Techstrong TV November 14, 2025

Hey, everyone. Are there some cracks appearing in the feed of the Kubernetes Megalith? You're watching Textron Gang.

Hey everyone, it's a Shimel and welcome to our Friday edition of Textron Gang. We're actually, as you know, we record Textron Gang the day before. So it's actually Thursday.

It's a little quiet here in the morning 'cause the place isn't open yet. But we're on the floor here at Cube Con and we're wrapping up our Cube con coverage On Thursday. Uh, you'll be able to start seeing a, well, if you didn't catch it live streamed.

You'll see it next week, probably on demand. Um, but for today's gang, I'm really introduced it, it pumped to introduce our latest guest here on the gang. He's becoming an official gang member.

He's someone I've known in the tech community for a long time, and I'll leave it at that. My friend Chris Short. Hey Chris, man.

Welcome. Thank you. Appreciate Chat.

We don't usually do this with me, we're on camera. We're on the side here and it's your first time on. Give people a little bit of you of who you are.

I'm Chris Short. I've been in tech pretty much my entire career. Uh, joined the Kubernetes project in 2017 and have been a member of the community ever since.

Uh, I'm currently the co-lead of Kubernetes contributor Communications and I also run the open source program office at a startup called CIQ. Excellent. Thank you Chris.

Thank you. And of course, joining Chris Sim me is the Dean, Mike Ard. He's been out doing interviews most of the night, he says, but you could guess.

Um, but he's here today and that's what's important. Mike, what do we got for today? Well, we're leading off with this article that you wrote over on Cloud Native now about the, uh, cracks maybe in the Kubernetes monolith community.

And the idea here is that, um, folks are getting a little frustrated maybe on one hand. And then the other side of it is that the foundation themselves have become, like most non-profits, they wind up kind of servicing their own internal political issues more than they actually wind drive driving the innovation maybe. Yeah.

So is mono monolith or megalith? I think it could be. You know, those might be synonyms if you would find Ary, Maybe.

Well, we'll Go with it. But, you know, here, here's, here's the thing though, and it just, it's why I love tech. Right?

Last year, the year before that we were writing about there is nothing on the horizon that's gonna stop this locomotive. Oh yeah. Right.

Maybe, yeah, Kubernetes mm-hmm. It's Pax Kubernetes, right. That that was the world we were looking at.

Mm-hmm. There's nothing, you know, to, to dim the, the bright lights. And I think 'cause we were looking for something that would replace Kubernetes, but instead, I think what we have found is something that Kubernetes may not be able to sort of internalize, and that's ai.

Mm-hmm. Right. We could use AI to try to make Kubernetes better, but can we use Kubernetes to make AI better?

And that fundamentally is gonna be the question of whether Kubernetes, we, we stay in a Pax Kubernetes kinda world. Mm-hmm. Or do we go into a dark ages after Rome fell or something like that, that Oh geez.

But, um, but on top of that though, there's some metrics that, that back it up. Chris and I were talking about it earlier. Micah and I spoke about it yesterday.

Look, they're crowing about 15 point something million members mm-hmm. In the cloud native community. Yeah.

That's fantastic. But when you look underneath that, this show has been flat even slightly down. I saw there was only about 9,000 and something people here.

Yeah. Not to 10,000. So it's been flat to slightly down.

60% of the people here are here for the first time. And that also is a, a kind of a trend steady statistic we've seen, at least in the us I, you Europe's still a little bigger and I don't know about the 60% number. Mm-hmm.

Um, Alright. So from the community side of things, right? If I look at the Kubernetes community and the, the AI things that are happening in the community itself, we see some work groups popping up, AI conformance a number of other AI working groups that the community as said, are needed now.

And that's because Kubernetes has kind of become the platform to do AI at scale. But at the same time, yes. Can we make AI better through Kubernetes?

That's where we're moving towards. Mm-hmm. Now it's a large community and consensus takes time sometimes.

Yeah. So we are seeing some things progress very quickly. We saw an announcement yesterday with my friend Mario Valant, uh, talking about what the AI performance group has done.

And I'm excited about that. But at the same time, I feel like the Linux OS and Kubernetes have a similar journey. We're getting to the point now where you no longer need a Kubernetes specialist on your team.

The skills are becoming more ubiquitous. Right. People understand containers like never before.

People understand, uh, pods, deployments, CRDs, the whole gamut of ways you can instrument something in Kubernetes. So with, with that kind of realization, you know, red Hat for example, you know, a lot of their revenue was from support for rl Sure. Long ago.

Now it's support for OpenShift. Yeah. Which is interesting to me.

Um, and we're seeing a lot more Kubernetes usage in general across industry, which is great to see. So at the same like, yes, there are some gaps showing. I think those are more so socioeconomic, oh, sorry.

Than actual like technology concern. I don't Know that I hear the tension along two paths. One is AI workloads are stateful and they need to scale really high.

And Kubernetes was never originally designed for that. And we run databases today on Kubernetes, but we don't run 'em at that level of scale. Right.

So people are saying, do we need to fix some of that engine and what goes into that work? Yes. Then the second thing that people are talking about is, you know, Kubernetes is not the only game in town, and you go talk to the data science community and they like a thing called S slm.

It's a job schedule that they're using as an alternative to Kubernetes orchestration because it's more accessible. And to them this is makes more sense. And they look at the Kubernetes thing and it's like, you know, a bunch of it guys telling them that there's this greatest thing over here to use data scientist looks over at Kubernetes and goes, Doesn't the law running jobs?

Well, yeah, exactly. There is that. Yeah.

There is that. Um, I know from my company's perspective where I work now, yes, slum is a big deal, but underneath that there's this technology called werewolf open source project widely used. It's stateless or state full cluster management at scale was designed for the HPC world.

Right. But is suddenly really relevant in the AI era. Yeah.

There you go. You heard it here on tech drug gag that's slim Comes from the same HPC community. Right, Right, right.

So like we're seeing the werewolf project itself, starting to adopt s LM standards and everything else so that you can actually drive those workloads in a more open manner. You know? But let me, let me not call bs, but let me just bring out statistics that I, or metrics that I've heard around Kubernetes adoption.

Mm-hmm. Yes. In Greenfield application deployments, Kubernetes and that whole cloud native stack mm-hmm.

Containerized application microservices probably represents 75, 80, 80 5% of, of Greenfield. Right. You know, critical mass.

Sure. It is the standard, it's the compute stack. Bingo.

But yet when we look at the entirety of applications that are running in the world mm-hmm. Not just in cloud Kubernetes, cloud native, you know, microservice architecture, they represent 15% of the existing applications. And I, and that stat stayed steady now for a while.

Mm-hmm. We have made no inroads into transforming, modernizing, I don't really care what you want to call it. Right.

This existing base, which is still larger mm-hmm. Than, you know, all of the new stuff we come out with all of this time. Well, I think what's happened is people have realized new workloads Yes.

Kubernetes, you want that ability to scale fast. You want that ability to just interact with APIs, but your previous legacy workloads aren't necessarily designed to work like that either. And because their legacy Right.

Re their legacy re-architecting isn't exactly gonna be high on the priority list. I think this Legacy to me means money maker. Right.

Absolutely. Right. I don't mess with, don't mess with Broke legacy means to me.

Yeah. It ain't broke don't. Right.

But I do think that That mentality is starting to change finally. I think. I think so.

Yeah. I mean, like we, we went through the DevOps era, right. I'm not saying DevOps is dead or anything like that, but we, you know, we've evolved s Well you're not saying that Chris.

Right. We still have to go back to those DevOps principles. Yes.

Because that's the problem. We need to be able to scale these legacy applications, but we're still managing them with proprietary network gear. We're not using open source load balancers or anything for that matter on those workloads, which is putting them at a disadvantage.

'cause open source is kind of the concrete foundation of a lot of these workloads. I don't think anybody really knows honestly how those workloads are actually constructed. That That's a problem too.

And so they're hoping maybe these AI tools will help with that. But if I don't know how the thing is constructed, I can't carve off a microservice off of this thing and start slicing it up. Uh, the only way I can get there is, you know, I gotta call consulting firm and then they show up with, you know, 50 kids in a bus who move in for a year and a half.

Agreed. Yeah. Agreed.

Yeah. And not cheap, but let, let me, let me call out an elephant in the room though. Sure.

They called this show Cube Con, the official name of course was Cloud Native Con. Right. But I think they stopped trying to correct people a few years ago and it's just CubeCon.

Yeah. But CubeCon has become a binary star system. Mm.

And right over there is open telemetry land or whatever they call it. Right, right. And when you take open tele hotel mm-hmm.

And you take Prometheus and you take some of these other observability projects that are in CNCF, you, you, you know, it kind of reminds me of the Arthur Clark 2001 where Jupiter becomes a star. Oh yeah. Right.

Yeah. You have a new star in this system and it, and it's, it rivals. Mm-hmm.

Yeah. The gold star. And is CNC is this town big enough to two star for two stars?

Yes. I think having two stars in the same foundation is a good thing, right? Mm-hmm.

Um, looking at it from the CCF f perspectives, you know, if I put that hat on, I see it as growth. Mm-hmm. Externally, it's not creating confusion, which I think is Yeah.

Some of the problems with a, you know, two star system, right? Like, which one do I choose? No, no, no, no one compliments the other.

So that's kind of set up well. Mm-hmm. But the thing that I've noticed is I'm starting to see things like Prometheus used in non-cloud native contexts.

So I'm starting to, you know, see exporters, um, literally. So is that a bad thing? That's Not a bad thing.

That's called maturity. Yep. And if, if someone's not using Kubernetes, but they are using some of the open and underlying components of Kubernetes, that's still a win in my book.

Right. Kubernetes is not the destination, it's a part of the journey. So if we're looking at higher level abstractions Yeah.

Hotel fits right in. And it makes a lot of sense to start using those things in non-cloud native workloads because they're more efficient, right? Like we've driven the efficiency into the underlying applications.

Yeah. So is there something to be done to jumpstart innovation a little bit in the Kubernetes TOC? Or is this just the nature of democracy as a messy system and it is what it, is?

It, Well, I can't comment on the TOC component of, or creating more innovation, but what we've seen is this very sharp uptick in AI investment. And that's kind of pulling some of those engineers and folks into those projects and not necessarily towards Kubernetes. Yeah.

So how do we make sure that the AI people are in the boat with us, is kind of how I'm seeing 2026 play out. Okay. More so than, um, Or Kubernetes gets in the AI boat Both, or, you know, very little Cross cross, uh, pollination.

Yes. Like, let's work together, let's push these things forward in a more collaborative manner than our normal company based silos. Yeah.

Which is good. Yeah. Let's, Let's last these boats together and have a party.

Yeah. Let's take a yacht outta all these Boat. That sounds Good.

You Know, you know what, we, we've gotta end this segment, but let me, I'll end it with this though. Let's keep an eye on Amsterdam and see what trends continue or what we can spot from there. So you'll have to wait until March on that.

But, uh, we're gonna be right back here in text Drunk gang, and we're gonna talk, what are we talking about next? Mike? We're talking about Kerv and open source project.

Hey, native, now moved into the CNCF. Very cool. You're watching Text Drunk Gang.

You've earned it. The spotlight, the responsibility, the weight of teams, companies, and entire industries fall on your shoulders. Lives depend on your decisions.

Your home life included that work. You are protected physically and digitally. Nothing gets through your team without a fight.

But in a globally connected world, everyone sees you, including those who mean to cause you and your organization harm. And now home your sanctuary attackers see an opportunity. Your digital front door is wide open.

And what compromises your home can breach your boardroom. Because the devil's greatest trick isn't targeting your workplace firewall. It's convincing you that your personal life isn't at risk.

Black cloak, digital executive protection, defending the new attack surface your personal life. Hey folks, we're back. And as promised, we're gonna talk a little bit about Kerv.

A project that was in a different Linux Foundation is now moving over to the CNCF. Kerv is a distributed inference server is one way of thinking about that. And a lot of more of the AI workloads are becoming more distributed.

So that's a good thing. I talked to the, at least one of the maintainers and you know, basically they said they just wanted to be hanging out with the cool kids because, you know, this is where they're gonna do a lot of the integration work going forward. Hmm.

And the other project feels like maybe it's starting to be more of a data science kind of project team, and it's an AI foundation within the Linux Foundation. And those guys are all focused on data and training. And the inference stuff is just hardware stuff that they're not interested in.

So maybe it is better to have this over here, but you're closer to this. What's your take? Uh, I think, so we've gone through some serious changes in the past, I think five years when it comes to open source or just community events in general.

So CubeCon is the one that's attracting a lot of attention. We talked about that in the last segment, but the, the everyone in the same pool model is kind of working, right? Like we've seen the CNCF landscape drive valuation in companies, and now we're seeing it draw in more tooling, which is pretty good thing.

I think CNCF landscape thing gives me a headache. I look, yes, They've made some improvements in the past year to make it less headache inducing, but I remember when they used to print those things, uh, and that required a mag. You needed, I was just So you needed a special printer to, you know, do those, what they call those things.

I slipped it out in sheets. Um, but, but here's the thing. So Mike, I think the proper nomenclature, it's not another Linux Foundation, it's a daughter foundation of the Linux, Linux found foundation.

Right? Same way. Theoretically.

CNCF. Mm-hmm. Um, and I'm just not sure how cool that is.

Is this a bunch of baby birds in a nest? And we just saw a baby bird eaters brother and sister. Hard to say.

But I think, um, er will probably get starved for engineering resources in, in that other daughter foundation, whereas it probably will be able to leverage up more on the core Kubernetes work here. I think, I hope I crossed my fingers. But, um, whether it's case server or not, we gotta figure out a way to make those AI workloads more distributed.

'cause we can't just keep scaling them up. We gotta scale 'em out. We gotta, and that's kind of the big challenge.

I don't disagree with that. Mm-hmm. I, I'm just saying at, at, at a higher level, what's healthier for the Linux Foundation.

Ah, okay. Right. That regard.

I would think, you know, fewer foundations, more concentrated work efforts is a good thing, I would think. Right? Yeah.

Because I've seen foundations come and go, so why they Keep giving up so many foundations. Gotta ask Jim. Yeah.

You gotta ask Jim on that one. But, but I think that's more so, uh, the business model of the Linux Foundation than it is the actual, like, industry. If that makes sense.

Yeah. I mean, there are, I don't know how many daughter foundations there Are in, I thought there was like 40 something. Is that all?

Maybe more than, I know the last I looked it was 42. Something like That. Okay.

Well that makes sense. 42. Yep.

The, the thing I think with K native serving is the K, right? Kubernetes is the underlying thing. Right?

So it belongs here. It should, it belongs here. And maybe it's a case where it was just put in the wrong place to begin with.

It Could've been. Yeah. And now it's a correction.

Well, 'cause here's the thing, right? Like Kubernetes and AI both kind of blew up at the same time, for lack of a better term. At least Nvidia, right?

Like, when I think about AI workloads, I think 70% of 'em are NVIDIA's. Sure. So Nvidia is scaling vertically, not necessarily horizontally right now, but their chips, they're saying, what was it?

The Jensen comment was like, a hundred x performance or something like that on their newest GPUs. Like, that's, that's huge. But that is a rip and replace operation, not necessarily making the most of the hardware you have plus.

Or you take the old ones and sell 'em to countries that can't buy the new ones. Well, no, actually you say that. But what those countries are actually doing is buying, putting all the data on hard drives, flying the engineers to a data center where they can churn through all that data and then bringing it back.

But, but that, and so that's very training specific. Mm-hmm. Hopefully as we move beyond training the inputs and other stuff, they won't be able to do that.

But who knows, We need to move the processing of the data and the, and the inference closer to the network edge where the data's being created and consumed. 'cause otherwise, is This another pitch for Waso? There's a, there's a thing called latency.

There's a thing called latency that gets in the way. Latency is always gonna be an issue, right? Like I, I actually talked to a, uh, new contributor at the Kubernetes SIG meet and greet yesterday.

And we were talking about the, the, the, the physics of networking Yeah. Are going to start getting in our way. So how do we work around those physics?

Every company supposedly has a solution to that. But what we're seeing now is more mergers and acquisitions than new companies spinning up. So as things become more pressed against the physical limits of like atoms and, you know, light and things like that, we're going to start seeing some better use cases for older GPUs, for older infra systems.

And then scaling them up for today's, you know, examinations and workloads is gonna be interesting. I Laugh 'cause we're going full circle. Yeah.

So when I was young, somebody once drove into my head, you should always bring the compute to the data. Nothing good happens when you move the data. So he, then we did the cloud and we moved data into the cloud.

Right? And now we're coming back full circle and saying, you know what? We gotta bring the compute back out to the data.

It's cyclic. This whole thing is sick. I Got a point of order.

I find that hard to believe someone told you that when you were young. Yeah. Yeah.

His definition of young, this is, This is, this is back when I was covering BDP elevens and Oh, vax. So you That's nice. You know, in, in the digital world, we had real computer science.

We had real computer science. Not these guys today. All right.

Fair enough. You know what though? Welcome the project to the, to the CNCF.

Yes. Come on in. There's 200 other people swimming in this pool.

Mm-hmm. And, uh, I, I do think it, this may be a case more of correcting Yes. A misplacement prior than, than anything else.

Let's take a break here on the gang. We'll come back and we have our third topic today, which is Google. Have they lost the love for CNCF or Linux?

I don't know. You're watching techron game. Discover Techron Group, the epicenter of tech innovation.

We are your go-to for reaching IT leaders and practitioners worldwide. Our secret impactful content that sparks awareness, engagement, and top quality leads with us. You'll access editorial websites, streaming videos, virtual events, custom content analyst research, and more.

Join our satisfied clients. Let's revolutionize your tech journey. Contact us today and tell your story to the world in the most powerful way with Textron Group.

Hey folks, we're back. And I've been at the show all week and there's a lot of things happening here, but I will say that this thing that we're about to talk about is probably the coolest thing I saw since I got here. Oh.

And, and basically what Google's announced is that they've created a sandbox for running your AI agents so that the AI agent can't go and, uh, wild and just start pulling stuff from all over the place. It creates a little bit of a, a barrier around the actual agent, which is good news because right now our biggest problem with AI agents is everybody's deathly afraid that these things are voracious and we don't know what they're gonna do when, so I kind of like this idea. It's built on that, uh, g visor thing that Google came out with a few Out Few years ago while Yeah.

Um, and so that to me was like, Hey, somebody's actually solving a problem we were all talking about. But yeah. But the issue then becomes, as I was walking away, I was like going, wow, Google's building a lot of cool stuff that they don't seem to be like giving back to the CNCF folks.

They seem to be either using that because they think it adds value to their model of GKE Or is it that they just got tired of the politics? Uh, well, yeah. I can't speak to the politics of it all, but the, I think what we're seeing is we're starting to see folks use AI and they're having interesting challenges with it, right?

Like no doubt. I know I had a conversation with a couple engineers a few weeks ago where it was like, yeah, I was using AI to build this tool and ended up deleting my cluster 'cause it thought it could do something better, which failed. So I reverted back, da da da da.

Luckily you were able to revert, you revert Thank you for using gi. Uh, I don't think I've ever said that before in my life, but the, the need for safety is very high right now. Right.

Because you can't think of all the edge cases to tell a, you know, put in a prompt, essentially. Say you're using the underlying infrastructure, don't change it. Right.

Make it work to your advantage, as opposed to starting from scratch and saying, I have this problem. Help me solve it. Go.

And then all of a sudden everything gets broken in the process of you just saying go. I mean, it's an interesting, there's other ways to get after this too. I was talking about this with Chronosphere and they've kind of taken a, a graph and wrapped it around the LLM.

And so the LLM only sees what the graph tells it it can see. Right. And so they're using the graph as a controlling function.

Yes. Um, but, you know, I think that's a, a good way of making it a little more deterministic. 'cause I'm only limiting the amount of data that I can show this thing, but I still need guardrails and policies.

Yes. I still have to figure out whether or not I trust the output or not, but Right. We seem to be starting to put together, you know, uh, a, a fabric or an ecosystem of things around the LLMs to kind of control the output better.

Mm-hmm. And I think there's good reason for that. Right.

We're all a little worried about that use case. I was speaking of actually like happening in production kind of thing, where you blow away my legacy infrastructure. 'cause you think you can do better, but you don't.

So when we look at systems built on safety, I know the CRA air traffic control network has had issues this week. We've all kind of been delayed getting here to the conference and everything. But that's a system built on safety first.

Yes. Where AI is a system built on innovation first. So it's, it's very easy to innovate, but it's harder to make it safe.

And you have to build for safety first. Mm-hmm. Before we had this air traffic control system, though, it was an innovation first system.

Right? Right. Right.

Yeah. You know, you gotta walk before your run and you can't make wine before it's time. Mm-hmm.

Right? So I think we're still in that barnstorming stage of ai, if you will, right. Wing walkers and everything else.

Yes, Yes, yes. In that aviation. We'll eventually get there.

But I guys, I think you're missing the story here. The story isn't about the tech. The story is about why isn't Google donating open source projects to foundations anymore, or literally not like what they were.

Mm-hmm. And that might be a political issue, but I don't, I don't know if it's truly it's political. It's financial.

It's, you know, it's not that Google's no longer not doing evil, it's also the Base, it's also the pace of innovation. Right? I mean, if I got a committee right, It, there might be that it, it, they feel it's slows down the innovation Consensus can take time.

And that's neither here nor there. That doesn't mean they can't release it later once they've got everything they need in a row. But Didn't, didn't you hear authoritarianism is cool again?

Come on. Ah, geez. Yeah.

There's that. Um, Stay two f shimmy says it two 30. I I talk about this, but no, it, it's, I mean, it's not authoritarianism in that look, hey, Google put the resources into developing this.

Mm-hmm. They're entitled to do what the heck they want with it. Right.

But when you look at, you know, this whole thing built on, on Kubernetes that Google donated and the, and the input and influence it had on even just establishing the CNCF. Yeah. Are we missing out on the next CNCF on the next community-wide industrywide revolution?

Yeah. Because Google's keeping this close to the chest to their breast. I think Google has learned some interesting lessons Yes.

When it comes to a cloud b, cloud native and then c open source versus closed source versus something else. Right? Yeah.

Like, they see an advantage in, in the middle, in, you know, it, I remember a thing called Anthos from Google that is basically become Google autopilot, GKE autopilot. Mm-hmm. And I think that serves their customers very well.

And some of that you can do with open source, but the actual getting it all the kinks ironed out is proprietary. Right. Which I think that's where folks are starting to find value is making resilience systems better and making them less error prone and more safe.

Going back to our, you know, original count there, Democracy is the most inefficient system of government. But it is, but it the best. It is the Best.

I Will tell you, I'm starting to hate this phrase that comes out of the valley about go fast and break things. 'cause if you're not on, if you're the guy on the plane, that's not what you want to hear. No.

No one wants to hear go fast and break things on a boat, a plane already. Yeah. Yeah.

I agree. I agree. But you know, it, I, I think so you, you, and you look at it from a historic Right.

You look at open source. Mm-hmm. So you had your sort of, you know, your, your cathedral and bizarre phase with Richard Storm and Dr.

Richard Stallman and stuff like that. Yeah. Where it was unrealistic Marxism Yeah.

Of, of, in a society of atheistic saints, right? Mm-hmm. Yeah.

Nice. But then, but then it went to like this big brother open source where a Google, an IBMA sun. Well, not sun.

They were good. Yeah. Um, whoever a, a company hp mm-hmm.

They, they did open source a particular tool or project, but they, they had their own opinions and ideas and wants for it. Right. So they retained control, but in doing so, it, it, it froze competition or froze competition out.

Yes. Right? So if you were, if you were IBM and I was HP or God darned, I'm not going to, uh, contribute to your success.

Right? Then we had this foundational era of open source mm-hmm. Where IBM and HP could work together along with Apple and Google and Meta and, and what have you.

Are we seeing that era now? Maybe. That's a great question.

And I think, you know, we were talking earlier, all the companies I've worked at that have told me not to work with other certain companies on a, you know, competition based thinking. Mm-hmm. I've worked with all those companies in the community.

Right? So we are now at a point where those companies are saying, okay, our teams are working with these other companies through in open source. Is that the best thing for us right now?

Yeah. And they're rethinking the now not the future. 'cause the future is open, let's face it.

Right? Yeah. Yeah.

So right now with belts tightening folks having to buy more GPUs and a lot of expenditures on infrastructure, yeah. You're gonna see things just not get the weight pushed behind them to make them popular and open source. They're gonna drive the bottom line to increase revenue.

True. Absolutely. Some of these things though, I mean, where they collaborate theoretically at least, least should be on something as non-differentiated value, right?

Mm-hmm. It's just an enabling tech. But I think to your point, it's getting harder to determine what's non-differentiated value.

Everyone's freaking out. Today's non is tomorrow's. Yes.

Right. There certainly is that there is that. See all these people taking pictures of us, I feel like, uh, I don't know.

We we're doing something wrong. Anyway, guys, we're about outta time here. We've, we've got the, the show floors open and we've got interviews and stuff to do.

Chris, man, thank you so much for coming in here and popping in. Appreciate it. We've gotta get you into the rotation.

Sure. We had a lot to the conversation. Thank you.

Mike, what do you got planned for the rest of today? I'm Gonna visit more boots and shake some hands and kiss some babies. I'm running for office.

Oh. But we need politician in the CNCF politician. Great.

Thanks. I hope you've enjoyed this text on gang. We will be back next week with our normal Dextron gang back in studio, but it's been a hell of a lot of fun doing it here.

I'm Alan Shimel, thanks for watching. You're out. We're out.

Hey everyone. Welcome back here to another Techstrong TV interview. Um, let me introduce you to my next guest.

He, it's his first time on Techstrong tv, so let's go easy on him, but his name is Lawrence Wong. Lawrence is with Cisco Lawrence. Welcome to Text Drunk tv.

Hey, thanks a lot, Alan. Happy to be here. We're happy to have yarn.

So, Lawrence, let's, let's start with you, if it's okay, kind of let's, we'll work backwards. Give us your role at Cisco now, and, and we'll work, we'll look at your career from going, you know, back from there. Yeah, sure.

Um, I'm the, uh, general manager of the Cisco Wireless and network platform business. And what that means is I'm responsible for the entire Cisco wireless portfolio, as well as how we actually bring together our network management platform for our campus and branch, uh, customers here. You know, prior to this role, I've held a variety of product management roles.

Uh, I, you know, have been the head of product, uh, for the Meraki portfolio. Uh, I started my product management career, uh, you know, being the product manager for switching and wireless. Uh, prior to all this, I was a electrical engineer, and I used to design, uh, RFIC circuits.

In fact, I, uh, built some of the first 8 0 2 11 ab, uh, g transceivers really? Back in the day. Really?

Yeah. Very cool. So, let me ask you the magic question.

How long have you been at Cisco? You know, it's kind of funny. Uh, you, you start losing track, but, uh, this stint has probably been the longest, uh, I've been here since, uh, Meraki was acquired by Cisco since 2012.

Really? And, and you were at Cisco before that? Again, I, uh, I, I was at Cisco where I had my first product management job.

So this is, uh, after my, uh, my career as an electrical engineer. You know, it's, it's funny, I, I've heard this, I've seen this pattern before. I have a lot of friends who like worked at Cisco, then went out and did the startup thing only to be reacquired by Cisco, right?

And, and continue along their Cisco career path. And it's, uh, you know, once Cisco always Cisco, I guess is the, is the word. Um, of course Lawrence, you know, I've been a follower customer of Cisco for 30 something years.

com, I helped start a company that was what we called an a SP application Service provider. And we were a Cisco powered network. Back then, everybody was, was the only game in town real.

I mean, there was Juniper, they were the new kids on the block then, right? And, um, but it was Cisco. It was a Cisco world.

Of course, today Cisco is very different than that Cisco. That Cisco was all about big honking routers and switches and smart switches and firewalls and, you know, hardware and, and so forth. Today, Cisco, it's a different animal.

It still does that hardware thing. It's probably still the leader in the world, I'm gonna assume. But there's also, there's the Meraki in the wireless unit, there's the software, there's security, there's observability.

How would you describe today Cisco is, especially from your perch, right? Being there as long as you have. Yeah, absolutely.

You know, look, I, I, as you stated, Cisco's been around for a long time. But, and I think one of the things that is amazing about this company is a continued transformation, especially as we think about, you know, how do we help our customers solve the biggest problems today? And the way I think about Cisco today is you we're helping to build the critical infrastructure for the AI era.

And so that spans, you know, several areas. I think first and foremost, how do we help our customers build AI ready data centers? It's not just for the workloads, uh, that you expect, but it's also workloads, uh, for ai.

How do we actually help our customers future proof their workplaces ca across, you know, carpeted, non carpeted, you know, enterprise networks across campus and branch networks, and of course, all of that backed by digital resilience, um, powered by our AI capabilities. So that is really how I think about Cisco in a nutshell. Today, it's a very different company than probably when, uh, you know, you were thinking about Cisco.

But yes, we still, obviously, you know, networking is a part of what we do, but increasingly it's how do we in, you know, think about infusing security into, uh, networking and everything that we bring up to the market today. Absolutely. And I just feel obligated to give a shout out to the WebEx.

And is WebEx, WebEx, uh, the group that WebEx is part of now is, uh, I, I don't remember the name of it, but, uh, my, my friends Jeff Schaffer and Arner Char are, are, uh, part of that group. And, uh, they do a great job too. So I just wanted to give them some Now I'll, I'll definitely let them know, yeah.

That, that's part of the future proof workplaces. Like, we're, we're really helping bring collaboration and networking together, uh, for those type of customers. So, uh, yeah.

That's great. So you mentioned this data center. Boom.

I haven't seen a data center boom like this since 1998, right? com bubble burst, it took us about 10 years to use up all that data center space and dark fiber that we had laid out. 8 trillion in AI factories or AI data centers.

I, I gotta imagine to a company like Cisco now that's, that's a market that could move the needle, right? Um, and, and so, you know, it pays to, to, to put some real resources behind that. Of course, these are next gen ai.

They don't even call 'em data centers. They call 'em AI factories. What's the challenge there for Cisco?

What are you going to do differently? What do you, what's new than, you know, just kind of what you've been doing, powering data centers for the last 25, 35 years? Yeah, I mean, I, I think you're right that, you know, right now there's a explosion in data center.

Data center is cool again. And I think that even some of the 3 trillion numbers, uh, that you shared, uh, other people may have even greater ambitions. But I think that the net of it is like the way that you think about building data centers and how you manage that infrastructure.

It has to move beyond just the, you know, traditional, Hey, I, I'm gonna run workloads in a, you know, tiny level server, uh, to a, you know, data center cluster gonna, you know, scale up from that and really moving towards world where you're really scaling out and really creating these like mega clusters across different data centers to, you know, the, you know, power, the workloads that, uh, you know, our customers and, um, you know, where the economy's heading right now. So I, I think the scale of the thing is just vastly different. So how you think about managing, how you secure it, like you just have to take a different approach these days.

I agree with you a hundred percent. You know, Lawrence, the other thing is a trillion here, a trillion there, you know, before, you know, you're talking real money, but for all the trillions of dollars that we're putting into these, let's call 'em hyperscale data centers, a lot of the action is gonna take place at the edge on the end point, right? In transit and everything else.

Um, how, and, and again, these are places where Cisco excels as well, right? They we're not putting all of our eggs in that big honking router at the data center space anymore. We, we need to go where the people are, and they're everywhere and anywhere.

How is, I mean, but that's, that's a different strategy. That's a different, you know, kind of view how, how's Cisco adapt to that? Yeah, absolutely.

You know, if you think about like this idea that AI is more than just in the data center and really out into the enterprise, it does mean how do you really think about everything from how you manage infrastructure to how you think about securing that infrastructure? You know, I'll give you a, you know, a an example in the old days, you know, you used to use command line interface to manage individual devices. And over time, that moved to a web UI interface ultimately to APIs.

And it's still very much a human-centered activity, if you think about it. But we think that the paradigm is shifting in a dramatic way. And that paradigm is shifting towards a world where it's human and AI agents working together, uh, to drive more automation and ultimately more hands-off operation of network infrastructure.

That also means that the way you think about security changes, if you think about this idea that you have agents, uh, you know, deployed in the enterprise, like you and I can have, you know, 10, 12 plus agents working on our behalf, you have to first start thinking about the security of it. Like, what is their identity? What resources are they allowed to access?

How do you know if you know, you actually are giving them the right policies? Like this is just a very different shift in mindset, uh, for the management layer as well, and security layer. Absolutely.

Absolutely. And, and it, you know, so I'm a security guy. I've been in security 25 years.

To me, it's just like we've blown up the attack surface exponentially, right? 'cause every one of these agents and 10 or 12 agents, to me, 10, 12 agents for you is probably conservative. We may, you know, you're looking at someone who has 250 passwords, Lawrence, You know, locked in my password manager.

So you could imagine how many agents, I'm gonna wind up with one deck. But that being said, you know, the, the, the, the mission of security here, and you mentioned security a few times in your answers already. We, you know, I don't think we were doing such a fantastic job in security to begin with, right?

And now here we are talking about, you know, a blown up attack surface at some level. I guess you need AI to fight ai, but, and this is a job for ai, if you would talk a little bit about how Cisco is scaling up to meet this challenge. Yeah, you know, I, I think in the part of the portfolio that, uh, you know, I focus in, I, I think there's a few different facets to this.

One is really understanding the identity and context of the human and agents. And, you know, there's a rich body at work happening here, but I also think there's like some more fundamental blocking tackling. If you think about the day in the life of a network, uh, administrator, let's say that they have a security vulnerability that comes in.

And then one of the things that they have to decide is, how much risk am I gonna take on? And of course, there's, you know, scores for these, the severity levels, uh, and sometimes they have to take action much sooner than they want to. That can mean upgrading, you know, the firmware on a device.

It's okay if you're talking about just a handful, but if you're managing hundreds of thousands of devices across a, you know, enterprise network, it, it, it becomes extraordinarily hard. And so at Cisco, we're investing capabilities. Uh, you know, one of the most recent ones we talked about here is Live Protect.

And what this does is it takes the investments that we made, uh, in EBPF and really hook this into the kernel of the embedded operating system so that we can actually apply compensating controls, which allows a administrator to give them more runway to do that, uh, you know, upgrade. So the compensating controls protects against, uh, some of these security vulnerabilities. Then you go in into other areas like, hey, you know, we're having conversations right now, uh, you know, with federal customers, with healthcare customers, uh, around getting ready for a post quantum compute era.

And what does that mean? You know, right now we know that, you know, the harvest now decrypt later, uh, is a real security threat. So with our, you know, most recent, uh, you know, uh, you know, hardware and software releases, we want to build, you know, these devices to be pqc ready.

And that means building it at all layers of the stack from how you sign the factory firmware, uh, to how you protect the embedded operating system to how you, uh, you know, harden the, uh, the control plane aspects and networking protocols for Maxi and IPSec and modernize it, uh, you know, for this era here. And these are conversations that, you know, I think a lot of customers, you know, know is out there, but they're struggling to understand how can we get ready for this right now? Absolutely.

You know, look, make no mistake, Q Day is coming, right? I've spoken to a lot of quantum and quantum security folks, and the thing I, I learned from them is that, you know, it, it, it's not gonna be like marked on your calendar. Oh, tomorrow's Q day, Q day is gonna happen.

It may take months, weeks, or months for us to realize it did, and we're probably gonna realize it did as a result of something bad, right? That, that, yeah. That, that happens from it.

But when you, when you take Quantum and you take ai and what we're doing with ai and even, you know, physical ai as we call it, robotics and stuff like that, the three of them together. I mean, you know, you wanna talk about Industrial Revolution 4 0 5, oh, whatever you want to call it. Um, I mean the, the, the, the promises, you know, astronomical.

But so is the risk, how, again, is Cisco, and we'll get a little specific, if you don't mind. How is specifically, what's Cisco doing to, you know, for this eventuality? Because though it may not be today, and it may not be tomorrow, it's not too much beyond that.

Yeah, and you know, I, I, I think the funny thing about all this is, if you look at the trends even before, uh, you know, this mega trend of ai, what we saw in the, uh, network infrastructure is the rise of non-traditional devices. So things beyond the mobile, uh, devices, the laptops, iot, like devices, things that are, you know, coming in terms of the factory floor, the distribution warehouse, uh, robotics, to your point, these non-traditional devices aren't always the easiest to identify, to profile, uh, to secure. And so I think a lot of the, you know, the work that we're doing is starting with some of the, the foundational aspects, which is a, how do you actually make sure that you're defining, you know, common security policies that can be applicable to your entire infrastructure, whether it is that carpet and non-car environment, how do you build the hardware and software so you could have the distributed enforcement to translate the policy intent into something that's actionable.

And then I think the third piece that's always been the struggle for a lot of customers is how do you do this at scale? How do you make it easy? How do you drive operational simplicity into this?

So, you know, even as an example here, a lot of the work that we're doing with our security, uh, peers around, you know, hybrid mesh firewall to, uh, you know, our access manager, uh, you know, solution that's basically a SaaS based, um, access control, you know, solution, uh, to help our customers, you know, really deploy security in a much more meaningful and simple way across the enterprise. These are things that we're doing, and then some to really help our customers get ready for this era. What advice would you give to IT leaders today who are listening to us talk here and saying, oh my God, I gotta do something this, but I thought AI alone was, was enough that quantum robotics, what we, what could they do, you know, real world to future proof their networks in ai?

It can't just be, buy more Cisco gear, right? Though, that might be an tion, but what would be your advice there? Yeah, I, I think, you know, just like we see with, uh, you know, a lot of, I mean, every company out there adopting AI tools, oftentimes the difference between simply trial and proof of concept to production level adoption is having clear goals of what problems you're trying to solve.

If you don't have clear goals, then everything is up for grabs, then you're gonna try, you know, the next, uh, shiny object here. And then I think at Cisco, fundamentally, you know, if I think about going back to what I said at the beginning, we're helping to build critical infrastructure for the AI era. I think that, you know, from our perspective, the way that we provide value, the way we deliver innovation starts with, you know, modernizing the infrastructure.

Starts with, you know, making sure that our customers have a platform that they can build on that can layer more capabilities, that as we deliver innovations, uh, to them, as our partners deliver innovations through our platform, they get the benefits of that. So I think a lot of it is like, you know, very foundational, moving away from the firefighting mode that many of our customers in to something that is much more, you know, strategic and, you know, thinking over the next, you know, 3, 5, 6, 7 years here. Absolutely.

Lawrence, we're running a little low on time. I wanted to pivot and talk a little bit about the partner summit that you guys had. I don't know if it was, well, by the time people see this after we record, it might be a week or two ago.

Um, and, and just a shout out to our Futurum sister company. I believe Tech Field Day was, was, uh, over at Partner Summit or, or more of the Cisco summit's, been doing a tech field day with Cisco on this stuff. But you guys announced the raft of, uh, enterprise connectivity solutions at the partner Summit Summit for those, and I'm sure a lot of people out here weren't able to be there.

For those folks who weren't kinda, what did they miss? Yeah. You know, and I think first and foremost, if you think back to what I just, uh, said before, how do we help our customers, uh, you know, really accelerate and deliver on the outcomes for their business?

It starts with the foundation of the platform. We've been working towards a unified management platform. And what we are doing at Partner Summit is taking that to next level.

How do we start providing greater options across all operational types? We don't care if you're cloud, if you're on-prem or everywhere in between. We want to be able to allow our customers, whether you're using Catalyst Center or Meraki, to, you know, get the innovations that we're building across AI system, AI canvas, and delivering that experience to all them through one single management experience.

That's number one. I think the second one is we wanna base, you know, continue to provide our customers more options, uh, for cloud management, uh, more hardware capabilities, modular switching as an example, but also just more, uh, software capabilities. How do you deliver on that operational simplicity while still delivering great security?

A cloud orchestrated, uh, fabric that's cloud managed allows our, you know, customers to achieve that. And ultimately, you know, radar automation, we've been working, uh, on capabilities with our AI assistant in AI canvas that is agent in nature. How do you use, you know, English as your foundational interface to these management platforms?

And these management platforms understand what you're asking for, and it's intelligent enough and, uh, to pull out the right workflows to kick off, uh, you know, to right pull in the right, uh, telemetry to help you troubleshoot issues at greater scale. These are the things that we're bringing to market in a way that, uh, I think our customers have been looking for from Cisco for a long time now. Absolutely.

Lawrence, we covered a lot of ground here today, specifically on the partner Summit announcements is, I realize you may not have to at your fingertips, but is there a particular place on the Cisco site people can go to kinda read up and dig deeper? You know, I, I'm terrible at that, so, uh, I'll have to, I get, you know what, I'm gonna ask if we could grab a URL or something and we'll put it in the notes on, on this interview. Of course, you can always Google it, and I'm sure with Gemini, you know, if you are using Google, uh, it'll, it'll probably pull it up for you too.

But that's a great thing about living in the world we live in now, isn't it? Oh, 100%. You don't Have have to.

Yeah. You don't have to remember those arcane URLs anymore. Um, well it reminds me of when you didn't have to take your calculator in a math class, right?

And then I wanted you, if kids could do the math. But anyway, Lawrence, I wanna thank you for coming on here on Techstrong tv. We'd love to see you back on here.

You know, I think there's so much going on in this whole space right now. And of course right now AI is the, uh, what's the, the, the spoon that's stirring the pot, right? Uh, to a large degree.

But, um, thanks for coming on and continued success at at Cisco. Oh, thanks for having me. It was great to be a first time guest.

Looking forward to talking again. We will for sure. Lauren Wan here from Cisco.

And, uh, we're gonna take a break on text. Drunk Gang, excuse me, tech Drunk tv. I just finished Text Drunk Gang on Tech Drunk tv, and we'll be back in a moment.

Hey folks, we're back at Atlassian Europe and we're here with my friend Shameek and we're gonna have a little chat about services and service collection in their portfolio. Shameek, welcome to the show. Thank you for having me.

One of the things you guys just announced at this show is that the customer service app is now generally available, but I wanted to ask you, is customer service and IT operations help desk, is all that starting to converge now on a single kind of platform? 'cause historically we always kind of had two different things, right? Yeah, But I, you know, across all kinds of service teams inside the company, we are seeing a lot more cohesion.

Um, so often, for example, when your customer service request comes in, the support desk is in it is, is in its own silo. With the existing tools, they're able to reach back out into other teams inside the company to be able to get the answers that they need to get back to their customers. So by having it part of the service collection, the customer service management app now is able to pull data from a common teamwork graph that we have and get the answer quickly and get the best answer back to the customer fast.

So having it part of the same collection allows us to pull all of that information together in a much better way. So it sounds like the primary mission is to not have the customer service person say, we'll get back to you. Exactly.

Right. And that's so frustrating for the end customer because when they get back, you have to again talk to somebody else, and then there's a whole cycle of repeating yourselves that we want to avoid. Yeah.

How is that whole service experience gonna change in the age of ai? Because for as long as I can remember, it was you know, somebody logs a ticket, somebody reviewed the ticket, we see if we escalated it, and then we close the ticket and rinse and repeat. Yeah.

Is that gonna be a different experience with all these AI agents running around? Yeah, Absolutely. Um, so firstly, there's a lot of, uh, queries that the AI agent can resolve automatically, right?

And the second thing it can do is that it can actually ask you clarifying questions that you don't have to repeat yourself every time. And it can put all of that information and connect it with the other information it already knows about the company to ask just the precise question that it needs, rather than having, uh, that rather than kind of circling around the question and again and again like you used to with human agents, right? So all of that is great, but the most interesting thing is that it keeps learning from both you, this particular interaction that it has with the customer, but also from its interactions with all other customers.

So it keeps getting better over time, right? So all of the training that it's getting, it's not just in one agent's head, it's now in that common robo customer service agent, so that it's learning from all the agent tech information that's coming in, all the customer support requests that are coming in, and it keeps getting better over time. Will that create a perception of memory?

My, in the service desk, and I'm asking this question in this regard, every time I call into some company somewhere, they, they never remember my last interactions. I mean, it's in there somewhere. Yeah.

But generally speaking, you know, it's a whole new experience and it's a whole different interaction. So will customer service have some level of, I guess we'll call it persistence, where they actually know me Yeah. And they know my last interactions and they have a better sense of my preferences?

Absolutely. So part of the reason why every customer interaction, um, today seems like it's disparate and no, the customers have the, the service has completely forgotten about you. It's not because the information is not there, it's just that it's so cumbersome for the customer support agent to pull all of that information back out in just the time to be able to respond to you quickly.

Right? But with VO and with ai, that becomes so much more automated and fast, right? So the RO customer service agent is able to pull together all your past history, summarize it in just the right way, whether it's resolving the problem or whether a human agent is actually resolving the problem, it knows and brings all of that data together in just the right, personalized way to be able to service you in a much better way.

Mm-hmm. Um, what does it take to put all this together? 'cause some folks would say, well, we're heavily invested in all these other platforms.

So if I was gonna migrate, what would that look like? And how big a how big is the lift? Yeah.

I mean, it's, uh, usually most customers have a customer service management system where they have all their customer records. So what you can do to get started is just point our customer service management app to your set of customer records and your set of order, um, picking systems and entitlement systems without replacing what you already have. You could say that, Hey, these kinds of queries are coming into our customer service management app, and thereby start incrementally, right?

And then as you see it performing better, and as it learns more and more about you, you can start expanding the number of queries that it gets to and the categories of queries that it's responding to. So I think we have designed it in a way that you can actually get started small and then expand over time, right? So it's a, it's a pretty easy lift in terms of how you get started over time.

Of course, you can start migrating more and more systems and customer support categories over into the CSM app, and the more you move in there, the more context it has, the better queries it can answer. In the age of ai, will we wind up restructuring many of these teams? 'cause right now I think that, you know, if there's level one, two, and three escalation, and it's like a pyramid at the bottom is mostly level one, and then it gets smaller and smaller.

But will a lot of the level one stuff be handled by an AI agent now and then I can reallocate my resources accordingly? Yeah, Absolutely. Um, there's a whole bunch of tedious queries that come in, right?

Which are mostly about just informational gathering and about, you know, where's my order, what happened to my, um, payment that got stuck and so on where the information is already there in the system, and then that just needs to be pulled out and given back to the customer, right? Um, a lot of that is already moving to self-serve as well, so customers can self-serve themselves, but whenever a customer needs a query that's slightly more concept that I would call tier one. That's where I think AI is making a lot of information, uh, dent right now.

And these are tedious tasks that no human really wants to solve because it's just a matter of looking up the data here and then answering it back. Um, those are areas where AI can do a fantastic job already. Um, and then for the more tier two and tier three category, um, queries there, the AI agent can provide an assistive capability.

It can summarize all the information of the past contacts with this customer and provide it to the human agent in a summarized form so that they can take action much more quickly. Mm-hmm. How do we maintain the personal touch?

Because sometimes you worry with AI that, you know, it all just becomes talking to a machine, but, um, is there a way to do this smartly so that people feel like, you know, somebody does still care? Yeah, So there's two things. Like one, as long as soon as we take all the drudgery out of the task, it aligned frees up the human agents to do all the more, um, the software aspects of the contact, right?

So what we wanna do is that when a person, when a customer is actually interacting with our customer service management app, we should be very clear about when are you interacting with our AI agent agent and when are you acting interacting with a human? So the AI agent looks at all the questions that are coming in and knows that this is a particularly very, um, a very tedious kind of an answer that it needed. And there it says, Hey, I'm answering this for you.

Do you want some more information? And if it sees that the human is looking for a more, um, complicated question, then it can easily figure out, or that the question is getting very sensitive, right? Or that, hey, it's going to, it's going to a loop with the human on the other side, then it can, um, escalate the problem to a human and be very clear to the customer that, look, now I'm not able to solve the problem for you.

I'm coming over to a human. And there, the human agent can come in and provide the software, touch the emotional, uh, support that the customer might need in that particular case. But, so this elevates the human agents to do the hard things, right?

And it leaves out all the tedious things for the AI agent to be able to solve. One of the things that is notorious about being in the service field is turnover is really high. Yeah.

Do you think that that will become, uh, less of an issue because we won't be maybe burning people out as quickly? Absolutely. I think that that's a, a pretty important part of this whole journey that this industry is going through, which is that, um, we really want to make sure that all the drudgery of that job is taken away so that the human agents can actually be working on the most, um, rewarding parts of the most value added part of, um, this particular role.

Yeah. So what's your best advice to folks today? You know, what do you see folks who are running service operations doing that makes you shake your head a little bit and go, folks, maybe we might wanna be a little bit smarter than that.

Yeah, I mean, I think, um, first of all, adoption of AI is, I think here, and people need to kind of embrace what's happening in the change that, um, AI is enabling because some of our customers are getting dramatic results by adopting ai, right? So just being more receptive to understanding what's happening and trying it out is, I think one thing that, you know, I would, would encourage all customers to do. The second thing is AI is only as good as the knowledge you have in the company.

So investing in more knowledge and putting all the information of your company, for example, um, what are my res, how do I respond back to a customer that has a payment failure? What are my processes for handling, um, a delayed order? Right?

These kinds of things are often not documented well, and there's no business processes that are well established. The more the companies invest in this, creating this kind of context and knowledge, the better. Not only do their AI agents become, but also the human agents become much more powerful.

I think a lot of folks would be concerned that the customer service agent might hallucinate. So are there guardrails that I can put in place to kind of prevent that from happening? Yeah, it's A very good question.

So two things like, number one, we encourage customers to start with the, the, the easier queries first, right? And to set and keep the settings so that, um, the more complex queries are going to the humans. And the second thing is that we provide a lot of control mechanisms so you can review all the answers that the AI agent is providing and coach it much like you would coach a new, um, human customer service agent to get better at their job, right?

So you can review all their answers and provide feedback on what went well, what didn't go well, and what a better answer would be. Thirdly, we provide what we call an evaluation system where even before you deploy it, you can, we, we provide a whole bunch of test queries and what are the suggested responses, and then we test the, uh, customer service management agent to see whether it's actually performing well and what the score is, right? So you would never deploy it unless you vanish to kind of tune it to get to a good score.

Very similar to how a human agent comes in and there's a training period, and you won't actually put them solo onto the, um, customer service, uh, task queue until they've actually met a certain threshold. Mm-hmm. In a lot of cases, people are using their customer service desk to upsell stuff to customers.

So would the AI agents be able to do that as well? Eventually? I think that's a place where we can get to, um, right now the, the focus of our app, and I think most of the industry has been to resolve the contact queries that are there, but upselling is definitely an area where I think, um, this whole field can get to.

One of the other issues that we have too is like, a lot of the times people get a call about something, but it's not really their issue, or it's related to some other company's thing that is dependent upon my thing, and then they all interact. Can the agents start talking to each other from different companies that are maybe part of the same solution and kind of resolve things? Yeah, Absolutely.

Um, there is obviously these, um, innovations that are happening in what's known as the MCP communications between agents, um, and also A two A, which allows agents to communicate with each other. These are areas that are still very relatively new and the, uh, connections between companies are still getting established in this area. But this is an area that absolutely we expect that if my company's service depends on another company's service downstream, then our agents should be able to talk to each other to resolve those issues.

We see that already to some degree in the observability space. Um, but in the field, operations, manufacturing, retail, and other spaces, this is still relatively new. We haven't seen a whole lot of that yet.

Yeah. So this sounds a lot better than the robotic AI type of experiences we've had so far with various chat interfaces that people have put together. Um, as you kinda look down the road a little bit, you know, what are you most excited about?

I think there's, uh, two, three things that are really exciting. Number one is, as you mentioned, agents working with other agents, whether it's inside your own company or whether it's outside, um, making that work well would really empower what we can do, because many of the issues are not dependent on what I know, but what other teams are de, you know, are doing as well, right? So that's one area where I think once we have that whole framework working will be even more powerful.

The second thing I think is really happening already, but can go, um, is likely to go even faster, is the quality of the response is getting a lot better. And what I mean by that is, it's not just text and chat, which used to happen before, but other forms of media, for example, voice. Um, so being able to talk to somebody and voice and get back a voice response that is easy to make sense of, there's no hallucinations.

The quality is so much better that, um, we are seeing significant improvements in the last year, and I expect that to continue getting even better, right? Mm-hmm. And the third thing is adding video and images to your answers also makes the, the answer so much more real and easier to understand that.

I think that's, um, another area where I expect a lot of innovation happening in the next year. Do you think that people will develop a relationship with the AI customer service agent because they'll perceive it as somebody who's regularly helpful, somebody who remembers them, and they'll start to, I don't know, assign personality traits to it? Yeah, I mean, I don't think that, um, I don't know if they'll be assigning personality traits to the customer service agent necessarily, but I do think that the trust and the expectations of what you can get from the, um, customer service agent on the other side is gonna go up significantly as they experience it more and more, and they get really high quality instantaneous results back.

Um, they would prefer getting that answer first and only when that fails would they fall back to a human agent. So that trust level, I think, inevitably is gonna go up and, uh, thereby their, uh, expectations of what can, what customer service is, goes up significantly over the next few years. Alright.

I want to come back to another point you were making about, um, making sure I have my knowledge bases in order to make the AI or work better. What from your perspective do organizations need to do to accomplish that? Because I think, you know, it's hit or miss sometimes in what's in those knowledge bases you.

So two or three things like, you know, one is being better about documenting your own business process and your policies and your, the, the way you respond to customers. What is the tonality you use? Just being more explicit about all of that, right?

Many companies have their own training manuals, not just for customer service, but even for employee service, right? But sometimes those documents and those policies are scattered, um, they need to be brought together, condensed, cleaned up, and so on. Having said that, not too many companies are going to go at it from scratch, right?

You know, if I have to build all of this knowledge from scratch, many companies are gonna just drop their hands and say, Hey, that's too much work, right? So we can apply AI even to that problem, which is to help suggest to companies, what are the questions that you're coming in from past responses that you're given to customers? Here's a likely set of answers and knowledge that we can generate for you and suggest for reviewing, right?

And then you can, with very little work, you can review it, clean it up, and then say, Hey, this is good to go, right? Because even though many companies don't have knowledge bases, they have a lot of history of past tickets that have come in how you've responded to them. Some of them might have been good answers, some of them might have been bad answers, but AI can help you summarize all of that, cleans it up, and also categorize into good, bad, you know, what are the right set of answers that I want to keep as, uh, templates for all future answers.

There's one school of thought that says every dollar spent on customer service is a dollar that doesn't go to the bottom line. And I guess, can we have a different attitude going forward where we think about investments now because the cost of the service is gonna drop dramatically? Yeah, Absolutely.

I think this actually, um, will encourage customers to actually spend more on customer support, because right now, as you mentioned the beginning, right? Customer support is not just a cost center, but also a place that of, uh, that leads a frustration for customers. Customers don't have a very good impression of their vendors because they feel like, you know, customer support just doesn't provide them the answers they're looking for, right?

So as customer support gets better by this combination of automation and humans, it's going to elevate the value that customer support is providing to every customer that every, uh, organization that has this capability, right? So it's actually gonna help improve your customer satisfaction, reduce your churn, and thereby lead to more revenue, right? So I, it's, it's not just with ai, even before, companies that have invested highly into providing better customer support have always had better customer satisfaction and therefore better retention rates, right?

And this is just, um, gonna improve that even further. All right. Hey guys, you heard in here, customer service is gonna get great soon.

Yeah, Absolutely. Thank you. Thank you.

We'll be back in a minute. Hey, welcome back to Atlassian, Europe, and we're gonna have a little chat now about AI regulations ethical use with my new friend, Stan, how you doing? Hey, Mike, great to meet you.

Thanks for having me on my online. My, my pleasure. I'm there are regulations all over the world, and you're the general counsel, and I'm sure you're keeping track of all this stuff, but different countries, different regions seem to have different attitudes.

So yeah. What's the current state of AI regulation? Because I know in the US we're kind of maybe anti-regulation, and in Europe they're pretty far ahead.

So yeah. How do I navigate all this stuff? Yeah, it's a tricky world.

Uh, it reminds me a little bit about where we were eight years ago with GDPR and privacy. Uh, so some analogies and some, some, some things that we can talk about that are different. Um, what I would say is that we here at Atlassian believe in smart regulation of ai.

Um, we believe that it's really a partnership between industry and lawmakers to, uh, I think create a regime that doesn't stifle growth, only encourages, um, you know, the development of technology, new technologies like ai, but also creates guardrails that, um, will produce AI that, um, is technology we all wanna live with. It creates more of a utopia rather than a a, a dystopia. Um, specifically when it comes to the geographic differences that you've talked about, Mike, um, right now, Europe is definitely leading the PACT with the EU AI Act.

Um, Atlassian signed on early to something called the EU Pact, um, which was sort of a, a, a, a lightweight regime that, um, we comply with today. And that's something we can offer our customers here, say, in Barcelona. Um, and then what I would say is that we're building towards that high watermark really sort of saying, okay, if the EU is leading the pack, let's go where the puck is moving.

Um, and then in the meantime, you know, if the US decides that it wants to move in a a different direction, then I think we'll remain agile and we can pivot, um, when it decides where it wants to go. The US is, uh, the United States of America and the various states have different attitudes towards AI as well. Correct.

We'll see things in California and New York, Colorado might in Texas, Colorado. Yep. Um, is there some sort of baseline standard that I can get to if I'm using AI that might be applicable to a broad number of these states and countries?

Yeah, It's a great question. Um, you know, what I would say is that, uh, in the US there's actually a government standard. It's called nist.

Um, and that's something that if you want to sell technology to the US government, um, you have to go through that checklist. And so for us, that's sort of been the, the closest industry proxy. There's also some, um, other industry standards that our customers are asking for in the United States, um, and elsewhere, it's an ISO standard, um, that's specific to ai.

And so that's something that we're also, uh, have in our roadmap to, to build towards. Um, but as far as anything that is necessarily required, um, you know, that is a much more of a matrixed, um, approach. And so what we're trying to do is really build for scale since we have customers globally, rather than try and get too specific, um, built again towards that high watermark that we see, um, you know, sort of the industry moving towards.

Um, but this conversation in a couple years could be very different, Mike, And there are other regulations that still apply, we'll sort say HIPAA and healthcare. Oh, yeah. Um, if I have an AI agent, it still has to comply with the HIPAA regulations.

Yep. And there's all kinds of other regulations. Yep.

So, um, do those need to be tweaked or can they just be applied as is to AI agents and we'll just treat them like any other end user? Yeah. So not only are there new laws, there's the existing laws that maybe have been around for, for, for many decades, privacy laws, data security laws, all like you just mentioned.

Um, and so absolutely those, um, I think are, are, are the laws today. And yes, they, they, they can apply to use cases. Um, specific to ai.

My sense is that they will also need to evolve, um, that lawmakers will need to keep up with the development of technology and make sure that those are rightly, you know, adjusted and applicable to, um, new, new use cases. Um, so all to say that this is something that takes a full legal team like mine to really stay out ahead of and make sure, um, that not only are we complying, but we're also leading and influencing. Um, and for example, we recently sent a team to Brussels here in Europe to help influence, um, the evolution of the law.

'cause again, this has to be this partnership between industry and lawmakers. Do you get the sense that the lawmakers are AI literate at this point, or are they, or is that still very much a work in progress? Yeah.

Well, I mean, most of them are not technologists by trade. Uh, some of them are, um, and they have a, a series, uh, and you know, a a bench of experts that they rely on. Um, but I think that's the opportunity that we feel at Atlassian.

And, you know, being tech lawyers on my team, we really can help bridge the two sides technology and law, and really, I think help influence the, the outcome. So I've been very pleased, um, in talking to lawmakers and their ability to be fast learners, to be able to understand, um, you know, new areas of technology and be open to, uh, curiosity and learning. Yeah.

Are you also working with other vendor partners who are also have similar interest in AI regulations? I mean, can the industry kind of speak with one voice, or is that always gonna be a hundred different voices? Yeah.

Um, so for me, you know, a simple, i I would say, um, you know, sort of model to look at within AI specifically are the fundamental LLM providers. So you're, you know, sort of anthropic, you're open ai, uh, your Gemini, um, and then you also have the deployers, which is more where Atlassian is, right? We're, we're taking the LLMs from the developers given all the, you know, incredible compute, um, infrastructure that it takes to stand up an LLM.

Um, and so maybe there, you know, I wouldn't say there's a, a schism or a divide, but I feel like when it comes to compliance and regulation, the laws today are looking a little bit more closely at the fundamental LLM providers saying, you know, do they have a kill switch? You know, are there things that are more consequential when you're actually developing the model than say, Atlassian, that's deploying the model for the end user? And so that's maybe where I see a little bit of the industry sort of, um, again, not a schism, but just sort of two different industry, uh, camps, um, and sort of how they're at least, um, looking at compliance and also what's the lift, um, in, in actually standing up a regime that that can comply.

Um, last time I checked, I don't think you can indict an AI agent, so we're, we're still responsible for what happens with this AI thing, but I don't know, do people really get that? I think are they gonna sit around and say, you know, well, they, AI did. It's not my fault.

Yeah, it's, it's great. I mean, it's absolutely a, a partnership between humans and ai, and at the end of the day, the humans have to be responsible, uh, for the actions that that, that are taken. I, I do think it will be interesting to see, uh, maybe as, uh, some litigation works its way through the courts, um, where the liability truly lies for an agent's behavior.

You know, was it the creator of the agent? Was it the, the user who, you know, gave, gave the command? Um, I, I, I think it's still too, too soon to tell.

Um, but at the end of the day, this really comes down to trust. And the, the only way that AI is really gonna be successful and is gonna be something that we want to use, um, and really fulfill, I think, the full capability and the full potential of AI will be if, if it is transparent. We know we're talking to ai, we're not talking to a human.

Um, do we understand how the models we're trained, how the biases we're either accounted for or not accounted for. All of that, I think is gonna be super, super important. And that's something that we here at Atlassian take very seriously.

You probably know we have a company value of, um, open company, um, no b******t. And so that transparency comes very naturally to us. Yeah.

So you guys have been using AI agents within your own practice, right? Yeah. So tell us a little bit about that.

How are you using these things and what surprised you? Yeah. I like to think, Mike, that we have the most innovative legal team, um, in, in any company out there.

We are not afraid to embrace new technology. We're curious. We like to experiment.

Uh, we work for a company that is agile. Um, and so very much that is in, in keeping with our, our legal brand as well. Um, we have, uh, identified two, we call them hero use cases for ai.

They're both related to ro uh, the Atlassian, uh, product. Um, and so the first one is about, uh, our service management. So we have a whole bunch of stakeholders internally within Atlassian who reach out to legal with questions, um, rather than get a whole bunch of emails and slacks, what we do is we funnel them in through what we call the legal one front door.

So we use Jira service management as that front door portal with VO powering, uh, all of the, the first line questions. So you might have a question and say, I wanna hire this new employee in this geo. Um, I wanna make some changes to the employment agreement.

You know, where should I go? Rather than have to have a human go in there and sort of point you, okay, here's the template and here's the page that tells you, you know, what changes are acceptable and which ones are not. RO can actually do the first line of defense on that.

Um, yes, you need a human behind the scenes, giving them the playbook, giving them, you know, doing the quality assurance to make sure that they're pointing, um, you know, the knowledge seeker in, in the right direction. But that's a really great example of how we can do more efficiency, um, and more throughput and not necessarily, um, you know, sort of have a, a, a human have to do that first pass. The second hero use case that we have, uh, using AI in, in, in the legal team that Atlassian, um, is around, um, knowledge extraction.

So think about, you know, in the old days you bought a company and they, uh, had a bunch of contracts and you had to hire a team of lawyers to go in and read those contracts. What are we buying? Vendor contracts and employment contracts, sales contracts, maybe some leases.

Well, rather than pay a law firm or have a team of, you know, five people, 10 people having to pour over these photocopies, you can feed those PDFs into roe, ROE will extract a summary, a high level accurate summary of all those documents, and give you a readout, say, in a confluence page that you can then sort and go through and sort of say, well, here's the ones that, uh, are highest risk. Here are the ones that we don't care about. Let those go through.

That's all the power of vo. Um, and so those are the two things that I'm really excited about that we've said tho, if, if we can focus on those and standing those up with my legal team, we can then focus on the more high value things that are really attorney work. Mm-hmm.

Are you at all worried that AI's gonna replace lawyers at some point? It's been, uh, I've, I've, I've had some, I've had some friends reach out to me and, uh, and, and ask me about that. You know, I think there's always going to be a need for judgment, that, that's the one thing that I think, um, the legal craft, um, needs humans to do, uh, and that AI cannot replace, is that there's always gonna be these, uh, I think very, uh, discreet edge cases that are gonna require really understanding the totality of facts precedent, um, being able to see shades of gray.

And I, I, I think maybe AI can get us 50, 60, 70% of the way there, but it's that last 30% that takes human judgment, human discretion, um, a lot of, I think just experience that perhaps, um, AI can cannot simulate. So maybe the legal craft evolves, um, but I don't think it ever replaces us. Mm-hmm.

Can we accelerate the legal process? I think if you ask most people who've ever been involved in the legal process, the one impression they got, it was, it takes a long time. Yep.

Can we like reduce this down to something that's more manageable? Absolutely. I think we can go much faster.

Um, and I think, you know, the business can run that much more efficiently, um, with, uh, AI helping, uh, lawyers, um, and part of what I was just describing of getting a box of 600 pages of photocopies that used to take, you know, a team of 10 people 24 hours overnight, pulling an all-nighter in a conference room to read, you can feed that into VO and you probably can get the readout in about 20 minutes, right? And just think about that. And there was some great examples in the, the keynote, uh, yesterday from re Rajeev around software coding.

Same thing. What used to take code review three days and a team of developers. I think you can do that now in a couple minutes.

So it's just gonna free us up to do better things with our time. That's, that's where I think the unlock is. One more question related to that.

Yeah. So when you take the bar exam, you're supposed to memorize all this stuff. Yeah.

Do I really need to memorize all that stuff if I have AI agents going forward? Good question. I, I feel like that comes back down to like the calculator of like, back in the, you know, we used to have to learn algebra, but now we have comp, you know, calculators and computers to do that for us.

Um, I think that there probably still will be some benefit in testing for knowledge and standardized testing. It's just gonna have to evolve. And maybe the things that we're testing for today are not the things we should be testing for in the future.

Maybe the test will be on how does the legal craft use ai, that that could be more of a skills-based test. There you go. Something to think about.

All right, folks, you heard it here. AI use it responsibly, but it can't do great things. Hey buddy, thanks for coming by.

Thanks, Mike. All right, thanks for a great conversation and We'll be back in a minute. Ransomware isn't paying so well and yet, NPM packages are full of ransomware.

Call Weave boosts AI development while Intel opens AI showrooms Platform nine celebrates an an anniversary and COMT makes data easier. All this on today's tech field day rundown. Welcome to the Tech Field Day rundown, where each time we meet, we run down the IT news of the week with variable degrees of snarkiness.

I'm your host Alistair Cook, and joining me on this special edition since well, my nor co-host Tom Hollingsworth is at Network Field Day. Joining me is Jim Ky. Hi everyone.

How are you doing today? Uh, good to see you all. And by the way, today is National Donut Day.

Quick fact. We used to have, uh, one of those Krispy Kreme donut shops right down the street from me here a few years back. It converted to a, uh, diabetic center.

So I presume there's some wisdom in that There's some poetic justice in the conversion. It's also American Football Day, which is a good way to burn off all of those excess calories. And exercise is one of the good preventative methods for, uh, diabetes and similar things.

Yes, that's true. Yes. Ransomware payments are dropping sharply with only 23% of breach companies paying hackers in the third quarter of 2025.

This is the lowest rate on record, according to co ware, stronger defenses, law enforcement pressure, and, uh, a revised corporate policies are reducing attackers leverage. While most ransomware now involves data theft and double extortion, um, companies are increasingly refusing to pay, causing average payouts to fall to just $377,000. Attackers like a carer and kien are shifting focus to medium-sized firms that are more likely to pay while remote access compromises and software vulnerabilities are becoming the main attack methods.

Are we actually getting better at it and better at resisting that there's tax jump? Um, you know, I certainly would hope that, uh, the numbers seem to reflect that, but still almost $400,000 per incident. Uh, um, one of the interesting things in the report, Alistair, that I didn't see was, was there additional pushback from insurance companies, right?

Uh, having, uh, have a bit of an insurance company background and having a spouse that started her second it career at one. Uh, you know, there was always the idea of managing risk. Uh, and maybe that's also part of it too.

The other interesting part of that article though, was that it's shifting towards much more remote access, uh, exploits. So, you know, we're still talking about people, uh, uh, not having good personal security, uh, with the whole idea of bring your own device, uh, still happening and with the, uh, on slots still of people doing AI on their own personal computers, perhaps, right? Um, maybe that's still a pretty rich, uh, target sphere and easier for people to penetrate.

So, yeah, just a fascinating, uh, down, you know, downturn and a good downturn for a change, but still kind of scary. NPM was hit by a phantom raven attack with 100 plus malicious packages being downloaded 86,000 times. So what happened is hackers have flooded the NPM repository with over 100 malicious packages that have been downloaded, as we said, 86 K times according to security form co.

The campaign called Phantom Raven Exploited an NPM feature called Remote Dynamic Dependencies, which lets packages pull code from untrusted sites without detection. The fake packages stole developer credentials and system data while hiding from security stands. Some of the malicious code even used AI generated names, oh boy, big surprise.

Um, that, uh, appeared to be legitimate. And about 80 packages are still active, exposing some pretty major risks in the open source software security community. This is pretty concerning, isn't it?

Because you think as you are building your application, you're controlling your dependencies inside your application, your CI CD platform is pulling just the versions of software that you think you should be using. Well, NPM allows these, uh, remote dynamic dependencies that are one software package that you've chosen to include pulling data out of some other location that is not the repository that you've originally looked at. And this is the, the mechanism for getting in here, uh, as, as soon as a package gets pulled in through these, uh, NPM repositories, uh, one of these a hundred malicious packages, uh, now there's only 80 of them, uh, that they're allowing Arbitrary code essentially to be pulled into your build environment and to run on your build servers and potentially then out into production servers.

Although the discussion here is primarily about attacking your build servers and extracting information, developer information access keys may be for cloud services, those kinds of, uh, credentials that are being used to access other remote systems. So pulling the contents of probably your software source code repositories out and, and sending them somewhere for an attacker to then use this is pretty horrific. Uh, really is something that is, uh, an enterprise security company.

Maybe I'm not wanting to use, uh, these public repositories where I have less control of where the, uh, the source code is actually being pulled from. Maybe I want to be doing my own governance and pulling all of my dependencies on premises, scanning them for the presence of these remote dynamic dependencies, and making sure that we actually remove those, uh, only allowing packages that don't rely on remote dynamic dependencies, or at least remapping those dependencies to only in my own source code repositories. It's a bit of a concern then that I'd need to be using a full set of source code repositories that are managed and governed on premises.

And this does put it out of reach for some smaller organizations that will wanna be very agile. It puts it into that space of the enterprise companies building larger applications. Hopefully, NPM themselves will address this as well, and will add some methodology for denying remote dynamic dependencies.

Um, that would help us to then mitigate the risk that future repositories get added, that, that, again, expose. Uh, 'cause of course, if you can add a hundred malicious packages into NPM repositories, you can probably add a hundred more and probably a thousand after that when you're good at, get good at hiding where your repositories are coming from. And when your AI does a better job of not just obfuscating the names of these, um, repositories, but also obfuscating the vulnerabilities within them while we're looking at cloud things.

Amazon's latest survey that they definitely want to tell you is, is leading to more cloud usage is expecting that, uh, cloud hosted application use will grow from 59% of enterprises to 75% in a year. Uh, most organizations see public cloud as safer and better for compliance than on-premises according to AWS and multi-cloud use is increasing it. Cloud and cybersecurity budgets are set to rise.

Our budgets never rise despite ongoing data breaches. Companies are confident in cloud platforms to support security compliance and AI initiatives. Jimma, are you gonna bit the, uh, bit your future on the cloud?

You know, it's interesting because I started out 25 years ago as a, a strict on-prem. This says, when I get that part in there, I, um, database administrator, right? Built the server on the ground up by myself.

This drives the whole nine yards, right? And when cloud really came about, I do a lot of work in, uh, you know, uh, not only Amazon, but also Oracle Cloud infrastructure, or OCI. Um, yeah, it's amazing to see the growth over time.

And, you know, security from my perspective is really one of the selling points, right? Um, the idea of having a centralized place to locate all of your security concerns, um, and to put all your eggs in that one basket and watch that one basket seems to make a heck of a lot more sense to me these days. Um, especially with all the exploits like we just talked about with MPM, right?

Uh, open source software, yeah, it's free, but who has control of it? Um, and depending on your vendor, your cloud vendor, generally, they've got a much better hand on what to look for security wise. So it's, it's not, I think, you know, Alistair not just, it makes sense to go to the cloud because it's easier, because it's, but also fewer potential people that could be compromised.

So, yeah, I, I, I certainly see that. And with the growth of ai, everything from generative to Ag agentic to whatever the next thing's gonna be, I think it's gonna make a heck of a lot more sense. Uh, I don't think it's gonna slow down.

Core Weave acquired a company called Marmo. Hope I'm saying that right. And open source Python notebook for AI and data work create a complete platform for AI development.

Marino's Notebooks will integrate with Core Weaves, cloud and tools like weights and biases, helping developers move from experimentation to large scale deployment. Uh, the Marino Notebook will stay open source and the acquisition builds on Core reef's, recent deals to expand its AI developer ecosystem. Fascinating company Core Weave, but interested, what's your take on this?

Well, firstly, I wanted to see a little bit more about Marmo and, and Marmo essentially is an alternative to using Jupiter. Uh, it is more recently developed, maybe a little more, uh, interesting as a, a reactive, um, more visual tool for development, but it's essentially about developing new AI applications in a, a nice, easy way. What I like in this is that core weaves objective here is to help companies move from the experimentation phase of building their first couple of AI applications to running dozens, hundreds, potentially thousands of AI applications, AI agents, uh, in, in production.

And, uh, Jim, you'll recall that AI infrastructure field, day three, the, uh, focus was very much on shifting from experimentation to production. And it is a difficult problem for organizations. I applaud, um, applaud Call we for addressing this, along with a bunch of the other acquisitions that they've had recently.

Uh, other acquisitions they've had included, uh, buying Monolith a ai, they're a company that specializes in, uh, bringing AI to the physical world of building machinery and, um, dealing with engineering and physics problems, uh, as well as the, uh, open pipe acquisition for reinforcement learning tools around their open weights and biases as well. So, uh, there's a whole lot of nice pieces being built into the wider core weave platform as they try and stake a larger and larger space in the AI application development. And then AI application hosting, that's where the payoff for them is make it easy to build these applications that will run on the call.

We platform. Let's hope that their commitment to keeping marmo open source and, uh, permissively licensed continues that other organizations will be permitted to use, uh, Mariama technologies as well as they wish to build out their own, uh, portfolios of AI applications. Uh, as always with these acquisitions, it will be interesting to see whether that permissive license remains, uh, or whether a more restrictive license appears over time.

It does seem like Core we've wants to have the majority of development of mamo remain community-based development, and that minimizes their cost, maximizes the possibility that things will remain open, uh, as soon as they pull to largely internal development, then their costs will go up, and that's the point at which you'll see them switch to a different licensing model. For the moment. The, their approach seems to be really nicely open source.

Intel is launching some pop-up AI experience stores all around the world, New York, London, Paris, Munich, even. So, to showcase AI powered PCs, uh, visitors can try laptops from major brands, uh, and see Intel's upcoming PET for Lake processes in action much focuses on gaming. While sole highlights AI applications in Asian non-English language, uh, in particular, the stores aim to boost brand visibility and demonstrate AI features and generate excitement ahead of Intel's next generation PC launch.

I guess these experiences are all about helping their partners get out and get the idea that an I PC is what you want. Jim, do you want an ipc? Uh, no, not particularly.

Uh, frankly, uh, I, I can, my first question is how do I turn it off? Probably, uh, but just my, as a creator, that's just my personal opinion. Uh, but seriously, the interesting thing about, uh, Intel is, you know, you think they're, well, you know, they make chips, right?

But, uh, the, the thing that was, uh, caught my attention was in Seoul, they actually are doing a Gangnam style type of rollout, you know, kind of to draw in perhaps a younger crowd or, you know, just even maybe being a little glitchy, if you will. Uh, but it was really interesting that Intel, uh, is starting to put, uh, Panther Lake in action. Uh, from everything I've read about them.

Uh, it's a, an intriguing idea, uh, in terms of being able to utilize what we typically think of as crucial for ai. But there may be other types of applications that may well run extremely well, as you mentioned, gaming especially, right? But we really don't know what the next generation of AI beyond generative and ag will bring, right?

Um, you know, you may, you may be seeing people doing, uh, much more interesting types of open source experiments or even, um, you know, um, I recall when the Planetary Society had people use their laptops screensavers to, you know, home through data to look for, uh, if I remember correctly, uh, lunar samples, or even they were look, uh, some, uh, uh, cosmic samples and things like that. So, uh, uh, you know, it's very interesting to see where all this is gonna definitely, uh, head towards. And the other part of Intel that I always found fascinating, I recently, I think within the past 15 months or so, they were at one of our Tech Field Day events, and they were basically saying, uh, you know, you don't necessarily need a GPU to do a lot of the modeling that people are trying to do.

So even though the GPU, like features of Panther Lake might be really good for gaming or for visualization, or for other things, a lot of times, uh, with smaller numbers of, uh, dimensions of attributes of features that you need for a typical model, for a lower powered model, they may be just fine. So it's an interesting, uh, idea that we see coming forward with that. Um, it should be interesting to see what that brings even for people like me who are not particularly AI driven.

Um, just my own perspective. Platform nine has, uh, marked its first year of private cloud director, which is offering VMware users a simple and cost-effective private cloud solution that runs on their existing infrastructure. Uh, interestingly, it was built by former VMware engineers.

It helps enterprises modernize cut costs and migrate quickly. And, uh, one, uh, particular example use case was a Fortune 500 firm that moved 40,000 VMs at a fraction of the typical costs using Platform nine's, v jailbreak tool, clever marketing there, um, supporting both on-prem and, uh, SaaS options. It's providing a trusted, familiar path for VMware customers looking to maintain control efficiency.

Alistair, here we go again with, you know, uh, the fallout, uh, the, the benefits, the right, whatever you wanna call it, from Broadcom's acquisition of VMware. Absolutely. There's been a lot of companies that have been looking at what are my alternatives to paying the increased price that VMware wants for VMware Cloud Foundation?

And my experience has been, if you're using the features of, of VMware Cloud Foundation, if you are using the, the full suite, you're still getting great value as, uh, you, you look at the VC suite as under, if you using all features, maybe that higher price for VCF isn't gonna suit you. And this is where companies like Platform Nine come into play. And Platform nine has, has been a long time friend of Tech Field Day, uh, cloud Field Day 21 is where they showed us private cloud director and VJL break for the very first time.

And it was, it was, and remains a pretty impressive tool. I've looked at previous migration tools or previous migration processes. I did a, a customer study of a, a, a company that migrated from VMware vSphere on premises to another virtualization platform.

Uh, a couple of the things that are crucial in there is having good tools for that migration. And that's where, particularly the VJ Break tool really is very effective, uh, scales out scales to moving large numbers of virtual machines. That Fortune 500 customer, they priced their migration of those 40,000 VMs at $35 per vm.

Now, this stands in contrast to the estimates that you see from, uh, other vendors or for, uh, in particular in the media. Somewhere between $303,000 per VM gets floated as the number, depending upon how much fear and uncertainty you want in there. Obviously the devil is in the detail of how you account all of these things.

If you want to inflate a number to say it's not worth moving, then you account a whole lot of things that maybe you wouldn't account if you wanted to say it was a good way of moving. I've always liked Platform nine. I've been a fan of them for many years as, uh, they've done managed OpenStack, managed Kubernetes, and now managed private cloud.

I really like their products, and we've, we've seen great uptake from, from customers. And that, that benchmark of moving 40,000 VMs at $35 per vm, that puts us very much in the wheelhouse of large organizations who have a lot of virtualization, but are just using vanilla virtualization from VMware. They would love to have, uh, the continuation of the, uh, vSphere alone licensing, but unfortunately, Broadcom is not looking to continue that.

So they're looking at alternatives and, uh, platform nine is probably one of the compelling alternatives we've seen. So great to see that success. And of course, I have a, a, a collection of friends who are part of the Platform nine team and, um, wishing them every continuing success.

Now it's time for us to take a closer look and we take a closer look at some innovations from Convault their data rooms, let data scientist teams quickly access and prepare backup data while using an AI powered interface that, that simplifies, uh, querying and managing this backup data for AI purposes. Using data that's already been classified as it was ingested into the backup system. Teams can reduce the manual preparation and apply role-based access control and sensitivity tags and shared data securely.

Uh, the tool aims to speed up AI model training by letting, uh, the IT team who are already managing the data, managing some of the management tasks of that data, managing management, uh, and freeing up data scientists to focus on the focus on the analytics rather than the data preparation side. Jim, you're very much in the working with data, doing good things with data. Would you like to be able to use the backup data repositories rather than the primary repositories For this?

I think it's a fascinating idea. I, I wonder why no one has thought of this before, right? If you have a database, you better be backing it up, right?

And you better be also, by the way, testing your backups. But that's another topic for another day. Uh, the key thing is, I'm surprised that, uh, it's, it's actually quite brilliant because you're gonna have to scan through that data, right?

Depending on the type of backup you're taking. Uh, but generally, you know, a lot of shops, if they have, um, structured data especially, right? Um, you're gonna be taking a full backup or some sort of, uh, you know, level zero incre, incremental level zero backup once a week anyway, you're gonna touch those blocks, right?

So, um, you might as well go ahead and capture that. And if you're backing up, uh, object or file, you know, again, similar concepts, but different, so you're gonna touch the, the, the data anyway. Um, one of the things that I still see organizations kind of struggling with is, well, how do we get value out of our structured data?

You know, there's metadata buried in there, we've got tables, we've got columns. Hopefully everything's described in some sort of master data catalog, right? Uh, guess where that's usually stored?

Yeah. Inside the database itself, you're backing it up. Why not take advantage of the bandwidth that you're gonna expel your network bandwidth and everything else, uh, to do that backup, right?

So this is a fascinating idea. Um, again, like I said, I'm really surprised no one's thought of it before. Um, it, it would be interesting to see just how well this works and what they're able to capture in terms of specialty.

I'm thinking the metadata more than anything else. What's your thoughts on that? Yeah, I think the, the, there's an element here of building a data catalog, and often we would see the data catalog being built in a data lake along with a, a read only repository of that data.

And so we would often see this being another reason to make another copy of our production data in some location where we're gonna, maybe, I mean, taking the, the classic cloud methodology, we're gonna dump this into a whole bunch of object storage. We're gonna have something crawl over that object storage and updated database with metadata every time we ingest new data. Well, I've already got a place where I'm storing a copy, a read only copy of all of this data, and that I also have version history over time because backup system does this data protection on a regular basis.

So this is absolutely, uh, the, a continuation of the discussion that, that these data protection companies need or are moving beyond simple backup, moving towards trying to extract as much value as possible out of the data that resides in this backup repository. That it's not just insurance, but it's actually usable data that we can then, uh, in this case expose through a model context protocol to allow easy consumption of that data in an AI application. Does then bring to, to mind that now my backup repository contain some primary data, not just protection copies, because of course, as soon as I'm grabbing that metadata, if that's stored inside my backup repository, how am I protecting that?

Whereas my second and third copies and one of them off site, I is that metadata actually being stored out somewhere else? Hopefully the metadata is being stored separately from the backup repository on a database server that I can then protect into that backup repository. Yeah, let's not get into a, a looping of our data as we're, uh, protecting it and then scanning the protection of our metadata about our protection um, somebody's slightly smarter than me has probably designed this, and so they've avoided that kind of looping problem.

But absolutely, I love when we see more value being delivered from the data that we're storing. One of the challenges here will be making sure that the backup systems performance suits the requirements of the application. That'll be consuming data from it.

Uh, it may well be that we'll see cing capabilities that layer on top of this, that allow, uh, the, the actual AI application to be working with a, a high speed copy of a subset of the data whilst the majority of the data stays on a, a lower cost tier within the storage system. Lots of implementation details that will be fascinating to look at. And of course, uh, tech Field Day will be at Commvault Shift in a couple of weeks.

And what I imagine we'll be hearing some more about that other places, the Tech Field Day will be, well, right now, this week, the reason Tom's not with us here today is that networking field, day 39 is happening. So today, November 5th and tomorrow November 6th, uh, Tom is hosting Networking Field Day with a full delegate panel and an interesting collection of presenting companies. So make sure you take a look on the Tech Field Day website or watch on along on LinkedIn.

If you are, um, watching there. The following week, I will be in Atlanta for Tech Field Day at Cobe Corn North America. So we'll be, uh, live streaming on November the 11th from the event.

Again, I have, uh, some great sponsors there. I have South Works, I have, uh, VMware by Broadcom, and I have, uh, traffic as well, uh, presenting lots to learn about those companies there. Uh, and again, tech show.

Other things that we will come up with. Well, tech Field Day Rundown will be here every Wednesday, so continue to, uh, catch us on YouTube or in your favorite podcast application. The Rundown is also streamed on Techstrong TV Z, and you can catch us on other Techstrong and Future and programs.

Uh, we'll be back next Wednesday to talk about all of the IT news of the week that was, and that's fit to print, or at least to talk about. Until then, for myself and for Jim and all of the tech fields, AAM is wishing you and yours a great day. Hi.

Welcome everyone. I am, I'm with K. And today we're going to explore how to build infrastructure that isn't just a capable, but also a ready.

There's a big difference between demoing something with chat GPT and actually deploying something at scale that is secure, observable, and also well wellover. So today we're gonna be diving into the cutting edge developments that are shaping AI infrastructure, a vertical component as organizations move from AI experimentation to production grade systems. The evolution of AI infrastructure is not just about new models or tools.

It's about creating a holistic, sustainable ecosystem that can adapt to emerging challenges, drive innovation and deliver business value reliable. Now, over the few slides, we're gonna highlight some key trends, best practices, and the patterns that define the state of AI infrastructure today, what it will look like in the future. So we will go over what production ready really requires some evolution about AI agents, APIs and agents.

And we'll talk about a little bit about how we can do that with Kong and how we are able to show you, um, some of these, uh, things in, in play. So, again, my name is Sugo. I'm part of the product team helping organizations to, uh, have a successful AI strategy using APIs, event-driven architecture and integration overall.

So let's get us started. Now, you will be asking like, you know, why are we talking really about AI ready infrastructure today, may or many organizations are eager to adopt generative ai, but the reality is that without clear definition of success, many projects risk failure. Today, we'll focus on wide defining success.

It's also essential. Now, let's go on, on some of the key stats. You know, a lot of companies are gonna be starting to use, uh, gen ai.

As I was mentioned, um, around 50% of the companies that have deployed AI strategies are working on that, and not just the side experiment, they're becoming central to their way business operate. Now, also, 70% of the increase of, uh, product workload will come from AI agents by 2028. In fact, like, you know, analysts like Garner is saying that 80% of the, uh, APIs from organizations will be consumed by, um, AI agents instead of, you know, traditional developers.

However, there's also some risk, and there has been a lot of studies, uh, and then, um, um, a flow of information regarding studies that says that at least 30% of generative AI projects will be abandoned by the end of this year. And this points to a significant issue. Many projects are set to fail due to clear objectives or ROI or not meeting business and technical needs.

Now, it's crucial to avoid jumping into projects without understanding what success look like and how you should be able to, you know, build enterprise ready, um, uh, infrastructure. So the high failure of rate of generative AI projects, coupled with a rapid growth of AI rolling products makes defining success a critical step. Now, while we're talking about this, well, again, we are moving past the AI experimentation phase, and we are ready to move into production ready AI systems.

So at the AI experimentation phase, you know, think of, uh, this as the testing real for AI organizations explore different AI technologies run pillars and test various models and control environments. It's about understanding what works and what doesn't, and learning how to best apply AI to their specific needs. Now, that's the experimentation part, but we were talking about production ready.

Now we're talking about AI that's been refined and optimize, that it's ready to be integrated into actual business operations, products or services. These AI systems are stable, scalable, and deliver consistent results without failures. They're designed to handle real world complexity and provide ongoing value.

Now, why does this transition matter? Well, moving to production means AI is no longer just an experiment. It's been used in a way that's reliable, secure, and able to scale with near the business.

This is a major milestone and AI that's ready to be, you know, prime time with all the complexities of daily operations in mind. You, as part of the DevOps team needs to be, uh, able to check this. So now, production ready ai, it's more than just, you know, working prototype.

It needs to be fully managed, secure, observable services that deliver reliable business value at scale. So there's a couple of, um, points here that will certainly help us to, you know, move and, and, and define really, you know, a, a production rate infrastructure. So we're talking about, you know, security and compliance, you know, and infrastructure pipelines, models that enforce that encryption access control, how it laws that also needs to meet industry's regulations.

Think about GDPR, uh, HIPAA and internal governance policies, or even your own internal compliance setup. It also needs to be reliable and robust. You know, production AI must deliver consistent results under all expected conditions.

We are adding new applications where new models, it needs to be consistent, needs to be stale, also needs to be, uh, able to, you know, do graceful error handling, retries, fallbacks, you know, um, behave without, uh, trying to void. You know, downtime also needs to be scalable and performing. It must be able to handle peak loads with low latency and predictable throughput.

It needs to be designed for horizontal scaling, you know, being able to use very well known techniques like caching, optimize inference, the possibility to be able to really handle the potential load of, you know, real work, real world workloads. Um, we mentioned about compliant. You know, we, we mentioned about having this, being able to handle all the information regarding your industry or your overall, um, policies, and also needs to be continuously improved.

We need to be able to have, uh, close, uh, feedback loops. We capture user intention in interactions, performance data. Also, we need to be able to have, you know, regular training, AB testing, model tuning, keep up, or, uh, keeping accuracy high all again with, you know, the possibility to continue, uh, keeping the human, uh, in the loop, being able to provide really, uh, resource.

Now we are talking about this production ready, and, and why do we mean by that when talking about, you know, DevOps? And here it's important because we are moving away from this experimentation phase where we are, um, just, you know, playing around, building, uh, prototypes and mocks that we want to move into the concept that we call the AI innovation factory. So there are many difficult, uh, difficulties organizations face when creating this systematic and repeatable process for developing and deploying AI at scale, what we call the innovation factor.

Remember, the innovation factor is this structural environment that consistently generates new AI solutions, models, applications, all designed to drive business value. They're going from artisan mode where we have a small team just gathering around trying to play and discover how this works into full mode production, where we have everything automated, applying DevOps principles. Essentially, it's about setting up a sustainable, efficient system that can keep producing impactful AI innovations over time in a way that is, as we were saying, again, you know, secure, compliant, scalable, and so on.

And this doesn't come, you know, without challenges. Like, for example, this graphic shows results from the 2024 Garner generative AI planning survey, which ask organizations about their biggest roadblocks to adopt, you know, g AI at scale, basically production. So there are a couple of, um, challenges that they highlighted.

Um, there were ones that went around data quality. You many GNI projects tell you, consistent, incomplete or unstructured data, poor quality data. Uh, second one was around privacy and security.

There are several concerns around explosive, exposing, sensitive or regulated data to LLMs to MCP servers, to agents in general. And organizations are trying to grab on how to sanitize inputs as well as outputs, you know, control the access to models and comply with data protection regulations. PII, you name it.

But again, the other, um, challenge on, on the top here is also regarding cost. Uh, you know, both, you know, compute infrastructure, if you're running your own model locally, UP coast as well as small licensee tokens can be expensive, especially at the scale. And if we are not, you know, having observability and tracing and tracking of where, you know, all those tokens are better spend.

So this challenge becomes more pronounced when teams overbuilt prototypes that aren't optimized for production as well. If I'm just running three requests, three prompts, I might be, you know, overusing some tokens. But when suddenly those 3, 5, 10 prompts become 10,000 prompts or 10,000 users using the same, uh, solution and hitting the mobile every single time consuming tokens, it can be costly.

And we, we don't have observability or, you know, at least an idea of where those tokens are exposed, you know, and, and consumed. There can be complicated, but the hyper wrong. G AI is real.

But organizations are hitting practical roadblocks, especially around data governance operation, operational efficiency. So addressing these challenges early is critical to unlocking sustainable gain AI initiatives. And as we were seeing, some of the things are just, you know, in track on how we're using.

But also there are some things that are common pitfalls we see when organizations rush to implement AI without proper architecture, without proper data pipelines, without involving the, uh, the mature, the SecOps teams, uh, for architecture governance and so on. Think about, you know, one model to rule the model approach. You know, when teams often rely on a single large model for everything, customer support, document optimization, co-generation, now this leads usually to poor performance, high cost, and lack of specialization.

Instead, we need to, you know, fit for proper models and agent like components that are able to get the most of each one of the capabilities. There's also, you know, no prompt or input ization ignoring input. Filtering opens the door to prompt injection bi leakage, security risks, sanitation.

It's key to save compliant G and AI deployments. So think about, you know, avoiding those kind of things. Or also add model deployment.

You know, models push directly to production without any kind of versioning testing or rollout strategies. 'cause suddenly we are just, you know, rushing to push everything to production, and we don't have really the pipeline to be able to deploy them safely. Being able to implement rollbacks to keep track of the, of those deployments we need CI CD style.

I also, for ai, it's another component of our infrastructure or another component of our architecture that needs to also be, versioned needs to be able to be managed and using practices. Uh, as part of CI CD GitHubs, it becomes critical. There's also, you know, data sum fittings, AI models, uh, fitting models on a structure, messy or redundant or redundant data impacts the quality.

You know, when you get garbage in, you will get garbage out. So you need to have curated government data sets if you're building or training or, you know, high tuning your own models. And that also requires specific kind of, you know, data pipelines that are being, um, um, reviewed, um, and, and mature.

Another kind of antipater we seen it's siloed teams and tooling, you know, ML engineer, data scientists, DevOps, and security, all working on isolation. DevOps teams, not talking with data scientists that suddenly are using, you know, um, continuing images or dependencies that have Thomas CCBs and then becoming, uh, you know, security nightmares and these kind of things. Slow down delivery and increases the risk, as I was saying.

'cause they're not talking to each other. Suddenly they're throwing things into production and it can, you know, increase only the surface of potential attacks when opening without, you know, the, uh, all the knowledge that the, uh, current teams handling applications can, um, implement. And for this not having an observability ledger, you know, without logging metrics tracing, it's really impossible to monitor the bulk AI behavior.

Serviceability, it's critical for the technique, drift hallucinations and obviously failures. And not only on the consumption, but only on availability of the infrastructure. As well as things like, you know, static prop engineering, you know, working prompts once hard coding them, uh, you know, knowing model evolution, changing inputs and, you know, reusing a, a ton of pre confi preconfigured, um, prompts, know prompts should be versioned dynamic, continuously improved based on feedback and context, avoiding these anti-patterns.

It's essential if you want to move beyond AI hype and into something that is really resilient, responsible, and production ready. And one of the points that we have seen, uh, is, is this, as organizations embed AI across systems, APIs become this connecting tissue that brings everything together. A strong API strategy is essential because, well, you know, it helps you expose AI capabilities.

All this kind of, um, uh, of endpoints are reusable, well documented. And it's basically the, uh, way to being able to access remote services across distributed, um, implementations. It allows teams to orchestrate AI models with other services.

Uh, thinking about authentication databases, business logic, and we will go on the details on how agents are working on that now and provides the, uh, foundation for governance monitoring version control of AI workflows. If you're not thinking about the API first mindset, when, you know, building your AI infrastructure architecture, you risk building AI solution in silos, it's gonna be hard to scale, test, or secure. You might lose the ability to abstract and swap models without breaking the client experience.

If you're embedding everything on the same solution models, serving framework, et cetera, it's gonna be complicated to really, you know, generate these applications that are distributed that can scale, and your applications will make it, you know, harder to integrate gen AI into existing ones, uh, applications as well as developer workflows. Now, APIs are also, you know, helping you enable modularity and flexibility so your teams can then experiment with different models behind the same interface. And we will see a little bit of how we can help you with that.

You can also being able to apply usage policies and quotas, you know, track usage and performance by your, um, endpoint, your model, your games, your applications, your final user, um, AI capabilities that, you know, are only as powerful as your ability to deliver them. And APIs are the deliver mechanism of today's, um, ai. If you really want to work on something that's scalable, secure, and really deliver something that's like compostable.

So we have seen what not to do, what we suggest, that will give you more challenges. So how does it looks like the things that we have seen that is make, uh, that helps organizations be more successful. So we have these patterns that, you know, represent these buildings building blocks of scalable, secure, and maintainable AI infrastructure.

So let's, let's go, uh, one by one. So the first thing is, as we were saying, thinking about model as a service, you know, treat your model, um, endpoint like APIs that are self-contained, scalable, accessible via standard interfaces. Think about microservices, where you are able to then scale them, distribute them, have them available everywhere.

This will help you support modular deployment, versioning, abstraction across model types. You're using open source, proprietary, custom made, uh, fine tuned ones on the cloud or on-prem. Think about this NCE of your architecture that will help you enforce control.

Think about LLM gateways, MCP gateways, AI gateways, these competencies between clients and models to handle things like routing, retries, catching rate, limiting observability. And these kind of components enables, you know, vendor flexibility, research, mentoring, policy enforcement, and route time and moves things away from the developer responsibility on the application side into a central point of governance, where you're then able to, um, configure and manage everything centralized. So think about this policy aware data governance layer, you know, government who can access what data, what model, and how it is being used.

So it's critical for privacy, compliance, and trust. You know, you can apply rules based on user roles, you know, sensitivity levels, data versus resiliency, uh, requirements where you can just block certain calls depending on the context of the, of the same prompt of the user call or where your, um, your information is flowing to. You certainly don't want to, you know, send just, you know, PI information directly to a cloud vendor that might be logging this information into their own systems.

And then suddenly you are, you know, um, out of compliance add model, app model monitoring and feedback looks. So you can track, you know, response quality, latency errors in real time, and then being alerted and take actions based on that. It will help you, you know, enable from turning, uh, prompt tuning, uh, retraining, you know, rapid rollback phase, uh, based on user feedback and for, and performance trends.

If you've seen that your latency is increasing, you need to be able to take actions to, you know, keep with the developer and the user experience. Another thing that it's being widely used and adopted as traditional models and chatbots use, as well as, uh, now with, um, uh, with agents, it's the use of vector databases and the rack path and the augmented generation that's combining semantic search with LMS to enhance context without the training. So you can use foundational models to be able to then enrich the information on, on, on every prompt.

So these retrieval generation ensures that you have grounded, accurate outputs, and especially for domain specific use cases or for information that it's, you know, proprietary as part of your organization. And again, think about these data mesh architecture, the centralized data ownership across different domains, while standardizing interfaces for AI consumption, being able to provide better quality to the, uh, to your models, to your, uh, your rack, uh, databases as well as your agents. You know, this will help certainly promote discoverability of data quality governance, really at scale.

Now, these patterns aren't just, you know, technical choices. There are strategic enablers that help you build AI systems that are flexible, you know, compliant, and again, production rate. Now this dig a little bit deeper on one of the, um, of the, uh, points and, and the companies that we were mentioning before, and where, uh, Kong really can help you with that.

Um, we're talking about AI gateways and, uh, according to Garner, by 2028, 70% of the organization will be building these multi LM applications that will require this kind of components. Hey, gateways are quickly becoming a critical path of GDG and AI stack. As organizations adopt multiple Ls, like, you know, open ai, cloud, Germany, mistral, they face growing complexity.

Like which model is best for a specific task? Um, how do you balance cost latency and output quality? How do you maintain governance and observability?

An AI gateway abstracts these concerns? And by doing dynamically routing requests based on policies like, you know, cost, performance risk, uh, help you of centralizing logging rate, limiting access, access, mobile failover, versioning testing. Know the shift is clear.

We just need, um, APIs to revolutionize microservices. So AI gateways will be foundational on a scalable volution engine AI deployments. Now we need to move away from just the gateway to this concept of the innovation factor, the AI gateway and soft, and the AI starting point provides, you know, with these basic pieces, routing, extraction, governance across models, but to drive ongoing value and differentiation, organization must evolve beyond that to what we call the AI innovation factory.

Again, the whole platform, the whole infrastructure. So an API platform connects everything models, data, business logic, events, into a cohesive, discoverable, and secure interfaces. So that's the idea, trying to standardize abstract, and add everything on a, uh, full view.

So the API platforms that is deliver ledger for your AI innovation factory, turning gen AI into visible, uh, reliable services. What you will be looking in this kind of, uh, of platforms, well, you will be, uh, searching for it to be compostable, uh, structure stack, individual components, you know, can be mixed and match. Um, it needs to be human-centric.

So designed for real users. Embedded AI since into existing workflows, uh, needs to be governed and observable, implementing policy informants at every Laker data access model, usage of filtering. So you can then meet privacy, security and compliance requirements.

And again, it needs to be rack aware, rack first and context aware where you can prioritize things like, uh, retrieval of many generation per LLM with vector databases to run responses. These elements of judge different patterns that help you compliment the, uh, this idea. Because, you know, while AI gateways play a critical role in managing and scaling gene AI workloads, they're just one piece of the puzzle.

The gateway may be the entry point, but true success comes from the entire ecosystem that surrounds and supports it. It's a foundation, but real success comes from the end-to-end AI and a p infrastructure route. And this is, you know, without considering this race of agentic AI workloads, welcome to the Gentech era.

Agents are just responding. They're acting, but to act, they need the infrastructure that connects them securely to real world. So this is why, you know, most agents today are just gimmicks.

You know, they're not really wired with APIs or with your own systems. You know, they can check calendars or sending emails. Uh, now they're getting more and more UpToDate.

But, um, the, the, the ones that you can see is just, you know, the weather applications. Something's, um, very smaller and, you know, AI, gen, AI agencies more than just models, because you might be thinking about, oh, I have a very good model. The model's gonna be calling, you know, uh, additional, uh, companies, but it's not like that.

And we have, um, we have z this in the past where there's, uh, when you are building a genetic AI gen in general, that they have a, you know, set of needs and, and, and, and, and requirements. You might be thinking about, you know, an l LM is enough, but it's not and enough. The, uh, l LM is just the base, the, the foundational layer.

But, you know, real agent needs, you know, prompt orchestration, augmented generation tool calling, authorization, authentication, orchestrations of all these steps. And this is where infrastructure really makes or breaks agents, they have been evolving. Most AI bots are sticks are still stuck on level one, you know, chat only.

That's what you most, most of the time see rather out in production. But the real value of these agents will become when they are merged into something that really, you know, moves into action, calling APIs, making decisions, doing things. But the important key thing here is that they need to do it securely.

So yes, RAG will give you more context. Tool calling will be doing the chat bot. When adding human into the loop, you are becoming more mature, more able into, you know, this fully target of, uh, fully autonom.

So again, AI agency more than just model. If the model is the brain, you know, APIs are the hands of your agent. Without tools, the agent simply connect and attack.

That's why your infrastructure must expose and secure those tools also at scale. So again, going back to the infrastructure, how we present things to models and to agents. Now without secure APIs, agents are just, you know, chatting interest with them.

They're really productive employees that will think of, we'll be talking about this more. Now, you might have heard about this, and it's, it was, it was all over the news. Know with agents ask, you know, the stakes go up.

Prop injection can delete databases. Uh, we must treat AI a actions with the same rigor as any privileged access control flow. This is why secure execution matters.

Jection is not funny in production. It's a breach. You know, red learned the hard way, and they were not even hacked.

They were just, you know, allowing the agent to, you know, to roll basically. And well, you know, what had happened, and this is where Kong can help you. This is where Kong Shines.

Kong sits between your models and your APIs, between your models and your agents, between your agents and your MCT servers, between your MCT servers and your uh, resources. LP you route enforce policies. You know, secure tools provide observability across every single step in this chain.

Um, you can expose your APIs as scalable resources, you know, let agents interact safely, quickly, and consistently. Um, Kong, uh, gateway helps you also being able to reroute and implement, you know, the multi LLM capabilities, um, through a single universal API Your applications can use a single SDK. And then being able to access the full, um, ecosystem of, uh, model vendors, providers that offer you different alternatives and con can, uh, help you to do that connection, um, without having to impact your applications.

Um, you also will need observability. And here's where Con can provide you with all these different style. Being able to go on the details, on debugging each one of your policies, one of the steps within the gateway, as well as the latency, the time that you will take.

And then being able to, um, serve directly on the, uh, uh, API platform or being able to export to your, uh, trusted, um, um, vendors for observability. You know, what was called, how it responded, and also what cost. It's also adding semantic capabilities.

Things like semantic caching will help you out to reduce the cost and avoiding having to hit the, uh, uh, model every single time. There's somebody asking, you know, what's the capital of France? Uh, if you have three requests, okay, but you have, again, 3000 requests doing the exact same thing.

You want to be able to improve the latency or reduced cost. Um, you can also do, um, for enforcing content moderation, um, doing semantic routing, taking decision based on what kind of, uh, problem you are getting to what model will be best suited to answer that. All built into calm.

You can have policies and capabilities like, you know, PI ization, where you have, um, the, uh, the gateway being able to sanitize inputs and outputs before or after the LLM system. But this is the way to really build enterprise grade trust when you are able to, you know, remove all these potential issues before they're, you know, getting out, uh, of your, um, of your organization. We were mentioning about the know, the LLM flexibility, you know, when you want to use the best model for the job, not always go with, you know, U PT four, you can route dynamically with Kong.

Uh, not all tasks needs the same thing. So you are able to then decide and design your, um, infrastructure to be able to then route to different models, uh, without having to affect the current application. You still use the same SDK, open AI called Gemini, and you can reroute, you can try then things like, um, 80 testing, you can just, you know, have failover being able to, uh, do load balancing and doing retries in case that you are either having your vendor or modeled down as it's happening right now with, uh, some of the vendors.

And then fall back to continue, you know, delivering experience to your users. Having a, um, an, uh, API platform for AI also allows you to implement things centrally like, you know, automatic, um, retrieval, augmented generation injection. You know, we were mentioned this is a powerful pattern, but sometimes it's hard to do the right.

And, you know, Kong enables this kind of context injection without bordering developers. Instead of having to make every single application implement the rack pattern, you can expose different endpoints where the virtual element generation is being applied by the fold, well, other pipelines. So you are, applications can get, its, you know, right away, you are adding more applications, will, you know, every, uh, instantly all these, uh, uh, configurations.

So it's a better way to handle it in central place. And we were talking about agents, right? So tool colleague without auther authentication or, you know, access control.

These dangers, as you know, replication did, did learn consu only the right agents can access the right tools on behalf of the right user. This is where you can then use, uh, the, uh, the, uh, central points of, uh, enforcement or, um, agent to m CCP call from MCP to applications or, um, agents to, to LLMs where we can then make sure that they are, uh, fulfilling the authentication, uh, and, um, and security mechanisms that you, that you require. Now, going a little bit deep on the MCP gateway, you want to have, you know, a gateway governance for MCP.

You can secure and govern your return on MCP servers, you know, APIs that enables structure action, being able to then have the gateway, being able to control how your agents are, you know, uh, discovering these servers, being able to access those servers, being to implement, you know, the developer productivity. And when you're talking about MCP servers and implement the MCP protocol that we are looking here, we also want to be able to leverage what you have already been building with your AI strategy and your build and your API resources and current, uh, assets, uh, in a way that is gonna be easier to then expose them as, um, as, uh, as part of your agent ecosystem, uh, making and turning every API into an entity server that is being then served agents and being exposed. If you already have, you know, develop, you know, series of APIs that are already available in your ecosystem already, you know, we will tracked production ready, then you're able to then just, you know, build this interface that will help your agents being able to then access them in your natural language, um, way.

So you will be able to turn these APIs into agent ready endpoints automatically. We can. That's where we are heading.

You can have then, you know, certainly, um, uh, production ready UI, where you're able to be enabled to see, discover, find out the models that you have, the agents that are available, the, uh, the, uh, uh, the MCP servers that are, uh, in your ecosystem, who's consuming, uh, what, what are the, uh, the, the, the biggest consumers, the expand that you're getting on, on cost and, and things like that. And the important thing here is that we have this, you know, unified infrastructure view. This is what modern infrastructure looks like, unified across APIs, events, LLMs, MCP servers, microservices, Kubernetes.

This is where Kong's AI infrastructure can help. You know, Hong's platform delivers the performance of observability and the enterprise SLA to do this today through this unified platform with alsos of features and one built on one of the most performing gateways over there in the market with everything. Uh, really, uh, I know this image, it's a little bit, you know, uh, crowded.

But this is the kind of things you need to, to be thinking of when, uh, you know, uh, when, uh, designing your infrastructure. All these places where you need to have LLM providers, AI governance from management, MCP, traffic, you know, um, uh, server generation, API calling tool calling authentication of agents. It's really a good landscape of what you will need to think about, implement, built on top of that.

Have these enterprise ready, uh, infrastructure pipelines that will help you out, deploy applications that are really, you know, what your business is expecting with the securities scalability. And compliance Con, uh, has this connect platform that is providing this unified API platform where you can build faster, cover your service easily with things like a developer portal, um, service catalog, where you can find not only from the external presentation of, um, models, agents and APIs, but also from internal, you know, reference for you to know exactly which, uh, grade your APIs or your models are being exposed with. Um, and regarding the security compliance.

You know, you're able to grade them, know information about where they're, uh, what's the team that is supporting them, full observability here and going the, the full range from building, um, APIs, AI models, agents, running them, secure leads, covering them, governing, and all the different kind of, uh, of, uh, of, uh, assets that your organization has. Um, APIs, ai, uh, events, uh, everything on a single platform, being able to have, uh, unified, uh, identities being able to consume. Now we're getting close to the end, and we want to, um, leave you with all this, uh, content.

Um, obviously the idea is that if you really want agility, you really need to have this model of kind of stacks and well Kong enables you to solve these models or tools, you know, agents, MCP servers without breaking your infrastructure, really leveraging what you have, the AI landscape just evolving so fast. Everything, there's, every single time there's something new that we need to be aware of, everything we discover, that there's gaps that we need to cover. If you do it yourself, you might be burning, uh, uh, burning your, uh, teams and your resources, uh, when you can, you know, support this kind of, uh, of, uh, of components with, uh, with things that have already been proven.

So you can avoid, obviously locking, increase your agility. Now, um, what you should take away, well, uh, there are three main things. You know, if you want really smarter, cheaper ai, you know, compare maybe smaller models with powerful, uh, APIs, um, Kong can help you make your stack AI ready, securely and scalable.

We went over more, different, uh, aspects of, uh, production ready, but the ones that we can help you, uh, regarding things like, you know, uh, security observability, uh, governance, um, uh, scalability is where, uh, you can certainly take a look at and more if you are building more complex things like agents. So, uh, take a look at this con, kind of provide you these foundational pieces to really scale your AI secure. Now, we have covered a lot of content.

Uh, we know this is just the, you know, uh, a high level, uh, overview. But if you want to continue, you know, con has this special offer for your audience, for our audience today. So for those that are interested in learning more, you know, about how Kong can support your organization with, you know, building this agenda ready infrastructure, you can get a free custom pair of Nikes.

So you can, uh, you know, uh, uh, scan the QR code, uh, you can see here to book, uh, a call, uh, with one of the, uh, our team members. And then, you know, the second call can help you all design your, your customized sneakers. Um, you can, uh, once you, you take the call, we will send you out this, uh, in thanks for your discussion on and for your time.

So we really hope to hear from you and well, uh, don't forget to learn more in our virtual booth. We have, uh, plenty more of content. com.

Uh, you can contact us, we're gonna be in the chat. We're gonna be, uh, able to, um, uh, to support you with, um, any more information. Um, but again, um, we'll be able check the time that you have with us, and thank you very much.

See you next time. Hey, everyone, are there some cracks appearing in the feed of the Kubernetes Megalith? You're watching Textron Gang.

Hey everyone, it's a Shimel and welcome to our Friday edition of Textron Gang. We're actually, as you know, we record Textron Gang the day before. So it's actually Thursday, it's a little quiet here in the morning 'cause the place isn't open yet.

But we're on the floor here at CubeCon and we're wrapping up our CubeCon coverage on Thursday. Uh, you'll be able to start seeing a, well, if you didn't catch it live streamed, you'll see it next week, probably on demand. Um, but for today's gang, I'm really introduced pumped to introduce our latest guest here on the gang.

He's becoming an official gang member. He's someone I've known in the tech community for a long time, and I'll leave it at that. My friend Chris Short.

Hey Chris, man, welcome. Thank you. Appreciate chat.

We don't usually do this, but me, we're on camera or, or on the side here, and it's your first time on. Give people a little bit of you of who, who you are. And So I'm Chris Short.

I've been in tech pretty much my entire career. Uh, joined the Kubernetes project in 2017 and have been a member of the community ever since. Uh, I'm currently the co-lead of Kubernetes contributor Communications, and I also run the open source program office at a startup called CIQ.

Excellent. Thank you, Chris. Thank you.

And of course, joining Chris Simmy is the dean, Mike Ard. He's been out doing interviews most of the night, he says, but you could guess. Um, but he's here today and that's what's important.

Mike, what do we got for today? Well, we're leading off with this article that you wrote over on Cloud Native now about the, uh, cracks maybe in the Kubernetes monolith community. And the idea here is that, um, folks are getting a little frustrated maybe on one hand.

And then the other side of it is that the foundation themselves have become, like most nonprofits, they wind up kind of servicing their own internal political issues more than they actually wind drive driving the innovation maybe. Yeah. So is it monolith or mega lift?

I think it could be, you know, those might be synonyms if you would find Ary. Okay. Maybe, we'll, We'll go with it.

But, you know, here, here's, here's the thing though, and it just, it's why I love tech, right? Last year, the year before that we were writing about there is nothing on the horizon that's gonna stop this locomotive. Oh yeah.

Right. Maybe, yeah, Kubernetes mm-hmm. It's Pax Kubernetes, right?

That that was the world we were looking at. Mm-hmm. There's nothing, you know, to, to dim the, the bright lights.

And I think, 'cause we were looking for something that would replace Kubernetes, but instead, I think what we have found is something that Kubernetes may not be able to sort of internalize, and that's ai. Mm-hmm. Right?

We could use AI to try to make Kubernetes better, but can we use Kubernetes to make AI better? And that fundamentally is gonna be the question of whether Kubernetes, we, we stay in a Pax Kubernetes kinda world. Mm-hmm.

Or do we go into a dark ages after Rome fell or something like that of that. But, um, but on top of that though, there's some metrics that, that back it up. Chris and I were talking about it earlier.

Micah and I spoke about it yesterday. Look, they're crowing about 15 point something million members in the cloud native community. Yeah.

That's fantastic. But when you look underneath that, this show has been flat to even slightly down. I saw there was only about 9,000 and something people here.

Yeah. Nine to 10,000. So it's been flat to slightly down.

60% of the people here are here for the first time. And that also is a, a kind of a trend steady statistic we've seen, at least in the us I, you Europe's still a little bigger and I don't know about the 60% number. Mm-hmm.

Um, Alright, so from the community side of things, right? If I look at the Kubernetes community and the, the AI things that are happening in the community itself, we see some work groups popping up, AI conformance a number of other AI working groups that the community as said, are needed now. And that's because Kubernetes has kind of become the platform to do AI at scale.

But at the same time, yes. Can we make AI better through Kubernetes? That's where we're moving towards.

Mm-hmm. Now it's a large community and consensus takes time sometimes. Yeah.

So we are seeing some things progress very quickly. We saw an announcement yesterday with my friend Mario Valant, uh, talking about what the AI Conformance group has done. And I'm excited about that.

But at the same time, I feel like the Linux OS and Kubernetes have a similar journey. We're getting to the point now where you no longer need a Kubernetes specialist on your team. The skills are becoming more ubiquitous.

Right. People understand containers like never before. People understand, uh, pods, deployments, CDs, the whole gamut of ways you can instrument something in Kubernetes.

So with, with that kind of realization, you know, red Hat for example, you know, a lot of their revenue was from support for rl Sure. Long ago. Now it's support for OpenShift.

Yeah. Which is interesting to me. Um, and we're seeing a lot more Kubernetes usage in general across industry, which is great to see.

So at the same like, yes, there are some gaps showing. I think those are more so socioeconomic than actual like technology concern. I Don't know that I hear the tension along two paths.

One is AI workloads are stateful and they need to scale really high. Mm-hmm. And Kubernetes was never originally designed for that.

And we run databases today on Kubernetes, but we don't run 'em at that level of scale. Right. So people are saying, do we need to fix some of that engine and what goes into that work?

Yes. And then the second thing that people are talking about is, you know, Kubernetes is not the only game in town. And you go talk to the data science community and they like a thing called S slm.

It's a job schedule that they're using as an alternative to Kubernetes orchestration because it's more accessible. And to them this is makes more sense. And they look at the Kubernetes thing and it's like, you know, a bunch of it guys telling them that there's this greatest thing over here to use data scientist looks over at Kubernetes and goes, Doesn't hold all running jobs.

Well, yeah, exactly. There is that. Yeah.

There is that. Um, I know from my company's perspective where I work now, yes, slum is a big deal, but underneath that there's this technology called werewolf open source project widely used. It's stateless or state full cluster management at scale.

It was designed for the HPC world. Nice. But is suddenly really relevant in the AI era.

Yeah, There you go. You herded Euros extra gag that's Slf comes from the same HPC community. Right?

Right, right. So like we're seeing the werewolf project itself starting to adopt slum standards and everything else so that you can actually actually drive those workloads in a more open manner. You know?

But let me, let me not call bs, but let me just bring out a statistics that I, or metrics that I've heard around Kubernetes adoption. Mm-hmm. Yes.

In greenfield application deployments, Kubernetes and that whole cloud native stack mm-hmm. Containerized application microservices probably represents 75, 80, 80 5% of, of Greenfield. Right.

You know, critical mass. Sure. It is the standard, it's the compute stack.

Bingo. But yet when we look at the entirety of applications that are running in the world mm-hmm. Not just in cloud Kubernetes, cloud native, you know, microservice architecture, they represent 15% of the existing applications.

And I, and that stat stayed steady now for a while. Mm-hmm. We have made no inroads into transforming, modernizing, I don't really care what you want to call it.

Right. This existing base, which is still larger mm-hmm. Than, you know, all of the new stuff we come out with all of this time.

Well, I think what's happened is people have realized new workloads Yes. Kubernetes, you want that ability to scale fast. You want that ability to just interact with APIs, but your previous legacy workloads aren't necessarily designed to work like that either.

And takes because their legacy Right. Rearchitect because their legacy rearchitecting isn't exactly gonna be high on the priority list. I think this Legacy to me means money maker.

Right. Absolutely. Right.

I don't mess with, don't mess With that broke legacy means to me. Yeah. And it ain't broke.

Don't. Right. But I do think That that mentality is starting to change finally.

I think You think so? Yeah. I mean, like we, we went through the DevOps era, right.

I'm not saying DevOps is dead or anything like that, but we, you know, we've evolved sre. Why you're not saying that Chris. Right.

We still have to go back to those DevOps principles. Yes. Because that's the problem.

We need to be able to scale these legacy applications, but we're still managing them with proprietary network gear. We're not using open source load balancers or anything for that matter on those workloads, which is putting them at a disadvantage. 'cause open source is kind of the concrete foundation of a lot of these workloads.

I don't think anybody really knows honestly how those workloads are actually constructed. That That's a problem too. So they're hoping maybe these AI tools will help with that.

But if I don't know how the thing is constructed, I can't carve off a microservice off this thing and start slicing it up. Uh, the only way I can get there is, you know, I gotta call a consulting firm and then they show up with, you know, 50 kids in a bus who move in for a year and a half. Agreed.

Yeah. Agreed. Yeah.

And not cheap, but let, let me, let me call out an elephant in the room though. Sure. They call this show Cube Con, the official name of course was Cloud Native Con.

Right. But I think they stopped trying to correct people a few years ago. It is just Cube Con, but Cube Con has become a binary star system.

Mm. And right over there is open telemetry land or whatever they call it. Right, right.

And when you take open tele o Hotel mm-hmm. And you take Prometheus and you take some of these other observability projects that are in CNCF, you, you, you know, it kind of reminds me of the Arthur Clark 2001 where Jupiter becomes a star. Oh yeah.

Right? Yeah. Yeah.

You have a new star in this system and it, and it's, it rivals. Mm-hmm. Yeah.

Gold star. And is CNC is this town big enough to two star for two stars? Yes.

I think having two stars in the same foundation is a good thing, right? Mm-hmm. Um, looking at it from the c ncf F perspectives, you know, if I put that hat on, I see it as growth.

Mm-hmm. Externally, it's not creating confusion, which I think is Yeah. Some of the problems with a, you know, two star system, right?

Like, which one do I choose? No, no, no one compliments the other. So that's kind of set up well.

Mm-hmm. But the thing that I've noticed is I'm starting to see things like Prometheus used in non-cloud native contexts. So I'm starting to, you know, see exporters, um, literally.

So is that a bad thing? That's not a bad thing. That's called maturity.

Yep. And if, if someone's not using Kubernetes, but they are using some of the open and underlying components of Kubernetes, that's still a win in my book. Right?

Kubernetes is not the destination, it's a part of the journey. So if we're looking at higher level abstractions Yeah. Hotel fits right in.

And it makes a lot of sense to start using those things in non-cloud native workloads because they're more efficient, right? Like we've driven the efficiency into the underlying applications. Yeah.

So is there something to be done to jumpstart innovation a little bit in the Kubernetes TOC? Or is this just the nature of democracy as a messy system and it is what it is? It, well, I can't comment on the TOC component of or creating more innovation, but what we've seen is this very sharp uptick in AI investment.

And that's kind of pulling some of those engineers and folks into those projects and not necessarily towards Kubernetes. Yeah. So how do we make sure that the AI people are in the boat with us is kind of how I'm seeing 2026 play out.

Okay. More so than, um, Or Kubernetes gets in the AI boat both, Or, you know, very little cross Cross, uh, pollination. Yes.

Like, let's work together, let's push these things forward in a more collaborative manner than our normal company based silos, which is good. Let's, let's last these boats together and have a party. Yeah.

Let's take a yacht out all That. You Know, You know what, we, we've got to end this segment, but let me, I'll end it with this though. Let's keep an eye on Amsterdam Yeah.

And see what trends continue or what we can spot from there. So you'll have to wait until March on that. But, uh, we're gonna be right back here in text Drunk gang.

And we're gonna talk, what are we talking about next? Mike? We're Talking about K serve and open source project.

Hey, native, now moved into the CNCF. Very cool. You're watching Textron Gang.

You've earned it. The spotlight, the responsibility, the weight of teams, companies, and entire industries fall on your shoulders. Lives depend on your decisions.

Your home life included that work you are protected physically and digitally. Nothing gets through your team without a fight. But in a globally connected world, everyone sees you, including those who mean to cause you and your organization harm.

And now home your sanctuary attackers see an opportunity. Your digital front door is wide open. And what compromises your home can breach your boardroom.

Because the devil's greatest trick isn't targeting your workplace firewall. It's convincing you that your personal life isn't at risk. Black clerk, digital executive protection, defending the new attack surface your personal life.

Hey folks, we're back. And as promised, we're gonna talk a little bit about Kerv, a project that was in a different Linux Foundation and is now moving over to the CNCF. Kerv is a distributed inference server is one way of thinking about that.

And a lot of more of the AI workloads are becoming more distributed. So that's a good thing. I talked to the, at least one of the maintainers and you know, basically they said they just wanted to be hanging out with the cool kids because, you know, this is where they're gonna do a lot of the integration work going forward.

And the other project feels like maybe it's starting to be more of a data science kind of project team. And it's an AI foundation within the Linux Foundation. And those guys are all focused on data and training and the inference stuff is just hardware stuff that they're not interested in.

So maybe it is better to have this over here, but you're closer to this. What's your tail? Uh, I think, so we've gone through some serious changes in the past, I think five years when it comes to open source or just community events in general.

So CubeCon is the one that's attracting a lot of attention. We talked about that in the last segment, but the, the everyone in the same pool model is kind of working, right? Like we've seen the CNCF landscape drive valuation in companies and now we're seeing it draw in more tooling, which is pretty good thing.

I think C CCF landscape thing gives me a headache. I like, Yes, they've made some improvements in the past year to make it less headache inducing, but I remember when they used to print those things. Oh, and that required a max.

You needed I was just So you needed a special printer to do those. Yeah. What they call those things.

I slipped it out in sheets. Um, but, but here's the thing. So Mike, I think the proper nomenclature, it's not another Linux Foundation, it's a daughter foundation of the Linux, Linux Foundation.

Right. Same way. Theoretically.

CNCF. Mm-hmm. Um, and I'm just not sure how cool that is.

Is this a bunch of baby birds in the nest and we just saw a baby bird eat his brother or sister? Hard to say, but I think, um, ser will probably get starved for engineering resources in, in that other daughter foundation, whereas it probably will be able to leverage up more on the core Kubernetes work here. I think, I hope I crossed my fingers.

But, um, whether it's case server or not, we gotta figure out a way to make those AI workloads more distributed. 'cause we can't just keep scaling them up, we gotta scale 'em out. We gotta, and that's kind of the big challenge.

Yeah. I don't disagree with that. Mm-hmm.

I, I'm just saying at, at, at a higher level, what's healthier for the Linux Foundation. Ah, okay. Right.

That regard. I would think, you know, fewer foundations, more concentrated work efforts is a good thing, I would think. Right?

Yeah. Because I've seen foundations come and go, so Why keep giving up so many foundations? Matt?

Gotta ask Jim. Yeah. You gotta ask Jim on that one.

But I think that's more so, uh, the business model of the Linux Foundation than it is the actual, like right. Industry, if that makes sense. Yeah.

I mean there are, I don't know how many daughter foundations there are In, I thought there was like 40 something. Is that all? Maybe more than, I know the last I looked it was 42, something like that.

Okay, well that makes sense. 42. Yep.

But The, the thing I think with K native serving is the K, right? Kubernetes is the underlying thing, right? So it belongs here.

It should, it belongs here. And maybe it's a case where it was just put in the wrong place to begin with. It could have, And now It's correction.

Well, 'cause here's the thing, right? Like Kubernetes and AI both kind of blew up at the same time, for lack of a better term. Yeah.

At least Nvidia, right? Like when I think about AI workloads, I think 70% of 'em are NVIDIA's. Sure.

So Nvidia is scaling vertically, not necessarily horizontally right now, but they're chips. They're saying, what was it? The Jensen comment was like, a hundred x performance or something like that on their newest GPUs.

Like that's, that's huge. But that is a rip and replace operation, not necessarily making the most of the hardware you have Plus, or you take the old ones and sell 'em to countries that can't buy the new ones? Well, no, actually you say that, but what those countries are actually doing is buying, putting all the data on hard drives, flying the engineers to a data center where they can churn through all that data and then bringing it back.

But, but that, and so that's very training specific. Mm-hmm. Hopefully as we move beyond training the inference and other stuff, they won't be able to do that.

But who knows, We need to move the processing of the data and the, and the inference closer to the network edge where the data's being created and consumed. 'cause otherwise, is This the pitch for Waso? There's A, there's a thing called latency.

There's a thing called latency that gets in the way. Yes. Latency is always gonna be an issue.

Right? Like I, I actually talked to a, uh, new contributor at the Kubernetes SIG meet and greet yesterday. And we were talking about the, the, the, the physics of networking Yeah.

Are going to start getting in our way. So how do we work around those physics? Every company supposedly has a solution to that.

But what we're seeing now is more mergers and acquisitions than new companies spinning up. So as things become more pressed against the physical limits of like atoms and, you know, light and things like that, we're going to start seeing some better use cases for older GPUs, for older inferencing systems. And then scaling them up for today's, you know, examinations and workloads is gonna be interesting.

I Laugh 'cause we're going full circle. So when I was young, somebody once drove into my head, you should always bring the compute to the data. Nothing good happens when you move the data.

So then we did the cloud and we move data into the cloud. Right. And now we're coming back full circling and saying, you know what?

We gotta bring the compute back out to the data. It's cyclic. This whole thing is cyclic.

I Got a point of order. I find that hard to believe someone told you that when you were young. Yeah.

Yeah. His definition of young, this is, This is, this is back when I was covering BDB elevens and Oh, vax. So you, that's nice.

You know, in, in the digital world, we had real computer science. We had real computer science. Not these guys today.

Alright, fair enough. You know what though? Welcome the project to the, to the CNCF.

Yes. Come on in. There's 200 other people swimming in this pool.

Mm-hmm. And, uh, I, I do think it, this may be a case more of correcting Yes. A misplacement prior than, than anything else.

Let's take a break here on the gang. We'll come back and we have our third topic today, which is Google. Have they lost the love for CNCF or Linux?

I don't know. You're watching Textron Gate. Discover Textron Group, the epicenter of tech innovation.

We are your go-to for reaching IT leaders and practitioners worldwide. Our secret impactful content that sparks awareness, engagement, and top quality leads with us. You'll access editorial websites, streaming videos, virtual events, custom content analyst research, and more.

Join our satisfied clients. Let's revolutionize your tech journey. Contact us today and tell your story to the world in the most powerful way with Techron Group.

Hey folks, we're back and I've been at the show all week and there's a lot of things happening here, but I will say that this thing that we're about to talk about is probably the coolest thing I saw since I got here. Oh. And, and basically what Google's announced is that they've created a sandbox for running your AI agents so that the AI agent can't go and, uh, wild and just start pulling stuff from all over the place.

It creates a little bit of a, a barrier around the actual agent, which is good news because right now our biggest problem with AI agents is everybody's deathly afraid that these things are voracious and we don't know what they're gonna do once. So I kind of like this idea. It's built on that, uh, g visor thing that Google came out with Yes.

With Avisor a few Few years ago while, yeah. Um, and so that to me was like, Hey, somebody's actually solving a problem we were all talking about. But Yeah.

But the issue then becomes, as I was walking away, I was like going, wow, Google's building a lot of cool stuff that they don't seem to be like giving back to the CNCF folks. They seem to be either using that because they think it adds value to their model of GKE or is it that they just got tired of the politics? Uh, well, yeah.

I can't speak to the politics of it all, but the, I think what we're seeing is we're starting to see folks use AI and they're having interesting challenges with it. Right? Like no Doubt.

I know I had a conversation with a couple engineers a few weeks ago where it was like, yeah, I was using AI to build this tool and ended up deleting my cluster 'cause it thought it could do something better, which failed. So I reverted back, da da da da. Luckily you were able to revert back.

You revert Thank you for using Git. Uh, I don't think I've ever said that before in my life, but the, the need for safety is very high right now. Right.

Because you can't think of all the edge cases to tell a, you know, put in a prompt, essentially. Say you're using the underlying infrastructure, don't change it. Right.

Make it work to your advantage as opposed to starting from scratch and saying, I have this problem, help me solve it. Go. And then all of a sudden everything gets broken in the process of you just saying go.

Right. I mean, Greg, it's an interesting, there's other ways to get after this too. I was talking about this with Chronosphere and they've kind of taken a, a graph and wrapped it around the LLM and so the LLM only sees what the graph tells it it can see Right.

And using the graph as a controlling function. Um, but, you know, I think that's a, a good way of making it a little more deterministic. 'cause I'm only limiting the amount of data that I can show this thing, but I still need guardrails and policies and yes, I still have to figure out whether or not I trust the output or not, but Right.

We seem to be starting to put together, you know, uh, a, a fabric or an ecosystem of things around the LLMs to kinda control the output better. Mm-hmm. And I think there's good reason for that.

Right. We're all a little worried about that use case. I was speaking of actually like happening in production kind of thing, where you blow away my legacy infrastructure.

'cause you think you can do better, but you don't. So when we look at systems built on safety, I know the CRA air traffic control network has had issues this week. We've all kind of been delayed getting here to the conference and everything.

But that's a system built on safety first. Yes. Where AI is a system built on innovation first.

So it's, it's very easy to innovate, but it's harder to make it safe and you have to build for safety first. Mm-hmm. Before we had this air traffic control system, though, it was an innovation first system.

Right? Right. Yeah.

You know, you gotta walk before your run and you can't make wine before it's time. Mm-hmm. Right?

So I think we're still in that barnstorming stage of ai, if you will, right. With wing walkers and everything. Yes, yes.

In that aviation. We'll eventually get there. But I guys, I think you're missing the story here.

The story isn't about the tech. The story is about why isn't Google donating open source projects to foundations anymore, or literally not like what they were. Mm-hmm.

And that might be a political issue, but I don't, I don't know if it's truly it's political. It's financial. It's, you know, it's not that Google's no longer not doing evil, it's also the Base, it's also the pace of innovation.

Right? I mean, if I got a committee right, It, there might be that it, they feel it slows down the innovation Consensus can take time. And that's neither here nor there.

That doesn't mean they can't release it later once they've got everything they need in row. But Didn't, didn't you hear authoritarianism is cool again. Come on.

Ah, geez. Yeah. There's that.

Um, Stay tuned for Shimmy says at two 30 I, I talk about this, but no, it, It's, I mean, it's not authoritarianism in that look, hey, Google put the resources into developing this. Mm-hmm. They're entitled to do what the heck they want with it, right?

Mm-hmm. But when you look at, you know, this whole thing built on, on Kubernetes that Google donated and the, and the input and influence it had on even just establishing the CNCF. Yeah.

Are we missing out on the next CNCF on the next community wide industry-wide revolution? Yeah. Because Google's keeping this close to the chest to their breast.

I think Google has learned some interesting lessons Yes. When it comes to a cloud b, cloud native and then c open source versus closed source versus something else. Right?

Yeah. Like they see an advantage in the middle in, you know, I remember a thing called Anthos from Google that is basically become Google autopilot, GKE autopilot. Mm-hmm.

And I think that serves their customers very well. And some of that you can do with open source, but the actual getting it all the kinks ironed out is proprietary. Right.

Which I think that's where folks are starting to find value is making resilience systems better and making them less error prone and more safe. Going back to our, you know, original count theory. Democracy is the most inefficient system of government, but it is, but the best, It is the best.

I will tell you, I'm starting to hate this phrase that comes out of the valley about go fast and break things. 'cause if you're not on, if you're the guy on the plane, that's not what you want to hear. No.

No one wants to hear go fast and break things on a boat, a plane already. Yeah. Yeah.

I agree. I agree. But you know, I, I think, so you, you, when you look at it from a historic right, you look at open source.

Mm-hmm. So you had your sort of, you know, your, your cathedral and bizarre phase with Richard Stallman, Dr. Richard Stallman and stuff like that.

Yeah. Where it was that unrealistic Marxism Yeah. Of, of, in a society of atheistic saints, right?

Mm-hmm. Yeah. Nice.

But then, but then it went to like this big brother open source where a Google, an IBMA son, well, not son. They were good. Yeah.

Um, whoever a, a company hp mm-hmm. They, they did open source a particular tool or project, but they, they had their own opinions and ideas and wants for it. Right?

So they retained control, but in doing so, it, it, it froze competition or it froze competition out. Right? So if you were, if you were IBM and I was HP or God darned, I'm not going to, uh, contribute to your success.

Right. Then we had this foundational era of open source mm-hmm. Where IBM and HP could work together along with Apple.

Yeah. And Google and Meta and, and what have you. Are we seeing that era now?

Maybe. That's a great question. And I think, you know, we were talking earlier, all the companies I've worked at that have told me not to work with other certain companies on a, you know, comp, competition based thinking.

Mm-hmm. I've worked with all those companies in the community, right? So we are now at a point where those companies are saying, okay, our teams are working with these other companies through in open source.

Is that the best thing for us right now? Yeah. And they're rethinking the now not the future.

'cause the future is open, let's face it. Right? No.

So right now with belts tightening folks having to buy more GPUs and a lot of expenditures on infrastructure, yeah. You're gonna see things just not get the weight pushed behind them to make them popular and open source. They're gonna drive the bottom line to increase revenue.

True. Some of these things though, I mean, where they collaborate theoretically at least should be on something as non-differentiated value, right? Mm-hmm.

It's just an enabling tech. But I think to your point, it's getting harder to determine what's non-differentiated value Is freaking out. Today's non is tomorrow's.

Yes. Right. It certainly is that there is that.

See all these people taking pictures of us, I feel like, uh, I don't know. We're doing something wrong. Anyway, guys, we're about outta time here.

We've, we've got the, the show floors open and we've got interviews and stuff to do. Chris, man, thank you so much for coming in here and popping in. Appreciate it.

Gotta get you into the rotation. Sure. We had a lot to the conversation.

Thank you. Mike, what do you got planned for the rest of today? I'm Gonna visit more boots and shake some hands and kiss some babies.

I'm running for office. Oh. But we need politician in the CNCF politician.

Thanks. I hope you've enjoyed this Textron gang. We will be back next week with our normal Textron gang back in studio, but it's been a hell of a lot of fun doing it here.

I'm Alan Shimel, thanks for watching. You're out. We're out.

Hey everyone. Welcome back here to another Techstrong TV interview. Um, let me introduce you to my next guest.

He, it's his first time on Techstrong tv, so let's go easy on him, but his name is Lawrence Wong. Lawrence is with Cisco. Lawrence, welcome to Textron tv.

Hey, thanks a lot, Alan. Happy to be here. We're happy to have ya on.

So, Lawrence, let's, let's start with you, if it's okay, kind of let's, we'll work backwards. Give us your role at Cisco now, and, and we'll work, we'll look at your career from going, you know, back from there. Yeah, sure.

Um, I'm the, uh, general manager of the Cisco Wireless and network platform business. And what that means is I'm responsible for the entire Cisco wireless portfolio, as well as how we actually bring together our network management platform for our campus and branch, uh, customers here. You know, prior to this role, I've held a variety of product management roles.

Uh, I, you know, have been the head of product, uh, for the Meraki portfolio. Uh, I started my product management career, uh, you know, being the product manager for switching and wireless. Uh, prior to all this, I was a electrical engineer, and I used to design, uh, RFIC circuits.

In fact, I, uh, built some of the first 8 0 2 11 ab, uh, g transceivers, really? Back in the day. Really?

Yeah. Very cool. So, let me ask you the magic question.

How long have you been at Cisco? You know, it's kind of funny. Uh, you, you start losing track, but, uh, this stint has probably been the longest, uh, I've been here since, uh, Meraki was acquired by Cisco since 2012.

Really? And, but, and you were at Cisco before that? Again, I, uh, I, I was at Cisco where I had my first product management job.

So this is, uh, after my, uh, my career as an electrical engineer. You know, it's, it's funny, I, I've heard this, I've seen this pattern before. I have a lot of friends who like worked at Cisco, then went out and did the startup thing only to be reacquired by Cisco, right?

And, and continue along their Cisco career path. And it's, uh, you know, once Cisco always Cisco, I guess is the, is the word. Um, of course Lawrence, you know, I've been a follower customer of Cisco for 30 something years.

com, I helped start a company that was what we called an a SP application service provider. And we were a Cisco powered network. Back then, everybody was, was the only gay in town.

I mean, there was Juniper, they were the new kids on the block then, right? And, um, but it was Cisco. It was a Cisco world.

Of course, today Cisco is very different than that Cisco. That Cisco was all about big honking routers and switches and smart switches and firewalls and, you know, hardware and, and so forth. Today, Cisco, it's a different animal.

It still does that hardware thing. It's probably still the leader in the world, I'm gonna assume. But there's also, there's the Meraki in the wireless unit, there's the software, there's security, there's observability.

How would you describe today Cisco is, especially from your perch, right? Being there as long as you have. Yeah, absolutely.

You know, look, I, I, as you stated, Cisco's been for a long time. But, and I think one of the things that is amazing about this company is the continued transformation, especially as we think about, you know, how do we help our customers solve the biggest problems today? And, and the way I think about Cisco today is you we're helping to build the critical infrastructure for the AI era.

And so that spans, you know, several areas. I think first and foremost, how do we help our customers build AI ready data centers? It's not just for the workloads, uh, that you expect, but it's also workloads, uh, for ai.

How do we actually help our customers future proof their workplaces co across, you know, carpeted, non carpeted, you know, enterprise networks across campus and branch networks, and of course, all of that backed by digital resilience, um, powered by our AI capabilities. So that is really how I think about Cisco in a nutshell. Today.

It's a very different company than probably when, uh, you know, you were thinking about Cisco. But yes, we still, obviously, you know, networking is the heart of what we do, but increasingly it's how do we in, you know, think about infusing security into, uh, networking and everything that we bring up to the market today. Absolutely.

And I just feel obligated to give a shout out to the WebEx. And is WebEx, WebEx, uh, the group that WebEx is part of now is, uh, I, I don't remember the name of it, but, uh, my, my friends Jeff Schaffer and Arner Char and are, are, uh, part of that group. And, uh, they do a great job too.

So I just wanted to give them some No, That I'll, I'll definitely let them know, yeah. That, that's part of the future proof workplaces. Like, we're, we're really helping bring collaboration and networking together, uh, for those type of customers.

So, uh, yeah. That's great. So you mentioned this data center.

Boom. I haven't seen a data center boom like this since 1998, right? com bubble burst, it took us about 10 years to use up all that data center space and dark fiber that we had laid out.

8 trillion in AI factories or AI data centers. I, I gotta imagine to a company like Cisco now that's, that's a market that could move the needle, right? Um, and, and so, you know, it pays to, to, to put some real resources behind that.

Of course, these are next gen ai. They don't even call 'em data centers. They call 'em AI factories.

What's the challenge there for Cisco? What are you going to do differently? What are you, what's new than, you know, just kinda what you've been doing powering data centers for the last 25, 35 years?

Yeah, I mean, I, I think you're right that, you know, right now there's a explosion in data center. Data center is cool again. And I think that even some of the 3 trillion numbers, uh, that you shared, uh, other people may have even greater ambitions.

But I think the, the net of it is like the way that you think about building data centers and how you manage that infrastructure, it has to move beyond just the, you know, traditional, Hey, I, I'm gonna run workloads in a, you know, tiny level server, uh, to a, you know, data center cluster. I'm gonna, you know, scale up from that and really moving towards world where you're really scaling out and really creating these like mega clusters across different data centers to, you know, the, you know, power, the workloads that, uh, you know, our customers and, um, you know, where the economy's heading right now. So I, I think the scale of the thing is just vastly different.

So how you think about managing, how you secure it, like you just have to take a different approach these days. I agree with you a hundred percent. You know, Lawrence, the other thing is a trillion here, a trillion there, you know, before, you know, you're talking real money, but for all the trillions of dollars that we're putting into these, let's call 'em hyperscale data centers, a lot of the action is gonna take place at the edge on the end point, right?

In transit and everything else. Um, how, and, and again, these are places where Cisco excels as well, right? They we're not putting all of our eggs in that big honking router at the data center space anymore.

We, we need to go where the people are, and they're everywhere and anywhere. How is, I mean, but that's, that's a different strategy. That's a different, you know, kind of per view, how, how's Cisco adapt to that?

Yeah, absolutely. You know, if you think about like this idea that AI is more than just in the data center and really out into the enterprise, it does mean how do you really think about everything from how you manage infrastructure to how you think about securing that infrastructure? You know, I'll give you a, you know, a an example in the old days, you know, you used to use command line interface to manage individual devices.

And over time, that moved to a web UI interface ultimately to APIs. And it's still very much a human-centered activity, if you think about it. But we think that the paradigm is shifting in a dramatic way.

And that paradigm is shifting towards a world where it's human and AI agents working together, uh, to drive more automation and ultimately more hands-off operation of network infrastructure. That also means that the way you think about security changes, if you think about this idea that you have agents, uh, you know, deployed in the enterprise, like you and I can have, you know, 10, 12 plus agents working on our behalf, you have to first start thinking about the security of it. Like, what is their identity?

What resources are they allowed to access? How do you know if you know, you actually are giving them the right policies? Like this is just a very different shift in mindset, uh, for the management layer as well, and security layer.

Absolutely. Absolutely. And, and it, you know, so I'm a security guy.

I've been in security 25 years. To me, it's just like we've blown up the attack surface exponentially, right? 'cause every one of these agents and 10 or 12 agents, to me, 10, 12 agents for you is probably conservative.

We may, you know, you're looking at someone who has 250 passwords, Lawrence, You know, locked in my password manager. So you can imagine how many agents I'm gonna wind up with one deck. But that being said, you know, the, the, the, the mission of security here, and you mentioned security a few times in your answers already.

We, you know, I don't think we would do such a fantastic job in security to begin with, right? And now here we are talking about, you know, a blown up attack surface at some level. I guess you need AI to fight ai, but, and this is a job for ai, if you would talk a little bit about how Cisco is scaling up to meet this challenge.

Yeah. You know, I, I think in the part of the portfolio that, uh, you know, I focus in, I, I think there's a few different facets to this. One is really understanding the identity and context of the human and agents.

And, you know, there's a rich body work happening here, but I also think there's like some more fundamental blocking tackling. If you think about the day in the life of a network, uh, administrator, let's say that they have a security vulnerability that comes in. And then one of the things that they have to decide is, how much risk am I gonna take on?

And of course, there's, you know, scores for these, the severity levels, uh, and sometimes they have to take action much sooner than they want to. That can mean upgrading, you know, the firmware on a device. It's okay if you're talking about just a handful, but if you're managing hundreds of thousands of devices across a, you know, enterprise network, it, it, it becomes extraordinarily hard.

And so at Cisco, we're investing capabilities. Uh, you know, one of the most recent ones we talked about here is life protect. And what this does is it takes the investments that we made, uh, in EBPF and really hook this into the kernel of the embedded operating system so that we can actually apply compensating controls, which allows a administrator to give them more runway to do that, uh, you know, upgrade.

So the compensating controls protects against, uh, some of these security vulnerabilities. Then you go in into other areas like, Hey, you know, we're having conversations right now, uh, you know, with federal customers, with healthcare customers, uh, around getting ready for a post quantum compute era. And what does that mean?

You know, right now we know that, you know, the harvest now decrypt later, uh, is a real security threat. Hmm. Yeah.

So with our, you know, most recent, uh, you know, uh, you know, hardware and software releases, we want to build, you know, these devices to be pqc ready. And that means building it at all layers of the stack from how you sign the factory framework, uh, to how you protect the embedded operating system to how you, uh, you know, harden the, uh, the control plane aspects and networking protocols for Maxi and IPSec and modernize it, uh, you know, for this era here. And these are conversations that, you know, I think a lot of customers, you know, know is out there, but they're struggling to understand how can we get ready for this right now?

Absolutely. You know, look, make no mistake, Q Day is coming, right? I've spoken to a lot of quantum and quantum security folks, and the thing I I learned from them is that, you know, it, it's not gonna be like marked on your calendar.

Oh, tomorrow's Q day, Q day's gonna happen. It may take months, weeks, or months for us to realize it did, and we're probably gonna realize it did as a result of something bad, right? That, that, yeah.

That, that happens from it. But when you, when you take Quantum and you take ai and what we're doing with ai and even, you know, physical ai as we call it, robotics and stuff like that, the three of them together. I mean, you know, you wanna talk about Industrial Revolution 4 0 5, oh, whatever you want to call it.

Um, I mean, the, the, the promise is, you know, astronomical. But so is the risk. How, again, is Cisco, and we'll get a little specific, if you don't mind.

How is specifically, what's Cisco doing to, you know, for this eventuality? Because though it may not be today, and it may not be tomorrow, it's not too much beyond that. Yeah, and you know, I, I, I think the funny thing about all this is, if you look at the trends even before, uh, you know, this mega trend of ai, what we saw in the, uh, network infrastructures, the rise of non-traditional devices.

So things beyond the mobile, uh, devices, the laptops, iot, like devices, things that are, you know, coming in terms of the factory floor, the distribution warehouse, uh, robotics, to your point, these non-traditional devices aren't always the easiest to identify, to profile, uh, to secure. And so I think a lot of the, you know, the work that we're doing is starting with some of the, the foundational aspects, which is, Hey, how do you actually make sure that you're defining, you know, common security policies that can be applicable to your entire infrastructure, whether it is that carpet in a non-comp environment, how do you build the hardware and software so you could have the distributed enforcement to translate the policy intent into something that's actionable. And then I think the third piece that's always been the struggle for a lot of customers is how do you do this at scale?

How do you make it easy? How do you drive operational simplicity into this? So, you know, even as an example here, a lot of the work that we're doing with our security, uh, peers around, you know, hybrid mesh firewall to, uh, you know, our access manager, uh, you know, solution that's basically a SaaS based, um, access control, you know, solution, uh, to help our customers, you know, really deploy security in a much more meaningful and simple way across the enterprise.

These are things that we're doing, and then some to really help our customers get ready for this era. What advice would you give to IT leaders today who are listening to us talk here and saying, oh my God, I gotta do something this, but I thought AI alone was, was enough to quantum robotics. What we, what could they do, you know, real world to future proof their networks in ai?

It can't just be, buy more Cisco gear, right? Though, that might be a solution, but what would be your advice there? Yeah, I, I think, you know, just like we see with, uh, you know, a lot of, I mean, every company out there adopting AI tools, oftentimes the difference between simply trial and proof of concept to production level adoption is having clear goals of what problems you're trying to solve.

If you don't have clear goals, then everything is up for grabs, then you're gonna try, you know, the next, uh, shiny object here. And then I think at Cisco, fundamentally, you know, if I think about going back to what I said at the beginning, we're helping to build critical infrastructure for the AI era. I think that, you know, from our perspective, the way that we provide value, the way we deliver innovation starts with, you know, modernizing the infrastructure.

Starts with, you know, making sure that our customers have a platform that they can build on that can layer more capabilities, that as we deliver innovations, uh, to them, as our partners deliver innovations through our platform, they get the benefits of that. So I think a lot of it is like, you know, very foundational, moving away from the firefighting mode that many of our customers in to something that is much more, you know, strategic and, you know, thinking over the next, you know, 3, 5, 6, 7 years here. Absolutely.

Lawrence, we're running a little low on time. I wanted to pivot and talk a little bit about the partner summit that you guys had. I don't know if it was, well, by the time people see this after we record it, it might be a week or two ago.

Um, and, and just a shout out to our Futurum sister company. I believe Tech Field Day was, was, uh, over at Partner Summit or, or more of the Cisco summit's. Been doing a tech field day with Cisco on this stuff.

But you guys announced a raft of, uh, enterprise connectivity solutions at the partner Summit Summit for those, and I'm sure a lot of people out here weren't able to be there. For those folks who weren't kinda, what did they miss? Yeah.

You know, and I think first and foremost, if you think back to what I just, uh, said before, how do we help our customers, uh, you know, really accelerate and deliver on the outcomes for their business? It starts with the foundation of the platform. We've been working towards a unified management platform.

And what we are doing at Partner Summit is taking that to next level. How do we start providing greater options across all operational types? We don't care if you're cloud, if you're on-prem or everywhere in between.

We want to be able to allow our customers, whether you're using Catalyst Center or Meraki, to, you know, get the innovations that we're building across AI system, AI canvas, and delivering that experience to all them through one single management experience. That's number one. I think the second one is we wanna base, you know, continue to provide our customers more options, uh, for cloud management, uh, more hardware capabilities, module switching as an example, but also just more, uh, software capabilities.

How do you deliver on that operational simplicity while still delivering great security? A cloud orchestrated, uh, fabric that's cloud managed allows our, you know, customers to achieve that. And ultimately, you know, radar automation, we've been working, uh, on capabilities with our AI assistant in AI canvas that is ag in nature.

How do you use, you know, English as your foundational interface to these management platforms? And these management platforms understand what you're asking for, and it's intelligent enough and, uh, to pull out the right workflows to kick off, uh, you know, to right pull in the right, uh, telemetry to help you troubleshoot issues at greater scale. These are the things that we're bringing to market in a way that, uh, I think our customers have been looking for from Cisco for a long time now.

Absolutely. Lawrence, this, we covered a lot of ground here today, specifically on the partner Summit announcements, is, I realize you may not have to at your fingertips, but is there a particular place on the Cisco site people can go to kinda read up and dig deeper? You know, I, I'm terrible at that.

So, uh, We'll have to, you know what, I'm gonna ask if we could grab a URL or something and we'll put it in the notes on, on this interview. Of course, you could always Google it, and I'm sure with Gemini, you know, if you are using Google, uh, it'll, it'll probably pull it up for you too. But that's a great thing about living in the world we live in now, isn't it?

Oh, 100%. You don't have to. Yeah.

You don't have to remember those arcane URLs anymore. Uh, well it reminds me of when you didn't have to take your calculator in a math class, right? And then I wanted to say kids could do the math, but anyway, Lawrence, I wanna thank you for coming on here on Techstrong tv.

We'd love to see you back on here. You know, I think there's so much going on in this whole space right now. And of course, right now AI is the, uh, what's the, the, the spoon that's stirring the pot, right, uh, to a large degree.

But, um, thanks for coming on and continued success at, at Cisco. Well, thanks for having me. It was great to be a first time guest and looking forward to talking Again.

We will for sure, Lauren Sw here from Cisco. And, uh, we're gonna take a break on Textron gang, excuse me, text drunk tv. I just finished Textron Gang on Textron tv, and we'll be back in a moment.

Hey folks, we're back at Atlassian Europe and we're here with my friend Shameek, and we're gonna have a little chat about services and service collection in their portfolio. Shameik, welcome to the show. Thank you For having me.

One of the things you guys just announced at this show is that the customer service app is now generally available, but I wanted to ask you, is customer service and IT operations help desk, is all that starting to converge now on a single kind of platform? 'cause historically we always kind of had two different things, right? Yeah.

But I, you know, across all kinds of service teams inside the company, we are seeing a lot more cohesion. Um, so often, for example, when your customer service request comes in, the support desk is in it is, is in its own silo with the existing tools, they're unable to reach back out into other teams inside the company to be able to get the answers that they need to get back to their customers. So by having a part of the service collection, the customer service management app now is able to pull data from a common teamwork graph that we have and get the answer quickly and get the best answer back to the customer fast.

So having it part of the same collection allows us to pull all of that information together in a much better way. So it sounds like the primary mission is to not have the customer service person say, we'll get back to you. Exactly.

Right. And that's so frustrating for the end customer because when they get back, you have to again, talk to somebody else, and then there's a whole cycle of repeating yourselves that we want to avoid. Yeah.

How is that whole service experience gonna change in the age of ai? Because for as long as I can remember, it was, you know, somebody logs a ticket, somebody reviewed the ticket, we see if we escalated it and then we close the ticket and rinse and repeat. Yeah.

Is that gonna be a different experience with all these AI agents running around? Yeah, Absolutely. Um, so firstly, there's a lot of, uh, queries that the AI agent can resolve automatically, right?

And the second thing it can do is that it can actually ask you clarifying questions that you don't have to repeat yourself every time. And it can put all of that information and connect it with the other information it already knows about the company to ask just the precise question that it needs, rather than having, uh, that rather than kinda circling around the question and again and again, like you used to, uh, with human agents, right? So all of that is great, but the most interesting thing is that it keeps learning from both you, this particular interaction that it has with the customer, but also from its interactions with all other customers.

So it keeps getting better over time, right? So all of the training that it's getting, it's not just in one agent's head, it's now in that common robo customer service agent, so that it's learning from all the agent tech information that's coming in, all the customer support requests that are coming in, and it keeps getting better over time. Will that create a perception of memory among in the service desk?

And I'm asking this question in this regard. Every time I call into some company somewhere, they, they never remember my last interactions. I mean, it's in there somewhere.

Yeah. But generally speaking, you know, it's a whole new experience and it's a whole different interaction. So will customer service have some level of, I guess we'll call it persistence, where they actually know me Yeah.

And they know my last interactions and they have a better sense of my preferences? Absolutely. So part of the reason why every customer interaction, um, today seems like it's disparate and no, the customers have, the, the service has completely forgotten about you, is not because the information is not there.

It's just that it's so cumbersome for the customer support agent to pull all of that information back out in just the time to be able to respond to you quickly, right? But with ro o and with AI, that becomes so much more automated and fast, right? So the VO customer service agent is able to pull together all your past history, summarize it in just the right way, whether it's resolving the problem or whether a human agent is actually resolving the problem, it knows and brings all of that data together in just the right personalized way to be able to service you in a much better way.

Mm-hmm. Um, what does it take to put all this together? Because some folks would say, well, we're heavily invested in all these other platforms.

So if I was gonna migrate, what would that look like? And how big a how big is the lift? Yeah, I mean, it's, uh, usually most customers have a customer service management system where they have all their customer records.

So what you can do to get started is just point our customer service management app to your set of customer records and your set of order, um, picking systems and entitlement systems without replacing what you already have. You could say that, Hey, these kinds of queries are coming into our customer service management app, and thereby start incrementally, right? And then as you see it performing better, and as it learns more and more about you, you can start expanding the number of queries that it gets to and the categories of queries that it's responding to.

So I think we have designed it in a way that you can actually get started small and then expand over time, right? So it's a, it's a pretty easy lift in terms of how you get started over time. Of course, you can start migrating more and more systems and customer support categories over into the CSM app, and the more you move in there, the more context it has, the better queries it can answer.

In the age of ai, will we wind up restructuring many of these teams? 'cause right now I think that, you know, if there's level one, two, and three escalation, and it's like a pyramid at the bottom is mostly level one, and then it gets smaller and smaller, but will a lot of the level one stuff be handled by an AI agent now and then I can reallocate my resources accordingly? Yeah, Absolutely.

Um, there's a whole bunch of tedious queries that come in, right, which are mostly about just informational gathering and about, you know, where's my order, what happened to my, um, payment that got stuck and so on where the information is already there in the system, and then that just needs to be pulled out and given back to the customer, right? Um, a lot of that is already moving to self-serve as well, so customers can self-serve themselves, but whenever a customer needs a query that's slightly more concept that I would call tier one, that's where I think AI is making a lot of informa, uh, dent right now. And these are tedious tasks that no human really wants to solve because just a matter of looking up the data here and then answering it back, um, those are areas where AI can do a fantastic job already.

Um, and then for the more tier two and tier three category, um, queries there, the AI agent can provide an assistive capability. It can summarize all the information of the past contacts with this customer and provide it to the human agent in a summarized form so that they can take action much more quickly. Mm-hmm.

How do we maintain the personal touch? Because sometimes you worry with AI that, you know, it all just becomes talking to a machine, but, um, is there a way to do this smartly so that people feel like, you know, somebody does still care? Yeah, So there's two things.

Like one, as long as soon as we take all the drudgey out of the task, it align frees up the human agents to do all the more, um, the software aspects of the contact, right? So what we wanna do is that when a person, when a customer is actually interacting with our customer service management app, we should be very clear about when are you interacting with our AI age agent and when are you acting interacting with a human? So the AI agent looks at all the questions that are coming in and knows that this is a particularly very, um, a very tedious kind of an answer that it needed.

And there it says, Hey, I'm answering this for you. Do you want some more information? And if it sees that the human is looking for a more, um, complicated question, then it can easily figure out, or that the question is getting very sensitive, right?

Or that, hey, it's going to, it's going to a loop with the human on the other side, then it can, um, escalate the problem to a human and be very clear to the customer that, look, now I'm not able to solve the problem for you. I'm coming over to a human. And there, the human agent can come in and provide the software, touch the emotional, uh, support that the customer might need in that particular case.

But, so this elevates the human agents to do the hard things, right? And it leaves out all the tedious things for the AI agent to be able to solve. One of the things that is notorious about being in the service field is turnover is really high.

Yeah. Do you think that that will become, uh, less of an issue because we won't be maybe burning people out as quickly? Absolutely.

I think that that's a, a pretty important part of this whole journey that this industry is going through, which is that, um, we really want to make sure that all the drudgery of that job is taken away so that the human agents can actually be working on the most, um, rewarding parts of the most value added part of, um, this particular role. Yeah. So what's your best advice to folks today?

You know, what do you see folks who are running service operations doing that makes you shake your head a little bit and go, folks, maybe we might wanna be a little bit smarter than that. Yeah, I mean, I think, um, first of all, adoption of AI is, I think here and people need to kind of embrace what's happening and the change that, um, AI is enabling because some of our customers are getting dramatic results by adopting ai, right? So just being more receptive to understanding what's happening and trying it out is, I think one thing that, you know, I would encourage all customers to do.

The second thing is AI is only as good as the knowledge you have in the company. So investing in more knowledge and putting all the information of your company, for example, um, what are my res, how do I respond back to a customer that has a payment failure? What are my processes for handling, um, a delayed order, right?

These kinds of things are often not documented well, and there's no business processes that are well established. The more the companies invest in these, creating this kind of context and knowledge, the better. Not only do their AI agents become, but also the human agents become much more powerful.

I think a lot of folks would be concerned that the customer service agent might hallucinate. So are there guardrails that I can put in place to kind of prevent that from happening? Yeah, it's A very good question.

So two things like, number one, we encourage customers to start with the, the, the easier queries first, right? And to sec and keep the settings so that, um, the more complex queries are going to the humans. And the second thing is that we provide a lot of control mechanisms so you can review all the answers that the AI agents is providing and coach it much like you would coach a new, um, human customer service agent to get better at their job, right?

So you can review all their answers and provide feedback on what went well, what didn't go well, and what a better answer would be. Thirdly, we provide what we call an evaluation system where even before you deploy it, you can, we, we provide a whole bunch of test, uh, queries and what are the suggested responses, and then we test the, uh, customer service management agent to see whether it's actually performing well and what the score is, right? So you would never deploy it unless you manage to kind of tune it to get to a good score.

Very similar to how a human agent comes in and there's a training period, and you wouldn't actually put them solo onto the, um, customer service, uh, task queue until they've actually met a certain threshold. Mm-hmm. In a lot of cases, people are using their customer service desk to upsell stuff to customers.

So would the AI agents be able to do that as well? Eventually? I think that's, uh, a place where we can get to, um, right now the, the focus of our app, and I think most of the industry has been to resolve the contact queries that are there, but upselling is definitely an area where I think, um, this whole field can get to.

One of the other issues that we have too is like a lot of the times people get a call about something, but it's not really their issue or it's related to some other company's thing that is dependent upon my thing and then they all interact. Can the agents start talking to each other from different companies that are maybe part of the same solution and kind of resolve things? Yeah, absolutely.

Um, there is obviously these, um, innovations that are happening in what's known as the MCP communications between agents, um, and also A two A, which allows agents to communicate with each other. These are areas that are still very relatively new and the, uh, connections between companies are still getting established in this area. But this is an area that absolutely we expect that if my companies service depends on another company's service downstream, then our agents should be able to talk to each other to resolve those issues.

We see that already to some degree in the observability space. Um, but in the field, operations, manufacturing, retail, and other spaces, this is still relatively new. We haven't seen a whole lot of that yet.

Yeah. So this sounds a lot better than the robotic AI type of experiences we've had so far with various chat interfaces that people have put together. Um, as you kinda look down the road a little bit, you know, what are you most excited about?

I think there's, uh, two, three things that are really exciting. Number one is, as you mentioned, agents working with other agents, whether it's inside your own company or whether it's outside, um, making that work well would really empower what we can do, because many of the issues are not dependent on what I know, but what other teams are de, you know, are doing as well, right? So that's one area where I think once we have that whole framework working will be even more powerful.

The second thing I think is really happening already, but can go, um, is likely to go even faster, is the quality of the response is getting a lot better. And what I mean by that is it's not just text and chat, which used to happen before, but other forms of media, for example, voice. Um, so being able to talk to somebody in voice and get back a voice response that is easy to make sense of, there's no hallucinations, the quality is so much better that, um, we are seeing significant improvements in the last year, and I expect that to continue getting even better, right?

Mm-hmm. And the third thing is adding video and images to your answers also makes the, the answer so much more real and easier to understand that. I think that's, um, another area where I, I expect a lot of innovation happening in the next year.

Do you think that people will develop a relationship with the AI customer service agent because they'll perceive it as somebody who's regularly helpful, somebody who remembers them and they'll start to, I don't know, assign personality traits to it? Yeah, I mean, I don't think that, um, I don't know if they'll be assigning personality traits to the customer service agent necessarily, but I do think that the trust and the expectations of what you can get from the, um, customer service agent on the other side is gonna go up significantly as they experience it more and more and they get really high quality instantaneous results back. Um, they would prefer getting that answer first and only when that fails would they fall back to a human agent.

So that trust level, I think inevitably is gonna go up and, uh, I'll thereby their, uh, expectations of what can, what customer service is, goes up significantly over the next few years. I want to come back to another point you were making about, um, making sure I have my knowledge bases in order to make the AI or work better. What from your perspective do organizations need to do to accomplish that?

Because I think, you know, it's hit or miss sometimes in what's in those knowledge bases. Yeah. So two or three things like, you know, one is being better about documenting your own business process and your policies and your, the, the way you respond to customers.

What is the tonality you use? Just being more explicit about all of that, right? Many companies have their own training manuals, not just for customer service, but even for employee service, right?

But sometimes those documents and those policies are scattered, um, they need to be brought together, condensed, cleaned up, and so on. Having said that, not too many companies are going to go at it from scratch, right? You know, if I have to build all of this knowledge from scratch, many companies are gonna just drop their hands and say, Hey, that's too much work, right?

So we can apply AI even to that problem, which is to help suggest to companies, what are the questions that you're coming in from past responses that you've given to customers? Here's a likely set of answers and knowledge that we can generate for you and suggest for reviewing, right? And then you can, with very little work, you can review it, clean it up, and then say, Hey, this is good to go, right?

Because even though many companies don't have knowledge bases, they have a lot of history of pass tickets that have come in how you've responded to them. Some of them might have been good answers, some of them might have been bad answers, but AI can help you summarize all of that, cleans it up, and also categorize into good, bad, you know, what are the right set of answers that I want to keep as, uh, templates for all future answers. There's one school of thought that says every dollar spent on customer service is a dollar that doesn't go to the bottom line.

And I guess, can we have a different attitude going forward where we think about investments now because the cost of the service is gonna drop dramatically? Yeah, Absolutely. I think this actually, um, will encourage customers to actually spend more on customer support, because right now, as you mentioned in the beginning, right, customer support is not just a cost center, but also a place that of, uh, that leads of frustration for customers.

Customers don't have a very good impression of their vendors because they feel like, you know, customer support just doesn't provide them the answers they're looking for, right? So as customer support gets better by this combination of automation and humans, it's going to elevate the value that customer support is providing to every customer that every, uh, organization that has this capability, right? So it's actually gonna help improve your customer satisfaction, reduce your churn, and thereby lead to more revenue, right?

So I, it's, it's not just with ai, even before, companies that have invested highly into providing better customer support have always had better customer satisfaction and therefore better retention rates, right? And this is just, um, gonna improve that even further. All right.

Hey guys, you're heard in here. Customer service is gonna get great soon. Yeah, Absolutely.

Thank you. Thank you. We'll be back in a minute.

Hey, welcome back to Atlassian, Europe, and we're gonna have a little chat now about AI regulations ethical use with my new friend Stan, how you doing? Hey Mike, great to meet you. Thanks for having me on my online.

My, My pleasure. Um, there are regulations all over the world, and you're the general counsel, and I'm sure you're keeping track of all this stuff, but different countries, different regions seem to have different attitudes. So what's the current state of AI regulation?

'cause I know in the US we're kind of maybe anti-regulation, and in Europe they're pretty far ahead. So yeah, how do I navigate all this stuff Yeah, it's a tricky world. Uh, it reminds me a little bit about where we were eight years ago with GDPR and privacy.

Uh, so some analogies and some, some, some things that we can talk about that are different. Um, what I would say is that we hear at Atlassian believe in smart regulation of ai. Um, we believe that it's really a partnership between industry and lawmakers to, uh, I think create a regime that doesn't stifle growth, only encourages, um, you know, the development of technology, new technologies like ai, but also creates guardrails that, um, will produce AI that, um, is technology we all wanna live with.

It creates more of a utopia rather than a a, a dystopia. Um, specifically when it comes to the geographic differences that you've talked about, Mike, um, right now Europe is definitely leading the PACT with the EU AI Act. Um, Atlassian signed on early to something called the EU Pact, um, which was sort of a, a a, a lightweight regime that, um, we comply with today.

That's something we can offer our customers here, say, in Barcelona. Um, and then what I would say is that we're building towards that high watermark really sort of saying, okay, if the EU is leading the pack, let's go where the puck is moving. Um, and then in the meantime, you know, if the US decides that it wants to move in a a different direction, then I think we'll remain agile and we can pivot, um, when it decides where it wants to go.

The US is, uh, the United States of America and the various states have different attitudes towards AI as well. Correct. We'll see.

Correct. Things in California and New York, Colorado. Yep.

And Texas. Colorado, yep. Um, is there some sort of baseline standard that I can get to if I'm using AI that might be applicable to a broad number of these states and countries?

Yeah, It's a great question. Um, you know, what I would say is that, uh, in the US there's actually a government standard. It's called nist.

Um, and that's something that if you wanna sell technology to the US government, um, you have to go through that checklist. And so for us, that's sort of been the, the closest industry proxy. There's also some, um, other industry standards that our customers are asking for in the United States, um, and elsewhere, it's an ISO standard, um, that's specific to ai.

And so that's something that we're also having our roadmap to, to build towards. Um, but as far as anything that is necessarily required, um, you know, that is a, a much more of a matrixed, um, approach. And so what we're trying to do is really build for scale since we have customers globally, rather than try and get too specific, um, build again towards that high watermark that we see, um, you know, sort of the industry moving towards.

But this conversation in a couple years could be very different, Mike, And there are other regulations that still apply, we'll say HIPAA and healthcare. Oh, yeah. If I have an AI agent, it still has to comply with the HIPAA regulations.

Yep. And there's all kinds of other regulations. Yep.

So, um, do those need to be tweaked or can they just be applied as is to AI agents and we'll just treat them like any other end user? Yeah. So not only are there new laws, there's the existing laws that maybe have been around for, for, for many decades, privacy laws, data security laws, all like you just mentioned.

Um, and so absolutely those, um, I think Are, are, are, are the laws today. And yes, they, they, they can apply to use cases. Um, specific to ai.

My sense is that they will also need to evolve, um, that lawmakers will need to keep up with the development of technology and make sure that those are rightly, you know, adjusted and applicable to, um, new, new use cases. Um, so all to say that this is something that takes a full legal team like mine to really stay out ahead of and make sure, um, that not only are we complying, but we're also leading and influencing. Um, and for example, we recently sent a team to Brussels here in Europe to help influence, um, the evolution of the law.

'cause again, this has to be this partnership between industry and lawmakers. Do you get the sense that the lawmakers are AI literate at this point? Or are they, or is that still very much a work in progress?

Yeah. Well, I mean, most of them are not technologists by trade. Uh, some of them are, um, and they have a, a series and, you know, a bench of experts that they rely on.

Um, but I think that's the opportunity that we feel at Atlassian. And, you know, being tech lawyers on my team, we really can help bridge the two sides technology and law, and really, I think help influence the, the outcome. So I've been very pleased, um, in talking to lawmakers and their ability to be fast learners, to be able to understand, um, you know, new areas of, uh, technology and be open to, uh, curiosity and learning.

Yeah. Are you also working with other vendor partners who are also have similar interest in AI regulations? I mean, can the industry kind of speak with one voice, or is that always gonna be a hundred different voices?

Yeah. Um, so for me, you know, a simple, i I would say, um, you know, sort of model to look at within AI specifically are the fundamental LLM providers. So your, you know, sort of anthropic your open ai, uh, your Gemini, um, and then you also have the deployers, which is more where Atlassian is, right?

Where we're taking the LLMs from the developers given all the, you know, incredible compute, um, infrastructure that it takes to stand up an LLM. Um, and so maybe there, you know, I wouldn't say there's a, a schism or a divide, but I feel like when it comes to compliance and regulation, the laws today are looking a little bit more closely at the fundamental LLM providers saying, you know, do they have a kill switch? You know, are there things that are more consequential when you're actually developing the model than say, Atlassian, that's deploying the model for the end user.

And so that's maybe where I see a little bit of the industry sort of, um, again, not a schism, but just sort of two different industry, uh, camps, um, and sort of how they're at least, um, looking at compliance and also what's the lift, um, in, in actually standing up a regime that that can comply. Um, last time I checked, I don't think you can indict an AI agent, so we're, we're still responsible for what happens with this AI thing, but I don't know, do people really get that? I think are they gonna sit around and say, you know, well, the AI did it, it's not my fault.

Yeah. It's, it's great. I mean, it's absolutely a, a partnership between humans and ai, and at the end of the day, the humans have to be responsible, uh, for the actions that that, that are taken.

I, I do think it will be interesting to see, uh, maybe as, uh, some litigation works its way through the courts, um, where the liability truly lies for an agent's behavior. You know, was it the creator of the agent? Was it the, the user who, you know, gave, gave the command?

Um, I, I, I think it's still too, too soon to tell. Um, but at the end of the day, this really comes down to trust. And the, the only way that AI is really gonna be successful and is gonna be something that we want to use, um, and really fulfill, I think, the full capability and the full potential of AI will be if, if it is transparent.

We know we're talking to ai, we're not talking to a human. Um, do we understand how the models were trained, how the biases were either accounted for or not accounted for. All of that, I think is gonna be super, super important.

And that's something that we here at Atlassian take very seriously. You probably know we have a company value of, um, open company, um, no b******t. And so that transparency comes very naturally to us.

Yeah. So you guys have been using AI agents within your own practice, right? Yeah.

So tell us a little bit about that. How are you using these things and what surprised you? Yeah.

I like to think, Mike, that we have the most innovative legal team, um, in, in any company out there. We are not afraid to embrace new technology. We're curious.

We like to experiment. Uh, we work for a company that is agile. Um, and so very much that is in, in keeping with our, our legal brand as well.

Um, we have, uh, identified two, we call them hero use cases for ai. They're both related to vo uh, the Atlassian, uh, product. Um, and so the first one is about, uh, our service management.

So we have a whole bunch of stakeholders internally within Atlassian who reach out to legal with questions, um, rather than get a whole bunch of emails and slacks, what we do is we funnel them in through what we call the legal one front door. So we use Jira service management as that front door portal with ro o powering, uh, all of the, the first line questions. So you might have a question and say, I wanna hire this new employee in this geo.

Um, I wanna make some changes to the employment agreement. You know, where should I go? Rather than have to have a human go in there and sort of point you, oh, okay, here's the template and here's the page that tells you, you know, what changes are acceptable and which ones are not.

Ro can actually do the first line of defense on that. Um, yes, you need a human behind the scenes, giving them the playbook, giving them, you know, doing the quality assurance to make sure that they're pointing, um, you know, the knowledge seeker in, in the right direction. But that's a really great example of how we can do more efficiency, um, and more throughput and not necessarily, um, you know, sort of have a, a, a human have to do that first pass.

The second hero use case that we have, uh, using AI in, in, in the legal team that Atlassian, um, is around, um, knowledge extraction. So think about, you know, in the old days you bought a company and they, uh, had a bunch of contracts and you had to hire a team of lawyers to go in and read those contracts. What are we buying?

Vendor contracts, employment contracts, sales contracts, maybe some leases. Well, rather than pay a law firm or have a team of, you know, five people, 10 people having to pour over these photocopies, you can feed those PDFs into vo, vo will extract a summary, a high level accurate summary of all those documents, and give you a readout, say in a confluence page that you can then sort and go through and sort of say, well, here's the ones that, uh, are highest risk. Here are the ones that we don't care about.

Let those go through. That's all the power of ro o. Um, and so those are the two things that I'm really excited about that we've said tho, if, if we can focus on those and standing those up with my legal team, we can then focus on the more high value things that are really attorney work.

Mm-hmm. Are you at all worried that AI's gonna replace lawyers at some point? It's been, uh, I've, I've, I've had some, I've had some friends reach out to me and, uh, and, and ask me about that.

You know, I think there's always going to be a need for judgment, that, that's the one thing that I think, um, the legal craft, um, needs humans to do, uh, and that AI cannot replace, is that there's always gonna be these, uh, I think very, uh, discreet edge cases that are gonna require really understanding the totality of facts precedent, um, being able to see shades of gray. And I, I, I think maybe AI can get us 50, 60, 70% of the way there, but it's that last 30% that takes human judgment, human discretion, um, a lot of, I think just experience that perhaps, um, AI can cannot simulate. So maybe the legal craft evolves, um, but I don't think it ever replaces us.

Mm-hmm. Can we accelerate the legal process? I think if you ask most people who've ever been involved in the legal process than one impression they got, it was, it takes a long time.

Yep. Can we like reduce this down to something that's more manageable? Absolutely.

I think we can go much faster. Um, and I think, you know, the business can run that much more efficiently, um, with, uh, AI helping, uh, lawyers, um, and part of what I was just describing of getting a box of 600 pages of photocopies that used to take, you know, a team of 10 people 24 hours overnight, pulling an all-nighter in a conference room to read, you can feed that into VO and you probably can get the readout in about 20 minutes. Right?

And just think about that. And there was some great examples in the, the keynote, uh, yesterday from Rajiv around software coding. Same thing.

What used to take code review three days and a team of developers. I think you can do that now in a couple minutes. So it's just gonna free us up to do better things with our time.

That's, that's where I think the unlock is. One more question related to that. Yeah.

So when you take the bar exam, you're supposed to memorize all this stuff. Yeah. Do I really need to memorize all that stuff if I have AI agents going forward?

Good Question. I, I feel like that comes back down to like the calculator of like, back in the, you know, we used to have to learn algebra, but now we have, you know, calculators and computers to do that for us. Um, I think that there probably still will be some benefit in testing for knowledge and standardized testing.

It's just gonna have to evolve. And maybe the things that we're testing for today are not the things we should be testing for in the future. Maybe the test will be on how does the legal craft use ai, that that could be more of a skills-based test.

There you go. Something to think about. All right, folks, you heard it here.

AI use it responsibly, but it can do great things. Hey buddy, thanks for coming by. Thanks, Mike.

All right. Thanks for a great conversation and we'll be Back in a minute. Ransomware isn't paying so well, and yet, NPM packages are full of ransomware.

Call Weave Boosts AI development while Intel opens AI showrooms Platform nine celebrates an anniversary, and Commvault makes data easier. All this on today's tech Field Day rundown. Welcome to the Tech Field Day rundown, where each time we meet, we run down the IT news of the week with variable degrees of snarkiness.

I'm your host Alistair Cook, and joining me on this special edition since well, my nor co-host Tom Hollingsworth is at Network Field Day. Joining me is Jim Ky. Hi everyone.

How are you doing today? Uh, good to see you all. And by the way, today is National Donut Day.

Quick fact. We used to have, uh, one of those Krispy Kreme donut shops right down the street from me here a few years back. It converted to a, uh, diabetic center.

So I presume there's some wisdom in that There's some poetic justice in the conversion. It's also American Football Day, which is a good way to burn off all of those ex excess calories. And exercise is one of the good preventative methods for, uh, you diabetes and similar things.

Yes, that's true. Yes. Ransomware payments are dropping sharply with only 23% of breach companies paying hackers in the third quarter of 2025.

This is the lowest rate on record, according to ware, stronger defenses, law enforcement pressure, and, uh, revised corporate policies are reducing attackers leverage. While most ransomware now involves data theft and double extortion, uh, companies are increasingly refusing to pay, causing average payouts to fall to just $377,000. Attackers like a carer and kien are shifting focus to medium-sized firms that are more likely to pay while remote access compromises and software vulnerabilities becoming the main attack methods.

Are we actually getting better at it and better at re resisting these tax jump? Um, you know, I certainly would hope that, uh, the numbers seem to reflect that, but still almost $400,000 per incident. Uh, um, one of the interesting things in the report, Alistair, that I didn't see was, was there additional pushback from insurance companies, right?

Uh, having, uh, have a bit of an insurance company background and having a spouse that started her second IT career at one. Uh, you know, there was the idea of managing risk. Uh, and maybe that's also part of it too.

The other interesting part of that article, though, was that it's shifting towards much more remote access, uh, exploits. So, you know, we're still talking about people, uh, not having good personal security, uh, with the whole idea of bring your own device, uh, still happening and with the, uh, on slot still of people doing AI on their own personal computers, perhaps. Right?

Um, maybe that's still a pretty rich, uh, target sphere and easier for people to penetrate. So, yeah, just a fascinating, uh, down, you know, downturn and a good downturn for a change, but still kind of scary. NPM was hit by a phantom raven attack with 100 plus malicious packages being downloaded 86,000 times.

So what happened is hackers have flooded the NPM repository with over 100 malicious packages that have been downloaded, as we said, 86 K times according to security form coy. The campaign called Phantom Raven Exploited an NPM feature called Remote Dynamic Dependencies, which lets packages pull code from untrusted sites without detection. The fake packages stole developer credentials and system data while hiding from security stands.

Some of the malicious code even used AI generated names, oh boy, big surprise. Um, that, uh, appeared to be legitimate. And about 80 packages are still active, exposing some pretty major risks in the open source software security community.

This is pretty concerning, isn't it? Because you think as you are building your application, you're controlling your dependencies inside your application. Your CICD platform is pulling just the versions of software that you think you should be using.

Well, NPM allows these, uh, remote dynamic dependencies that are one software package that you've chosen to include pulling data out of some other location that is not the repository that you've originally looked at. And this is the, the mechanism for getting in here. Uh, as, as soon as a package gets pulled in through these, uh, NPM repositories, uh, one of these hundred malicious packages, uh, now there's only 80 of them, uh, that they're allowing arbitrary code, essentially to be pulled into your build environment and to run on your build servers and potentially then out into production servers.

Although the discussion here is primarily about attacking your build servers and extracting information, developer information access keys may be for cloud services, those kinds of, uh, credentials that are being used to access other remote systems. So pulling the contents of probably your software source code repositories out and, and sending them somewhere for an attacker to then news. This is pretty horrific.

It really is something that is a, an enterprise security company. Maybe I'm not wanting to use, uh, these public repositories where I have less control of where the, uh, the source code is actually being pulled from. Maybe I want to be doing my own governance and pulling all of my dependencies on premises, scanning them for the presence of these remote dynamic dependencies, and making sure that we actually remove those, uh, only allowing packages that don't rely on remote dynamic dependencies, or at least remapping those dependencies to only my own source code repositories.

It's a bit of a concern then that I'd need to be using a full set of source code repositories that are managed and governed on premises. And this does put it out of reach for some smaller organizations that will wanna be very agile. It puts it into that space of the enterprise companies building larger applications.

Hopefully, NPM themselves will address this as well, and will add some methodology for denying remote dynamic dependencies. Um, that would help us to then mitigate the risk that future repositories get added, that, that, again, expose. Uh, 'cause of course, if you can add a hundred malicious packages into NPM repositories, you can probably add a hundred more and probably a thousand after that when you're good at, get good at hiding where your repositories are coming from.

And when your AI does a better job of not just obfuscating the names of these, um, repositories, but also obfuscating the vulnerabilities within them while we're looking at cloud things. Amazon's latest survey that they definitely want to tell you is, is leading to more cloud usage is expecting that, uh, cloud hosted application use will grow from 59% of enterprises to 75% in a year. Uh, most organizations see public cloud as safer and better for compliance and on-premises according to AWS and multi-cloud use is increasing it.

Cloud and cybersecurity budgets are set to rise. Our budgets never rise despite ongoing data breaches. Companies are confident in cloud platforms.

You know, it's interesting because I started out 25 years ago as a, a strict on-prem. This is when I get that part in there, I, um, database administrator, right? Built the server on the ground up by myself.

This drives the whole nine yards, right? And when cloud really came about, I do a lot of work in, uh, you know, not only Amazon, but also Oracle Cloud infrastructure, or OCI. Um, yeah, it's amazing to see the growth over time.

And, you know, security from my perspective, is really one of the selling points, right? Um, the idea of having a centralized place to locate all of your security concerns, um, and to put all your eggs in that one basket and watch that one basket seems to make a heck of a lot more sense to me these days. Um, especially with all the exploits like we just talked about with NPM, right?

Uh, open source software, yeah, it's free, but who has control of it? Um, and depending on your vendor, your cloud vendor, generally, they've got a much better hand on what to look for security wise. So it's, it's not, I think, you know, Alistair not just, it makes sense to go to the cloud 'cause it's easier because it is, but also fewer potential people that could be compromised.

So, yeah, I, I, I certainly see that. And with the growth of ai, everything from generative to Ag agent to whatever the next thing's gonna be, I think it's gonna make a heck of a lot more sense. Uh, I don't think it's gonna slow down core.

We acquired a company called Marmo. Hope I'm saying that right. And open source Python notebook for AI and data work create a complete platform for AI development.

Moreno's Notebooks will integrate with Core weaves, cloud and tools like weights and biases, helping developers move from experimentation to large scale deployment. Uh, the Marino Notebook will stay open source and the acquisition builds on core reef's, recent deals to expand its AI developer ecosystem. Fascinating company Core Weave, but interested, what's your take on this?

Well, firstly, I wanted to see a little bit more about marmo and, and Marmo essentially is an alternative to using Jupyter. Uh, it is more recently developed, maybe a little more, uh, interesting as a, a reactive, um, more visual tool for development, but it's essentially about developing new AI applications in a, a nice, easy way. What I like in this is that core weaves objective here is to help companies move from the experimentation phase of building their first couple of AI applications to running dozens, hundreds, potentially thousands of AI applications, AI agents, uh, in, in production.

And, uh, Jim, you'll recall that AI infrastructure field, day three, the, uh, focus was very much on shifting from experimentation to production. And it is a difficult problem for organizations. I applaud, um, applaud call we for addressing this, along with a bunch of the other acquisitions that they've had recently.

Uh, other acquisitions they've had included, uh, buying Monolith ai. They're a company that specializes in, uh, bringing AI to the physical world of building machinery and, um, dealing with engineering and physics problems, uh, as well as the, uh, open pipe acquisition for reinforcement learning tools around their open weights and biases as well. So, uh, there's a whole lot of nice pieces being built into the wider core we've platform as they try and stake a larger and larger space in the AI application development.

And then AI application hosting, that's where the payoff for them is make it easy to build these applications that will run on the call Weave platform. Let's hope that their commitment to keeping Marmo open source and, uh, permissively licensed continues that other organizations will be permitted to use, um, Remo technologies as well as they wish to build out their own, uh, portfolios of AI applications. Uh, as always with these acquisitions, it will be interesting to see whether that permissive license remains, uh, or whether a more restrictive license appears over time.

It does seem like cowe wants to have the majority of development of mamo remain community-based development, and that minimizes their cost, maximizes the possibility that things will remain open, uh, as soon as they pull to largely internal development, then their costs will go up, and that's the point at which you'll see them switch to a different licensing model for the moment. Their, their approach seems to be really nicely. Open source Intel is launching some popup AI experience stores all around the world, New York, London, Paris, Munich, even.

So, to showcase AI powered PCs, a visitors can try laptops from major brands, uh, and see Intel's upcoming pet. The lake processes in action much focuses on gaming while so highlights AI applications in Asian non-English language, uh, in particular, the stores aim to boost brand visibility and demonstrate AI features and generate excitement ahead of Intel's next generation PC launch. I guess these experiences are all about helping their partners get out and get the idea that an AI PC is what you want.

Jim, do you want an AI pc? Uh, no, not particularly. Uh, frankly, uh, I, I can, my first question is how do I turn it off?

Probably, uh, just my, as a creator, that's my personal opinion. Uh, but seriously, uh, Intel is, you know, you think they're, well, you know, they make chips, right? But, uh, the, the thing that was, uh, caught my attention was in Seoul, they actually are doing a Gangnam style type of rollout, you know, kind of to draw in perhaps a younger crowd or, you know, just even maybe being a little glitchy, if you will.

Uh, but it was really interesting that Intel, uh, is starting to put, uh, Panther Lake in action, uh, from everything I've read about them. Uh, it's a, an intriguing idea, uh, in terms of being able to utilize what we typically think of as crucial for ai. But there may be other types of applications that may well run extremely well, as you mentioned, gaming especially, right?

But we really don't know what the next generation of AI beyond generative and AgTech will bring, right? Um, you know, you may, you may be seeing people doing, uh, much more interesting types of open source experiments or even, um, you know, um, recall when the Planetary Society had people use their laptops to, you know, home through data to look for, uh, if I remember correctly, uh, lunar samples, or even they were look, uh, some, uh, uh, cosmic samples and things like that. So, uh, uh, you know, it's very interesting to see where all this is gonna definitely, uh, head towards.

And the other part of Intel that I always found fascinating, I recently, I think within the past 15 months or so, they were at one of our Tech Field Day events, and they were basically saying, uh, you know, you don't necessarily need a GPU to do a lot of the modeling that people are trying to do. So even though the GPU like features of Panther Lake might be really good for gaming, or for visualization, or for other things, a lot of times, uh, with smaller numbers of, uh, dimensions of attributes of features that you need for a typical model, for a lower powered model, they may be just fine. So, it's an interesting, uh, idea that we see coming forward with that.

Um, it should be interesting to see what that brings even for people like me who are not particularly AI driven, um, just my own perspective. Platform nine has, uh, marked its first year of private cloud director, which is offering VMware users a simple and cost-effective private cloud solution that runs on their existing infrastructure. Uh, interestingly, it was built by former VMware engineers.

It helps enterprises modernize cut costs and migrate quickly. And, uh, one fortune thousand VMs at a fraction of the typical costs using Platform Nine's, v jailbreak tool, clever marketing there, um, supporting both on-prem and, uh, SaaS options. It's providing a trusted, familiar path for VM where customers looking to maintain control efficiency.

Alistair, here we go again with, you know, uh, the fallout or the, the benefits, the right, whatever you wanna call it, from Broadcom's acquisition of VMware. Absolutely. There's been a lot of companies that have been looking at what are my alternatives to paying the increased price that VMware wants for VMware Cloud Foundation?

And my experience has been, if you're using the features of, of VMware Cloud Foundation, if you're using the, the full suite, you're still getting great value as, uh, you, you look at the new price, the, the VCF suite as a whole was underpriced previously, but if you're not using all the features, maybe that higher price for VCF isn't gonna suit you. And this is where companies like Platform nine come into play. And Platform nine has, has been a long time friend of Tech Field Day, uh, cloud field Day 21 is where they showed us private cloud director and v jailbreak for the very first time.

And it was, it was, and remains a pretty impressive tool. I've looked at previous migration tools or previous migration processes. I did a, a customer study of a, a, a company that migrated from VMware vSphere on-premises to another virtualization platform.

Uh, a couple of the things that are crucial in there is having good tools for that migration. And that's where, particularly the VJ tool really is very effective, uh, scales out scales to moving large numbers of virtual machines. That Fortune 500 customer, they priced their migration of those 40,000 VMs at $35 per vm.

Now, this stands in contrast to the estimates that you see from, uh, other vendors or for, uh, in particular in the media. Somewhere between $303,000 per VM gets floated as the number, depending upon how much fear and uncertainty you want in there. Obviously, the devil is in the detail of how you account all of these things.

If you want to inflate a number to say it's not worth moving, then you account a whole lot of things that maybe you wouldn't account if you wanted to say it was a good way moving. I've always liked Platform nine. I've been a fan of them for many years as, uh, they've done managed OpenStack, managed Kubernetes, and now managed private cloud.

I really like their products, and we've, we've seen great uptake from, from customers. And that, that benchmark of moving 40,000 VMs at $35 per vm, that puts us very much in the wheelhouse of large organizations who have a lot of virtualization, but are just using vanilla virtualization from VMware would to have, uh, continuation of the, uh, vSphere alone licensing. But unfortunately, Broadcom is not looking to continue that.

So they're looking at alternatives and, uh, platform nine is probably one of the compelling alternatives we've seen. So great to see that success. And of course, I have a collection of friends who are part of the Platform nine team and, um, wishing them every continuing success.

Now it's time for us to take a closer look and we take a closer look at some innovations from Convault their data rooms, let data scientist teams quickly access and prepare backup data while using an AI powered interface that simplifies, uh, querying and managing this backup data for AI purposes. Using data that's already being classified as it was ingested into the backup system. Teams can reduce the manual preparation and apply role-based access control and sensitivity tags, and share data securely, uh, the times to speed up AI model training by letting, uh, the IT team who are already managing the data, managing some of the management tasks of that data, managing management, uh, and freeing up data scientists to focus on the focus on the analytics rather than the data preparation side.

Jim, you're very much in the working with data, doing good things with data. Would you like to be able to use the backup data repositories rather than the primary repositories for this? I think it's a fascinating idea.

I, I wonder why no one has thought of this before, right? If you have a database, you better be backing it up, right? And you better be also, by the way, testing your backups.

But that's another topic for another day. Uh, the key thing is, I'm surprised that, uh, it's, it's actually quite brilliant because you're gonna have to scan through that data, right? Depending on the type of backup you're taking.

Uh, but generally, you know, a lot of shops, if they have, uh, structured data especially, right? Um, they, you're gonna be taking a full backup or some sort of, uh, you know, level zero incremental, level zero backup once a week anyway, you're gonna touch those blocks, right? So, um, you might as well go ahead and capture that.

And if you're backing up, uh, object or file, you know, again, similar concepts, but different, so you're gonna touch the, the, the data anyway. Um, one of the things that I still see organizations kind of struggling with is, well, how do we get value out our structured data? You know, there's metadata buried in there, we've got tables, we've got columns.

Hopefully everything's described in some sort of master data catalog, right? Uh, guess where that's usually stored? Yeah, inside the database itself, you're backing it up.

Why not take advantage of the bandwidth that you're gonna expel your network bandwidth and everything else, uh, to do that backup, right? So this is a fascinating idea. Uh, again, like I said, I'm really surprised no one's thought of it before.

Um, it, it would be interesting to see just how well this works and what they're able to capture in terms of, especially, I'm thinking the metadata more than anything else. What's your thoughts on that? Yeah, I think the, the, there's an element here of building a data catalog, and often we would see the data catalog being built in a data lake along with a, a read-only repository of that data.

And so we would often see this being another reason to make another copy of our production data in some location where we're gonna, maybe, I mean, taking the, the classic cloud methodology, we're gonna dump this into a whole bunch of object storage. We're gonna have something crawl over that object storage, an updated database with metadata every time we ingest new data. Well, I've already got a place where I'm storing a copy, a read only copy of all of this data, and that I also have version history over time, because the backup system does this data protection on a regular basis.

So this is absolutely, uh, the, a continuation of the discussion that, that these data protection companies need, or are moving beyond simple backup, moving towards trying to extract as much value as possible out of the data that resides in this backup repository, that it's not just insurance, but it's actually usable data that we can then, uh, in this case expose through a model context protocol to allow easy consumption of that data in an AI application. Does then bring to, to mind that now my backup repository contains some primary data, not just protection copies, because of course, as soon as I'm grabbing that metadata, if that's stored inside my backup repository, how am I protecting that? Whereas my second and third copies and one of them off site i'd, is that metadata actually being stored out somewhere else?

Hopefully the metadata is being stored separately from the backup repository on a database server that I can then protect into that backup repository. Yeah, let's not get into a, a looping of our data as we're, uh, protecting it and then scanning the protection of our metadata about our protection. Um, somebody's slightly smarter than me has probably designed this, and so they've avoided that kind of looping problem.

But absolutely, I love when we see more value being delivered from the data that we are storing. One of the challenges here will be making sure that the backup systems performance suits the requirements of the application. That'll be consuming data from it.

Uh, it may well be that we'll see caching capabilities that layer on top of this, that allow, uh, the, the actual AI application to be working with a, a high speed copy of a subset of the data, whilst the majority of the data's stay on a, a lower cost tier within the storage system. Lots of implementation details that will be fascinating to look at. And of course, field Day will be at Commvault Shift in a couple of weeks, and I imagine we'll be hearing some more about that other places.

The Tech Field Day will be, well, right now, this week, the reason Tom's not with us here today is that networking field, day 39 is happening. So today, November 5th and tomorrow November 6th, uh, Tom is hosting Networking Field Day with a full delegate panel and an interesting collection of presenting companies. So make sure you take a look on the Tech Field Day website or watch on along on LinkedIn.

If you are, um, watching there. The following week, I will be in Atlanta for Tech Field Day at Kbn North America. So we will be, uh, live streaming on November the 11th from the event.

Again, I have, uh, some great sponsors there. I have South Works, I have, uh, VMware by Broadcom, and I have, uh, traffic as well, uh, presenting lots to learn about those companies there. Uh, and again, tech Field Day website will show you all of these things.

Other things that we will come up with, well, tech Field Day Rundown will be here every Wednesday, so continue to, uh, catch us on YouTube or on your favorite podcast application. The Rundown is also streamed on Techstrong tv, and you can catch us on other Techstrong and Future and Programs. Uh, we'll be back next Wednesday to talk about all of the IT news of the week that was, and that's fit to print, or at least to talk about.

Until then, for myself and for Jim and all of the tech fields, a team is wishing you and yours a great day. Done. Hi, welcome everyone.

I am, I'm with K. And today we're going to explore how to build infrastructure that isn't just a capable, but also a ready. There's a big difference between demoing something with Chat GPT, and actually deploying something at scale that is secure, observable, and also well governed.

So today we're gonna be diving into the cutting edge developments that are shaping AI infrastructure, a vertical component as organizations move from AI experimentation to production grade systems. The evolution of AI infrastructure is not just about new models or tools. It's about creating a holistic, sustainable ecosystem that can adapt to emerging challenges, drive innovation and deliver business value reliable.

Now, over the few slides, we're gonna highlight some key trends, best practices, and the patterns that define the state of AI infrastructure today, what it will look like in the future. So we will go over what production ready really requires some evolution about AI agents, APIs and agents. And we'll talk about a little bit about how we can do that with Kong and how we are able to show you, um, some of these, uh, things in, in play.

So, again, my name is Sugo. I'm part of the product team helping organizations to, uh, have a successful AI strategy using APIs, event driven architecture and integration overall. So let's get us started.

Now, you will be asking like, you know, why are we talking really about AI ready infrastructure today? May many organizations are eager to adopt generative ai, but the reality is that without clear definition of success, many projects risk failure. Today, we'll focus on why defining success.

It's also essential. Now, let's go on some of the key stats. You know, a lot of companies are gonna be starting to use, uh, gen ai, as I was mentioned, um, around 50% of the companies that have deployed AI strategies are working on that, and not just the side experiment, they're becoming central to their way business operate.

Now, also, 70% of the increase of, uh, product workload will come from AI agents by 2028. In fact, like, you know, analysts like Garner saying that 80% of the, uh, APIs from organizations will be consumed by, um, AI agents instead of, you know, traditional developers. However, there's also some risk, and there has been a lot of studies, uh, and then, um, um, uh, flow of information regarding studies that says that at least 30% of generative AI projects will be abandoned by the end of this year.

And this points to a significant issue. Many projects are set to fail due to clear objectives or ROI or not meeting business and technical needs. Now, it's crucial to avoid jumping into projects without understanding what success look like and how you should be able to, you know, build enterprise ready, um, uh, infrastructure.

So the high failure of rate of generative AI projects, coupled with a rapid growth of AI rolling products, makes defining success a critical step. Now, while we're talking about this, well, again, we are moving past the AI experimentation phase, and we are ready to move into production ready AI systems. So at the AI experimentation phase, you know, think of, uh, this as the testing real, or AI organizations explore different AI technologies, run s and test various models and control environments.

It's about understanding what works and what doesn't, and learning how to best apply AI to their specific needs. Now, that's the experimentation part, but we were talking about production ready. Now we're talking about AI that's been refined and optimized, that is ready to be integrated into actual business operations products.

Our services, these AI systems are stable, scalable, and deliver consistent results without failures. They're designed to handle real world complexity and provide ongoing value. Now, why does this transition matter?

Well, moving to production means AI is no longer just an experiment. It's been used in a way that's reliable, secure, and able to scale with near the business. This is a major milestone and AI that's ready to be, you know, primetime with all the complexities of daily operations in mind, us, part of the DevOps team needs to be, uh, able to check this.

So now, production ready, ai, it's more than just, you know, working prototype. It needs to be fully managed, secure, observable services that deliver reliable business value at scale. So there's a couple of, um, points here that will certainly help us to, you know, move and, and, and define really, you know, a, a production rate infrastructure.

So we're talking about, you know, security and compliance, you know, and infrastructure pipelines, models that enforce that encryption, access control, audit loss, that also needs to meet industry's regulations. Think about GDPR, uh, HIPAA and internal governance policies, or even your own internal compliance setup. It also needs to be reliable and robust.

You know, production AI must deliver consistent results under all expected conditions. We are adding new applications where new models, it needs to be consistent, needs to be stale, also needs to be, uh, able to, you know, do graceful error handling, retries, fallbacks, you know, um, behave without the trying to avoid, you know, downtime also needs to be scalable and performing. It must be able to handle peak loads with low latency and predictable throughput.

It needs to be designed for horizontal scaling, you know, being able to use very well known techniques like caching, optimizing inference, the possibility to be able to really handle the potential load of, you know, real work, real world workloads. Um, we mentioned about compliant. You know, we, we mentioned about having this, being able to handle all the information regarding your industry or your overall, uh, policies, and also needs to be continuously improved.

We need to be able to have, uh, close, uh, feedback loops, capture, user intention, interactions, performance data. Also, we need to be able to have, you know, regular training, AB testing, model tuning, keep up, or, uh, keeping accuracy high all again with, you know, the possibility to continue, uh, keeping the human, uh, in the loop, being able to provide really, uh, resource. Now we are talking about this production ready.

And, and why do we mean by that? When talking about, you know, DevOps? And here it's important because we are moving away from this experimentation phase where we are, um, just, you know, playing around, building, uh, prototypes and mocks that we want to move into the concept that we call the AI innovation factory.

So there are many difficult, uh, difficulties organizations face when creating this systematic and repeatable process for developing and deploying AI at scale, what we call the innovation factor. Remember, the innovation factor is this structural environment that consistently generates new AI solutions, models, applications, all designed to drive business value. They're going from artisan mode where have a small team just gathering around trying to play and discover how this works into full mode production, where we have everything automated, applying DevOps principles.

Essentially, it's about setting up a sustainable, efficient system that can keep producing impactful AI innovations over time in a way that is, as we were saying, again, you know, secure, compliant, scalable, and so on. And this doesn't come, you know, without challenges. Like, for example, this graphic shows results from the 2024 Garner generative AI planning survey, which as for organiz about their biggest roadblocks to adopt, you know, G and AI at scale, basically production.

So there are a couple of, um, challenges that they highlighted. Um, there were ones that went around data quality. Many HG and I projects tell you, consistent, incomplete or unstructured data, poor quality data.

Uh, the second one was around privacy and security. There are several concerns around explosive, exposing, sensitive or regulated data to LLMs to MCP servers, to agents in general and organizations and trying to grab on how to sanitize inputs as well as outputs, you know, control the access to models and comply with data protection regulations. PII, you name it.

But again, the other, um, challenge on, on the top here is also regarding post, uh, you know, both, you know, compute infrastructure. If you're running your own model locally, UPO cost as well as small licensee tokens can be expensive, especially at the scale. And if we are not, you know, having observability and tracing and tracking of where, you know, all those topics are better extend.

So this challenge becomes more pronounced when teams overbuilt prototypes that aren't optimized for production as well. If I'm just running three requests, three pumps, I might be, you know, overusing some tokens. But when suddenly those 3, 5, 10 prompts become 10,000 prompts or 10,000 users using the same, uh, solution and hitting the model every single time consuming tokens, it can be costly.

And we, we don't have observability or, you know, at least an idea of where those tokens are exposed, you know, and, and consumed. There can be complicated, but the hyper wrong. GN AI is real.

But organizations are hitting practical roadblocks, especially around data governance operation, operational efficiency. So addressing these challenges early is critical to unlocking sustainable gain AI initiatives. And as we were seeing, some of the things are just, you know, in track on how we're using, but also there are some things that are common pitfalls we see when organizations rush to implement AI without proper architecture, without proper data pipelines, without involving the, uh, the mature, the SecOps teams, uh, for architecture governance and so on.

Thinks about, you know, one model to rule the model approach. You know, when teams software rely on a single large model for everything, customer support, documents, optimization, code generation, you know, this leads usually to poor performance, high cost and lack of specialization. Instead, we need to, you know, fit for proper models and agent like components that are able to get most of each one of the each same capabilities.

There's also, you know, no prompt or input sensitization ignoring input filtering opens the door to prompt injection, bi leakage, security risks, San sanitization, it's key to save compliant GN AI deployments. So think about, you know, avoiding those kind of things. Or also ADOC model deployment.

You know, models push directly to production without any kind of versioning testing or rollout strategies. 'cause suddenly we are just, you know, rushing to push everything to production, and we don't have really the pipelines to be able to deploy them safely. Being able to implement rollbacks to keep track of the, of those deployments we need CI ICD style.

I also, for ai, it's another component of our infrastructure or another component of our architecture that needs to also be, versioned needs to be able to be managed and using practices. Uh, as part of ci cd, GitHubs, it becomes critical. There's also, you know, data sum fittings, AI models, uh, fitting models on a structure, messy or redundant or redundant data impacts the quality.

You know, when you get garbage in, you will get garbage out. So you need to have curated governance data sets if you're building or training or, you know, fine tuning your own models. And that also requires specific kind of, you know, data pipelines that are being, um, um, reviewed, um, and, and mature.

Another kind of anti-patent we've seen it's siloed teams and tooling, you know, ML engineers, data scientists, DevOps, and security, all working on isolation. DevOps team is not talking with data scientists that suddenly are using, you know, um, container images or dependencies that have Thomas CBEs and then becoming, uh, you know, security nightmares and these kind of things. Slow down delivery and increases the risk, as I was saying.

'cause they're not talking to each other. Suddenly they're throwing things into production and it can, you know, increase only the surface of potential attacks when opening without, you know, the, uh, all the knowledge that the, uh, current teams handling applications can, um, implement. And for this not having an observability ledger, you know, without logging metrics tracing, it's really impossible to monitor the bulk AI behavior observability.

It's critical for the technique, drift hallucinations and obviously failures. And not only on the consumption, but only on the availability of the infrastructure, as well as things like, you know, static prop engineering, you know, working prompts, ones hard, coding them, uh, you know, knowing model evolution, changing inputs and, you know, reusing a, a ton of pre confi preconfigured, um, prompts. The prompts should be virtually dynamic, continuously improved based on feedback and context, avoiding these anti patterns.

It's essential if you want to move beyond AI hype and into something that is really resilient, responsible, and production already. And one of the points that we have seen, uh, is, is this, as organizations embed AI across systems, APIs become this connecting tissue that brings everything together. A strong API strategy is essential because, well, you know, it helps you expose AI capabilities.

All this kind of, um, uh, of endpoints are reusable, well documented. And it's basically the, uh, way to being able to access remote services across distributed, um, implementations. It allows teams to orchestrate AI models with other services.

Uh, thinking about authentication databases, business logic, and we'll go on the details on how agents are working on that now, and provides the, uh, foundation for governance monitoring version control of AI workflows. If you're not thinking about API first mindset, when, you know, building your AI infrastructure architecture, you risk building AI solution in silos, that's gonna be hard to scale, test, or secure. You might lose the ability to abstract and swap models without breaking the client experience.

If you're embedding everything on the same solution models, serving framework, et cetera, it's gonna be complicated to really, you know, generate these applications that are distributed that can scale, and your applications will make it, you know, harder to integrate gen AI into existing ones, uh, applications as well as developer workflows. Now, APIs are also, you know, helping you enable modularity and flexibility so your teams can then experiment with different models behind the same interface. And we'll see a little bit of how we can help you with that.

You can also being able to apply usage policies and quotas, you know, track usage and performance by your, um, endpoint, your model, your teams, your applications, your final user, um, AI capabilities that you know, are only as powerful as your ability to deliver them. And APIs are the deliver mechanism of today's, um, ai. If you really want to work on something that's scalable, secure, and really deliver something that's like compostable.

So we have seen what not to do, what we suggest that we will give you more challenges. So how does it looks like the things that we have seen that is make, uh, that helps organizations be more successful. So we have these patterns that, you know, represent these buildings building blocks of scalable, secure, and maintainable AI infrastructure.

So let's, let's go, uh, one by one. So the first thing is, as we were saying, thinking about model as a service, you know, treat your model, um, endpoint like APIs that are self-contained, scalable, accessible, BI standard interfaces. Think about microservices where you are able to then scale them, distribute them, have them available everywhere.

This will help you support modular deployment, versioning, abstraction across model types. You're using open source, proprietary, custom made, uh, fine tuned ones on the cloud or on-prem. Think about this competence of your architecture that will help you enforce control.

Think about LLM gateways, MCP gateways, AI gateways, these competencies between clients and models to handle things like routing, retries, catching rate, limiting observability, and, and these kind of components enables, you know, vendor flexibility, research, mentoring, policy enforcement and route time and moves things away from the developer responsibility on the application side into a central point of governance, where you're then able to, um, configure and manage everything centralized. So think about this policy aware data governance layer, you know, government who can access what data, what model, and how it is being used. So it's critical for privacy, compliance and trust.

You know, you can apply rules based on user roles, you know, sensitivity levels, data resiliency, uh, requirements, or you can just block certain calls depending on the context of the, of the same prompt of the user call or where your, um, your information is flowing to. You certainly don't want to, you know, send just, you know, pi information directly to a cloud vendor that might be logging this information into their own systems. And then suddenly you are, you know, um, out of compliance add model, app model monitoring and feedback looks so you can track, you know, response quality, latency errors in real time, and then being alerted and take actions based on that.

It will help you, you know, enable from turning, uh, prompt tuning, uh, retraining rapid rollback phase, uh, based on user feedback and for and performance trends. If you've seen that your latency is increasing, you need to be able to take actions to, you know, keep with the developer and the user experience. Another thing that it's being widely used and adopted as traditional models and chatbots used as well as, uh, now with, um, uh, with agents, it's the use of vector databases and the rack path and augmented generation that's combining semantic search with LMS to enhance context without the training.

So you can use foundational models to be able to then enrich the information on, on, on every prompt. So these retrieval generation ensures that you have grounded, accurate outputs, and especially for domain specific use cases or for information that it's, you know, proprietary as part of your organization. And again, think about this data mesh architecture, decentralized data ownership across different domains while standardizing interfaces for AI consumption, being able to provide better quality to the, uh, to your models, to your, uh, your rack, uh, databases as well as your agents.

You know, this will help certainly promote discoverability of data quality governance, really at scale. Now, these patterns aren't just, you know, technical choices. There is strategic enablers that help you build AI systems that are flexible, you know, compliant, and again, production rate.

Now let's dig a little bit deeper on one of the, um, of the, uh, points and, and the companies that we were mentioning before, and where, uh, Kong really can help you with that. Um, we're talking about AI gateways and, uh, according to Garner, by 2028, 70% of the organizations will be building these multi LM applications that will require these kind of companies. Gateways are quickly becoming a critical path of GDG and AI stack.

As organizations adopt multiple Ls, like, you know, open ai, cloud, Germany, Mistral, they face growing complexity. Like which model is best for a specific task? Um, how do you balance cost latency and output quality?

How do you maintain governance and observability? An AI gateway abstracts these concerns and by doing dynamically routing requests based on policies like, you know, cost, performance risk, uh, help you centralizing logging rate, limiting access control, mobile failover versioning testing. Now the shift is clear.

We just need, um, APIs to revolutionize microservices. So AI gateways will be foundational on a scalable efficien engine AI deployments. Now we need to move away from just the gateway to this concept of the innovation factor, the AI gateway and soft, and the AI starting point provides, you know, with these basic pieces, routing, extraction, governance across models, but to drive ongoing value and differentiation, organization must boil beyond that to what we call the AI innovation factory.

Again, the whole platform, the whole infrastructure. So an API platform connects everything models, data, business logic, events into a cohesive, discoverable, and secure interfaces. So that's the idea, trying to standardize, abstract, and add everything on a, uh, full view.

So the API platforms that is deliver ledger for your AI innovation factory, turning gen AI into visible, uh, reliable services. What you will be looking in this kind of, uh, of platforms, well, you will be, uh, searching for it to be compostable, uh, structure stack, individual components, you know, can be mixed and match. Um, it needs to be human-centric.

So designed for real users. Embedded AI since into existing workflows, uh, needs to be governed and observable, implementing policy informants at every Laker data access model, usage of filtering. So you can then meet privacy, security and compliance requirements.

And again, it needs to be rack aware, rack first and context aware where you can prioritize things like, uh, retrieval of many generation per LLM with vector databases to run responses. These elements of judge different patterns that help you complement the, uh, this, because, you know, while AI gateways play a critical role in managing and scaling gen AI workloads, they're just one piece of the puzzle. The gateway may be the entry point, but true success comes from the entire ecosystem that surrounds and supports it.

It's a foundation, but real success comes from the end-to-end AI and a p infrastructure route. And this is, you know, without considering this race of argentic AI workloads, welcome to the Gentech era. Agents are just responding.

They're acting, but to act, they need the infrastructure that connects them securely to real world. So this is why, you know, most agents today are just gimmicks. You know, they're not really wired with APIs or with your own systems.

You know, they can check calendars or sending emails, so now they're getting more and more up to date. But, um, the, the, the ones that you can see is just, you know, the weather applications. Something's, um, very smaller and real, you know, AI gen, AI agencies more than just models because you might be thinking about, oh, I have a very good model.

The model's gonna be calling, you know, uh, additional, uh, companies, but it's not like that. And we have, um, we have seen this in the past where there's, uh, when you're building a gen ai gen in general, that they have a, you know, set of needs and, and, and, and, and requirements. You might be thinking about, you know, an l LM is enough, but it's not and enough.

The, uh, l LM is just the base, the, the national layer. But, you know, real agent needs, you know, prompt orchestration, augmented generation tool calling, authorization, authentication, orchestrations of all these steps, and this is where infrastructure really makes or breaks agents. They have been evolving.

Most AI bots are sticks are still stuck on level one, you know, chat only. That's what you most, most of the time see rather out in production. But the real value of these agents will become when they are merged into something that really, you know, moves into action, calling APIs, making decisions, doing things.

But the important key thing here is that they need to do it securely. So yes, RAG will give you more context. Tool calling will be doing the chat bot.

When adding human into the loop, you are becoming more mature, more able into, you know, this fully target of, uh, fully autonom. So again, AI agency more than just model. If the model is the brain, you know, APIs are the hands of your agent.

Without tools, the agent simply connect and attack. That's why your infrastructure must expose and secure those tools also at scale. So again, going back to the infrastructure, how we present things to models and to agents.

Now without secure APIs, agents are just, you know, chatting interest with them. They're really productive employees that will think of, we'll be talking about this more. Now, you might have heard about this and it's, it was, it was all over the news.

Know with agents acts, you know, the stakes go up. Prop injection candidly, databases, uh, we must treat AI a actions with the same rigor as any privileged access control flow. This is why secure execution matters from rejection is not funny in production.

It's a breach. You know, rabbi learned how the hard way, and they were not even hacked. They were just, you know, allowing the agent to, you know, the role basically.

And well, you know, what had happened, and this is where Kong can help you. This is where Kong Shines. Kong sits between your models and your APIs, between your models and your agents, between your agents and your MCP servers, between your MCT servers and your uh, resources.

LP you route enforce policies. You know, secure tools provide observability across every single step in this chain. Um, you can expose your APIs as scalable resources, you know, let agents interact safely, quickly, and consistently.

Um, Kong, uh, gateway helps you also being able to reroute and implement, you know, the multi LLM capabilities, um, through a single universal API your applications can use a single SDK. And then being able to access the full, um, ecosystem of, uh, model vendors, providers that offer you different alternatives and con can, uh, help you to do that connection, um, without having to impact your applications. Um, you also will need observability.

And here's where Con can provide you with all these different style. Being able to go on the details, on debugging each one of your policies, one of the steps within the gateway, as well as the latency, the time that you will take, and then being able to, um, serve directly on the, uh, uh, API platform or being able to export to your, uh, trusted, um, um, vendors for observability. You know, what was called, how it responded, and also what cost.

It's also adding semantic capabilities. Things like semantic caching will help you out to reduce the cost and avoiding having to hit the, uh, uh, model every single time. There's somebody asking, you know, what's the capital of France?

Uh, if you have three requests, okay, but you have, again, 3000 requests doing the exact same thing. You want to be able to improve the latency or reduced cost. Um, you can also do, um, for enforcing content moderation, um, doing seman routing, taking decision based on what kind of, uh, problem you're getting to what model would be best suited to answer that all built into com.

You can have policies and capabilities like, you know, PI ization, where you have, um, the, uh, the gateway being able to sanitize inputs and outputs before or after the LLM system. But this is the way to really build enterprise grade trust when you are able to, you know, remove all these potential issues before they're, you know, getting out, uh, of your, um, of your organization. We were mentioning about, you know, the LLM flexibility.

You know, when you want to use the best model for the job, not always go with, you know, U PT four, you can route dynamically with Kong. Uh, not all tasks needs the same thing. So you are able to then decide and design your, um, infrastructure to be able to then route to different models, uh, without having to effect the current application.

You still use the same SDK OpenAI, cloud Gini, and you can reroute, you can try then things like, um, 80 testing, you can just, you know, have failover being able to, uh, do load balancing, doing retries in case that you are either having your vendor model down as it's happening right now with, uh, some of the vendors, and then fall back to continue, you know, delivering experience to your users. Having a, um, an, uh, API platform for AI also allows you to implement things centrally like, you know, automatic, um, retrieval, augmented generation injection. You know, we were mentioned this a powerful pattern, but sometimes it's hard to do, right?

And, you know, Kong enables this kind of context injection without bordering developers. Instead of having to make every single application implement the rack pattern, you can expose different endpoints where the element generation is being applied by the fold. Well, the pipeline, so your applications can get it, you know, right away.

You are adding more applications. It will iterate every, uh, instantly, all these, uh, uh, configurations. So it's a better way to handle it in central place.

And we were talking about the agents, right? So tool colleague without authentication or, you know, access control is dangers as you know, the REPLIC team did. They did learn ERs only the right agents can access the right tools on behalf of the right user.

This is where you can then use, uh, the, uh, the, uh, central points of, uh, enforcement or, um, agent to MCP call from MCP to applications or, um, agents to, to LLMs where we can then make sure that they are, uh, fulfilling the authentication, uh, and, um, security mechanisms that you, that you require. Now, going a little bit deep on the MCP gateway, you want to have, you know, a gateway governance for MCP. You can secure and govern your internal MCP servers, you know, APIs that enables structure action, being able to then have the gateway, being able to control how your agents are, you know, uh, discovering these servers, being able to access those servers.

We to implement, you know, the developer productivity. And when you're talking about CP servers and implement EMC protocol that we are looking here, we also want to be able to leverage what you have already been building with your AI strategy and your build and your API resources and current, uh, assets, uh, in a way that is gonna be easier to then expose them as, um, as, uh, as part of your agent ecosystem, uh, making and turning every API into an entity server that is being then served agents and being exposed. If you already have, you know, develop, you know, series of APIs that are already available in your ecosystem already, you know, we will tracked, um, production ready, then you're able to then just, you know, build this interface that will help your agents being able to then access them in your natural language, um, way.

So you will be able to turn these APIs into agent ready endpoint. So automatically we can, that's where we are heading. You can have then, you know, certainly, um, a production ready UI where you're able to be enabled to see, discover, find out the models that you have, the agents that are available, the, uh, the, uh, uh, the MCP servers that are, uh, in your ecosystem, who's consuming, uh, what, what are the, uh, the, the, the biggest consumers, the expand that you're getting on, on cost and, and things like that.

And the important thing here is that we have this, you know, unified infrastructure view. This is what modern infrastructure looks like, unified across APIs, events, LLMs, MCP servers, microservices, Kubernetes. This is where Kong's AI infrastructure can help.

No Kong's platform delivers the performance of observability and the enterprise SLA to do this today through these unified platform with s of features, and one built on one of the most performing gateways over there in the market with everything, uh, really, uh, enterprise ready. I know this image, it's a little bit, you know, uh, crowded, but this is the kind of things you need to, to be thinking of when, uh, you know, uh, when, uh, designing your infrastructure. All these places where you need to have LLM providers, AI governance from management, MCP, traffic, you know, um, uh, server generation, API calling tool, calling authentication of agents.

It's really a good landscape of what you will need to think about, implement, build on top of that. Have these enterprise ready, uh, infrastructure pipelines that will help you out, deploy applications that are really, you know, what your business is expecting with the securities scalability. And compliance Con, uh, has this connect platform that is providing this unified API platform where you can build faster, cover your services easily with things like developer portal, um, service catalog, where you can find not only from the external presentation of, um, models, agents and APIs, but also from internal, you know, reference for you to know exactly which, uh, grade your APIs or your models are being exposed with.

Um, and regarding the security compliance, you know, you're able to bring them no information about where they're, uh, what's the team that is supporting them, full observability here and going the, the full range from building, um, APIs, AI models, agents, running them, secure leads, covering them, governing, and all the different kind of, uh, of, uh, of, uh, assets that your organization has. Um, APIs, ai, uh, events, uh, everything on a single platform, being able to have, uh, unified, uh, identities being able to consume. Now, we're getting close to the end, and we want to, um, leave you with all this, uh, content.

Um, obviously the idea is that if you really want agility, you really need to have this model of kind of stacks and well Kong enables you to solve these models or tools, you know, agents, MCP service without breaking your infrastructure, really leveraging what you have, the AI landscape just evolving so fast. Everything, there's, every single time there's something new that we need to be aware of, everything we discover, that there's gaps that we need to cover. If you do it yourself, you might be burning, uh, uh, burning your, uh, teams and your resources, uh, when you can, you know, support this kind of, uh, of, uh, of components with, uh, with things that have already been proven.

So you can avoid, obviously locking, increase your agility. Now, um, what you should take away, well, uh, there are three main things. You know, if you want really smarter, cheaper ai, you know, compare maybe smaller models with powerful, uh, APIs, um, Kong can help you make your stack AI ready, securely and scalable.

We went over more, different, uh, aspects of, uh, production already, but the ones that we can help you, uh, regarding things like, you know, uh, security observability, uh, governance, um, uh, scalability is where, uh, you can ly take a look at and more if you're willing more complex things like agents. So, uh, take a look at this con kind, provide you these foundational pieces to really scale your AI secure. Now, we have covered a lot, lot of content.

Uh, we know this is just the, you know, uh, a high level, uh, overview. But if you want to continue, you know, Kong has this special offer for your audience, for our audience today. So for those that are interested in learning more, you know, about how Kong can support your organization with, you know, building this agenda ready infrastructure, you can get a free custom pair of Nikes.

So you can, uh, you know, uh, uh, scan the QR code, uh, you can see here to book, uh, a call, uh, with one of the, uh, our team members. And then, you know, the second code can help you all design your, your customized sneakers. Um, you can, uh, once you, you take the call, we will send you out this, uh, in, thanks for your discussion on and for your time.

So we really hope to hear from you and well, uh, don't forget to learn more in our virtual booth. We have, uh, plenty more of content. com.

Uh, you can contact us. We're gonna be in the chat. We're gonna be, uh, able to, um, uh, to support you with, um, any more information.

Um, but again, um, we'll beate the time that you have with us, and thank you very much. See you next time.