 Hello and welcome to CUBE's SuperCloud 4. I'm Rob Streche, Managing Analyst with theCUBE Research. And today we are going to talk to Igor Yablokov, CEO of Pryon. Welcome Igor, glad to have you on here. Somebody who's an OG in this and we're going to be talking not only just gen AI but AI in general and kind of looking at where it came from and where it's going. Thanks for having me. So why don't we kick off, I think in the pre-session we were having some discussions and both of us I think come out of the same feeling which is this isn't totally new where we are and from a technology perspective. There are new applications of what's going on with AI and machine learning and different algorithms but really this has been around for quite some time and that's actually where your company has come from. A lot of that history as well. You're exactly right. This has been decades in the making in some ways. So the old guard typically was working on AI for improving accessibility, improving safety if you think about your texting while driving scenarios to bridge cultural divides with early machine translation and the like. So many of us have been working on the soup for a long while. Yeah, I think it makes sense and I think what's interesting is your company is really focused on how do you make knowledge management more accessible and how do you connect disparate drives and boxes and drop boxes and bring all of those documents together so hey, if I'm looking at HR or something like that I can go in and say, and I think I like the demo on your site there, you know, I feel sick what should I do kind of scenario? Yeah, and this has been a long time in coming. So I used to lead the multimodal research team at IBM where we discovered the baby version of Watson and when they didn't want to commercialize it many of us departed and stood up our last company which a year into it was started secretly working with Apple on Siri. This is well before they acquired the SRI team. And then a half a decade later it ended up becoming Amazon's first AI related acquisition that birthed what many people know as Alexa. So Alexa's my older sister's name which is a complete coincidence but the code name for that was Pryon and that's why we ended up reusing it for this company. No, that makes total sense. And I think being around this for quite some time and having my friends who are out there in fact some will be on later on in the actual super cloud event that we're having but one of the things is, you know so many organizations that I talked to today that aren't, I guess you could say haven't been around this are sitting there and trying to understand where do they get started? What is the first thing? How do they look at this? What's your advice to those companies and where should they really, you know how should they think about it? How should they get started? Yeah, and I guess the first step was when we got started many years ago there was just our dearth of enterprise AI there was just nothing in our respective workplaces and it didn't matter if you were attending a manufacturing plant a hospital, you were in a government agency there was just no enterprise AI to speak of over half a decade ago. So we decided to catch her on football and the paradox with these style of systems is when you walk into a Best Buy and you buy an Amazon Echo or HomePod and you bring it home and you plug it in it has a couple of different dozen domains essentially language models ready to go out right out of the box, right? So news, weather, sports, music and things of that sort the paradox to bringing this into your workplaces is that I would have to show up with a blind trust essentially an empty vessel that has no idea what content you're gonna put into it and yet you're going to be expecting top marks on accuracy, scale, security and speed especially if you're going to be introducing it into critical infrastructure. Yeah, no that takes, that totally makes sense and I think a big piece of it is looking at okay, where do I start, you know and I, some of the companies to your point it's what data are you really trying to enable and what are you trying to get at and what's the use case behind that or the business reason for doing it. One example that I like to use is hey I'm the CFO and I need people to come up to speed on how we actually organize our financial returns and how we actually build them out and what our jargon is around that and our financial jargon. So maybe I build an LLM or an SLM a segmented language model that is really specific to that set of content. Is that where you see companies that you're working with? I mean again, obviously from a knowledge management that's a very segmented, you know hey I'm going after HR or I'm going after the CFO's office or maybe even customer support or customer success. Well, unfortunately there's about 17 different risks associated with adoption of LLM style technologies in these enterprises. And so Brian was actually designed differently. So while you're hearing a lot from the picks and shovels crowd, so hey let's construct your own LLM or let's fine tune a model that's available to you from open source land we actually took a third approach which is to only leverage these style of technologies to model language but keep the enterprises content from not mingling with that core itself. And that way you don't have the issues which from model drifting to hallucinations that you've seen discussed all over the press. So does that also mean that do you get pressure from others that are say hey well I just heard about LLM2 being open sourced by Meta and hey what is the underlying models that you're using can I bring my own model and do you get pushed on that? LLMA is basically the equivalent of Lucy pulling the football from Charlie Brown, right? Because when you actually read the terms and conditions it doesn't allow you to use it in critical infrastructure which many industries that many of you work in could fall under that umbrella as well. So it almost is a marketing gimmick that hey look at this we open sourced it but then on the flip side you can't actually use it for anything real that supports our actual communities. So that's what that represents. Look there's not enough data scientists and software engineers and AI practitioners to go around. So while some of you know how to construct LLMs and some of you know how to fine tune models the vast majority of our communities don't know how to do that. And the same way that most of us don't know how to be chefs or cook food. So what do we do? We drive into McDonald's and we just need something you know that has rapid time to value. In that case it's chomping on a hamburger or in our case it's you know within minutes can you ingest an enterprise's existing content and immediately start retrieving you know positive intelligence that could help you run your environment. How do you deal with just the data scale aspect of it because you know when you start to look at this and you know again garbage in, garbage out how do you deal with that? You know just not only the scale but the quality aspects of data. Well I mean we're glens for punishment so we actually started with the hardest which is unstructured data right? So these are your documents these are audio, video, images, text right? Web pages, PDF files, PowerPoints, Word files and the like so it automatically has to cleanse. It connects to your existing systems or record you don't have to replace or displace any of your existing investments. It should automatically ingest it should automatically cleanse and it should automatically prepare it and curate it so that it's available to you as your own private incarnation. Look because we're OG practitioners we don't even need a single API call out to anything that the hyperscalers have. So this is a fully self-contained platform that's gonna allow these enterprises to choose how they want to consume such a platform whether it's public cloud multi-tenant which is how a number of our existing clients use it or before the end of this year private cloud and on-prem. Wow okay so on-prem which is I think again to me that to me makes a lot of sense where a lot of people are looking at it going I want to control my costs. You know hey I already went and invested in buying some NVIDIA gear or I have some really good top of the line CPUs. How do you help them? And as you move towards that private cloud as we say cloud is an operating model it's not a place. So as they bring cloud closer to them are you seeing that they have to they can run this on standard equipment or is it hey you have to go get an eight way NVIDIA 100 machine and you know be able to support this kind of throughput. Yeah your intuition is spot on because again of our previous experiences this inference is so compute efficient it can run on x86 if it had to. So that way it has high degree of resiliency to changing conditions because look the one universal constant is there's always gonna be scarcity you know in A accelerators in water and electricity and food there's always scarcity in the economic environments that we operate in. And so if we build to that scarcity that gives you a lot more flexibility to deliver services you know where they're most needed and it also the flip side of being able to operate in with such compute efficiency as it gets you to much larger scale. So when you look at some of the enterprise search cognitive search capabilities that some of the hyperskillers are throwing out we don't operate one order of magnitude larger than them we operate multiple orders of magnitude higher than them in terms of capability. Yeah no I think to me that is especially where you want inference potentially at the edge and as we go to like you said hey having inference in the car especially with critical infrastructure I think you know as people look at auto driving cars and things of that nature you know these cars are not running you know 10s and hundreds of Nvidia cards or GPUs or what have you in there they're running x86 for the vast majority of what they're doing with their sensors and what they're what they're doing but they've trimmed the model down to be able to be efficient for that set of tasks. And I think that sounds like a lot of what you're doing is you know again where you're bringing a purpose built model to go and do this and do it close to where the data is. You're exactly right and as they wrap their minds around the power of this new form of knowledge management the closer we drift towards their crown jewels and the closer you drift towards that the more scarcity there is the more security there is the more you know the more failure not being an option it becomes. Yeah and is today are you in their own VPC or is it truly just a multi-tenant deployment right at the moment or you know we're we're definitely headed towards bare metal for them. Our stuff's already deployed in nuclear reactor sites which is a surprise to most folks. That's you know we've been you know getting let's say privileged access to certain styles of content because of the way that this thing has been designed with a pretty watertight security model. Yeah so let's kind of unpack the security model a little bit and not with getting into the technical bits and bytes but what does it mean to be secure when you're talking about this kind because like you said if it's documentation and things like that of you know critical infrastructure you want to keep that you know private for that matter. So what does it mean to be secure? Yeah there's many definitions it means that they run the platform on resources that they control it means connecting to existing constructs like there's single sign on it means allocating the right users to the right knowledge collections. It means presenting the answer only with certain contexts like you're in a certain location and you're allowed to see certain things. It means whatever document level controls that you have are preserved by the system when it gets ingested into the system. You're also preventing spills right? So if one individual is not allowed to see export control one which is uranium enrichment you need to make sure that as you've now transposed that underlying content into this interactive format that they still can't see that knowledge as well. So those are some of the different dimensions when you use that as a use case. Yeah no I think that makes a lot of sense in the fact that you know security can mean a lot of different things and especially as you get into a lot of you know hopefully the US will get on boarding we'll have some privacy laws that help a little bit more like GDPR and other things around the world and was a pipa up in Canada and what have you but I think that becomes really important that information doesn't leak out as well and you know you're looking at it in aggregate not in individual and that stuff there's not ways to prompt engineer around particular safeguards. I would assume that that's a lot of what goes into your core IP is how do you really do this prompt engineering? How do you make sure that the security is what the security is at different levels and there has to be a huge amount of architecture that goes around that. Yeah but think about there is a lot of work associated with it but on the other side of this they get to Nirvana. So look the internet and the web as we know it died in 2022 right because we were used to doing Google searches and seeing you know the creativity of our fellow humans right? You had music, you had videos, you had things that were reading, poetry whatever all of that was human originated well before the end of the decade it's going to be generative nonsense it'll be a hall of mirrors. And so what's going to end up happening in the Fortune 500 is there is going to be a retreat to safety where they're going to have to have the equivalent of their own version of a library of Congress of things that they trust as truth so that they can continue operating their respective businesses and they're going to be careful what partners they allow into that right and so that's what we're foreseeing and why knowledge management is going to be a critical asset to these individuals because they're not going to have anything else they can fall back to so that they can continue operating their businesses. I think that actually leads to a really good final question which is trust and understanding here's where that data came from we were kind of riffing beforehand is crypto people investing in this now now they're all on board with AI and dot AI everything and you start to look at where blockchain is even being positioned as kind of being the copyright. What do you see from helping organizations really understand that source of truth because I totally agree that as companies look at it they want to know where did the data come from is it reliable? How do you have that transparency? You're exactly right so think of this way every organization out there has these assets strewn about all over the organization they're stuck in applications in Salesforce and service now in SAP, DocuSign, so on and so forth they're stuck in repositories such as Confluence and SharePoint so they have all of these critical resources already there but there's a point of friction in terms of getting this to the right people to perform the right workflow so how do you reduce the distance between knowledge and people? It's transforming those assets into this interactive experience that becomes the refinery that eventually powers the cars that are their workflows if we continue following that metaphor so what does that mean? What does that get them? Well now instead of hunting for a few hours you can get a sub second response now know this there's never hallucinations from Pranin's platforms every solution that comes out every answer card you can click on it and I'll show you the exact page and highlights exactly where it learned it from so it's always anchored to an enterprise's fundamental assets so that way they say all right I trust this publicly available information because it came from a regular use website I trust this because it came from some published content that we've licensed into the org I trust it because it's proprietary or I trust it because it came out of my personal storage as well the big vision here is for the first time ever we're going to have this new knowledge fabric in the center of all Fortune 500 companies that's going to be agnostic of human languages that things are all authored or recorded in it's not going to matter where it's stored and it's not going to matter what the underlying object type is and then you get a choice of blending internal and external assets as well maybe from trusted partners trusted government agencies and the like everybody that's listening to this should be getting goosebumps because for millennia we've been hearing about the Tower of Babel and yet it's literally within our grasp now that you will be able to take knowledge from any culture and any language and use it to benefit your communities through whatever workflows you're attending to and we've never had that opportunity before and we're on the cusp of it I think that's a great place to call it a interview here I really appreciate it Igor for coming on board here and really sharing it one of the OGs of AI and ML and the modeling and all of this NLP and other stuff that's under the hood there and that we're going to be really digging down deep into so thank you for coming on board thanks for having me and thank you for watching SuperCloud 4 and stay tuned we have a jam-packed day of this and really appreciate you hanging in here to watch take care