SAS Innovate 2024 | Udo Sglavo | SAS

Clips
More from SAS Innovate 2024

Udo Sglavo

VP of Applied AI and Modeling

SAS

play_circle_outline Buyers, users, and builders of AI in the market

play_circle_outline Unlocking Interactive Communication: Democratizing AI with Out-of-the-Box Models

play_circle_outline Marketplaces and Self-Service Experimentation: Addressing Concerns and Attracting Companies Reluctant to Invest in AI

play_circle_outline Unlocking Innovations: Collaborative Ecosystem Expansion and Containerization for SAS and Partner Growth

play_circle_outline SAS accepting the challenge of going public and building new sales strategies

play_circle_outline Focus on success stories and showing proof of real results with customers

Info

Udo Sglavo | SAS

Udo Sglavo

VP of Applied AI and Modeling SAS

search

>> Welcome back everyone to theCUBEs Live coverage here in Las Vegas for SAS Innovate 2024. I'm John Furrier, host of theCUBE. Dave Vellante, my co-host, co-founder, as well as he runs theCUBE research. As we wind down two days of wall-to-wall coverage, we got the best for last. We got the star of the show, Udo Sglavo, VP of applied AI and Modeling with SAS, the key player in a lot of the AI that's going on in their strategy, what's under the hood and the models. I think one of the most important things here is the introduction of the model. So great to have you back. Good to see you.Thanks for coming on.Good to see we're all going to get the energy. This last segment, we're going to amp it up.

Udo Sglavo

>> Wonderful.

>> Tell us how are you feeling?

Udo Sglavo

>> I'm excited, but I'm also a little bit exhausted because there was a lot of interest at this event about the announcements we made, releasing models as a product and just of course a major step in my mind forward for us as a company and the industry, if I may say.

>> So I want to just give everyone a quick reset. You guys introduced the SAS, get the three things, productivity, performance, trust, three pillars. You introduce the concept that there's buyers of AI, users of AI, buyers of solutions. Basically, I'm using AI and then you've got the people who build AI. There are certain companies that have to get the hardcore machine learning, hardcore ops, hardcore AI, and then you got the subscribers of model people who are going to use models into the stuff, whether they're building AI or using it. I think the first company, other than OpenAI that had that little library or marketplace, you guys are bringing SAS models that you guys have vetted into one place in use cases that are known and well understood, applying generative AI to fraud. Take us through what that is because with doing that, people in theCUBE have come in and said it from the show. That gives you some advantages, put governance around it. Take us through what is actually in that model package that people can subscribe to. Take us through it.

Udo Sglavo

>> Absolutely. And I mean you basically gave a wonderful overview of what we are doing already. The point is that we've been in the market for a long time, almost 50 years. Over this 50 years. We solved a lot of individual customer problems, very specific, unique challenges of customers, mostly using our consulting resources. And now we are basically saying, "Why don't we take all that IP, which we gained over the years, combine it with the algorithms which we deliver, and package that up as a put into production systems because it's all sitting in a container and it's basically accessible through APIs." So you were wondering about what's inside of that container? Well, obviously it's the logic. There is some math, there is some way how we solve these problems, but on top of that, we also have metadata about the model. So we also launched this concept of a model card. So the model will also say, "I have these ingredients, these KPIs, which can then be fed into a model card." And to make things even more exciting, we also include prompts so that when you deploy the model, you can have actually a discussion with the model. So you can say, "What can you do?" And the model will say, "I'm a model for fraud detection. These are the questions which you can ask me."You

>> actually made the comment that you wished we heard more about democratization of AI, but in a lot of ways when you're offering these out of the box AI models, you're dramatically lowering the technical expertise required to exploit them, and that is a form of democratization.

>> Well, before you ask him, I wanted to say for clarity, I challenged the democratization on the opening keynote because the focus was on the developer side of it. So the data science developer piece of it, where they had that red cube.

>> You think the understated-

>> No, the Viya Workbench is the democratization angle. So they have two things going on here.

>> As is the model out of the box model.I

>> mean, that's democratization. That's bringing new capabilities and faster time to get stuff done or figured out or discovered.

Udo Sglavo

>> Also, putting it into hands of people who are not used to use this models. So we had this example in the opening session where we worked with a large CPG company and their business case is that they have several inventories where they can store items. Once they want to use these items in the production line, they will basically pull an inventory out of the warehouses. So the question they have is, well, if a truck driver arrives in the morning, where should the truck driver go to deliver his or her goods? So in the past, the planner would probably pull out a piece of paper and say, "Well, yesterday we determined that there's space." With the model which we are introducing, an optimization algorithm will run behind the scenes, it'll take all the constraints into account and it'll then communicate with a truck driver in language, please go there, find this person and deliver your goods there. And this is almost like Google Maps, isn't it? Where you're basically saying, I want to go from A to B. I not need to know how the root calculation is handled behind the scenes. I just need the answer to my specific problem.

>> The prompt libraries are critical, and as a nuanced part of the announcement. I'd love to get your thoughts on this. We'll come back to the models, the prompt storage because that seems innocent on the surface. Yeah, I got some stored prompts that like a canned prompt. But I think where it leads us is more prompts and ultimately a prompt-less environment where now under the covers, the system can work on your behalf. Am I reading that right?

Udo Sglavo

>> Yes, absolutely. That's exactly the idea that it's more driven in the communication with the model rather than a person dreaming up Bronx. What may I say to the LLM, which is of course different how we are using LLMs today in our personal lives where you just go there and you may ask random questions. The model has a specific charter and a specific task. That's why we envision that. The first question is what can you do for me? It's almost like... But imagine now you have a library of models and-

>> Before you leave, Dave and I wants have one more thing. The other thing I want to point out that you guys use the word lightweight, and I want you to explain what that means because we've been saying we were the first ones to put out the power law of models a year ago. First research firm to do that. We were right. A little self-congratulatory, humble brag of course, we're usually right most of the time. But the power law, our thesis was the large language got all the attention, but the value was in the specialty models where the IP was. A model can be small, but high-Optane model, very powerful. So that happened, that's what you guys are basically pointing out here. So the payment fraud detection, the welfare fraud detection, document vision, all digital assistance. You guys have data and pre-canned, I won't bad word, but pre-packaged the models for use in well-known use cases.

Udo Sglavo

>> And the lightweight angle comes that we want to be able to plug in eventually into any IT system. I mean obviously initially it will be designed for the SaaS user base and we will target our own market, but a model can potentially run even outside of the SaaS ecosystem because it's mass, it's large language model, prompts, it's metadata. So we can even imagine that you can basically deploy our containers, let's say inside of a database vendor. And in fact, we are already doing some experiments there that with the more modern database vendors that we can say, why don't we put the container inside of your ecosystem so there's no data movement, and then basically, it's all happening inside of the database. Does that make sense?

>> Yeah, absolutely. No data movement got my attention. Now you mentioned, I think it was in the media briefing that LLMs are actually... Foundation models are a small part of the equation here. We were talking earlier about Llama 3 was just introduced and or announced, and we get excited about these things, but so how are you actually using LLMs and what are the most salient aspects of these models?

Udo Sglavo

>> So two things about that. So first all currently we are using LLMs mostly for the interaction with the model. The model itself basically uses whatever algorithm we believe is appropriate. There is this misconception that you can all of a sudden use LLMs for any problem of the world. And it's very interesting to see that when we have discussions with customers about their use cases for LLMs, the first reaction is almost always, but that's an optimization problem or this is a forecasting problem, why don't you use the appropriate algorithm. So that's step one. But our vision is that we want to be able, and this is more targeted at a use case where you build the model in collaboration with the user. So the LLM will then ask questions in a similar fashion like you would ask questions to a data scientist. So it'll basically try to determine what is your business question, in which category of analytics does that fall? And then once it knows that, it'll ask questions, which will help to construct the model on the fly. And this is pretty exciting research that we are doing. Of course, it's not ready yet, so this will not be part of the first.

>> So I got to ask you about some of the trends we're seeing and get your thoughts on it and get your reaction. I mean, that's all what I love about this market. It's unfolding as we see it in front of us. You guys are doing a good job with it as well, and you're digging ahead and get the products. Cross modality reasoning is becoming a big topic. Let me explain what that mean by that. You probably know what I mean, but . Its reasoning is a hot topic. So prompt response, you get an answer prompt, get canned, certainly promptless, everyone sees that with ChatGPT and other prompts. Reasoning means I got to figure out stuff. Basically solve something, figure reasoning and understand it, the context, all that stuff. So reasoning has become a hot topic. Now when you have images and text, maybe DNA sequencing, I mean if it's not text, it's another mode. So multi-modal means not just text. Text is actually easy. So text is great, and that's why all the use cases are text-based. Images are huge. And you gave a demo today of what we all see in old school government contracts, paper with handwritten stuff on it. There's a lot of that stuff around. That's an image, that's computer vision. So as the modals come in, it's hard to do cross model modality if it's not all in one place. You're smiling because you're like, you agree?

Udo Sglavo

>> I fully agree.

>> And this solution is what? Put it all in one place, centralize it or-

>> Federate it.

>> Tell us your vision on this because I think if you get this right, you can save a lot of efficiencies and latency and get more time to value.

Udo Sglavo

>> I think we discussed that earlier. In my mind, the one thing which separates SAS as a company from a lot of our competitors out there is that we like to be output driven. So we think about the output, the decision first, what are we trying to accomplish? And then we think about, well, which technology, which mode from this different universes, text streaming data images is required to address that problem. And then we basically build the pyramid button up. So it's a top-down button up approach, if that makes sense. Again, we want to deliver these models as lightweight containers, and the only thing you need to know are the APIs, which you can use to input information to the container and the APIs, which will return the information. So eventually we want the users just to focus on that, and basically what happens in the guts, well, that's our business. Does that makes sense?

>> Yeah, totally does. So then governance becomes a big important piece.

Udo Sglavo

>> Wonderful, yes. 100%.You

>> can't get AI without governance solved first.

Udo Sglavo

>> Yes. You see, this is why we also want to focus initially on our own SAS ecosystem because we also have tools like Model Manager, which help us to determine is the model still valid? And that's something we should typically do over time. And the thing we are concerned about is what's called model decay. Basically the accuracy of the model gets worse and worse and worse. The model itself will not know that, of course. So we are also thinking about, well, how does this work if you're not in the SAS ecosystem? So somehow we need to introduce some persistence layer to the model, so it can basically store information over time and then it can assess itself. How am I doing?

>> It's like observability for LLM.Yes,

Udo Sglavo

>> 100%Whole

>> other category of observability created in the market.

>> Can I ask you a question about the document vision demo that you showed on stage?

>> That was awesome.It

>> was awesome. My question was how is it differentiated from some of the things we've seen like in document understanding from say some of the RPA vendors? Is this different? Is it new tech capabilities?

Udo Sglavo

>> Let me say that there are similar offerings in the market. This is not a unique new offering. What makes this unique is that this is based on our consulting experience. So this is stress-tested technology, which we implemented several times for very complex PDF formats. Maybe I should explain what we are doing. So the question of documentation, which it tries to address is you have unstructured data maybe in a PDF format. And the PDF document is not just text. It's all over the place, handwritten notes, tables, images, maybe some key relationship which you're interested in. And we are using a mix of AI and NLP after we process the PDF document with an OCR service. And for that, we are using Microsoft at this point in time, but it could be any OCR service to add additional information, which we can get out of the document. Does that make sense?

>> Yeah.

Udo Sglavo

>> Absolutely. So the idea is then basically you're turning very unstructured data into structured data. So what is really new in my mind is this is, and I repeat myself, stress tested. It's working. We have done this many, many times for very complex customer engagements, and now we are basically packaging that up to provide it to other customers.

>> So that means higher yields, better outcomes that make Six Sigma for business applications where you wouldn't normally have that level of quality.

Udo Sglavo

>> Yes. That's what I'm thinking.

>> Yeah, makes sense. Okay. So when does that ship?

Udo Sglavo

>> Our goal is to release the first models in the fourth quarter of this year. And of course, we want to be modest in the way how many models we roll out because we want to assess the market as well. Where should we be heading and what models are required? But our vision is of course, to scale this up. I always think about this as a model marketplace, which you need to introduce.

>> Certain you're going to have app marketplaces because if developers get their hands on this with Viya Workbench, you're going to start to see a lot of experimentation. And then people will accidentally on purpose discover things through their ease of use of self-service.

Udo Sglavo

>> There you go.

>> I mean that's a console that looks like Amazon Web Services 2010. Okay, it's been up from EC-2. Now you got cloud, compute, you got the models all there. That's going to create more developers.

Udo Sglavo

>> And you see also an aspect which we haven't discussed at all is we are also going after companies who are so far reluctant investing in AI because they feel like, "We don't have the people, we don't have the environment and

>> how-We don't want to give our skis and pay all the service....

Udo Sglavo

>> and how can we show fast return on investment?" Now with models, you take not much of risk. You pick one use case, you basically get a quick results because everything is already ready for you. Just plug in the container and then you can basically grow your appetite for more AI while you're actually seeing value. And this is different to how we approached the market before, where we would basically say, "Well, buyers or builders of AI, you're either a buyer where you basically build a ready-made solution from us, or you are a builder, then you have to do these things yourself." Now with models, you don't have to be in the other two categories.

>> So it's because our data is consistent as we show just under 20% of the firms, and this is a decent sized survey, 18, almost 2000 say they're not pursuing GenAI because moving too fast, it's too risky, too complicated, hallucinates, we don't trust it. So they say, "We'll wait."Yes.

>> But their experimentation is with their own data. That's why RAG is popular because it's easy to do.

>> These guys, they claim they're not doing. So I reached out to several and said, "Are you really not doing anything? Yeah, it's moving too fast and we don't trust it and we're just going to wait." Now you can go to them and say, "You don't have to wait. You've got these pre-built models." You must've heard this.

Udo Sglavo

>> Yes, several times. And also a little bit tongue in cheek. They don't even need to know that there's AI in the model package. Because the discussion is different. The discussion is, "You've got all these PDF files with unstructured data. Don't you need it for reporting or for analysis, even if the analysis is just some sums and means and whatever. We help you with that." So you see, we are lifting the discussion away from the technology discussion.

>> Yeah. It becomes invisible in way. It's an-

>> You're providing a service. I'm not going to compare SAS to the hyperscalers or Meta, but if you look at what Meta's doing with LLlama 3 right now, what they're essentially are doing, this is a genius strategy move for them, and it's similar to what you do. I'm going to get your reaction to this because I'm going to tie it to something. They're putting in hundreds of billions of dollars of investment from a major tech player for developers with Llama. So that is just basically giving developers free open source software and infrastructure to run it on to start. Now that's hundreds of billions dollars. You guys are putting 2 billion into AI development that you're going to offer to your customers. So in a way, there's a parallel there. So we're seeing the new branding or brand value in targeting, not the tech nerds per se, the developers who were creating value. So we're shifting to seeing the entrepreneurial and the innovation people. Stars aren't the hardcore developers, they still are. But the new stars, the new rock stars, the new people that are going to, could be all fans for is going to be business people or anyone because now you have a developer using all that infrastructure for LLlama 3 with Meta, you're bringing SAS's DNA to the table, Google next last week, they're bringing all their AI chops to the table. That's what you guys did. And then said, "Hey, let's put 2 billion into it and then it out to the market. And that's our competitive strategy."That's

Udo Sglavo

>> exactly right. And you summarized it very nicely. But I also want to say that as part of our vision, and this will not happen right away, as part of our vision, we want to open that ecosystem also to partners. So we can say, you have the intellectual property, we have the technology. Why don't we go jointly to the market? We as the trusted provider of analytics and technology, you as the trusted advisor of business knowhow, that's exactly what we want to do. And then the idea is basically that we could even say, give us your models and we'll turn it into software behind the scenes.

>> Let me ask you a question.

>> So can I follow up on that?

Udo Sglavo

>> Yes.

>> I want to pick up something you said earlier. You can containerize these models and bring them to data platforms. So I can bring the models to the data and that's part of the vision? Or is it you actually have plans to do that in various-

Udo Sglavo

>> We are researching that today. So we have concrete discussions with some of the modern database vendors.

>> There's three hyperscalers and there's Databricks and there's Snowflake. I mean, those would be the obvious places to go. So some combination of those guys.

Udo Sglavo

>> Read his eyes. Dave, is asking the question I'm like studying the end of it. It has a yes in there.Then

>> beyond that, how do you 10 x or maybe even a hundred x this. Okay. That's what you're talking about now, as developers. So maybe they get a little taste in one of those modern data.

Udo Sglavo

>> You mean that's the show floor?

>> This exhibitor in ecosystem, because you mentioned scale before. Is that how you scale this?

Udo Sglavo

>> That's how we want to scale it. And the plan is to first scale it within SAS, because our current modeling team is US centric. So if we come up with a tax fraud model as an example, which we'll work very well for the US market, it won't work for Germany, which has a completely different text work. So we will basically pull in our consultants from Germany because SAS is a global company. So we have modelers around the globe and basically say, "Well, why don't you take our model, adjust it for your market needs," and then we create a specialized German version of the tax fraud model. So that's step number one. Step number two is then the partners and basically say, "Why don't you do that?" And we become more of the brokers of the models.

>> So developers and partners... You remember a ServiceNow when we first started doing the ServiceNow shows like 2013, it was actually here at the Aria and it was about this size of an ecosystem. And then they announced the CreatorCon, I think it was called, and they started to build a developer platform. It was very raw at first and then exploded. Now look at ServiceNow. This is similar path.

Udo Sglavo

>> That's our hope, of course. If you ask me, currently, of course we are humble and we are nimble with the models, so I just release them. I wouldn't be surprised that in five years the models may be even the most important market which we are addressing. If we get this right-

>> It's a huge lever.

Udo Sglavo

>> It's huge. Because now we can go literally to every company, regardless of their technology expertise-

>> Plug this in....

Udo Sglavo

>> plug this in, make it happen.

>> And then you have the performance side on the platform and the trust and the security and governance built in, and it's smart and intelligent and fast. That does a lot of the underhand of the

>> platform.It's a platform to add value on top of it. Things that you might not necessarily do yourselves.

Udo Sglavo

>> Yes, that's exactly right. So they can do it. So they can also shine internally.

>> Okay. Coming into focus, that's exciting.

>> So I have to ask you as a culture. First of all, we'd love the SAS culture, lines with theCUBE, very strong people, high integrity, worked hard, play hard, and are family-oriented like you guys are. Now you're going to go public. Okay. You don't have to comment on the public offering. I know you won't. But now you have to build an ecosystem. You have to do things you haven't done before. You got to build a different sales strategy and direct sales. You got to have an ISV partner network. You have to have customer success integrated into that. All new mechanisms. The good news is on the contracting side, you're used to service. So the product, customer side, revenue, I think that's we solid. But the partner, that's going to probably be a big part of the future revenue. You got to get that right. What's it like internally? Are you guys feel like you're comfortable with there? Are you going to take your time? What's the

Udo Sglavo

>> instinct?No, we accept the challenge. I think Brian Harris said, we are a 47-year-old startup. Of course, we are not a startup when it comes to size and revenue-

>> And mentality,

Udo Sglavo

>> But the mentality is like this. And if you look at SAS like maybe 10 years ago, it's nothing like the SAS, which we talk about today. So we had this wave of evolution, let's say, and now we are basically saying, "Okay. We are ready for the next challenge." And there are a lot of dance smart people on our campus, and I'm sure we can figure this out.Well,

>> you guys are packaging, you're seeing models as a standalone product. That's the big news here. Obviously this is the future of data plus AI. I love that equation that you had. I'll pull it up in my notes here. I'll end with this. I thought it was very clever. Data plus AI plus decisions plus outcomes equals learning rate. Now, what's interesting is that we debated this learning equals... Usually outcomes is the equal part. It's on the wrong side of the equation, but it's on the other side. It's on the left side. The equals learning rate implies always changing, hence degenerative. What do you mean? What does that mean? Can you share inside the why you say learning rate because you see it never stopping or is that what-

Udo Sglavo

>> It's basically also reflecting the maturity of our customer base. Some of them, even if they have all the ingredients right on the left side, the amount of analytics they apply may still be limited to a certain areas. So their learning rate may be lower, but then they may challenge themselves and say, we want to become more AI driven and maybe even eventually a completely AI centric company, where AI is a focus area of us as a company, as a corporate strategy. And that's where then the learning rate of course increases. I mean, new gentlemen know that much better than me. We live in such volatile days. It is almost like every day we are waiting for the next thing to happen.

>> Crazy world right now.

>> But we're getting smarter and smarter and smarter as collectively. Nobody wants to go backwards.

Udo Sglavo

>> No, I hope so.Thank

>> you for coming on. Great to see you. And congratulations. As we wind down our model decay is kicking in. As we try to finish strong, final question for you to close us out. Next year, we're sitting down talking. Where will we be? What will be talking about then?

Udo Sglavo

>> Okay. So my expectation for our team is to be ready to show success stories for the modeling pieces. So I want customers to go on stage and basically proof that what we talked about now is real. And of course, we are all excited about this and we put a lot of energy into this now, but the proof is in the pudding. Eventually, it's not us who decide success, it's our customers. And I want customers to stand up maybe on the main stage and say, "The stuff they talked about last year, it saved us X million dollars. And guess what? We implemented it in two days."The

>> Georgia-Pacific example was very, very strong.

>> That was solid.

>> So if you can show many, many more of those applied to the models, wow.

>> More production workloads. Our team said at the beginning of the year, this year, we're looking to score people up on their AI chops by the judgment of the performance. And then the KPI for us is how many AI workloads are in production? And that's going to be the beginning. And then you look at, "Okay. What's happening there?" So if it's not in production, that's still cool, but it ain't the top of the mountain. That's where we look at it.

Udo Sglavo

>> That's exactly right.

>> Great to have you on.

Udo Sglavo

>> No, thank you

>> so much having-Thanks for helping us wrap up the show here at SAS Innovate. I'm John Furrier with Dave Vellante for the whole team, the whole crew here. Brendan, Noah, Rob and the team, thanks for all your support and thanks for watching. And go to thecube.net for all the videos, obviously is real-time here. And siliconangle.com with John Furrier with Dave Vellante. I'm John Furrier. Thanks for watching.

Udo Sglavo | SAS

search

>> Welcome back everyone to theCUBEs Live coverage here in Las Vegas for SAS Innovate 2024. I'm John Furrier, host of theCUBE. Dave Vellante, my co-host, co-founder, as well as he runs theCUBE research. As we wind down two days of wall-to-wall coverage, we got the best for last. We got the star of the show, Udo Sglavo, VP of applied AI and Modeling with SAS, the key player in a lot of the AI that's going on in their strategy, what's under the hood and the models. I think one of the most important things here is the introduction of the model. So great to have you back. Good to see you.Thanks for coming on.Good to see we're all going to get the energy. This last segment, we're going to amp it up.

Udo Sglavo

>> Wonderful.

>> Tell us how are you feeling?

Udo Sglavo

>> I'm excited, but I'm also a little bit exhausted because there was a lot of interest at this event about the announcements we made, releasing models as a product and just of course a major step in my mind forward for us as a company and the industry, if I may say.

>> So I want to just give everyone a quick reset. You guys introduced the SAS, get the three things, productivity, performance, trust, three pillars. You introduce the concept that there's buyers of AI, users of AI, buyers of solutions. Basically, I'm using AI and then you've got the people who build AI. There are certain companies that have to get the hardcore machine learning, hardcore ops, hardcore AI, and then you got the subscribers of model people who are going to use models into the stuff, whether they're building AI or using it. I think the first company, other than OpenAI that had that little library or marketplace, you guys are bringing SAS models that you guys have vetted into one place in use cases that are known and well understood, applying generative AI to fraud. Take us through what that is because with doing that, people in theCUBE have come in and said it from the show. That gives you some advantages, put governance around it. Take us through what is actually in that model package that people can subscribe to. Take us through it.

Udo Sglavo

>> Absolutely. And I mean you basically gave a wonderful overview of what we are doing already. The point is that we've been in the market for a long time, almost 50 years. Over this 50 years. We solved a lot of individual customer problems, very specific, unique challenges of customers, mostly using our consulting resources. And now we are basically saying, "Why don't we take all that IP, which we gained over the years, combine it with the algorithms which we deliver, and package that up as a put into production systems because it's all sitting in a container and it's basically accessible through APIs." So you were wondering about what's inside of that container? Well, obviously it's the logic. There is some math, there is some way how we solve these problems, but on top of that, we also have metadata about the model. So we also launched this concept of a model card. So the model will also say, "I have these ingredients, these KPIs, which can then be fed into a model card." And to make things even more exciting, we also include prompts so that when you deploy the model, you can have actually a discussion with the model. So you can say, "What can you do?" And the model will say, "I'm a model for fraud detection. These are the questions which you can ask me."You

>> actually made the comment that you wished we heard more about democratization of AI, but in a lot of ways when you're offering these out of the box AI models, you're dramatically lowering the technical expertise required to exploit them, and that is a form of democratization.

>> Well, before you ask him, I wanted to say for clarity, I challenged the democratization on the opening keynote because the focus was on the developer side of it. So the data science developer piece of it, where they had that red cube.

>> You think the understated-

>> No, the Viya Workbench is the democratization angle. So they have two things going on here.

>> As is the model out of the box model.I

>> mean, that's democratization. That's bringing new capabilities and faster time to get stuff done or figured out or discovered.

Udo Sglavo

>> Also, putting it into hands of people who are not used to use this models. So we had this example in the opening session where we worked with a large CPG company and their business case is that they have several inventories where they can store items. Once they want to use these items in the production line, they will basically pull an inventory out of the warehouses. So the question they have is, well, if a truck driver arrives in the morning, where should the truck driver go to deliver his or her goods? So in the past, the planner would probably pull out a piece of paper and say, "Well, yesterday we determined that there's space." With the model which we are introducing, an optimization algorithm will run behind the scenes, it'll take all the constraints into account and it'll then communicate with a truck driver in language, please go there, find this person and deliver your goods there. And this is almost like Google Maps, isn't it? Where you're basically saying, I want to go from A to B. I not need to know how the root calculation is handled behind the scenes. I just need the answer to my specific problem.

>> The prompt libraries are critical, and as a nuanced part of the announcement. I'd love to get your thoughts on this. We'll come back to the models, the prompt storage because that seems innocent on the surface. Yeah, I got some stored prompts that like a canned prompt. But I think where it leads us is more prompts and ultimately a prompt-less environment where now under the covers, the system can work on your behalf. Am I reading that right?

Udo Sglavo

>> Yes, absolutely. That's exactly the idea that it's more driven in the communication with the model rather than a person dreaming up Bronx. What may I say to the LLM, which is of course different how we are using LLMs today in our personal lives where you just go there and you may ask random questions. The model has a specific charter and a specific task. That's why we envision that. The first question is what can you do for me? It's almost like... But imagine now you have a library of models and-

>> Before you leave, Dave and I wants have one more thing. The other thing I want to point out that you guys use the word lightweight, and I want you to explain what that means because we've been saying we were the first ones to put out the power law of models a year ago. First research firm to do that. We were right. A little self-congratulatory, humble brag of course, we're usually right most of the time. But the power law, our thesis was the large language got all the attention, but the value was in the specialty models where the IP was. A model can be small, but high-Optane model, very powerful. So that happened, that's what you guys are basically pointing out here. So the payment fraud detection, the welfare fraud detection, document vision, all digital assistance. You guys have data and pre-canned, I won't bad word, but pre-packaged the models for use in well-known use cases.

Udo Sglavo

>> And the lightweight angle comes that we want to be able to plug in eventually into any IT system. I mean obviously initially it will be designed for the SaaS user base and we will target our own market, but a model can potentially run even outside of the SaaS ecosystem because it's mass, it's large language model, prompts, it's metadata. So we can even imagine that you can basically deploy our containers, let's say inside of a database vendor. And in fact, we are already doing some experiments there that with the more modern database vendors that we can say, why don't we put the container inside of your ecosystem so there's no data movement, and then basically, it's all happening inside of the database. Does that make sense?

>> Yeah, absolutely. No data movement got my attention. Now you mentioned, I think it was in the media briefing that LLMs are actually... Foundation models are a small part of the equation here. We were talking earlier about Llama 3 was just introduced and or announced, and we get excited about these things, but so how are you actually using LLMs and what are the most salient aspects of these models?

Udo Sglavo

>> So two things about that. So first all currently we are using LLMs mostly for the interaction with the model. The model itself basically uses whatever algorithm we believe is appropriate. There is this misconception that you can all of a sudden use LLMs for any problem of the world. And it's very interesting to see that when we have discussions with customers about their use cases for LLMs, the first reaction is almost always, but that's an optimization problem or this is a forecasting problem, why don't you use the appropriate algorithm. So that's step one. But our vision is that we want to be able, and this is more targeted at a use case where you build the model in collaboration with the user. So the LLM will then ask questions in a similar fashion like you would ask questions to a data scientist. So it'll basically try to determine what is your business question, in which category of analytics does that fall? And then once it knows that, it'll ask questions, which will help to construct the model on the fly. And this is pretty exciting research that we are doing. Of course, it's not ready yet, so this will not be part of the first.

>> So I got to ask you about some of the trends we're seeing and get your thoughts on it and get your reaction. I mean, that's all what I love about this market. It's unfolding as we see it in front of us. You guys are doing a good job with it as well, and you're digging ahead and get the products. Cross modality reasoning is becoming a big topic. Let me explain what that mean by that. You probably know what I mean, but . Its reasoning is a hot topic. So prompt response, you get an answer prompt, get canned, certainly promptless, everyone sees that with ChatGPT and other prompts. Reasoning means I got to figure out stuff. Basically solve something, figure reasoning and understand it, the context, all that stuff. So reasoning has become a hot topic. Now when you have images and text, maybe DNA sequencing, I mean if it's not text, it's another mode. So multi-modal means not just text. Text is actually easy. So text is great, and that's why all the use cases are text-based. Images are huge. And you gave a demo today of what we all see in old school government contracts, paper with handwritten stuff on it. There's a lot of that stuff around. That's an image, that's computer vision. So as the modals come in, it's hard to do cross model modality if it's not all in one place. You're smiling because you're like, you agree?

Udo Sglavo

>> I fully agree.

>> And this solution is what? Put it all in one place, centralize it or-

>> Federate it.

>> Tell us your vision on this because I think if you get this right, you can save a lot of efficiencies and latency and get more time to value.

Udo Sglavo

>> I think we discussed that earlier. In my mind, the one thing which separates SAS as a company from a lot of our competitors out there is that we like to be output driven. So we think about the output, the decision first, what are we trying to accomplish? And then we think about, well, which technology, which mode from this different universes, text streaming data images is required to address that problem. And then we basically build the pyramid button up. So it's a top-down button up approach, if that makes sense. Again, we want to deliver these models as lightweight containers, and the only thing you need to know are the APIs, which you can use to input information to the container and the APIs, which will return the information. So eventually we want the users just to focus on that, and basically what happens in the guts, well, that's our business. Does that makes sense?

>> Yeah, totally does. So then governance becomes a big important piece.

Udo Sglavo

>> Wonderful, yes. 100%.You

>> can't get AI without governance solved first.

Udo Sglavo

>> Yes. You see, this is why we also want to focus initially on our own SAS ecosystem because we also have tools like Model Manager, which help us to determine is the model still valid? And that's something we should typically do over time. And the thing we are concerned about is what's called model decay. Basically the accuracy of the model gets worse and worse and worse. The model itself will not know that, of course. So we are also thinking about, well, how does this work if you're not in the SAS ecosystem? So somehow we need to introduce some persistence layer to the model, so it can basically store information over time and then it can assess itself. How am I doing?

>> It's like observability for LLM.Yes,

Udo Sglavo

>> 100%Whole

>> other category of observability created in the market.

>> Can I ask you a question about the document vision demo that you showed on stage?

>> That was awesome.It

>> was awesome. My question was how is it differentiated from some of the things we've seen like in document understanding from say some of the RPA vendors? Is this different? Is it new tech capabilities?

Udo Sglavo

>> Let me say that there are similar offerings in the market. This is not a unique new offering. What makes this unique is that this is based on our consulting experience. So this is stress-tested technology, which we implemented several times for very complex PDF formats. Maybe I should explain what we are doing. So the question of documentation, which it tries to address is you have unstructured data maybe in a PDF format. And the PDF document is not just text. It's all over the place, handwritten notes, tables, images, maybe some key relationship which you're interested in. And we are using a mix of AI and NLP after we process the PDF document with an OCR service. And for that, we are using Microsoft at this point in time, but it could be any OCR service to add additional information, which we can get out of the document. Does that make sense?

>> Yeah.

Udo Sglavo

>> Absolutely. So the idea is then basically you're turning very unstructured data into structured data. So what is really new in my mind is this is, and I repeat myself, stress tested. It's working. We have done this many, many times for very complex customer engagements, and now we are basically packaging that up to provide it to other customers.

>> So that means higher yields, better outcomes that make Six Sigma for business applications where you wouldn't normally have that level of quality.

Udo Sglavo

>> Yes. That's what I'm thinking.

>> Yeah, makes sense. Okay. So when does that ship?

Udo Sglavo

>> Our goal is to release the first models in the fourth quarter of this year. And of course, we want to be modest in the way how many models we roll out because we want to assess the market as well. Where should we be heading and what models are required? But our vision is of course, to scale this up. I always think about this as a model marketplace, which you need to introduce.

>> Certain you're going to have app marketplaces because if developers get their hands on this with Viya Workbench, you're going to start to see a lot of experimentation. And then people will accidentally on purpose discover things through their ease of use of self-service.

Udo Sglavo

>> There you go.

>> I mean that's a console that looks like Amazon Web Services 2010. Okay, it's been up from EC-2. Now you got cloud, compute, you got the models all there. That's going to create more developers.

Udo Sglavo

>> And you see also an aspect which we haven't discussed at all is we are also going after companies who are so far reluctant investing in AI because they feel like, "We don't have the people, we don't have the environment and

>> how-We don't want to give our skis and pay all the service....

Udo Sglavo

>> and how can we show fast return on investment?" Now with models, you take not much of risk. You pick one use case, you basically get a quick results because everything is already ready for you. Just plug in the container and then you can basically grow your appetite for more AI while you're actually seeing value. And this is different to how we approached the market before, where we would basically say, "Well, buyers or builders of AI, you're either a buyer where you basically build a ready-made solution from us, or you are a builder, then you have to do these things yourself." Now with models, you don't have to be in the other two categories.

>> So it's because our data is consistent as we show just under 20% of the firms, and this is a decent sized survey, 18, almost 2000 say they're not pursuing GenAI because moving too fast, it's too risky, too complicated, hallucinates, we don't trust it. So they say, "We'll wait."Yes.

>> But their experimentation is with their own data. That's why RAG is popular because it's easy to do.

>> These guys, they claim they're not doing. So I reached out to several and said, "Are you really not doing anything? Yeah, it's moving too fast and we don't trust it and we're just going to wait." Now you can go to them and say, "You don't have to wait. You've got these pre-built models." You must've heard this.

Udo Sglavo

>> Yes, several times. And also a little bit tongue in cheek. They don't even need to know that there's AI in the model package. Because the discussion is different. The discussion is, "You've got all these PDF files with unstructured data. Don't you need it for reporting or for analysis, even if the analysis is just some sums and means and whatever. We help you with that." So you see, we are lifting the discussion away from the technology discussion.

>> Yeah. It becomes invisible in way. It's an-

>> You're providing a service. I'm not going to compare SAS to the hyperscalers or Meta, but if you look at what Meta's doing with LLlama 3 right now, what they're essentially are doing, this is a genius strategy move for them, and it's similar to what you do. I'm going to get your reaction to this because I'm going to tie it to something. They're putting in hundreds of billions of dollars of investment from a major tech player for developers with Llama. So that is just basically giving developers free open source software and infrastructure to run it on to start. Now that's hundreds of billions dollars. You guys are putting 2 billion into AI development that you're going to offer to your customers. So in a way, there's a parallel there. So we're seeing the new branding or brand value in targeting, not the tech nerds per se, the developers who were creating value. So we're shifting to seeing the entrepreneurial and the innovation people. Stars aren't the hardcore developers, they still are. But the new stars, the new rock stars, the new people that are going to, could be all fans for is going to be business people or anyone because now you have a developer using all that infrastructure for LLlama 3 with Meta, you're bringing SAS's DNA to the table, Google next last week, they're bringing all their AI chops to the table. That's what you guys did. And then said, "Hey, let's put 2 billion into it and then it out to the market. And that's our competitive strategy."That's

Udo Sglavo

>> exactly right. And you summarized it very nicely. But I also want to say that as part of our vision, and this will not happen right away, as part of our vision, we want to open that ecosystem also to partners. So we can say, you have the intellectual property, we have the technology. Why don't we go jointly to the market? We as the trusted provider of analytics and technology, you as the trusted advisor of business knowhow, that's exactly what we want to do. And then the idea is basically that we could even say, give us your models and we'll turn it into software behind the scenes.

>> Let me ask you a question.

>> So can I follow up on that?

Udo Sglavo

>> Yes.

>> I want to pick up something you said earlier. You can containerize these models and bring them to data platforms. So I can bring the models to the data and that's part of the vision? Or is it you actually have plans to do that in various-

Udo Sglavo

>> We are researching that today. So we have concrete discussions with some of the modern database vendors.

>> There's three hyperscalers and there's Databricks and there's Snowflake. I mean, those would be the obvious places to go. So some combination of those guys.

Udo Sglavo

>> Read his eyes. Dave, is asking the question I'm like studying the end of it. It has a yes in there.Then

>> beyond that, how do you 10 x or maybe even a hundred x this. Okay. That's what you're talking about now, as developers. So maybe they get a little taste in one of those modern data.

Udo Sglavo

>> You mean that's the show floor?

>> This exhibitor in ecosystem, because you mentioned scale before. Is that how you scale this?

Udo Sglavo

>> That's how we want to scale it. And the plan is to first scale it within SAS, because our current modeling team is US centric. So if we come up with a tax fraud model as an example, which we'll work very well for the US market, it won't work for Germany, which has a completely different text work. So we will basically pull in our consultants from Germany because SAS is a global company. So we have modelers around the globe and basically say, "Well, why don't you take our model, adjust it for your market needs," and then we create a specialized German version of the tax fraud model. So that's step number one. Step number two is then the partners and basically say, "Why don't you do that?" And we become more of the brokers of the models.

>> So developers and partners... You remember a ServiceNow when we first started doing the ServiceNow shows like 2013, it was actually here at the Aria and it was about this size of an ecosystem. And then they announced the CreatorCon, I think it was called, and they started to build a developer platform. It was very raw at first and then exploded. Now look at ServiceNow. This is similar path.

Udo Sglavo

>> That's our hope, of course. If you ask me, currently, of course we are humble and we are nimble with the models, so I just release them. I wouldn't be surprised that in five years the models may be even the most important market which we are addressing. If we get this right-

>> It's a huge lever.

Udo Sglavo

>> It's huge. Because now we can go literally to every company, regardless of their technology expertise-

>> Plug this in....

Udo Sglavo

>> plug this in, make it happen.

>> And then you have the performance side on the platform and the trust and the security and governance built in, and it's smart and intelligent and fast. That does a lot of the underhand of the

>> platform.It's a platform to add value on top of it. Things that you might not necessarily do yourselves.

Udo Sglavo

>> Yes, that's exactly right. So they can do it. So they can also shine internally.

>> Okay. Coming into focus, that's exciting.

>> So I have to ask you as a culture. First of all, we'd love the SAS culture, lines with theCUBE, very strong people, high integrity, worked hard, play hard, and are family-oriented like you guys are. Now you're going to go public. Okay. You don't have to comment on the public offering. I know you won't. But now you have to build an ecosystem. You have to do things you haven't done before. You got to build a different sales strategy and direct sales. You got to have an ISV partner network. You have to have customer success integrated into that. All new mechanisms. The good news is on the contracting side, you're used to service. So the product, customer side, revenue, I think that's we solid. But the partner, that's going to probably be a big part of the future revenue. You got to get that right. What's it like internally? Are you guys feel like you're comfortable with there? Are you going to take your time? What's the

Udo Sglavo

>> instinct?No, we accept the challenge. I think Brian Harris said, we are a 47-year-old startup. Of course, we are not a startup when it comes to size and revenue-

>> And mentality,

Udo Sglavo

>> But the mentality is like this. And if you look at SAS like maybe 10 years ago, it's nothing like the SAS, which we talk about today. So we had this wave of evolution, let's say, and now we are basically saying, "Okay. We are ready for the next challenge." And there are a lot of dance smart people on our campus, and I'm sure we can figure this out.Well,

>> you guys are packaging, you're seeing models as a standalone product. That's the big news here. Obviously this is the future of data plus AI. I love that equation that you had. I'll pull it up in my notes here. I'll end with this. I thought it was very clever. Data plus AI plus decisions plus outcomes equals learning rate. Now, what's interesting is that we debated this learning equals... Usually outcomes is the equal part. It's on the wrong side of the equation, but it's on the other side. It's on the left side. The equals learning rate implies always changing, hence degenerative. What do you mean? What does that mean? Can you share inside the why you say learning rate because you see it never stopping or is that what-

Udo Sglavo

>> It's basically also reflecting the maturity of our customer base. Some of them, even if they have all the ingredients right on the left side, the amount of analytics they apply may still be limited to a certain areas. So their learning rate may be lower, but then they may challenge themselves and say, we want to become more AI driven and maybe even eventually a completely AI centric company, where AI is a focus area of us as a company, as a corporate strategy. And that's where then the learning rate of course increases. I mean, new gentlemen know that much better than me. We live in such volatile days. It is almost like every day we are waiting for the next thing to happen.

>> Crazy world right now.

>> But we're getting smarter and smarter and smarter as collectively. Nobody wants to go backwards.

Udo Sglavo

>> No, I hope so.Thank

>> you for coming on. Great to see you. And congratulations. As we wind down our model decay is kicking in. As we try to finish strong, final question for you to close us out. Next year, we're sitting down talking. Where will we be? What will be talking about then?

Udo Sglavo

>> Okay. So my expectation for our team is to be ready to show success stories for the modeling pieces. So I want customers to go on stage and basically proof that what we talked about now is real. And of course, we are all excited about this and we put a lot of energy into this now, but the proof is in the pudding. Eventually, it's not us who decide success, it's our customers. And I want customers to stand up maybe on the main stage and say, "The stuff they talked about last year, it saved us X million dollars. And guess what? We implemented it in two days."The

>> Georgia-Pacific example was very, very strong.

>> That was solid.

>> So if you can show many, many more of those applied to the models, wow.

>> More production workloads. Our team said at the beginning of the year, this year, we're looking to score people up on their AI chops by the judgment of the performance. And then the KPI for us is how many AI workloads are in production? And that's going to be the beginning. And then you look at, "Okay. What's happening there?" So if it's not in production, that's still cool, but it ain't the top of the mountain. That's where we look at it.

Udo Sglavo

>> That's exactly right.

>> Great to have you on.

Udo Sglavo

>> No, thank you

>> so much having-Thanks for helping us wrap up the show here at SAS Innovate. I'm John Furrier with Dave Vellante for the whole team, the whole crew here. Brendan, Noah, Rob and the team, thanks for all your support and thanks for watching. And go to thecube.net for all the videos, obviously is real-time here. And siliconangle.com with John Furrier with Dave Vellante. I'm John Furrier. Thanks for watching.