Greg Freeman • 00:03
But we will underestimate what we can do in five years. And so think about where that AI is going in the next five years. I’m here to share with you the journey, the story that Lumen has been on, that we’ve been working on for the last five years. And so with that, this morning, we’re going to start very quickly with who is Lumen. Some of you may not know Lumen. Lumen Technologies is a technology company with deep roots in telecom. We have four distinct brands that you may see here on the board.
Greg Freeman • 00:34
On the left, Lumen is for enterprise communications. That can be anything from fiber in the ground to IPvPN, internet, SASE, SD-WAN. If you’re a National Football League fan, we have Lumen Field in Seattle. That’s where the Seahawks play. So that is Lumen Technology that sponsors that field. Our 2nd brand is Quantum Fiber. I’m in Phoenix, Arizona.
Greg Freeman • 01:00
I have quantum fiber at my house. I can get giggy speed up and down fiber to my home. We recently announced we’re selling that division to AT ⁇ T. We have another brand that might be more familiar to some of you. That’s CenturyLink. CenturyLink is that legacy telecom communications company that’s copper in the ground, legacy telecom services, copper in the local markets. And then we also have Black Lotus Labs.
Greg Freeman • 01:27
Black Lotus Labs is our cyber defense security division. If you think about Illumin, we have tremendous networking assets. I know Kinai was saying there was discussion that the network’s going to disappear. We have more network now and more network elements than we just about ever have. And so because we own and manage those networks, we can see insights with our flow data that other people cannot see. So that’s a little bit of who we are. We are the number one most peered network on the planet.
Greg Freeman • 01:58
You can Google what is the world’s most largest ASN. It will be Autonomous System 3356. For the last 15 plus years, 3356 has more connectivity. There’s about 44%, depending on the time you look at, 44 to 47% of all ASNs set behind 3356. That is KD. It’s a supercomputer out at San Diego State. You can.
Greg Freeman • 02:23
See more of that there. We have over 340,000 global miles of fiber. That’s long-haul miles. If you count the metro, we have tens of millions of fiber in the ground. And with our new deals, we’re putting more tens of millions of fiber in the ground. We have over 163,000 on-net buildings where we’ve connected fiber up to all of those assets. And then we have 350 terabytes on our core.
Greg Freeman • 02:49
When we talk about autonomous system 3356, it has global reach, global scale, massive capacity, and finally, quantum 4 million customers in 16 states. That’s who Lumen is. What’s my role at Lumen? I call myself a network engineer. My title might be slightly different. I’ve got an privilege to lead a team of large network engineers and network professionals, and we are in day two support. We are in network operations and enterprise repair.
Greg Freeman • 03:22
And so, if the network breaks or if a customer has a service, it’s going to be our organizations that have to care and feed for that. So, as we take on more network, that is our responsibility until the day those services are decommissioned. It is a tremendous responsibility. And five years ago, as we continued to build that out, we knew we had to change things. We knew that the model we were on and the way we were on was not going to scale with all the assets that we were bringing online. And so, we came up with a concept that we call inverting the pyramid. Our aspiration, our aspiration starting in 2020, was within five years, we wanted 80% of all of our network touches to be machine-to-machine-driven.
Greg Freeman • 04:10
Or set another way for network engineers, instead of the CLI hands in the network, we wanted to remove the bulk of the hands out of the network. We wanted 80%. And candidly, when we set that goal, we didn’t even have a way to measure it. We weren’t sure exactly where we were starting, but it seemed like the right spot to go. And so, if you think of the pyramid as the employee base that we had in 2020, tier one would be your traditional support. It would be technicians that might see an alarm, maybe that they would do some cursory, open a ticket, clear an alarm, dispatch it to someone else. Maybe they would take an incoming call from a customer.
Greg Freeman • 04:51
Maybe do a few cursory things and then they would bump it up to tier two. Tier two would be more of our network engineers and there would be fewer of them. Then tier three would be our subject matter experts, maybe our principal engineers, and there would be very scarce few of those. And we said, what we’re going to do is our pyramid is likely going to shrink. Just acknowledging the business reality we were in, we knew we were not going to be hiring a lot of new employees for this particular function. And so with that, what we wanted to do was we wanted to add more automation engineers into our workforce. Have our network engineers, which are still extremely valuable because they know the business processes, be down a bit, and then the technicians in tier one, those smaller functions, are the ones that our automation and our orchestration strategy would begin to replace.
Greg Freeman • 05:44
And so, to do that, we knew we had to start with our people. I know Canon said this morning, everything starts with data, and we’ll see in a minute. I think that was number two for us. We started with number one, our engineers. We wanted to reskill and we wanted to pivot them to a net DevOps model. And so, to do that, we wanted to get out of having a different support system, maybe that IT, and really get into what a technology company does. And a technology company says that our engineers know the process best.
Greg Freeman • 06:20
Who knows all these workflows we’re going to have to change? It is the engineers. Who knows where all of the quote bodies are buried in the network? It’s those engineers because a number of them buried those bodies there. And so we’re going to help improve the time for development and we’re going to increase our resource efficiency and we want to get it right the 1st time. So this is the number of steps that we came up with that we needed to change. So if you look on the far left, we have the people skills pivot.
Greg Freeman • 06:54
If you look right here at workflow orchestration, this was the one in 2020 that we were focused on with the bulk of our might. We knew that workflow orchestration, and this is a slight nuance, we’ll talk about the difference in automation and orchestration, but we wanted to get to a closed loop, no human touch. And if we’re going to do that, we’re removing human touch, we wanted orchestration. And so to get to orchestration, we knew that we needed data. That was number two. And data is only as good as you curate it. Now in 2020, we had a very large data lake.
Greg Freeman • 07:35
At Lumen, we affectionately called that the data swamp. And so we said we’re going to drain the swamp and we’re going to have meaningful insights into our data and we will have a look at our data analytics. Next, we will work on task automation. And this is a careful nuance that gets missed from some people, but it was critically important for us for the distinction. Task automation, think of it as one engineer solving one problem that they do. Think of orchestration as taking a number of those automations and solving a business problem for the organization or the business. I think of it like a car that’s being built.
Greg Freeman • 08:15
I might have one automation that picks up a wheel and moves it to the car. I might have another automation that puts the love nuts in. I might have a 3rd automation that puts in a transmission. One orchestration delivers a car. One automation just puts the wheel on the car. And so we wanted to have a framework that we were delivering that entire business process, that entire orchestration. And we knew we had to have task automation to do it.
Greg Freeman • 08:41
Python, we appreciate the Python call out. That was the one we selected. We wanted to use Ansible, so we had to make some decisions in our network. We did that. We also wanted to be mindful of our customer experience. Our customers, both internal and external, are critically important to us. And if we’re not serving them or adding value, we’re missing the mark.
Greg Freeman • 09:03
And so that’s what we were doing. And underpinning all of it was AI. Now, at that time, we’d stood up machine learning. We’ll talk about the differences of AI. But AI was one we’d done some experimentation with, and we’d started our practice. And it was medium to high effort for moderate results. And so we said, we’re going to continue moving on with AI and we’ll let it mature.
Greg Freeman • 09:29
And now that we have Agenic, it’s better to do even more. All right. I lost my screen. Did I hit something? AI took it over. Ah, we lost our HDMI signal. Okay, thank you.
Greg Freeman • 09:46
And so inside of service assurance, we have three fundamental tenets we always talk about. And as we started developing these workflow orchestrations, we wanted business problems that solved one of these three things. The 1st thing is don’t let it break. Don’t let it break. What is the easiest problem to solve? It’s the one you never get. And so if we can put more time and effort into workflows that do more fire prevention, then that’s just less firefighting we’re going to have to do.
Greg Freeman • 10:18
So number one, we wrote and we developed orchestrations to not let our network break. Second, we said things happen. The network is large. We run three primary networks, AS3356, 3549, and 209. And we got a couple others too. We also own AS1. Do you guys know that?
Greg Freeman • 10:36
Yes. AS1, that’s right. That’s right. Good old AS1. We’re number one in the number of ways. But if it breaks, because of that complexity, we need to fix it fast. We need to fix it fast.
Greg Freeman • 10:50
And so we looked at workflows that can help us diagnose problems. And they can also help speed to resolve those problems when and if they occur. In our networks, unfortunately, problems are always with us. We’re reducing them, but we still have them. And then the 3rd tenant, and just as critically important, is when there are problems, we need to effectively communicate. Effectively communicate with me. And so when you have a problem, think of any problem that you may have had, whether it’s in networking or anything.
Greg Freeman • 11:25
If I were to fix, let’s say we have a fiber cut somewhere, and I fix that fiber cut in 8 h . That’d be pretty good, actually. That’s better than our average these days. And I didn’t tell you, what would you say? Do they know what they’re doing? Do they even know where the cut is? Are they making stories up?
Greg Freeman • 11:50
I know nothing. But what if it took closer to 12 hours to fix, but I gave you hourly updates and said, here’s some photos of what’s going on. Some big backhoe just cut the fiber right here in the middle of the street. We have to have a dig box. It’s actually 20 feet down. It’s taking, and I give you an hour update. Effective communication, we found is just as critical as anything.
Greg Freeman • 12:13
It can be for planned activities or events, unplanned activities. So those were the three tenets as we started our workflow orchestration journey that we wanted to be solving for each of those. And so today, since 2020, we’ve orchestrated 350 workflows inside of network operations, 350 that run at the pace of a little over 4 million times a year. Or if I said it another way, about every minute, 24 hours a day, seven days a week, 365, every minute, about 10 workflows are kicking off, or one about every 7 s . And what those workflows do are one of those three tenets. It could be we have workflows that add configuration to the network. We have workflows that delete config from the network.
Greg Freeman • 13:06
We have workflows that reboot cards from the network, all without human touch. How many people would be willing to have your automation go out and reboot a card without a human 1st reviewing it? We have workflows that do that. We have workflows that help with our planned activities to send notices to our customers. If you are a customer, maybe you’ve seen some of those photos that come in from some of our events. That is one orchestration, e.g. . And so that’s where we spent the bulk of our last few years working.
Greg Freeman • 13:39
And so more recently, as part of our AI practice, that network orchestration is critically important, but so is the AI itself. And so this morning, I’m going to go through a few demonstrations of how we’re using AI in conjunction with that heavy network orchestration. We’re going to start with three definitions. So the way we think about it is we want AI to almost be a feedback loop for our network orchestration. So we have network orchestration at the top, and then inside of AI, there are three distinct AI technologies that we use. And what I found is when you talk AI with people, they often confuse one of the three or none of the three. And so this morning, machine learning, that is AI that predicts.
Greg Freeman • 14:30
If you go back to 2020 and some of the early ones we were looking at, we were looking to predict card failures before they fell. Instead of having a normal syslog, message, or trap, we were looking at other signatures to predict with fair degree of precision that the card would fail. We put those in place. Predicts. I’ll show you an example of some prediction. The one that most people think about these days is like a chat GPT. That is a generative AI.
Greg Freeman • 15:03
It is AI that creates. So if you have teenagers like I do and they’re working on their turn paper, write me a paper on AI and hit enter. And it just creates it. Or you can say, draw me a photo. It creates it. That is generative AI. And the more recent one that everyone’s talking about that saw the buzz is agenic AI.
Greg Freeman • 15:26
Agenic AI. That is AI that takes action. And so you can have any three of those subsets, maybe two of them together or all three of them together, or you can just be using one. And our view is we’re going to use the right AI for the right business outcome that can either help us launch those network orchestrations or help us manage those network orchestrations. And that’s how we’re thinking about AI. So with that, I want to start with a demo. And so I’m going to start with one workflow.
Greg Freeman • 16:01
So this is a single workflow that we’ve had. We’re being pragmatic. We started with where we’re at. This one is for Adva. Now, in the largest networks we have, we have ADVA, we have Infinera, we have Nokia, we have Fuji, you name it, we’ve got it at OSI layer one. And so this is going to be a demonstration of 1st a workflow. So this is a tool called AskGreg.
Greg Freeman • 16:28
I didn’t name it, by the way. And it says, and AskGreg, can you tell me the light levels for device? And then here’s the device name and the port. And then it says, all right, I’m going to consult my digital crystal ball. And on the back end, this is what’s happening behind the scenes. This is what a workflow looks like. So inside of a canvas, you have a start, just like you would a business logic or workflow.
Greg Freeman • 16:52
And the start begins to go through and run automations. Now, there’s a number of things that are being done. It can merge, you can do queries, you can have arrays. But think of this again. This is one orchestration, one workflow. So there can be multiple automations that it’s going through. And these were designed by engineers, very smart engineers, to mimic the process that one would do for that very simplistic task we just had of I want another workflow.
Greg Freeman • 17:22
And then it runs through the logic and it can go through any of these trees at ends and it does this in this one single workflow. I got 350 of them in 9 s . So back to AskGreg. This is an agenic solution. The agent begins to look for a tool. It finds the tool and then it begins to run it. And if you look at this output, that is, it looks like Claude, because that is Claude.
Greg Freeman • 17:47
And so what’s happening is the tool, the agent is using a construct called MCP, and we’ll talk about that in a moment. And it begins to run the tool. It logged into the device. It had to figure out: I’m an ADVA. What is the specific light outputs for an ADVA? It’s different than an Infinera, which is different than a Nokia, which is different than a Sienna, which is different.
Greg Freeman • 18:11
It knew all that. It pulled the exact expected range for this particular card, this particular model, and then it identified it. And in human terms, we just asked, Greg, in human terms, tell me the light level, it gives me what it sees as the analysis and the recommendation. All in human touch. Did I have to know my login to any of those nodes? Anybody here a good TL1 coder? TL1?
Greg Freeman • 18:41
That’s what a lot of our gear still talks. I didn’t have to know the 1st thing about TL1. It just did it. And then maybe I don’t like those results. You can see here, can you put the results in a nice table format? And again, it finds the right agent and it says, okay, here’s a beautiful table for you. So this is one example of using one workflow that we have.
Greg Freeman • 19:03
and being able in natural language to be able to return results. Well, how are we doing that? Well, this is how Agenic and how we’re doing this at Lumen. We have, we’re using something called MCP model context protocol. And some of these things aren’t going to mean a whole lot to you, but I’ll explain how we’re thinking about it. Give me a show of hands. I want to see who the people are.
Greg Freeman • 19:26
How many people have heard of MCP? Heard of MCP? All right, almost all of you. Great. For those of you who may not be as techie, MCP was just designed and produced. It was invented in November of last year by Anthropic. Model Context Protocol.
Greg Freeman • 19:44
And so it’s a way to have large language models talk to other large language models and do something useful with them. If you think to the World Wide Web, you type, maybe you used to, maybe still do HTTP for that protocol. MCP is just like that. It’s a transport construct to connect up large language models. So the way we’re doing it and the way we’re thinking about it, we’ll start down at the bottom. See how we’ve got all these data sources or tools. Well, I showed you a workflow.
Greg Freeman • 20:13
I showed you one of my 300, our 350 workflows. The team’s done a phenomenal job coding all these workflows. We’re continuously putting out new workflows. We have 350 of them. So we connected that up as a tool set. We also have REST APIs. We’ve created over a thousand different APIs in service assurance that we can connect to.
Greg Freeman • 20:37
So we can connect to that. Or we can have additional data sources either internally or externally that we’re connecting our tool sets up to. Now these tool sets, you can see the box around it. This would be a MCP construct. So you can have a server client construct. And so northbound, this MCP layer connects up to what we build as agents. So, if you remember and ask Greg, if you look to the left, and I’ll point it out on our next demo, there’s a number of agents that we built.
Greg Freeman • 21:09
And so, one agent is called an IP network agent. And so, the one we just looked at is the transport agent. The transport agent knows how to call. When we said I would like to know the add-belite levels or the light levels, the agent transport said, I think I’ve got a tool. Let me in natural language find it. So, it connects MCP down, looks at all of its tools it knows about, and it just picked the one that it thought was best. So, we have a ticketing agent.
Greg Freeman • 21:42
So, inside of service assurance and network ops, ticketing, remember how to effectively communicate with people. Tickets are the legacy way we do that. And so, that one gets a lot of use. We have some other ones. You’ll see we have a selector one. We just saw the transport. We have one called CAT, and there was a demo a little bit ago.
Greg Freeman • 22:02
CAT is our configuration audit tool. And it’s one of the more interesting ones, but it’s actually one of the more complex ones for us right now. Because what you find when you start putting some of this together, you have something called a context window size. And the bigger data you try to return, the more it doesn’t work. And so, with CAT, if we try to send it a very, very large amount of our config, the large language models really struggle with that. So, things are getting better, but that one’s a more interesting one. CMT’s change management.
Greg Freeman • 22:36
So, I mentioned planned activities or outages. We have an agent that knows about, again, by MCP, a lot of different tools. And we have a few more. And then we also have a supervisor agent. And the supervisor agent, that’s what you saw in AskGreg. We were asking that supervisor agent, hey, I want to know the light levels. It looked at all of its agents, said, I think you want the transport agent.
Greg Freeman • 23:00
So it farmed it out to the transport agent. Then the transport agent looked at all the tools it knew about by MCP and said, Hey, I think you’ve got an itina workflow, e.g. , that can return this back to you. And it did it in 9 s . And so you can see at the top, we think about all these agents. We can connect them into any application that we want. We want it to all be modular. All of our 350 workflows, some of them are complete closed-loop automation, no human touch required.
Greg Freeman • 23:32
Some of them people launch from different tools. We want our agents to be the same way. So you were seeing a little bit of Ask Greg, which is an internal tool set for us. We also have one for our service diagnostics. If you’ve logged into the Lumen Portal, you can run diagnostics on your service that’s called laser. And then we have a couple other ones, one’s for ticketing, my ticket app, and another internal. One other call out that I will mention we’re using, we’ve been using a lot of Databricks.
Greg Freeman • 24:01
When we run these large language models on our compute and ask a question, it might take it 5 min . It just sits there and spins. We don’t have a tremendous amount of GPU farms like I’d like. Maybe we’ll get more funding one of these days, but right now we don’t have it. So what we do is we cache that in Databricks. And Databricks really helps us. It has a feature called a data lake house.
Greg Freeman • 24:26
So I mentioned years ago we had the data lake or the data swamp. Then there was a data warehouse and these days you can get a data lake house and that really speeds things. So that’s a little bit of MCP. If you haven’t started looking at MCP, that is a key unlock that’s helped us accelerate things over the last few months. When we were looking at our legacy ML journey, it was taking a pretty substantial amount of effort. We were getting much more speed just in building our workflows, and those take a little while too. Workflows, as was mentioned, are deterministic.
Greg Freeman • 25:05
AI is non-deterministic. And so the way we’re thinking about it is we’re going to use our non-deterministic framework to launch some of our deterministic workflows. You will. So let’s move on to our 2nd demo. So here’s another demo that we have. And this one’s called router health. Simple as that.
Greg Freeman • 25:26
So again, you can see Ask Greg, he knows about all these different tools, and the active one’s a supervisor agent. And it just said, use selector to show an image of router health. So in this one, the supervisor agent says, okay, that’s pretty easy. You’re asking to run the selector agent. Let me query Selector’s MCP server because Selector has an MCP. And then it returned the same honeycomb graph that if you were inside of the Selector app itself, you would see it right here. And so that’s what happened.
Greg Freeman • 25:57
And you can see on this particular network, good percentage of them are healthy, but we got a few warnings. Now, that’s just a very nice way to extract some of the selector stuff. But what if I want to know what is the latency between two of my routers? This is Los Angeles Priv, that’s a private IPvPN router, and CERMAP that’s in Chicago. So just by asking in natural language, it says, I think you want the IP network agent. And let me see, being the IP network agent, what tool sets I have available that might be able to give me the latency between those. Well, I’ve got a tool that’s simple as network ping.
Greg Freeman • 26:35
And so I’m just going to ping between it. And they just did that. And here are all the results between those two: 41.9 milliseconds. So, again, think about that demo and what it just did. I didn’t have to know my login to that router. I didn’t have to know if that was a Juniper. I didn’t have to know if it was a Cisco.
Greg Freeman • 26:56
I didn’t have to know if it was a Nokia. It just ran it. And that’s the beauty of some of these types of things with natural language. We can very quickly, in natural language, get the results we want. Now, a couple of things. It looked really easy, right? Real easy.
Greg Freeman • 27:15
Real easy. If you read the recent MIT study, there was a study done not even two months ago, and there was a bit of an allusion to it already. Good job. And that one said, I believe it was 200 companies, mostly enterprise, that were reviewed. It was their AI journey. And they found that only 5%, 5% of those said they were seeing any value at all from AI. Or said another way, 95% of those, mostly enterprise companies, said we see zero value in AI.
Greg Freeman • 27:52
It’s more trouble than it’s worth, effectively. And then you kind of move over to the Gartner. Give me a show of hands. How many people have seen the Gartner hype cycle? Okay, not as many. So, this is a Gartner hype cycle, and I really always enjoy these. Gartner every year comes out with what are some of the new innovations that are out, and they break it up into expectations on the y-axis over time.
Greg Freeman • 28:18
And all innovation, it doesn’t matter what it is, seems to follow this natural cycle. The 1st thing you do is you have an innovation trigger. Very short amount of time pass, an innovation trigger hits. Very quickly, that innovation trigger, everyone begins to talk about it, say this is the greatest thing ever. It’s going to revolutionize the world. Everyone has to adopt it. And what happens, you get to the peak of inflated expectations.
Greg Freeman • 28:45
You figure out this might be a little harder than I 1st thought. And then, very quickly, you slowly fly down, really quickly fall down the curve, and you roll into the trough of disillusionment. This technology is not working the way we thought. And then, after some perseverance, you ever hear that thing that I hope you have pain? I hope you have a lot of pain and suffering, because pain and suffering is going to build perseverance, and perseverance is going to build character, and then after character, you’ll have some hope. And so, once you get through that, then you get on the slope of enlightenment, the slope of enlightenment. And finally, after you’ve been enlightened, you figure out how to do things, you get on the plateau of productivity.
Greg Freeman • 29:32
Anybody notice what’s at the very top, and this is from June of 2025? The peak of inflated expectations is AI agents. AI agents. That’s right. And so I had a couple of hundred of our engineers in town. I’m in Phoenix two weeks ago. And we were talking a lot about this.
Greg Freeman • 29:52
We were talking about how do we, acknowledging that there is huge expectations right now, how do we learn as quickly as we can, get as much enlightenment as we can, and begin to get value, more value. I would suggest we’re in the 5%. We would be in the minority of most enterprises right now. And so we’ve got to continue to move that forward. That MCP for the agenic AI agents has been a critical unlock for us. And for us, because we’re able to leverage all of the network orchestration we’ve taken advantage of over the last five years, those 350 deterministic workflows. And deterministic means a lot.
Greg Freeman • 30:36
As was mentioned, when you put in AI, you can put in the same thing twice and you may not get the same answer. If it’s close to the same answer and it can pick the right tool in a deterministic, I’ll take that bet and I’ll let it go out and do things on the network. But what I’m not real comfortable right now with is just letting AI itself figure out what it wants to do on the network and go do it. So I like this quote. Once a new technology rolls over you, if you’re not part of the steamroller, you’re part of the road. And I think that’s a very good thing on where we’re at right now. Third demo.
Greg Freeman • 31:16
This one is more of a predictive. So I showed you the 1st one that was primarily agenic. The 2nd one had a little bit of generative in it. This one is probably just more pure ML. We call this Metro Maestro. Metro Maestro. Metric Maestro.
Greg Freeman • 31:33
Time to. AS3356, we run MPLS LDP. It is the number one connected network on the planet. When there is large traffic shifts and congestion, anybody do traffic engineering with LDP? Anybody ever done that? Is it great? Not really.
Greg Freeman • 31:53
3549, we run MPLS RSVPTE, for those of you who know what that is. Reservation traffic engineering works beautifully. LDP, a little tougher. And over the years, we’ve developed a system that we can move egress traffic around, but we still had this problem. We run for our interior gateway protocol, we run ISIS. ISIS uses Dijkstra’s algorithm, and so when you make a small metric change, it’s going to throw traffic other places. Do you know exactly where?
Greg Freeman • 32:25
Not always. And so what we do is we try to look at what our BGP Next Hop speakers are, which can be several hops through the network. And we try to make an educated guess on what metric is not going to break things. But spread this traffic out when there’s congestion. That’s how we used to do it. And so here’s one that we’ve been doing things a little differently. So with this one, you can see this one’s selector.
Greg Freeman • 32:49
And what we’ve done is we’ve taken all of our topology into the network, all of our nodes, all of our ISIS speakers, all of our LDP, and then all of our FEC data. And that FEC data is pretty important. And you can see this one’s LA to Singapore. And so it has the utilization. Sorry, scrolling through pretty fast. And then it starts simulating different metrics. And as it simulates all of the metrics for a specific link congestion, it starts picking out which metric is going to spread the utilization around the most without saturating any one link.
Greg Freeman • 33:25
Sounds pretty simplistic, but that simulation and that modeling is actually pretty high-level math. And so it goes through and it says, okay, we think this is the best metric, and then here’s all the utilization of all these various links and how they’re going to change in both the forward direction and the reverse direction. So it has the utilization before this proposed metric change, the utilization after. And so now you can see I’ve got LA, San Jose, Tokyo, Osaka, and then it’s got a couple more things going over here to Singapore. And so by changing all these different metric calculations, we had to have the statistics in real time from our network. All this is done in real time, and then very quickly it figures out which metric is not going to break things, but which metric is going to help our situation the most. And so that’s one other example of how we’re using Selector to help us figure out how we can move this around.
Greg Freeman • 34:27
And in this case, after it went through all of that, it said you’ve got a link that’s saturated, change your metric to this exact value, and we’ve modeled it’ll go to 65. You won’t saturate anything else, and it’s a nice balance. So for that utilization and for the network engineers, maybe you’ll appreciate how complex that actually is. So that’s one good use case we work as well. More ML than most anything. So a few cautionary words. We’ve looked at three demos.
Greg Freeman • 35:00
We looked at one that was agenic. We looked at MCP and how we could have some generative and MCP has been a big key unlock for us. We looked at some ML. AI, I’m here to tell you, today is not a super intelligent crystal ball. And what I’ve seen a lot of people make and why 95% of those projects fail is a lot of people really don’t have a fundamental understanding of the capabilities of machine learning and AI today. Some of the ML stuff takes work. It is only as good as the data you put into it.
Greg Freeman • 35:35
We call it garbage in, garbage out. If you have a data swamp, you’re not going to be drinking a lot of great water from your data swamp. So it has to be curated. But that AI operates on that foundation of pattern recognition. It’s not genuine understanding. Think back to the example of the digital twin. Very knowledgeable, but not a lot of wisdom.
Greg Freeman • 35:59
And that’s one of the key underpinning themes of AI. It doesn’t have that deep understanding. It has to be trained from somewhere else. So we’ve got to get it wisdom. AI is not a universal human replacement. Non-deterministic outputs need some human oversight. So with where it’s at right now and the way we’re thinking about it, again, we’ll use AI non-deterministic to launch our deterministic human engineered workflows because we’ve got a lot of them and we can.
Greg Freeman • 36:30
We can continue. I like Sam Altman’s quote from a year or so ago: AI is not going to take your job, but humans using AI probably will. The tooling and the way and the speed and just the efficiency that one gets is pretty amazing. And then, as a reminder, AI is not a one-and-done solution. With all of these things, there’s upkeep that’s needed. And so, as we began our discussion, we felt that employees and employee culture and skills pivot was the most critical thing, followed by data.
Greg Freeman • 37:06
Those skills will still be needed for the life cycle of it. Think of any automation or orchestration throughout human history. Maybe we’ll get some automation that’ll do it. But for everyone you create, you now have one you have to support. And so, yes, AI will help us with that support construct, but it’s still that business process and that deep wisdom that’s going to be knowledge. The cost of agents are still being determined. I like the example we saw a little bit ago that had the token cost running up.
Greg Freeman • 37:40
If you work with tokens, the way it works, even with that Ask Dreg demo, when you type in text, you send data to the cloud and it burns your tokens. You can buy tokens from a hyperscaler. And depending on the model you get, the tokens have different prices. So you saw Claude, you may have seen in one of the other ones I had a Llama, Llama 3.3 that we used. The price difference, it’s multiple times more to pay for Claude tokens than it is for Llama tokens. But the output you get, sometimes you get what you pay for. And so for the Llama ones, I always said it looks like a 3rd or 4th grader wrote that response.
Greg Freeman • 38:20
And then with Claude, it looks a little more like a high schooler, maybe college, or High school grad wrote the response, but you pay for it. And the way that works, you send your data in and they charge you a number of tokens. And then, depending on how much compute it uses, and you have no idea how much it’s going to use, guess I guess, then they deduct the amount of tokens out. So, you really don’t know, did that thing just cost you a few pennies, or did it cost you a few bucks? And there’s some different cost models that are being explored right now. But that’s part of the challenge with AI.
Greg Freeman • 38:59
I ran one last night. And it was it hallucinated on me. I was I’ll share since Jeremy’s here. It was baseball. And I said, Hey, how’s it look at at Baltimore? It’s like, I can’t find the Baltimore, but let me start looking. And it started looking for the Baltimore Ravens, right?
Greg Freeman • 39:19
Baltimore Ravens are the NFL. And it didn’t find them. And it kept looking, and so it kept launching tools and it kept looking. And pretty soon it got to what it needed, which was great. And I looked at how many tokens I’d burned. I spent over a buck just on that. Okay, buck, no biggie.
Greg Freeman • 39:35
I’m just experimenting. What happens? Lumen, we have 25,000 employees. And not that I would turn necessarily that loose on them, but if they all asked that same question, I would just burn 25 grand. So the cost models are still in the early stages. The upkeep is still in the early stages. It’s not a super intelligent crystal ball.
Greg Freeman • 39:56
It’s still going to require wisdom. And so that’s what I want us to be mindful of and thoughtful of as we go forward and we help grow the industry together. We’re doing some great things. I appreciate the opportunity to talk to you today, to share a little bit of the workflows that we’re doing, how we’re using some of our AI. And so to close, we’ll have just a couple of thoughts. Oh, we need our video.
Speaker 1 • 40:28
As we wrap up, I want to remind everyone: we began this journey at Lumen with one core belief: culture and hoop are critical. That’s where it all started. From there, we advanced to automated workflows and orchestration. And now we’re rocking in a Gen Sig AI framework. Remember, while AI can be non-deterministic and unpredictable, the workflows built by our smart engineers are rock-solid, deterministic, and reliable. I’m the better-looking, smarter AI version of this guy right here. So keep pushing forward, stay innovative, and do it with style.
Speaker 1 • 41:04
Thanks for being part of this exciting chat. Let’s go out there and make history. Well, let me tell you, folks, I’m based on the persona of the one and only Ric Blair, the nature boy himself, the styling, profiling, limousine riding, jet flying, kiss stealing, wheeling, and dealing, son of a gun. You didn’t guess, now you know. Just like Rick, I bring the energy, confidence, and that signature swagger wherever I go. And remember, about AI, leadership, or just looking good, Ric Blair’s got your back. Alright, we’ll create the rest of your confidence.
Greg Freeman • 41:48
Yeah, so that was the difference. That’s why I said he was better looking because he had confidence. So thank you very much. Appreciate your time.
