 Welcome back everyone to the LiveCube coverage here in Las Vegas for SAS Innovate 2024. I'm John Furrier, host of theCUBE. My co-host Dave Vellante heads up CUBE Research, also co-founder, co-CE with me, founder of theCUBE. Brian Harris is here at CTO of SAS, multiple times. Second time on this event, we're coming back to go do a deeper dive on some of the tech and really what's driving the AI is the technology but also the business model workflows and data and the three things we talked about last today was performance, productivity, performance and trust. Great to see you. Good to see you, John. Good to see you, David. Always a pleasure. Hey, Brian. So for the folks that are watching, you might not have seen yesterday, you did the hot wings, what was it called? Hot ones. Hot ones. Where you go down. Yeah, they're Sean Evans and start from one wing to ten with a progressive heat going on. It was insane. One of the greatest experiences of my life. Amazing. They show the replay today, the keynote. You were dripping. I was like a tomato. I was still hot out there. Out of body experience. Under duress. Under duress. It was so good. As I said, a lot of hot takes going on there here in theCUBE. So we want to get into it. So I want to get into the AI impact. We're going to go on to the hood a little bit, but we want to talk about performance. So the big thing with AI is you need more horsepower. Yep. I mean, generalizing AI as XPUs, which is GPUs, TPUs, what do you want to call it? And quantum's in the mix there. You brought up in your keynote. And so you're starting to hear quantum in the same conversations now when you talk about these advanced AI demos and presentations, because there's a performance angle there. What do you mean by quantum? When you bring it up on the keynote and with all the goodness you guys are announcing, a lot of accomplishments, a lot of tech, customers with their AI stacks. Why quantum? Well, I think that when we look at the, because obviously generative AI is getting a lot of the headlines these days, but it really doesn't have a strong story in the quantitative space. It's really around reasoning of facts. And obviously they can do things with multimedia. We get that. But when you talk about dealing with numbers, it's actually very, very weak, right? And so you got to start asking yourself, where are we going to start seeing acceleration on quantitative reasoning? It's really what I would look at it, right? And so when you start looking at that, obviously quantum, because of its architecture and ability to have the superposition, allows you to kind of have multiple states at once, which means you can do some really incredible, almost futuristic things. So one of the big problems that we see is you got to find the right problem with quantum, for it's not for every problem. I mean, GPUs exist because they're great at doing linear algebra and multi-threaded linear algebra for deep learning. But there are optimization problems that require combinatorial trial and error type scenarios. And these are some of the hardest problems in the world, and they're very impactful for four industries. Life sciences, banking, right? In the government space, as well as in the, you talk about, like I mentioned life science, but pharmaceutical industry, insurance. And so if you can find ways to compute these really, really hard, combinatorial problems faster, you can really drive the cost of compute down and solve problems that were unsolvable before. The basic example is this traveling salesperson problem, which is like, for instance, you got a salesperson's in a city, right, center, they have to visit a bunch of cities, right, and only visit one of those cities once and get back to home. And they got to do that in a way that minimizes, minimizes distance, and there's no algorithm to do this. You just have to try them all out, evaluate the distance, and then pick the right one. Well, in real world problems, that becomes exponentially complicated. And so quantum allows us to search that entire space of combinations and get to the right solutions faster. I think you said, I think it was you, and your keynote yesterday, it was a great keynote. You said within a year and a half, two years, your quantum is going to have its gen AI moment. That's correct, I believe it. The market is showing a lot of progression on it, so I think it's like $35 billion as the market size right now, but it's projected to go to $1 trillion by 2030. So you got to imagine what's happening in the next six years, the leaps we're going to make in this. And one of the big things that's happening in quantum is this idea of, well, qubits. And this is very analogous to what was going on with traditional computing. We had eight-bit computers, 16-bit computers, 32-bit computers, and 64-bit computing. So scalability of arithmetic is the big issue with quantum right now. So as we go through the qubits, scaling up on that, we'll be able to scale into more sophisticated problems in the market. And I think that when you look at what the demand of complexity of problems that are in the financial world, the hyperconnectivity of all these financial systems right now is creating combinatorial problems that need to be explored differently, and quantum can help solve these things. So, I mean, I really hadn't paid much attention to quantum until recently, because it looks like it's finally starting to get real. And there's a lot of trade-offs and a lot of different techniques. You could scale, but then the qubits aren't stable, or you can get the qubit stable, but it's really slow, which defeats the purpose of quantum. From a technology standpoint, it's, correct me if I'm wrong, but it seems like we're solving those problems. Do you have sort of line on site on maybe which of those techniques, or have you sort of chosen a path that works for your industry? There's two approaches that quantum computers are approaching. And again, we're not in the business of building them, we're in there just to leverage them, but there's quantum annealing, which is, and these strategies are really about, quantum annealing allows to maintain state of a one or a zero, or in the case of quantum, you get this kind of idea of a superposition where you can actually be in two different states at once, that's why it has incredible power. It's like mind-bending. Yes, and so when you think of that, though, like the other one is quantum gates, which is where they're trying to mimic the traditional semiconductor like in transistor. And so those are the two kind of big areas where people are pushing and research on of those. And those different quantum computers have different problems that are good at solving. I think there's a pursuit right now of quantum gates that we could just kind of approach traditional problems one for one, because it looks like what a semiconductor would do where it transitions and things like that. Whereas quantum annealing, I think, is actually really stable right now in many areas of optimization problems, which I said is there's no algorithmic efficiency for that. So just one more follow-up. When you talked about the 816-64, when we went to 64, you had to make some application changes to take advantage of it. And that's what you've got to do with quantum, right? So you're in the process of doing that or not necessarily? It's less so, actually, so it's more around, I mean, yes, there's going to be part of that, but a lot of these, when we leverage them, we're just sending them off as APIs to these things, and that's all kind of hidden behind the scenes, right? It's just a matter of scale of how much can a process of one time based on, and then what's the efficiency of doing that? So, but we see that's just a natural progression in that industry that will be solved, and when it's solved, SAS wants to be ready to be leveraging these things. And what we're doing is we're putting them in this hybrid architecture of traditional computing and quantum. And I would say this, you know, GPUs didn't take over the CPU, they're just a specialized class of XPU, as you called it, right? And I think quantum's going to have its space, right? So when you're talking about GPUs, it's great for linear algebra and, you know, matrix multiplication, and quantum's going to be great at search space optimization for how do I discover solutions in a very complicated problem? Now, that will be presented as compute in the cloud, right? Or as a service to a quantum computing company, and you're just going to leverage that service into your existing data and AI lifecycle, and leverage incredible outcomes with it. I have to ask you, you mentioned that you're not building quantum, but you're leveraging, you mentioned it again. How are you leveraging it specifically? Is it in research stage right now, or is it you're moving the needle on product integration? Talk about what you're doing with that from a leverage standpoint, where you see it leveraged, being leveraged for the benefit of the customers, and then what does the customer benefit? Yeah, the, if you look at, like I said, I mentioned that traveling, sales from person problem, right? Where you have to try all the combinations. If you were to run that in traditional cloud computing, you could, it could take you days, maybe months to run those trial and error scenarios, and then finally get a best solution. And that is expensive, especially in the cloud. What quantum will give you the abilities, instantaneously search that entire space because of its unique characteristics and qualities, and then find the best solutions that can be put back into traditional computing. And it's like almost like shortcuts to good enough answers that can be put back into traditional computing, which means all that cloud spend collapses to a much shorter amount of time. So you see it leveraging customers where they can use cases where unattainable outcomes or ideas or unknown thoughts about what they can leverage their data. Like drug discovery, I mean there's a lot of things like molecular compounds, there's chemistry, right? Understanding the physics and the interactions of molecules at the lowest level is a very complicated task, exhaustive computationally. And you can imagine every drug company out there, pharmaceutical company, wants to build a killer pipeline so they can get ahead of the FDA submissions and then that's one of the key areas that quantum can have a huge help then. So I'm going to play the skeptic, okay? Brian, hey, I hear you in all this, you know, quantum is a science project. I know if they're smoking in those labs, entanglements, qubits, whatever, HPC and AI are going to solve this problem. High-performance computing's already there. They can do a molecule stuff. Cloud, they're slow, but HPC and AI, GPU clouds, they're the rage. Is that just kind of the longer distance? I think you got to, I think I appreciate that perspective and I think that's, you got to be careful with the hammer or nail scenario. If someone's making GPUs, every problem's going to look like a GPU should solve it, right? If someone's making CPUs, everyone looks like a CPU. I think it's something that SAS does well is that we're looking at all the technology market, figure out what's the right use of that technology for the problem. And that's why, that's the area of research we're doing. We're doing things like, what are the classifications of problems, what are the industries that's going to have an impact in? It just so happens, it coincides with our top four strategic industries, banking, insurance, life sciences and government, right? There's a lot there. Security, big deal, right? The whole shop 256, right? Everyone's worried about Bitcoin being uncracked open because quantum can basically, you know, try all the ways to go and identify a key, a private key and then unlock, you know, wallets. So there needs to be a lot of research here. So to answer your question, it's not just research for us. We are actually looking and working with customers who want to solve problems. And we're exploring that. You've identified use cases specifically like the crypto hacking or cracking the wallet as a realistic possibility. Yeah, this is a known concern. That's why there's now quantum proof encryption out there because this is emerging at a point where people are concerned about this. So I think for us, we are a company that always wants to deliver results, not hype. We want it, we cut through the hype and look at how we can apply our technology in ways that is groundbreaking but pragmatic for businesses. So what we're doing is when we interact with quantum computers, we bring it back into our SAS Viya ecosystem and then they get all the governance and explainability and transparency, they will get it. It's no different if we were to leverage a GPU for a deep learning model in our model studio. So it's us, we see it as normalizing the way we integrate it into our stack and then the customer benefits from that and they don't have to change their workflow with our software. And so, I mean, quantum computers, they're huge. They're like bigger than that, right? And they're all liquid cooled and then they're loud. High capital, high CapEx. Right, so where are these things going to live? Obviously all the cloud guys are doing it. Satya showed it, that Ignai, IBM's doing it, et cetera. So they're going to be in the cloud but you see all these specialized GPU clouds popping up all over the place, which is kind of interesting. You know, some people are doing their AI on-prem. Where are these things going to live? I think with quantum computing, I mean, it's clear that I think it's going to take the same path as GPUs. So just like, let's say, NVIDIA is a provider to the cloud providers, right? You've got Super Micro, right? These are all supply chains to the cloud providers. I think you're going to see the same thing. I think you're going to see the emerging quantum computing companies become just supply chains to the other cloud providers and then they will present those things, right? Cloud providers will present that as, you know, scale up, scale up, down, compute for your environment. That's why I said, really, we expect quantum, right? Our ability to leverage quantum will be really the GPU for optimization problems in back to your original question. When we have problems that make sense for quantum, we're going to leverage that to accelerate our compute, just like a GPU does. And some of your customers might build their own quantum, I doubt that. Data centers or no? Like even the government? I think, well, the government could get there, but I think that, I don't think that- Yeah, the banks aren't going to do that, right? Yeah, I think that's just, there's too much capital. It's not quarter of their business. And I think that, honestly, there's so much, there is so much still, progression needs to happen on that space that you've got to be focused on that. Again, you guys have been very loud and clear about use cases here, pragmatic, but also high impact. And real quick too on this, we are numbers company. So if there's any company in the world that should be an expert in this, we should be. Because we have been, from the beginning, a statistics company now, an AI company, but so all of that is foundational to all of that is our ability to do quantitative reasoning and quantitative analysis. So quantum is just an extension of us doing that better and faster. We shouldn't think that quantum computers are going to replace traditional computers or should we? Now Jensen would say that GPUs are going to replace, every workload is going to be accelerated so there's really no need for, he implies, there's no need for traditional CPUs, but we know that he's going to be. That's the hammer and nail. Exactly, we know there's going to be a lot of diversity out there, but help us understand that. Help the audience understand that we're not talking about a wholesale replacement of traditional computing, von Neumann architectures. I mean, where are we? The GPU came about because we had the offload graphics off of the CPU, right? And so once we could do that, then we realized the same math that's in GPUs can be applied for AI, right? For gaming, right? Can be done for AI. And crypto mining. And crypto mining, right? And so I think you're just going to have the same thing here. I think the difference with quantum is because it's so big and so nebulous in a sense that people don't feel it as finite as, like, say, a GPU is. We can all put a GPU card in our computer, but quantum is just to see that quantum computer as just another GPU in your options. I mean, SaaS, your model is just a lot of love is whatever computer, XPUs I can get for as much horsepower for the use case that it needs. In some cases, you need monster and quantum and HPC with AI could work too. By the way, Dave, HPC and AI over the past two years, we've been going super computing for the past couple of years with theCUBE. I mean, that's the show that's been around since I graduated college in 1988. They haven't moved a needle. It's an inch by inch. Hyper-portion piece of this niche thing. They're doing stuff high performance now, like wing construction for Boeing. It takes literally minutes now. It used to take hours and hours of compute. So there's advancements on HPC, but that's not quantum. No. That's a huge difference between HPC. Well, in fact, Viya, for SaaS Viya inside of SaaS Viya is something called the Cloud Analytics Service, which is our in-memory analytic engine. And some, you know, we can actually run a multi-node, a very large scale and we have one of the top many semiconductor companies in the world runs a whole automated model tournament using Viya's in-memory compute that behaves very much like HPC. It's holding data in memory and then doing multi-pass calls against it at high, high rates. And they can run thousands of model scenarios to then optimize their yield of their semiconductors. So we do this today, and we're, I mean, this is like, you know, incredibly important to their business because that's how they improve their entire yield, which improves their margins and quality. And those are, they're not huge data sets, are they? They're quantitative data sets. I mean, it's 20 terabytes in size, but you're holding 20 terabytes in memory. Right, it's not 10 bytes, it's exabytes. It's quantitative data sets that are terabytes. Yeah, yeah, it's not, yeah. I mean, I think at some point you have to allow data to sit on disk because it gets so expensive to put it all in memory. So one of the highlights yesterday of the demo was Dr. Goodnight was on stage doing a demo. And a couple of comments jumped out at me. Obviously I'd like to admire the entrepreneur companies doing this for decades. He made a comment, we're not an interpreter, we actually compile our code. Yeah. And because he was referencing the performance. That's correct. Okay, so performance is huge. Let's riff a little bit here, put you on the spot. You presented yesterday, performance is huge, LLMs, you got people who use AI, they'll build their own, they'll subscribe to models. Will all this overhead that will come out of the opportunity to make that faster, compiling LLM, not compiling a little bit, you have to start putting this stuff together. Yes. Data's involved. You guys have good experience with data and making a high performance. Where's the work need to get done on the LLM foundation models on the performance side? Is it integration? Is it mashing up models? Are we going to have a mash up culture soon? Or is it already here? Where's the pressure point? What's the constraints? Not besides power and shit, like at the software layer. Well, let's talk about post model training, right? Into the scoring side, which we've called scoring, but I think one of the big things at LLMs is the latency. I mean, customers want to put these things into scenarios where there has to be some second latency, right? Or maybe a second latency. If you can't approach those responses on that, it takes it out of the entire possibility of being used for- It's not that good. It's not that good. It's not inside the decision window. And you heard, with some of the work we did with Georgia Pacific, they had like, half a second latency to make decisions on these things. So I think that there's just latency as a concern. And then when you think of latency, because the latency is due to a couple of reasons, these rag architectures, while cool, it's just a lot of orchestration across, okay, I'm going to ask an LLM a question. Before I do that, I'm going to go query my internal knowledge into a vector data store. And that's been basically chunked out, right? So they got to do this massive similarity search, bring that data back, present it back to the LLM and say, here's a few shot examples, now reason over this. And then by the way, I need to call an enterprise system over here to go get some real numbers, because LLMs can't deal with that well. Let's pull that stuff back in and put it in there and then present the result back. So there's a huge opportunity for that, the performance of that orchestration, you know? That sounds like an operating system to me. You're linking, you're scheduling, you're loading. That's a great point. I mean, you're basically runtime, because generative AI, generating, it's a new category. You don't know what you're going to get until the prompt or thing happens. That's correct. It's a nondeterministic model. So as a result, we got to react to it, right? Or we're going to have an AI operating system that's going to not look like Linux or anything else. It's going to be some other knowledge graph, neural network. I think as in any technology, any transformative phase of technology, once the design pattern emerges, then people optimize the execution of it, right? You see that, like I said, CPU. How do we offload the CPU to do graphics work? Let's build a GPU, right? And so you can imagine there's probably similar strategies going to emerge in RAG architectures and all of them interactions that you're going to see. I mean, look at even Apple's system on a chip. I mean, the whole point of system on a chip is that they lowered the power needs and the cooling needs because they've got everything integrated in the same silicon. And that's amazing. With the big shared SRAM, it is amazing. AI is a great opportunity for your company. The last question as we wrap up for the folks that are customers and prospects, what is AI doing for your company and where do you see that going? What's it going to turn SAS into? With all that installed base, all those assets you have, you're bringing to the table, what does SAS transform into with this industry shift? I think, you know, for me, I think people should see us as one of the most trusted companies. I mean, we, to deal with AI, to cut through the hype and deliver real results of AI. At the end of the day, as you know, you guys have been, you go to many conferences, you meet with many customers. They're all the words are being said, right? There's only so many words we can all say. So what does it come down to? It comes down to execution, right? How well do you execute, right? And how well do you perform? How does the software perform and how well does the project execute for that customer? And then there's the post-sale experience which is how good is your support? I heard a customer today, two customers, two banks say, SaaS communities is unbelievable. They said that to me directly, unsolicited. And so that's part of the experience, the brand experience of SaaS. So for me, we are excellent with AI and we have the most experience in the world. We've been doing, you know, neural networks in production for 25 years. We talked about that before. People don't know that, but we've been doing it. We've been writing some of the books on these things. We're going to continue to be us, be thoughtful, cut through the hype, deliver real results, and ultimately execute policy with customers. And if you hit those marks, the revenue is just the result of that execution. That's the lagging indicator. You guys are good at like hard math and stuff like that too. But I got to ask Brian about our bet. Did you see my Twitter poll? I did. Okay, so I got to ask you about AGI. Where do you stand on this? So John and I made a bet yesterday. Don't tell him what I did, so don't prompt him. No, I'm not going to bias him. Don't lead the witness. And we bet dinner, like loser buys dinner, as John said, as long as a robot makes it. This is funny. The question is in the poll, when do you think we'll see machines display broadly adaptable intelligence, similar to human cognition? By 2030, many decades from now, centuries from now or never? How do you see it? I want to say never. I think this is like, it's like, as we said, there's a lot of power being used and compute being used just to predict the next word right now. And I know that at the attention strategies we've got with transformers are really powerful, but the reality is that humans draw from so many data points in their experiences in life to make a decision. And what we're really asking is for computers to help scale that. Just like your social media has been an extension of your memory bank, right? Your pictures, your friends, your networks, right? We can scale our network of people because I can keep in touch with them in different ways. But that doesn't replace the human experience. And so I believe this whole, this movement towards like AGI is just, I think it's like guys, to what reason here? To what extent are we trying to do this here? There are a lot of problems we need to solve right now. I think there's reasons to do that, to say we should pursue it, but I think we should be reasonable that that's asymptotic, that I think you don't really ever get there. You just improve the way to scale human reasoning and decision-making. You're walking halfway to this. So you're a human optimist. Yeah, I like it. Okay, I got to ask you a follow-up question. Then where are you on FSD, full self-driving? Are you similar? I have a real problem actually with some of this. I actually think that there's a whole perspective in the past where you build bridges. There's like 11 people that die when you build a bridge and people just accepted that, that should happen. My issue is that where is the transparency to make sure that when people are driving on a road that they are aware there are full self-driving vehicles on the road completely, right, on that? And without that, I don't think we've been transparent enough with society to really make sure that could be done effectively because as we know, software has bugs, right? We all know that. Anyone says that the design of bugs is ridiculous. And when someone's loved one is killed by an FSD car, they're not going to care about the stats that someone says like, well, full self-driving is basically 60% better at avoiding accidents. It doesn't matter when your loved one is in the, even the small minority percentage of people, they just, they're upset that their loved one was injured or hurt. And I think that we have to remember that. Like no one's going to care about the aggregate statistics when your loved one is injured. So we really have to be careful and transparent with this type of technology. So that's, I like to ask technologists that's really, that's a good answer. And I agree with it. It's not really a technical issue you have. It's more like the regulators don't know who's accountable for this stuff. Right, well, that's the thing. I mean, you're going to have lawyers packed. Yeah. I mean, the courtrooms are going to be packed to technologists over the next decade because those are, the people are all going to be trying to figure out who's responsible, who's suing who. And Reggie Townsend on your team, he was on stage, Kara Swisher and the AI was on there. This is going to have to be transparent. Yeah. And because there's a lot of work to do. And as they pointed out, and we did on theCUBE yesterday, there's a lot of good work going on with AI right now. Yeah, yeah. And that's got to get done and the foundation's got to be built. Yeah. Well, I mean, that's the work we all got to do. I mean, that's why I said like it's, I am a human optimist because I believe that we are building AI for us. It's not the other way around, right? I mean, we're building it for our benefit. And so we should be thinking about how we're going to use it for ultimately how we want a better society. Are you happy with the event so far? Oh, this has been unbelievable. The customers, I mean, it's packed. The action, it's so great. I mean, the work that my R&D team has done is incredible. The work that the events team has done is incredible. Our, you know, support Dr. Goodnight getting on stage and just still coding, does it every day. You know, it's just the best. As like I said, on my tweet, Shannon and your team retweeted it. You had the sizzle and the steak. And you guys, and we were at Explore. We said, you said, this is what we're going to do. Yeah. And you guys deliver it. So, you know, when you see production workloads, that's what we said last year going into this year. That's going to be the benchmark for AI reality in the enterprise. That's right. Show me production workloads. You guys have them on stage. So congratulations. We had one, I don't want to say to us all killer, no filler. That's what we do. Frank, thanks for coming on. All right. Thank you. All right. John Furrier with Dave Vellante. We'll be right back after the short break.