Guy Daniels, TelecomTV (00:13):
Hello, you're watching the AI Native Telco Summit on TelecomTV. I'm Guy Daniels and in today's program we are going to look at how AI capabilities are being applied to telecom's networks, particularly 5G. And joining me now to explain more is Paul Miller, Chief Technology Officer of Wind River. Hi Paul. Good to see you again. Now lemme start by asking you, how is Wind River utilizing AI to address the challenges of this increased complexity we see in 5G Networks?

Paul Miller, Wind River (00:47):
Well Guy, thanks for having me here. Great. Looking forward to this conversation. I think AI has really exploded on the scene with respect to 5G networks. We're seeing it used in a lot of different ways, whether you look at the radio interface and doing complicated AI based functions like dynamic beam forming and energy management. These are things that are really helping the OPEX for a service provider as they look at energy consumption across their network and the ability to support high density devices. Our company, however, is more focused on O Ran and O ran infrastructure. That's kind of the role that we play in 5G deployments globally. Whether we talk Verizon or Vodafone or many other carriers and we're seeing me, we're seeing the real challenges with managing a disaggregated network. You have many different vendors coming together in a modern 5G virtualized architecture from the hardware platforms that we run our operating systems and virtualization technologies on to then multiple applications from the core of the network to the edge of the network and managing all these vendors software solutions, the disaggregated aspect of the network and then the geod nature of it where you have upwards of tens of thousands of sites being globally distributed across carrier's footprint.

(02:02):
You can see the picture here. It's an incredibly complex network with multivariate problems in it and to operationally manage that is a significant challenge. And so that's one area that we're seeing AI start to take hold. We obviously have the radio functions as I mentioned earlier, things like the real-time Rick and the SMO doing that, but in network operations is where we're seeing a tremendous benefit. One of the things that really slowed the RAN adoption curve down was the complexity that I just mentioned and AI is having a significant impact. We're very hopeful for what it will do, enabling lower skilled labor to manage such a distributed network with such complexities and that's really enabled by AI really presenting for us a solution for automation. This is really about software automation. That's what AI does for you. It simplifies the interaction with these complex network elements and acts as an automation engine to interact with systems that are below the user and typically in the past these would be engaged with things like scripts and dynamically building tools and automation within the service provider's catalog of assets, very complex, taking days to weeks and many times even months to build these things to see what's going on.

(03:19):
AI can do this in seconds to minutes, so we think it has a tremendously positive impact to network operations from the perspective we have in O Ran and 5G

Guy Daniels, TelecomTV (03:28):
Now there's huge interest around ai, but why is it important to distinguish between AI and machine learning for telecom use cases?

Paul Miller, Wind River (03:38):
Yeah, I think it's really about the application and what customer problem you're solving and also where in the network different functions are happening. If you look, machine learning has been used for years in the far edge of networks, and this is really around functions like predictive outage avoidance or predictive analytics, things like event correlation or root cause analysis. These are areas that have been used for some time where machine learning has tremendous power. If we talk about anomaly detection, for example, where you may have a network operating for weeks, and we all know in a service provider network you have time of day and day a week behaviors, right? Particular network bursts during busy hour, particular quiet periods in the network, the difference on a day-to-day basis, weekends versus weekdays for these type of time cycles. These create patterns in the network with respect to traffic and systems operation and machine learning can be used to look at these logs and events and identify when an anomaly occurs, and these anomalies can be used to predict network outages before they ever occur.

(04:47):
So that obviously is tremendously attractive to a service provider because they can potentially avoid problems rather than reacting to a problem when there's an outage, they can prevent it from occurring. So machine learning is a great tool, but people often confuse that with artificial intelligence. Artificial intelligence is a step above that. Many people have seen chat GPT and the ability to interact with an AI large language model. This is more on the AI front where you have an artificial intelligence trained on a particular model and that training typically occurs in the center of a network or in a public cloud based on a specific data set. And then AI inference is used to use that model and leverage it in a live manner in the deployed network, oftentimes done with a human language interaction with the AI to ask it questions about things and have it automate and report back on network events. That's a bit different than a machine learning tool that has a much more specific purpose and may be trained with a neural net but does not use a large language model or real AI functionality. They're kind of two separate technologies, but they get confused for people.

Guy Daniels, TelecomTV (05:59):
Well, let's go deeper into AI then, Paul. Tell us more about natural language interaction and gen AI and how Wind River proposes to use these for network management.

Paul Miller, Wind River (06:11):
Yeah, thanks guy. I think we've actually got a really interesting demo that I would encourage people to look at on the telecom TV website where we actually show a live demonstration of interacting with a large language model. AI in an O ran 5G network and it's pretty exciting to see what's possible here. We do use oftentimes specialized hardware technology from Intel such as their Intel Xeon family, as well as Intel gouty for AI acceleration. These type of things are really critical at the core of the network In that demonstration, one of the things, we did a pretty novel idea. We took the AI and we trained its model on a set of APIs, and the reason why this is innovative is typically when you interact with an AI LLM such as chat GPT, it's actually a relatively static dataset that it's operating on. It's created a model based on, in the case of chat, GPT being trained on the corpus of data that has come in from the internet that can be publicly available and then answers and creates answers for you based on the questions you ask from that model.

(07:19):
The difference that Wind River has done with their novel approach is we've taken and trained the AI on APIs themselves, and therefore when you ask it a question, what it's actually doing is constructing a set of API calls that are fired down into the live network to gain the data back. Now, what this means is rather than talking to an AI LLM on a specific fixed model or dataset, which is at a point in time now, the AI is used to construct a dynamic actions in the network that can give us information back from that live network. So say for example, I've got a distributed cloud system, I'm curious if I have any Kubernetes certificates expiring on the east coast of the us. You can literally ask at that, do I have any certificates expiring on the east coast? The AI will then construct the API calls, launch them with other open source components down into the running systems in the network, gather that information back, and then respond to the user.

(08:18):
Yes, you've got three expiring in Boston or three expiring in the uk, right? These type of answers and automations would've taken somebody hours or days to create scripts and tools around the various APIs to make the same thing happen. But because the AI is so fast and it has a pre-trained knowledge base running on accelerated hardware that allows it to dynamically create these APIs on the fly, it provides an ultra powerful automation for somebody trying to operate these networks. And in particular, the AI doesn't seem to care how many APIs you train it on, right? You can train it on a corpus of 20, 30 different system types, and it will automatically construct given you've trained it and created the software around it correctly, those dynamic APIs to go down into the running systems and get information. And as we go forward in time, eventually as people become more comfortable with AI, will allow AI to make changes in the network to dynamically correct problems as it detects 'em.

(09:18):
And that's really called closed loop automation where we're using AI to not only learn and present and alert, but using it to take actions down into the network. Now, we're not at that point yet, right? I think people aren't comfortable yet to have an AI actually run their networks, but using AI in the early days to provide information to provide learning and reduce the opex and effort that it takes a operations team to run a network is an incredibly valuable asset. And that's where we see it happening first. That's where our investments are pointed as a business.

Guy Daniels, TelecomTV (09:50):
Great. Thanks, Paul. It's incredible progress we're seeing at the moment, but can I ask you what potential drawbacks there may be with AI and how does Wind River propose to overcome these?

Paul Miller, Wind River (10:03):
Well, it's pretty interesting. Some of the odd things we've experienced, right, both on the operations and on the development side, the obvious one is something called AI hallucinations, right? And this is if you don't thoroughly train the model correctly and cover the scope of the type of questions that can be asked, it's interesting to watch an AI engine attempt to keep the human that's talking to it happy, and oftentimes it will construct answers to try to satisfy the human being without having a full set of data to base that conclusion on. So one of the things that we have as a strong caution for these early days of AI tools is watch for hallucinations. In many cases, if you're going to take an action based on data that's come back to you, you may want to ask a second question from a different perspective to validate the answer that it's just given you.

(10:53):
The other option is running multiple ais and correlating the data that they're providing. That's another option that we're looking at. But fundamentally, we've been able to work around these problems to date by thorough training. You have to have very comprehensive training about, in our case, the APIs and actions that they can take and the order and sequence in which they should be fired, and creating boundaries to prevent hallucination in the AI responses. That's a particular area that you have to be cautious of as you look at these technologies. And we are, of course, we're in early days for ai. I don't believe that it's in common use, for example, in any telco networks. We're kind of in the early days of product development and bringing these tools to the service providers. But in the next three to five years, I think we're going to see an explosion of the use of this as we've seen in other industries that are now quite pervasively using ai, but obviously a little more careful in the telco, given the safety critical nature of 9 1 1 services, emergency services, et cetera, that are hosted on that network, uptime and availability is critical.

(11:55):
So we expect the customers to be cautious as they move forward and adopt the technology.

Guy Daniels, TelecomTV (12:00):
Okay, thanks Paul. Well, final question for you. What advice do you have for telco operators and vendors as the industry looks to integrate AI capabilities?

Paul Miller, Wind River (12:10):
Yeah, there's a few things I'd comment on. Obviously a lot of the vendors in the ecosystem that provide product and technologies to the service providers are driven by solving service provider problems and having the service providers open early proof of concepts enable the vendors to have access to the data from their running networks. This is going to be critically important for the adoption of ai, and as a service provider can see the attraction of the things we've been talking about here today. Well, sure, I would love to have lower opex. I'd love to be able to use less skilled personnel to run my network. I'd love to have faster response time and predict my outages. These are very attractive things. They're only going to get enabled if the service providers welcome in and drive the adoption of this technology in their networks. And that's really through proof of concepts and early working with vendors, making sure data is fed in.

(13:02):
Obviously a lot of these AI and machine learning systems are dependent on a large corpus of data, and that's really going to only work once we have it. Well-trained on a service provider network. So I think there's a strategic alliance or partnership that needs to happen between the service providers and the vendors to make sure that this technology is adopted. And as I mentioned a moment ago, I think you also have to understand that it is early days, right? As we've talked about hallucinations and other artifacts, we are seeing some incredible power in the type of tools that are capable of being built here, even with today's technology. And these things can be used pretty much immediately, as you've seen from the demo that I mentioned earlier. We'd like to see that, right? It's through the adoption of the technology and partnering with the vendors that you'll truly be able to realize the materialization of these benefits that we've been talking about today.

Guy Daniels, TelecomTV (13:53):
Yeah, thank you very much, Paul. It's much more to come. We look forward to seeing and hearing more, but for now we've got to leave it there. Good talking with you again, and thanks so much for sharing your views with us today. Thank you guy. Great to be here.