MWC Barcelona 2026 | Molham Aref, RelationalAI & Mark Austin, AT&T

Clips
More from MWC Barcelona 2026

Molham Aref

CEO

RelationalAI

Mark Austin

VP Data Science

AT&T

play_circle_outline Open Telco Initiative Unveils 30 Industry-Trained Models; Phone-Runable Edge AI Beats Frontier Models for Field Techs

play_circle_outline Full-weight fine-tuning enables adapters, classifiers, root-cause analysis, and extensibility.

play_circle_outline Embedding models and re-rankers optimized for document retrieval and answer localization.

play_circle_outline GPU-agnostic training/inference stack supports NVIDIA and AMD (ScalarLM support mentioned).

play_circle_outline Use of relational knowledge graphs and 4D digital twin (people, places, things, processes).

play_circle_outline Training + inference integrated loop for reinforcement learning and continual model updates.

Info
Transcript

Molham Aref, RelationalAI & Mark Austin, AT&T

Molham Aref

CEO RelationalAI

Mark Austin

VP Data Science AT&T

In this interview from MWC Barcelona 2026, Mark Austin, vice president of data science and AI at AT&T, and Molham Aref, chief executive officer of RelationalAI, join theCUBE’s John Furrier and Dave Vellante to explore how specialized AI is transforming telecommunications infrastructure. Austin reveals the strategic thinking behind AT&T’s recent release of 30 open-source, open-weight models specifically trained to “speak Telco.” Highlighting the limitations and high costs of general frontier models, Austin dives into how these specialized, industry-specific mo... Read more

explore Keep Exploring

How is the availability of more data and new AI capabilities changing analytics in telecom — what is different from before and what new use cases does it enable? add

Why are open model weights (open-source trained model weights) important to you as a customer? add

How do you measure and validate the performance of your document question-answering/retrieval system? add

How are you using open-source model architectures and ScalarLM together with AT&T data to optimize the network, and what is the AT&T guarantee? add

How can an organization train or customize a language model on its own data, semantic framework, and tools so the model understands company context and can perform decision-making and optimization across structured and unstructured data? add

How does the system enable combined training and inference for domain-specific reinforcement learning, and what hardware/software support does it offer? add

bolt Powered by CUBE AI

Molham Aref, RelationalAI & Mark Austin, AT&T

search

John Furrier

>> Welcome back over to The Cube's live coverage here in Barcelona for MWC26. This show is going to be continuing to talk about AI, but really the infrastructure of telecom, the operators. Its AI infrastructure meets the data. I'm John Furrier with Dave Vellante. We have Molham, Cube alumni back, CEO of RelationalAI. Good to see you. Thanks for coming back on.

Molham Aref

>> Good to see you, John.

John Furrier

>> And Mark Austin, VP of Data Science and AI at AT&T. Big name in telecom. So guys, thanks for coming on. Talk about data, AI impact to the network.

Mark Austin

>> Yeah, thanks for having us.

John Furrier

>> It's an overwhelming event. Every year it's always talking about, hey, look at the business model transformation, technology advancement, but it seems like there's so much stuff going on around the infrastructure, and AI is at the center of it. A little bit of quantum conversations, but mainly it's AI. How does it impact the operations? There's so much new innovations coming on. Molham, talk about the key AI story lines. Outside of the hype side, what's the real deal here with AI?

Molham Aref

>> I think we are very excited to be here to make some industry impacting announcements around Open Telco Initiative, and releasing a new generation of open models, open weight models that are been trained on the industry. And so we're very, very excited about that because we think that's obviously a very important component. And as people start to unlock value from frontier models, they're starting to realize that those frontier models are very expensive, and they take time, and the latencies are high and so on. And so, when you have use cases that are high value, you can actually get a better return on investment if you're actually using specialized models, industry specialized models and so.

John Furrier

>> I want to dig into that, 'cause Mark, I want to also come to you. Data science and AI, we've seen this movie before in the analytics game. So analytics has been around for a while, but now the AI side of it comes in. What's different from your perspective now that you got the data, you got now more things to look at and analyze, and actually enables new use cases? What are you seeing?

Mark Austin

>> So I've been at this thing for years as well. So I remember, I always tell a story 12 years ago, I was working on self-optimizing networks. So that was some of the first intelligence we were kind of introducing there. It was phenomenal. We took out 40% of the drop calls at the time. But I'll tell you a story of like how we got started on these 30 models that we open sourced here. A guy walked in my office, and I run Ask AT&T at AT&T. It's 27 billion tokens a day that are processed through that. It's all across the company. So HR, finance, network, you name it, we're using Ask AT&T. And I have all the models, and I thought I had all the models. I have OpenAI, I have Claude, I have all the open source, Llama, Mistral, Gemma, you name it. And he walks in and he goes, "Mark, I need another model." I go, "What do you mean another model?" "I need Grok." I didn't have Grok yet. "So why do you need Grok?" "'Cause it can't answer this Telco question, and I don't want to pay to give this to the vendor to kind of... So I want to solve it ourselves, 'cause every time we don't solve, we have to give it to somebody else." He goes, "But Grok knows the answer." I go, "Show me that. How does Grok know the answer to that?" And he showed it to me. It was pretty complex of how the RAN is operating with the core. And sure enough, it knew the answer. And I was saying, "You know what? I've heard that frontier language models are not great at Telco, and GSMA has been talking that for some time. And if you grade them, they have all sorts of these-

John Furrier

>> They don't speak Telco.

Mark Austin

>> They don't speak Telco. So they're like 60, 70% they get the answer right. And sure enough, this was an example of that. I said, "We're going to fix this." And the nice thing is the reason we work with them is how do you fix it? Well, you have to train models. You have to get millions and millions of pages of documents. You got to get all the standards, you got to get all the publications. And fortunately, they brought 40 companies together to get these millions and millions of pages of documents and they gave them to us. And the other nice thing is AMD pitched in hundreds of GPUs that were able to put it at scale. And over a few months, we've unlocked it. We released, 2 days ago, 30 open source models that are better than the frontier, sometimes a lot better. And the cool thing is they're all sizes. So we even have some that you can run on your phone. So, let's say that you have an outage or let's say you have planned maintenance, and you're out there, a technician trying to figure something out, if you don't have connection, what are you going to do? Well, you can pull up your phone and you could run this small model. You got documents, so we have a small vector store on the phone, you can get an answer.

John Furrier

>> It's a quasi agent really, and specialized with that use case.

Dave Vellante

>> Okay. So I got to ask you, explain why open weight. Let's start there. A lot to unpack there, John. Yeah. Why is open weight so important to you as a customer?

Mark Austin

>> Yeah. So, this is the starting point. And the reason we went with open source train models, one, small and cost-effective, low speed, but you want to train... And by the way, we trained all the weights. And we didn't just do like the fine-tuning with Laura and you just do a fraction of the weights. If you fine tune all the weights, it makes it very easy to add adapters. Adding adapters lets you do classification. So now, I can classify and I can put all these things to do root cause analysis, which is ultimately what we want to get, right? But I want to understand the basics. I want to answer the question and I want to be able to put classifiers on top of them and extend them.

Dave Vellante

>> So 27 billion tokens a day, that's got to get expensive. You mentioned that these LLMs, the frontier models are expensive. Why are they so expensive and what makes these specialized models better than the frontier models? Give us an example. How are you measuring that?

Mark Austin

>> Yeah. So, we measure it on a variety of things. You have the GSMA benchmarks, but on this one, we basically took hundreds of thousands of questions that we knew the answer 'cause we had all the documents. That's what we trained them on. So you always do what you always do. You have a training set, you have a holdout set. Okay? "Here's all the things I'm training you on. Now I'm going to ask you this over here, and does it answer the question?" So that's how we did it. And we have models that are really good at search. And what I mean by that, these are, you call them the embedding models. "So which document, which paragraph, which page has the answer to this question?" And it retrieves that very well. So we have the embedding models, we have the re-rankers, and then we have the large language models themselves.

Molham Aref

>> Yeah. We were getting world-class results with four billion parameter models, which are two or three orders of magnitude smaller than the frontier models. And because we are focused on this domain, we don't need to know chicken soup recipes and other things. We can really prioritize the kind of content that matters here.

Dave Vellante

>> There is though a school of thought that says these four frontier models, and there really are four. You mentioned Llama, maybe put them as five-ish kind of in the back. We'll see what happens. They got to get a little more compute. They're working on that. But there's a school of thought that says that over time, the cost per token will go down so much because they have access to compute because NVIDIA is coming out so fast that ultimately they will be much less expensive and more capable. Do you buy that or do you see that sort of delta continuing in the future? Or it doesn't even matter?

Mark Austin

>> I'm cheering for that, but I have a problem now, and I can solve it now. So we're taking care of it. And you can't run 27 billion tokens a day unless you start doing some of these things. So we've been kind of going after fine-tuning models across lots of use cases. We target the top ones, which ones are using the most, and we train the models there.

Dave Vellante

>> But to the other point, you do use the frontier models?

Mark Austin

>> Of course. We're loving the frontier models.

Dave Vellante

>> Explain how you use them and how they compliment your specialized models.

Mark Austin

>> I think advanced reasoning. So you got the expert. When do you go to the subject matter expert? You use them at that, and use these guys for the normal stuff. And then you train them to be specialized. When you go specialize, then they can get really good on the specialized stuff. If you go to generalist, you got to get it together to general-

Dave Vellante

>> And I want to get into the innovation that these guys have, but I know you want to jump in here, so please do. And I want to come back to what you guys are doing that is groundbreaking.

John Furrier

>> Well, I can tie into that, but I think this is a great example of what we've been saying about specialty models. They're not mutually exclusive. In this case, you have, is it public information that you had?

Mark Austin

>> This all public.

John Furrier

>> Is there any internal AT&T stuff?

Mark Austin

>> Not in this.

John Furrier

>> Not yet?

Mark Austin

>> But I mean, to your point-

John Furrier

>> It's an unlock 'cause you're filling a gap.

Mark Austin

>> And by the way, we trained every architecture we could find on open source. So we have a great variety now. We have all sorts of sizes. And thanks to ScalarLM, which RelationalAI supports, it's GPU agnostic. So now, we have the same software, we can go NVIDIA, we can go AMD, and we have the expertise. And what are we doing? We're taking that same concept and we're adding the AT&T data that we don't share outside because there's a bit of special sauce in terms of, "How do we set the parameters? How do we optimize that?" And by the way, we've gotten so good at this, I think they're the only company that offers the AT&T guarantee, which means that we put our money where our mouth is. If you have a problem with the network, we pay. And using these things and using the AI we've been doing, we've been able to do that.

John Furrier

>> It's a great success story. And I think it's just the beginning of what's going to happen in the real time network with more AI intelligence. When you start thinking about bringing that intelligence to operations, that's a broad picture. You guys nailed this piece. That's the starting point. The flywheel kicks in. I'm sure it's going to translate to sitting at the edge of the towers, RF management. It's limitless.

Mark Austin

>> I don't have to tell you guys or any of the people here. If you're talking about applying some of the AI at scale with network data, you have to be efficient. So, we will take those models, we'll put adapters on them, we'll classify them, and then we'll pop them up for reasoning when we see things. Now, I'll just give you one example. When you do have the human involvement, the humans the hardest stuff. Sometimes these problems are so hard, it takes four to six hours, and it's them looking at charts and graphs and data and everything. So you take this thing, you put it in the agentic system, and you arm it with tools, and the tools can see your desktop now. They can see your charts and graphs and they can reason over things. So that's the really hard stuff, but we're going to unlock the small language models to go at scale and go process all those things.

Dave Vellante

>> So Molham, you and I and George and some others-

Molham Aref

>> We go back.

Dave Vellante

>> We do go back. And early on, we saw the potential of what you guys were doing. And at the time, we said we're moving from an application-centric world to a data-centric world that's going to feed intelligence and deliver these tokens, and support all this new cool stuff that we're talking about all the time. Now, what we observed at the time is that what enterprises want, ultimately, is this sort of digital twin of their enterprise in real time. And to do that, they need a 4D map. And we talked about people, places, things, and activities or processes. Talk about what you guys are actually building to enable that. And I'm curious as to how AT&T is taking advantage of it.

Molham Aref

>> Yeah. So I appreciate the question. So, I think we all realize now what's good for the human is also good for the language model, because the language model is trained on the kinds of stuff that humans produce, and what do humans want? They want data, they want semantics and context, and they want tools. I'm much better at doing anything if I have access to the data and the library, and the books, and the calculators, and the computers and so on. And the language model's exactly the same. So, we've had the tools, we've had the semantics, and what we've added recently is the ability to train a language model, like you train a human, on that data estate, on that semantic estate, and on the tools.If you have a frontier model, you're hiring a very, very smart person who's never worked with you before. And then every time you want them to do something for you, you got to explain who you are and the context, "I'm AT&T. I have 100 million tables, I have all of these capabilities, et cetera." And then you ask the question, and then it goes and thinks about it, and then comes back to you with the answer. Now, what if I could take maybe a less expert person, but I can train them?I send them to training for some time at AT&T and they come back and now they know AT&T. And every time I ask them a question, I don't have to tell them and the context and the prompt, "I'm AT&T, I have all these tables, I have all these columns and so on." And so adding this ability to train, to post-train an open model, and the open models are already pretty incredible, and of course they're free. And being able to post-train them on, as Mark said, a variety of GPU architectures so that you're not locked into premium priced compute options, compliments very nicely the context that we create through the relational knowledge graph and the semantics. And that all the decision intelligence reasoners that we have for rule-based reasoning, predictive reasoning, prescriptive reasoning, graph reasoning, because now you're giving the language models, whether it's a big one or a small one, the tools it needs to do the job.

Dave Vellante

>> And you're able to do that irrespective of data type and structured data, unstructured data, transaction data, analytic data. Explain.

Mark Austin

>> I got a favorite one. We were just talking about it right before we came up here. What I'd like to do, and this takes to in account the small language models and the higher reasoning models too, I'd like to give them a problem. "Here is a cluster of sell sites. I want you to tell me all of the optimization opportunities I could do and you tell me what is best to increase my revenue." So you're combining business laws, along with the relation of how these sell sites are kind of working together. And some of them might be, there's a parameter change here, I might have a congestion problem here, I might have to build a sell site here, but it's going to reason through all of those things that's going to create the business case and tell me what I should do. And that's the relationship of 4D.

John Furrier

>> And you guys have so many sell sites, got central offices all over the world-

Mark Austin

>> We serve 100 million people.

John Furrier

>> You have projects ongoing, upgrades always happening. You have power, you have dormant power, in some cases, probably hanging around. So is there more analytics around, "Okay, want to make money. How about leveraging some of that energy"?

Mark Austin

>> Absolutely. That's what I love about tying the cross domain. So, I'm tying the network domain with the business domain, maybe even the customer care domain, you're tying to, "Are these getting customer care calls that I need to take care of as well?" Tying those all together and coming up with an overall business sensor.

John Furrier

>> You can slice the salami, as Dave always says, many different ways looking at what's coming across the network, put policy to that. That's where these AI factory nodes become very interesting because you got the data work done.

Molham Aref

>> Exactly. And what Mark is really fundamentally saying is these shouldn't be silos anymore. We have the overarching semantics over all of it. And so you make engineering decisions with customer satisfaction and profitability in mind.

Dave Vellante

>> And what you just described was you have a set of top-down objectives, revenue. You could even, I presume, add in, "But we don't want to have less margin." And so optimize that plan.

John Furrier

>> Whatever you want to optimize.

Dave Vellante

>> But order to do that-

John Furrier

>> AI Optimization, this dev agent. I love that one. I'm going to write that down.

Dave Vellante

>> But in order to do that, you have to have the context, you have to be able to harmonize the data, all different data types, and essentially have real time. I mean close to real time.

Molham Aref

>> One of the open models that know how to do that, like models trained on how to run a telco business.

Mark Austin

>> I think your models are the experts. So you have this overall orchestration agent, and it's calling on the experts. "You go tell me, what are all the ones that have congestion? And what are all the ones that..."

John Furrier

>> It's super refreshing because most people just have this answer of hand waving, "We'll just throw compute at the edge." Actually, it's multiple factors. There's orchestration going on. It's the platform component. So I have to ask you, 'cause we just had a session, and we were just covering and predicting on the Cube research, that the traffic's going to be three to four X in the next three years on the uplink. And then you got radios talking to each other, meshing up. So you now have new traffic patterns. This isn't moving packet from point A to point B and throwing compute at it. You're talking about AI analytics, having all these agents. So with all that new information, it's new data. It's not like it's been trained. You're getting real-time information coming in. So I love this idea of training and inference kind of a software defined. Can you explain more, 'cause I think this is compelling? Because all the new data coming inbound is new information. It's not been trained yet. Do you send it back? Do you do it at the edge?

Molham Aref

>> Yeah, this is, I think we're working with AT&T and a group like Mark's is very important, because the traditional sort of setup is that you have an inference stack and you have a training stack and they're independent of each other. But if you want to be able to do reinforcement learning to really learn a domain, a data state of semantics and industry, you need these things working together where the training calls the inference, writes a query or runs a integer program, or runs a simulation, and then gets that back and feeds that back into the weights. You want the industry stuff in the inner loop of that.

John Furrier

>> And they know their network. So it's not like they have to go to college, AI college to figure out what the traffic is.

Molham Aref

>> Exactly.

John Furrier

>> They could use context. Right?

Molham Aref

>> Yeah, exactly. So you want all of that in the training. And so what we provide is an open source system that works with open weight models that does training and inference together on NVIDIA and AMD GPUs. It's all in the open. And-

John Furrier

>> Software's independent of whatever hardware?

Molham Aref

>> It's independent of the GPUs, which is-

Dave Vellante

>> It has a little nugget that you threw in there, 'cause you've got the software that you've... Is that you guys develop that? Is that to basically make it transparent?

Molham Aref

>> Right.

Dave Vellante

>> And what? It taps CUDA and it taps-

Molham Aref

>> We support that and we use it together to build these 30 models.

Dave Vellante

>> And when do they shift?

Molham Aref

>> We've extracted that complexion either way.

Mark Austin

>> he's talking about as open source, abstract the software that'll train and inference NVIDIA or AMD. So that's very nice.

John Furrier

>> Just to get the date out there, when did you guys ship the models to open source recently?

Mark Austin

>> Oh, we announced it March 2nd.

John Furrier

>> And they're there?

Molham Aref

>> They're already out there.

John Furrier

>> They're already out there? Okay. All right.

Mark Austin

>> You can go get them on Hugging Face. By the way, right after this, we're running to a masterclass where we're teaching people how we did it.

John Furrier

>> That's awesome. Very cool.

Mark Austin

>> And what's really exciting is for some of the GenAI kind of folks out there, we have the lead author of Attention is All You Need, Ashish Vaswani is going to be joining us as well. They've taken notice. We trained his model. He runs Central AI. He's the CEO of that now.

Dave Vellante

>> Yeah, yeah. Of course. Right.

Mark Austin

>> He's going to be there. And then Greg Diamos on Molham's team. Some fantastic-

John Furrier

>> Give us the TLDR of that masterclass. Go.

Mark Austin

>> Yeah. That's right after this. Back to your other point, I think that your decision depends on how fast you could take the action, too. So some of these things, depending on what the fix is, if it's a sell site, hey, that could be a year maybe sometimes. So, I think teeing all those up and getting the right kind of loop and the right actions in place of when you can do it makes a lot of sense.

Dave Vellante

>> How do you embed governance into all this? What does that look like?

Mark Austin

>> So, I like tying governance with the business side as well. So everything that we do, we run it through business and governance, simultaneous. And the way that we work at AT&T, which is why we've been at... Our first year GenAI two years ago was 2x ROI, in year. Last year, was 5x ROI, in year. And it's one of the few, I haven't really heard others do that, but what we do is this is a formula. It starts off with the CFO cutting everybody's budget, so everybody kind of is motivated. And then we kind of take all the innovation money, you pool it in one place and everybody competes for that money. And you come up with the best ideas. And we have no shortage of ideas at AT&T. Hundreds and hundreds of ideas. I opened it up for two weeks. I had 260 submissions for ideas, and we stack rank those with ROI and we go after the highest ones first, and those go into the governance piece. And governance, we have legal and we have compliance, and we have chief security officer there. We're there from the architecture point of view. You put it on a common platform, so you get lots of reuse, so you're not doing one-off things, reinventing the wheel all the time, and that's how you do it.

Dave Vellante

>> What are you optimizing for in that ROI? Is it an absolute dollar? Is it a timeframe for break even?

Mark Austin

>> Good questions? Okay. So this ROI is real. Okay? So this is free cash flow. Okay? So, it's either somebody's budget went down or there's real revenue there. We don't count the productivity. You just got twice the amount of things done.

Dave Vellante

>> It's hard money.

Mark Austin

>> It's hard dollars.

Dave Vellante

>> Okay. So it's hard dollars.

Mark Austin

>> Hard dollars.

Dave Vellante

>> And how important is the timing of that breakeven in that equation?

Mark Austin

>> The first time we did it in year, we're getting a little bit more relaxed because this thing kind of rolls, and it keeps growing and accumulating, but it's still mostly in year, but we're funding some stuff that's two to three years out as well.

Dave Vellante

>> Some of it's inside of 12 months. What about the-

Mark Austin

>> All the ones were inside 12 months.

Dave Vellante

>> Okay. So you said the first year was 2x ROI. So there was a conventional wisdom, Mark, that early on, Molham in this AI thing, don't waste your money on AI until you get your data in order. And what we've found, and I've personally talked to folks like JPMC, Lilly, Walmart, and others to try to understand this dynamic. And what they said was, "Yeah, we thought that, too." But what we realized is that getting our AI infrastructure, the database, the vectors, all that stuff, get that in order, do that, and then just dive into AI and let the AI help harmonize that data and you'll get on the flywheel much, much faster.

Mark Austin

>> You got to do both. But the cool thing is, remember, the gen AI unlocks a whole set of data we didn't have access to before. What does getting unstructured data in place mean? It's already there. You just feed it in.

Dave Vellante

>> It's unstructured.

Mark Austin

>> There's nothing to do. It's unstructured by definition. So that was like all the low hanging fruit. All those were first year, all the customer care, summarizing a call summarizing at the beginning, intent, all that stuff. It was all low hanging fruit.

Dave Vellante

>> Did you find that the learnings from those first few projects could have applied to the next ones and that you were able to speed them up?

Mark Austin

>> That's what I went with the platform. The platform is critical 'cause then you get the reuse. Okay, "This HR has a question and answer of HR documents. This one has a question and answer or customer care documents." It's the same exact pattern. And what we do is, "Hey, Joe, you go get with Bob. We're going to build one thing for both of you." And that's the flywheel.

John Furrier

>> That's awesome. It's great success. And you get the data and you got the AI doing the work. It's just the beginning. Any headroom, in your mind, what's next? You're obviously emerged and the flywheels kicking in. What are you seeing as dots to connect? As other new, bigger things to work on?

Mark Austin

>> Yeah. Personally, there's some stuff we're doing with Rational AI, some really cool stuff. I'll let him talk about it. But getting agents right is the next thing. And what I mean by that is agents for process optimization. And I've come up with three things that we want to do. And if I go back to the ROI, it's earnings, efficiency, and effectiveness. So earnings is the ROI piece. That's one's pretty easy to understand. Efficiency is maybe not as understandable, but what you want is you want something that's efficient, and there's different ways you could do the agent. So think of it as if it's a process, it's a flow chart. I don't want to hard-wire every single one of these workflows as a flow chart. I want what we call a super agent. And what I mean by that, a super agent is you just give it a configuration, it could do that workflow and this workflow and this workflow. So the metric for efficiency is workflows per agent. The more workflows per agent you have, the more efficient you are. And then effectiveness is, does it actually work? So you got to get the evals around it and look at your accuracy, but getting agents right. And then if I go back to the cost thing, it's these fine-tuned models to help us keep growing, 'cause these... I'm 5x, the 27 billion versus last year on tokens per day.

Dave Vellante

>> And you're maximizing the value of your tokens that you're paying for?

Mark Austin

>> Absolutely.

Dave Vellante

>> If you can put more through that pipe. Right?

Mark Austin

>> Yeah.

Dave Vellante

>> That's instant.

Molham Aref

>> If I can just add one more thing. I think one of the really exciting things that we're working on is leveraging GenAI to solve some of the problems that you talked about, Dave. So messy data, little semantics. These were long polls in the tent that prevented people from doing the harmonization and the stuff that we've been working on for some time now. And so GenAI is part of the solution as well because we have agents now that you can point at a data estate and they'll create a starter semantic model, an ontology, a relational knowledge graph that gets you going. And once you have that, you can attack the other long pole in the tent, which is creating an eval that you can use to train a language model or assess a language model. And so once you have a semantic model, an ontology that describes the domain, you can now start generating queries and questions, and effectively making it so that we can create an infinity of question and answer pairs that generate reasoning traces that you can use to train a language model to be really good at that data estate. And so we're getting to the point now, we're close to push button, "Point me at a data estate, and I'll build you a model that understands that deeply."

Dave Vellante

>> And that was the aha moment of my research, John, is that you can use the AI and tools like relational AI to actually help you harmonize and cleanse that data and shape it.

Molham Aref

>> That's right. And so-

Mark Austin

>> The very interesting thing is asking questions of unstructured data was relatively easy with the Vector database. Asking questions of structured data, tables and sometimes you're joining 10, 12 tables together. Getting that right and having these GenAI understand it, "What does this column really mean? What does it mean by join? How do I calculate this other thing that the finance guy is like?" Getting that right in the past has been very difficult. And by the way, we've won the two competitions, Text2SQL 10 times. And this over Google and Alibaba, and with Spider Board, with RelationalAI. So we know it's very hard, but this push button thing, I don't take that lightly. That is a big deal if we can get that right, 'cause this takes a long time to train these things.

Dave Vellante

>> And you get the expressiveness of a knowledge graph with the simplicity of a SQL query. Yeah.

John Furrier

>> Super agent in action, basically. That's a super agent right there. You want that button. "Hey, super agents go work on that," and it pops out. Guys, this is awesome. Well, great to see you again. Mark, John. Great success and continued, fly moment. We'll keep checking in with you.

Mark Austin

>> Yes, please do.

John Furrier

>> Pushing the envelope. Love the open source.

Mark Austin

>> Awesome.

John Furrier

>> 30 new models. Hey, they speak Telco now, these big models. So get in there, check it out. Good contribution to the community. It's Cube doing our part. I'm John Furrier with Dave Vellante. Thanks for watching.