fi controllers. There was Airspace, Aruba, Trapeze. There were a whole bunch of us trying to do that. That's what I call timing.
(Bob Friday at 00:34:23) Emerging market's a bet on market timing. When you make an emerging technology bet, you're betting that some catalyst is gonna... In my case, it was Microsoft. When Microsoft put Wi-Fi into Centrino back then, that was the catalyst. That kicked the market off.
(Bob Friday at 00:34:43) Also, Wi-Fi showed up in all our laptops.
(Joel Beasley at 00:34:46) Yeah, I remember my dad's first laptop. It didn't have Wi-Fi. There was an extra card.
(Bob Friday at 00:34:51) A card put in there, using many cards. You'd plug it into your laptop.
(Bob Friday at 00:34:56) Yep. Mist was more of what I call the architectural... If you're going to disrupt, like, the CRM space... What you find is most of these industries get disrupted every ten to fifteen years.
(Bob Friday at 00:35:12) Some new technology will disrupt an existing technology or not... Something will happen. Cloud disrupted a whole bunch of areas. Uber, I mean, Salesforce. Salesforce disrupted the whole CRM market back in its time.
(Joel Beasley at 00:35:31) That's right. They were so different too. It was great. I got to meet Parker. I'm sure you're out there in the space, in the San Fran... You're in the California area, right?
(Bob Friday at 00:35:42) I'm in Bay Area.
(Joel Beasley at 00:35:44) What area are you in? I'm in Nashville.
(Bob Friday at 00:35:46) Ah, okay.
(Joel Beasley at 00:35:47) Yeah, so I'm in country music city, but I spent a lot of time out there in San Francisco and all those spots. Oakland and...
(Bob Friday at 00:35:56) I think HPE has a tech jam. We have our sales engineering tech jam out there this year.
(Joel Beasley at 00:36:03) Oh, really? In Nashville?
(Bob Friday at 00:36:05) Nashville. All the... A bunch of HPE sales engineers are coming to Nashville.
(Joel Beasley at 00:36:12) There's another group of people I've been seeing really take with the vibe coding—people that I know that are more sales engineers and then getting into building solutions. I've thought that's fascinating.
(Bob Friday at 00:36:24) Yeah, I mean, this is the AI... It's gonna... The other discussion I have with people is, is this really just another technology that's going to create more jobs and replace us, or is this more of the industrial revolution? Is it just a seismic change? Now we're going to automate everything to the point that we're going to finally redefine work.
(Joel Beasley at 00:36:50) I think the second one. So I look at it as we just invented the excavator, but we have majority of the economy right now doing manual labor, lifting the boulders. And now this... You get one excavator and you can do the work of 10,000 men. And it's...
(Joel Beasley at 00:37:10) And some people are like, not true, it's not happening. And others are like, it is happening. And some people are like, let's buy an excavator just to play with it.
(Bob Friday at 00:37:18) Well, I mean, the other thing I tell people is, over my career, I've seen... I started with the slide rule. I saw the calculator. When I started high school, we were still using slide rules.
(Bob Friday at 00:37:33) Calculator, computer, cell phone came out, Internet came out, Wi-Fi came out. How many things changed over a career now? How fast...
(Bob Friday at 00:37:48) Back in the old days, right, usually things didn't change that fast. For thousands of years, your life was not much different than your parents' life.
(Joel Beasley at 00:37:57) Well, the rate of change is definitely increasing from my personal experience. I don't have any math on it, but, man, the number of products and services and technology and advancement that we have is, like, every day there's something new.
(Bob Friday at 00:38:10) Yeah, and look how fast you picked it up, right? This agentic stuff is only two years old. When did all this MCP agent stuff come out?
(Joel Beasley at 00:38:18) Yeah, within the past two to three years. And then it's gotten so much better because, like I said, I was a software engineer for twenty years. And then when I played with it, like, two years ago, I was like, okay, this is slightly better than autocomplete. I would still have to watch every method, function, review every piece of code. Now I'm at the point where I just... I don't.
(Joel Beasley at 00:38:38) I just... I feel almost ashamed to say it, Bob. I feel almost dirty when I'm saying it. I'm like, I don't look at every line of code written.
(Bob Friday at 00:38:46) Because you're trusting GitHub Copilot, right?
(Joel Beasley at 00:38:50) And it's so easy to figure...
(Bob Friday at 00:38:51) Out a new piece of code, right? Just have GitHub Copilot tell you, can you please explain this to me?
(Joel Beasley at 00:38:58) Yeah. Yeah, I just give it... I can... I've gotten to the point where...
(Joel Beasley at 00:39:02) And it does take some skill to learn how to prompt it correctly. But I have gotten to the point where I'm pretty good at giving an outcome-driven request that I wanted to achieve. And then I'll...
(Joel Beasley at 00:39:17) I do a dry run on it where I always make it put a plan together because I have all of my experience, and sometimes it'll do things that I disagree with or that I know what's coming three steps down the road for what I want this application to do, and I want that model to be polymorphic, right? Or it's gonna make it not.
(Joel Beasley at 00:39:38) And so I do have it make plans, and then I approve the plans and let it build. And then I run the code, and I test it myself. And if it does what I want, I move on to the next thing.
(Bob Friday at 00:39:52) Yeah. Well, I mean, that's why I'm still here because I think this generative stuff is going to change... Get us to self-driving now. It's like a new tool in the toolbox.
(Bob Friday at 00:40:05) Yeah, that is so cool.
(Joel Beasley at 00:40:07) And it's so cool that you're smart enough to see that, make that connection, and then design the product team and the way that you interact and experience the product in a way where that is an option where they can be partially involved and as trust builds become less and less involved on their own choice.
(Bob Friday at 00:40:25) Yeah, let me, as I tell people, things worked out better than expected. It's like being at the right place at the right time. For the agentic stuff, it's making sure you had the right tools, which are really just APIs, right?
(Bob Friday at 00:40:40) You can make sure we have the right APIs and tools available. And then the skills is more about making sure we have the right domain experts to help create all these skill files.
(Joel Beasley at 00:40:50) That's right. So let's put yourself five years into the future. Where do you think things are going to be in five years?
(Bob Friday at 00:40:59) I'll be retired.
(Joel Beasley at 00:41:01) Do you have... You picked a date already, by the way? Do you know that you want to retire?
(Bob Friday at 00:41:05) I think... I think... But I think before I retire, I think we are going to see enterprise IT guys trust more of us to do at least half their things. They'll start offloading, you know, please replace missing VLANs.
(Bob Friday at 00:41:26) There'll just be a whole bunch of little tasks that they'll just start trusting Marvis or their AI assistants to do. They'll get to the point where, hey, it's not worth having an enterprise IT person do it. Now, I would tell you what I found is people who actually have gone down this path, typically, what they...
(Bob Friday at 00:41:46) If you talk to the enterprise IT team, they're very reactive. Usually, they're always coming in firefighting. Once they start moving into this cloud AI paradigm, it becomes much more proactive. They're not firefighting as much.
(Bob Friday at 00:42:04) And then they take on more responsibilities. They just found out they don't have to be firefighting, logging in to CLI, solving problems anymore. They basically have this new tool taking care of the fires for them.
(Joel Beasley at 00:42:15) For CTOs who are just starting to think about autonomous infrastructure, where should they even begin? What's 101 for them?
(Bob Friday at 00:42:22) You know, for me, when I think about agentic, I think harness, frontier models, you know, models here, skills and tools down here, right? And it usually starts with the harness.
(Bob Friday at 00:42:39) In my case, we built a harness for networking. There's plenty of harnesses out there. You pick one. I don't know. Which... What's your favorite agentic framework now?
(Joel Beasley at 00:42:49) Me? Uh, I'm actually not entirely sure of the word harness in the context, but I use Cursor for... Is that... Would that be considered a harness?
(Bob Friday at 00:42:58) Well, I think of a harness, you gotta have short term, long term context memory. You gotta have guardrails. You gotta have...
(Joel Beasley at 00:43:05) Cursor.
(Bob Friday at 00:43:06) Yeah. Tools. You gotta have something to select tool... The skills tool selection.
(Joel Beasley at 00:43:11) Yeah. I have a GitHub repo for... All my projects are in a repository. And then inside the repository, I have all the different folders for skills and the hard rules and every... In my...
(Joel Beasley at 00:43:22) I call them my dot files. I have all of those there.
(Bob Friday at 00:43:27) And so that's... That would be my guidance is, think through those components, and then I would tell them to think through the efficacy. If you're gonna... In your case, you're building something small for your family. You are the domain expert.
(Bob Friday at 00:43:39) My other reason is, like, benchmarking is fine, but, ultimately, the correlation between benchmarking and actual domain experts is not totally correlated.
(Joel Beasley at 00:43:53) Okay. I don't have a lot of experience there. So... But I believe what you're saying too.
(Joel Beasley at 00:43:58) And outside of just the networking and IT and the Marvis stuff that we've been discussing, what's the one thing in AI that's very exciting for you? Like, do you use it in your personal life at all?
(Bob Friday at 00:44:09) So I also make a barrel of wine every year, and I thought this would be a great problem for AI because I don't know if you're a wine connoisseur.
(Joel Beasley at 00:44:21) I'm not.
(Bob Friday at 00:44:22) Okay. So you think about wine. What I found is every year I put half of it in a French barrel, in American... Can people really tell the difference between French oak and American oak? And the answer is yes.
(Bob Friday at 00:44:37) But that's like an organic chemistry problem. How come I can't just add a little something? You know, how come I have to use a barrel to make this happen?
(Bob Friday at 00:44:50) That seems like the perfect problem for AI. We should be able to sort this out. And that's what I'm saying. Organic chemistry and agentic programming are looking very similar.
(Bob Friday at 00:45:03) How can I give the same tools to two agentic programmers and I get different answers? Then I have to dig into their skill files to find out exactly what they do different to get different results.
(Bob Friday at 00:45:17) Yeah. So that's kind of my AI outside of networking.
(Joel Beasley at 00:45:22) If you were to start your career today, what is the one thing that you would work on? Brand new startup, what would you do?
(Bob Friday at 00:45:29) I would get very good at agentic programming.
(Joel Beasley at 00:45:33) For the CTOs out there, as we start to close, what's your next for HPE networking? Like, anything that you can share that's coming down the pipeline?
(Bob Friday at 00:45:42) I mean, I think for me personally here, HPE networking is really the self-driving vision, right? I mean, I think you're going to see us... Marvis get very close to this Marvis trust list. You know, we're going to see more enterprises trust Marvis to do more on their network.
(Bob Friday at 00:46:02) So that's what the... That's the mission I'm on, and that's what I think we're going to see happen more at HPE and in any industry in general.
(Joel Beasley at 00:46:12) Thank you so much for listening. And if you found this episode useful, please share it with a friend or colleague who you think would get value from it. And if you have topics that you would like to hear discussed on the podcast, either add me on LinkedIn or send me an email, [email protected]. Every time I get an email or LinkedIn message, it absolutely makes my day and inspires me to keep going.