GTT Fletcher Keister on the Three Pillars Needed to Achieve AI Goals

ai Leadership Insight series. I'm your host, Mike Baard. Today we're with Fletcher Kester, who's chief product and Technology Officer for GTT.

And we're talking about the three pillars of AI readiness because, well, it looks like maybe we finally learned a thing or two. Fletcher, and welcome to show. Thank you very much for having me, Mike.

Happy to be here. All right. In some ways, uh, we, we seem like we're incapable of learning.

We keep making the same mistakes over again. And I got a feeling AI's not that much different, but it seems like, well, yes, we need infrastructure, we need data, and we need some actual software. But everywhere you turn it looks like people are struggling with these issues or they're not quite prepared.

So where are we on this journey right now? And have we got to the point now where maybe everybody's kind of figured out, well, at least I got the core components of something that I need to build with? Yeah, it's a great question and well, when I step back even a little bit further from that and think about it, where we are in the journey, you know, I think about the, you know, you think about any new technology that's come out or any new promise of a new technology to make life better in some way, shape or form.

I think we do fall into that trap of forgetting about, again, those foundational things you have to do to be prepared to leverage that technology to get the outcome you're looking for. And, you know, a key couple things in that is the technology, you know, the AI technology in and of itself doesn't solve problems from my perspective. It helps accelerate our ability to get to an answer faster, but it still requires all the things you mentioned, uh, you know, that live underneath that in terms of, you know, infrastructure and data.

And probably most importantly that I see, uh, we forget most of the time is what are we actually trying to accomplish with this technology? And if I were to really simplify, hey, that gets lost in the excitement, uh, of the new technology. And you have to stay focused on what outcome you're trying to drive and what's the business value you're trying to create.

Hmm. And to your point, I've heard these tales where somebody has decided to go build an AI model, they'll automate a process and then they determine if they put it in production, it's gonna cost millions and millions of dollars to replace a function currently handled by two people who are making, you know, 80 grand a piece. So I guess the question is, is um, do we really understand the cost of these models and what it's gonna take to run them in production environments?

'cause I think maybe we need to work back from that to get through our use cases. Yeah, it, it's a great point and as we've thought about it, uh, and I've thought about it some is the creation of the models themselves I think are best left to other people to build and develop and to really think about things for us in terms of a framework of capability to where we can use and leverage the models that other people are building. Again, for use case specific business outcome specific drivers that we have where we're trying to produce a very specific business outcome.

And I think, you know, our intention to stay incredibly focused on, I'll, I'll say over and over the business outcome rather than a science experiment because all of those millions and billions of dollars going into the creation of the next new model, next new capability, right? It still has to serve its purpose and serve its function in terms of driving value for somebody. And is, correct me if I'm wrong here, but I also feel like there's a, a new respect for data management.

It's not like we haven't been doing it for three or four decades now, but people seem to have a greater appreciation for what it takes to get the right data to the right place and into the right AI model to drive the right output. So maybe one of the benefits is we're finally learning this lesson. Yeah, I absolutely think so.

And you know, back to the point you made earlier, it's like the lessons that we thought that we learned but somehow we forgot, uh, and not having to relearn again. And you know, we've always known the importance of data and there's a lot of work over the course of the last, you know, 10, 15, 20 years to better organize our data and to get them into data lakes to get them into, you know, into the cloud where we can do more things with our data. You know, but again, with the advent of AI and the models that have been created, like we're just talking about, is I think it takes an even el more elevated view of data engineering.

Not just do you have someone managing your data, but you're actually thinking about your data and how you're engineering your data and how you're structuring your data and how you're relating your data to other data to give that opportunity for the model to actually go and look for and get the right the answer that you wanted to produce. And I think that's again, more of an elevated function as we think about data and data engineering versus data management. It also seems to be a fierce debate these days about where these models are gonna run.

And on the one hand, we seem to be training the models in the cloud 'cause that's where there's GP and makes it reason assumptions. But uh, some folks are saying the cost of running the inference engines in the cloud is just too expensive and you're better off doing it on premise. 'cause if you're running in the cloud, you're gonna get hit with token costs on the input and the output and before you know it, uh, things spiral outta control.

Yeah. So it's the, we call it the magnified cloud challenge, meaning, you know, again, for the last 10, 15 years, everyone's put, put their data application into the cloud only sometimes to experience a elevated cost. That was more than expected.

Now just think about all of that activity that AI is gonna generate. And the movement of data just really compounds that problem statement. And so, you know, really take a view, and I've got a perspective that, you know, it's going to be both the best place to run some of these models, you know, from inferencing to training is going to be, you know, whether the application specific than some of it will be better served running on premise, some will be better served, and what's now the growing private cloud function that, you know, carriers like GTT are starting to, you know, bring up, bring to market as well as some things better served running in public cloud or even some of the neo clouds, you know, um, like openwave, uh, and companies like that who are doing more with GP farming.

Right? And I think as we, as we learn, just like we've gone from LLMs to SLMs, and they'll probably be extra small LLMs, you know, at some point in time is that, you know, the application's gonna drive, how much compute does it need, how much stores it need, how much, you know, heavy weighted model does it need? And what data is required and where's the best place to run that?

And, and I think it's gonna be distributed just like cloud is becoming distributed or an AI model is best run is going to be, you know, outcome specific from premise to hosted to public. We also seem to be suffering a little bit from, uh, being overly attached to the latest and greatest GPU processors. And what seems to be happening is everybody says, Ooh, I gotta have the next one.

This is awesome. And then they forget about the previous generations, but I can't find the latest generation 'cause well, there's just not enough of 'em being made and everybody wants one. So, um, do we need to get smarter about what GPUs and for that matter, other processors we might be using to run what?

Yeah, I think so. And that kind of ties into how we thought about it and how we are implementing our, our AI set of capabilities at GTT is really think about things as frameworks and not explicit decisions and a framework being, you know, we talked about the three pillars of, of data, uh, of AI frameworks and of AI factory. And all of those really at the end of the day are, are simply models and constructs.

And the idea around that you're bringing up relative to GPUs is like, again, what are we trying to accomplish? What are we trying to do as we've deployed an AI factory meaning just, you know, infrastructure and GPUs in a couple of our data centers. We've done it in a way where the architecture is not dependent upon a particular GP provider, right?

Because the very reason you mentioned, you know, some of the higher ends, they're hard to get hold of, right? Because people are, you know, gobbing 'em up and they're not them on the market. And other companies outside of Nvidia are bringing more and different chip sets to market.

And we want to have the ability and flexibility to say, okay, we're going to use this particular type of GPU for this workload. But you know, guess what? For some of these smaller, you know, smaller, uh, scale applications or inferencing, we don't need all that horsepower, right?

So we can plug in, you know, a different GPU into the, into the at AI factory, uh, and to be able to manage costs that way from an infrastructure perspective. Mm-hmm. The other thing that seems to be an obsession too is, you know, we gotta go big and everybody wants to build LLMs and I can't help but wonder, sometimes I look at some of these smaller language models and they seem to be more accurate, consume less processing.

So maybe we should be thinking maybe small is beautiful. I definitely agree with you in that regard. And I think about the work it takes to build an LLM and for most enterprises, you know, 'cause we, we may be a service provider, but we are also a large enterprise.

You know, one is the, the cost and the time it takes to build an LLM and the expertise it takes in a, in a world where offers for that type of talent resource or off the charts in some cases, right? And it again, overweighting the problem statement, uh, or overweighting the, the answer to the problem statement by building large complex models where, you know, again, when you think about it in terms of frameworks, which is how we've tried to, again, to build our, our ai ai uh, architecture is a framework that we can plug in and pull out LLMs or S SLMs or whatever size they take, you know, with more, with like a, you know, really and LLM gateway sitting on top that really what transacts with it touches that first before it touches an LLM so that you can have flexibility and you're not stuck, oh no, I picked a large scale model and I really just needed a small one. Or I can plug in four more small ones because they're more fit for purpose what I'm trying to accomplish.

And I think we all need to be open-minded that just because the first things we saw come out were LLMs doesn't mean that's the, that's always going to be the best tool in the toolkit to pull out to solve a problem. I also feel like we're somehow back in the mainframe era of AI and everybody seems to think that we gotta build some massive data center, but, um, we invented distributed computing for a reason. So do you think that as we go along, we might make greater use of distributed computing both, and not just for training, but for inference and we can get smarter about deployments?

I absolutely believe that to be the case, and it goes back to what we were talking about a few minutes ago, is we are going to learn over time called the next six months, 12 months, 24 months, that, you know, there are far, there are more places we can deploy these models and run and run our use cases than in, you know, the large, you know, public cloud data centers, you know, at scale. And that we will see more and more deployments inside of enterprise networks at their premises, in their own data data centers or somewhere that the places that work in between the customer premise and public cloud. I think that will take a little bit of time and a little bit of learning and some trial and error.

Uh, I think that's what we, we've seen over and over because, uh, the pendulum swinging from distributed to centralized back to distributed has taken many forms over the course of, you know, the last few decades. I think we probably are in that the pendulum swinging one direction and we'll probably come back a little bit more. So what's that one thing you see people doing today when they make these AI projects and initiatives go, um, that makes you shake your head and just go, folks, I wish we could be a little bit smarter than that.

Uh, I'm gonna answer that question a little bit tongue in cheek, uh, and then I'll try to try to be a bit more serious about it. And I think the mistakes that I see is are when leadership in, in the enterprise, and I say leadership from board level to CEO to executive suite, you know, see all that's happening or, or they hear what's happening in the world of ai and the reaction is we need to have that here, but there's not a clear sense of, well, what are you going to do with it and what outcomes are you trying to drive? I don't know, but we need it, right?

And it can be that, that I would say knee jerk reaction, but that quick reaction to knowing that there's something there, but not having a clear path and a plan, I think no matter what new technology emerges, you know, what, you know, we're now in a, in a, you know, you know, pretty significant seed change of capability and what, and something else will come after this probably. 'cause it always does, you know, it's really staying grounded and focused on, or where people make mistakes, I should say, is they don't stay grounded on those very simple things of what's my vision, what's my strategy, and what's my plan execution for the purpose of the business that they're in, versus simply bringing in a new technology, a new capability. I think that's the mistake of, you know, you know, of having, of having an intention before having a strategy.

Mm-hmm. And then I'd love to get your opinion on this, but who's in charge of these AI projects these days? And I asked the question because originally it felt like, you know, somebody put together a TIGER team and they said we're gonna run ai, but increasingly, especially with inference engines, and as we try to do this stuff in scale and data engineering, I wonder if this whole thing is just gonna move back to traditional IT and CIOs because well, they have the experience and the knowledge to make it work.

Yeah. This could be a whole hour of conversation in that one question. Uh, and it's a great one of like, you know, how is this going to change the nature of organizational design and structure, right?

Is a really interesting question because, and I'll try not to jump around too much, but we'll try to keep it to the point. But, uh, I've always had this belief, you know, just personally is that the most successful future business leaders are going to be the ones who can straddle the worlds of operations and technology, right? And operations being whatever, whatever functional role that may be from sales to product, to marketing, to, you know, operations to other functions, even, you know, legal and finance, but didn't know how to bring those two worlds together and then can use the technology to drive a business outcome.

I've always, I've had that view for a long time and even more so now, and I think if we, we broadly allow it to move back into just the technology organization, we will miss the opportunity that this really represents. Because it's really more about, and I I could also say where I see people making mistakes with this technology is just seeing it as how do I make my function better and how do I improve and make it, you know, more efficient, more effective, produce more value when in reality as I, as I learn and understand more things about, especially like agenda AI, is we should really be rethinking the entire, the entire notions of how organizations are constructed and moving more towards broader generalists that 'cause when you have a world where agents can take action, you know, and are not, and don't have to live inside the boundaries of a organization structure, that opens up your mind to thinking about very different ways you can deliver your service, your product, uh, what you bring to market differently than what you're doing before. So, um, I co-mingled a couple of different ideas or to, so to bring it back to your question is, you know, what I'm hopeful of is that people will be able to work more collaborative collaboratively across functions in an organization to get the most value out of the, the AI capabilities that they're deploying for the company.

And to be open-minded enough and mentally agile enough to think about how things work differently. I think if the companies that can do that, I think we'll go a lot further and a lot faster in terms of getting value from, uh, the new tools and toolkit that we all have. All right, folks, you heard it here.

Two things can be true at the same time, even if they're fundamentally opposites. One is AI is gonna change everything out related, and we need to rethink our structure. Two, there's no substitute for remembering your IT fundamentals.

Hey Fletcher, thanks for being on the show. Well, you're welcome. Thanks for having me.

All right. Thank you all for watching the latest episode of the Techron AI Leadership Inside series. You can find this episode and others on our website.

We invite you to check them all out. Until then, we'll see you next time.