theCUBE + NYSE Wired present the East Coast AI Leaders Executive Series | Eiman Ebrahimi, Protopia AI

Clips
More from theCUBE + NYSE Wired present the East Coast AI Leaders Executive Series

Eiman Ebrahimi

CEO

Protopia AI

play_circle_outline Protopia AI: Revolutionizing Data Use in Financial, Insurance, Healthcare, DOD, and Federal Government Industries

play_circle_outline Protopia.ai CEO discusses private AI and machine learning with historical barriers

play_circle_outline Company formation, focus on scalable solutions for machine learning platforms

play_circle_outline Company founded three and a half years ago based on complex machine learning platforms

play_circle_outline Revolutionizing Privacy in AI: Protopia's Use of UC San Diego Tech in Machine Learning Evolution

play_circle_outline Privacy-Preserving Data Utilization: Innovations from UC San Diego

play_circle_outline Unlocking Innovation: Customer Reactions to Private Machine Learning Challenges and Enterprise Value Initiative

play_circle_outline Emphasis on protecting data even with open source models and third-party model hosting opportunities

play_circle_outline Maximizing Flexibility and Security Through Open Source Models While Addressing Deployment and Data Privacy Concerns

play_circle_outline Culture and core competency focused on enabling AI in a safe manner

play_circle_outline Focus on data security and encryption in AI

play_circle_outline Systems thinking approach with focus on hardware independence, framework compatibility, and minimal latency impact

play_circle_outline Progress of private AI adoption and deployment

play_circle_outline Future vision of Protopia: enabling privacy layers for models on any creation environment

Info
Transcript

Eiman Ebrahimi, Protopia AI

Eiman Ebrahimi

CEO Protopia AI

Protopia.ai, led by CEO Eiman, focuses on enabling the use of private data with machine learning and AI without traditional barriers. The company, established for three and a half years, recognizes the increasing complexity of machine learning platforms. Protopia addresses how enterprises can efficiently and scalably utilize their data with shared infrastructure. Eiman's background at NVIDIA informs the company's approach, which involves a post-training step to create a transformation layer called stained-glass. This layer allows for secure use of private dat... Read more

explore Keep Exploring

What is the vision of protopia.ai in enabling the use of private data with machine learning and AI, and how does it aim to address historical barriers in this area? add

What problem is the company aiming to solve with their foundational technology based on understanding the complexities of machine learning platforms and infrastructure? add

What is the challenge currently facing use cases for data implementation in various industries, and how is Protopia addressing this challenge with their technology? add

What are some potential advantages of open source models for enterprises in terms of flexibility and deployment options, while also considering the challenges of protecting information from exposure to infrastructure and the potential need for service providers to host and manage models on behalf of the enterprise? add

What is the core focus and drive in enabling AI use cases in a safe manner without requiring high-cost private, physically separated infrastructure environments? add

What factors are critical for ensuring that investments made in AI by enterprises become self-sufficient and start returning value? add

What is the potential impact of integrating stained-glass transforms into model creation environments for Protopia? add

bolt Powered by CUBE AI

Eiman Ebrahimi, Protopia AI

search

>> Hi, welcome back to theCUBE's coverage here, day two of Media week. We are at the NYSE. This is theCUBE's new East Coast studio on the balcony overlooking the show floor of the New York Stock Exchange. A opening bell just kicked off. Eiman is here. He's the CEO of protopia.ai. Great to see you. Great meeting last night, networking event in New York City. Top founders were here. We had a nice chat. Good to see you. Thanks for coming in this morning.

Eiman Ebrahimi

>> Thanks for having me.

>> So I really like the vibe here in New York. A lot of founder, a lot of action. The tech scene is booming mainly because there's a concentration now of opportunity and talent, entrepreneurial talent, but also with AI coming on board, you've got a lot more disruption in all facets. It's not just, "Hey, I'm a tech company and I'm going to start, I'm going to hit all the SASS vendors," usually the early adopters. Now all companies are adopting. It's been a really interesting dynamic because everyone's now available to be a customer, and so the tech's also moving at Rapid Space. So you're at the center of it. Talk about what you guys do, then we'll get into some of the things that are going around your company.

Eiman Ebrahimi

>> Yeah, of course. So protopia.ai is a vision for enabling the use of private data with machine learning and Ai, but without a lot of the historical barriers that addressing this problem has had. So when you think of historically when we've talked about private machine learning, a lot of the approaches have centered around pretty costly solutions, whether that be in terms of using new platforms, using new hardware, or even not being able to use shared environments at all and making everything private. And so what we're addressing are those bottlenecks in making private AI a reality in the shorter term for essentially the businesses that do exist in places like New York, so financial services, insurance. These are places with a lot of private information that do want to take advantage of AI and how they do that privately is what we focus on.

>> I want to get into the whole data is an actual property angle because that's super important to what you do, but first talk about the company formation, when you founded it, what's the status? Inside the numbers, give us a taste of where you guys are at.

Eiman Ebrahimi

>> Yeah, the company's been around for about three and a half years now, and we founded it really based on this understanding that machine learning platforms and infrastructure are increasingly getting complicated. And to that extent, it becomes really difficult to imagine a world in which these infrastructures are going to be owned and retained at every center where the data lives. Now, even when you have on premises infrastructure, there are a multitude of different data silos that wants to use that infrastructure. Even then there is a sharing happening in some form of multi-tenant sharing of the system. And so with that in mind, which comes partially from the time that I spent at NVIDIA prior to Protopia, understanding those problems was very key in also understanding the opportunity around building a scalable solution to that problem of how can enterprises scalably use their data with infrastructure that ultimately does need to be shared on some level, even if it's within the organization? And so that's been the problem that we've been attacking, but there's some foundationally new technology that we are building upon that came out of University of California San Diego, and it had been developed there for a few years before we started the company, and we've been building upon that foundational technology that changes some of the fundamental assumptions about how data is exposed in machine learning.

>> You brought up NVIDIA. Talk about your background. There's a very interesting background how you got here. You were at NVIDIA, pre-big wave that's there now, but it was the precursor because we all know NVIDIA was working on this for, CUDA was started what, eight years ago? I mean-

Eiman Ebrahimi

>> Probably much longer than that.

>> Much longer than that. I mean, we hear that. Talk about your background NVIDIA, how did you get to where you are?

Eiman Ebrahimi

>> Yeah, so I initially joined NVIDIA as a systems architect. My background is in high-performance computing and computer architecture. And I worked on discrete GPU memory systems for a number of years there. But like many others at NVIDIA kind of evolved into looking at the deep learning machine learning space and how systems could be best built for this emerging application space. During the time that I was at NVIDIA, I was fortunate enough to work on a project that really helped to understand a lot of these future problems that are going to exist, especially as part of a project called Megatron that NVIDIA started building their own language model basically. We were able to understand a lot of these challenges that are going to exist in the future and that became part of what the understanding around this evolving problem space would be, which caused the ability to see an opportunity around that as well.

>> Yeah, Eiman, it's interesting. When I was in college in the '80s, when I would do interns and hopefully get a full-time job being the technical person, I was in the DP department, which stood for data processing. And you mentioned those projects. I mean we are seeing major data processing improvements with the system architecture. We saw ACI come on, hyperconvergence built hyperscaler, but now we're seeing the system architecture in these clustered systems completely being reset, reset and re-specced as a system architecture to handle large-scale data processing. That's what we're talking about here.

Eiman Ebrahimi

>> Correct.

>> And that data is public data, private data, which you're working on, and so you're starting to see that the old server model was put some servers on a rack and put a top of rack switch, okay, that's not any more relevant. You're seeing this new look at all these servers connected together as one thing.

Eiman Ebrahimi

>> That's right.

>> This is the sea change. This is going to increase the capability to enable new applications. And we saw this wave with the PC and PC servers, which was as it got faster, smaller, cheaper, the apps got better.

Eiman Ebrahimi

>> Absolutely.

>> And when the apps get better, entrepreneurs say, "Hey, there's a new app. I can do something new I couldn't do before." This is, to me, the real critical point around this wave is because all the yesterday apps are going to work, especially if they have a lot of data want, and then now new things are emerging. What are you seeing there? Because there's low hanging fruit. "Okay, I'm a big bank in New York and I got a lot of data. By the way, it's my data. I really don't want to blend it in with a large language model or a foundation model, I want to interact and take advantage of it, but at the same time I want to leverage it." It's my leverage for property, so what's my server architecture look like? So now this is kind of where it comes together. I think this is where I think you're at the center of the act.

Eiman Ebrahimi

>> Absolutely.

>> You're on the front wave of this. Hey, you need to bolt on some capabilities or redesign from scratch foundationally. Take us through your thoughts on this.

Eiman Ebrahimi

>> Yeah, so there's something very fundamental about how data is exposed to these systems. That has always been the cause for concern by the data owner of where does my data actually get exposed? And when we talk about exposure, data is generally encrypted at rest, it's encrypted when you transmit it somewhere, but for a very large majority of the applications out there, almost all of them in the deep learning and AI space, as soon as the data gets to the implementation, it needs to be decrypted in order to be processed. And that's where the assumption has generally been that the exposure of data to the application and the infrastructure it runs on is going to be plain text and so decisions are made based off of that. Which means that historically and even now, one approach to solving this problem has been, well, if I can control the system, if I can bring the system, that cluster, that new system architecture, however big or small it is, if I can bring it to where the data lives now I have more control over what the exposure is, which is an approach. But how scalable that is and how long it takes to actually now start deriving value from these applications becomes a big question. Because if every time a data owner in the organization needs an entire one of those now new system architectures is a lot of either GPUs or any other compute substrate to come to where the data is for the people that know how to use that and maintain it, it's to come to where the data is. Now you're talking about longer and longer time to actual value. And this is causing a lot of these use cases that people are interested in implementing, whether it be in finance or elsewhere, to ultimately keep getting pushed out and maybe never even see the light of day. What we're challenged with is changing this fundamental assumption that data needs to be exposed in that manner. The technology that we've built at Protopia over the past three and a half years, and prior to that, the research that preceded it is focused on changing that assumption. What if you didn't have to expose your data in plain text? What if you decoupled the notion of data ownership from what gets exposed on the infrastructure? Then suddenly the set of things that you can think of in terms of which infrastructure can I use right now to start taking advantage of my data without exposing it rapidly changes, and that's what we're focused on.

>> That's interesting because that also brings up the opportunity to say, "Hey, at time of decrypt, that's when you execute the policy of the gen AI requirement," which is interesting. I mean this brings up the fundamental point about, okay, I'm a company. Okay, you got me on the, I like the topology, I love the architecture, how do I implement that? Because I got my data, I don't want to get... Private AI is all about IP. How do I implement this? What's the workflow look like? Take us through a day in the life of a customer.

Eiman Ebrahimi

>> Absolutely. So when we think about historically, again, what has challenged techniques like homomorphic encryption for instance? It's been that when it comes to deep learning style applications, the algorithms are sophisticated enough that if you try to operate just on encrypted data, it becomes very, very time-consuming and latency basically kills many of the use cases at that point. What we've done in order to not fall into that problem has been to make very deliberate trade-offs about how we identify the entropy or the stochasticity that needs to be added to the data, much like encryption does, but we do it in a very, very targeted manner. So you asked about workflow. Our product consists of two stages of software being run. The first is a post-training step that gets abutted to what training is already happening. It can happen on the same infrastructure that the model is being trained on, where the transformation layer that we call stained-glass transform gets created for a target model in that post-training step. Because this is a post-training step and it's abutted to training, it's next to an already very long-running process. Compared to that very long-running process, the creation of a stained-glass generally takes less than 5% of the time that it would've taken to create the model itself. So then you can now just make it part of the CI-CD pipeline that runs right after training. But what comes out of that without changing or altering the foundation model is that now you have a foundation model, but alongside it you also have this transformation layer that can now run on much, much less capable hardware. You can even run these transformations on CPUs. Now on GPUs, yes, they will be faster, but what we're doing as a result of that is we're saying wherever you hold the data, where your sensitive private information is, the smallest amount of compute that you have next to that data can transform the representations in a one-way manner into these stochastic representations that can then be computed upon by the algorithm. The algorithm does not, the model need not change, but now you've got transformed data that you're sending in. So there's two steps. The post-training step that creates the stained-glass transform and then the application of the stained-glass transform to the data that happens at runtime during inferences.

>> So I'm a developer. I'm in the CI-CD pipeline. I'm basically dealing with data already prepared for me. This almost guardrails, kind of like security, gets set up and then on the distributed computing architecture, I don't need to worry about form factor. I've just got a little bit of compute on a device, maybe it's a small form factor, there's a little bit of compute against the data. I'm there. You're there processing and transforming on the node.

Eiman Ebrahimi

>> Absolutely.

>> That's the benefit because the alternative is to move the infrastructure equivalent to the edge.

Eiman Ebrahimi

>> That's correct. That's correct. Which is generally not very scalable. It's useful for some places, but it's definitely... Generally it's not scalable, mostly not even necessarily as a function of cost. Cost is very, very important because at the end of the day we're trying to improve efficiencies in some way in the organization, but more so availability. We live in a world where compute is just very, very scarce. So when we're talking about even beyond GPUs, other types of compute, the notion of we're going to have enough compute right next to every single data holding, right, that becomes very challenging to imagine. And so what we are saying is that there is the ability to now expose representations of the data that do not expose plain text to those environments. And were they to leak, you are not leaking plain text information, which is what happens today. And that changes the game in terms of the cost, in terms of what use cases you can imagine, and that's where we've been seeing a lot of interest from places like financial services, insurance and healthcare and also the kind of the DOD and federal buildings.

>> It's a secure edge basically. You're building security into the transformation-

Eiman Ebrahimi

>> Yes....

>> so that you have essentially reliable and resilient architecture. Resilience is a hard problem to solve.

Eiman Ebrahimi

>> Absolutely. Without adding a lot of latency and maintaining the accuracy of the target models, which is super, super important. Because that's the other part of the story that the use of private data is mainly to reduce things like hallucinations from the models and make predictions more accurate. That's the entire purpose. But today that private data ends up not getting used, and so your accuracy with respect to that data is nothing because you're not actually deploying those use cases. We are an unlock to efficient infrastructure for the use of private data by way of, again, changing this fundamental assumption about what the exposure is.

>> As an interesting solution, what's the reaction from customers? You walk in, are they skeptical? "Whoa, whoa, there's a lot of overhead involved here." Are they more like, "Oh, finally a solution? Take us through some examples of deployments you guys have been working through."

Eiman Ebrahimi

>> Yeah, I think there's two segments to the market. There's parts of the market that have understood this problem for a long time, right? The problem of private machine learning existed before LLMs and AI, right? There's been things like federated learning or secure multi-party computation that people have worked on for a number of years. So there's parts of the market that understand the problem very well, and there's parts of the market that have understood the problem very well in the past year and a half because of the proliferation of LLMs. But when we go into customer conversations, we tend to mostly engage with customers that are a little farther along in their journey where it's not just public nonsensitive information that they're using in their use cases, where they're actually trying to take advantage of enterprise data or they're using use cases that actually need to use the sensitive information. Because when it comes to solutions that have existed in the market around private data, most solutions over the past year and a half have focused on how to stop leakage, how to stop information that the organization has decided this information shouldn't go outside the walls of the data owner. And that's been great and is important to make sure nothing is used when you don't want to use it. But a lot of the value and arguably most of the value that comes to the enterprise is in that private data. So when you do decide to actually use it, how do you do that? That's the question that we're trying to answer. And for customers that are already there, we see a lot of pull. Now, who our customers are are not just the end enterprises that have the private data. We also interact with the layers of the stack underneath that, whether that be application providers that are building applications based on these foundation models or whether it be the infrastructure providers that are allowing for customers to, for instance, have foundation model endpoints that they can hit on those infrastructures. In the latter two cases, we embed our product into what their offerings are into the application or into the infrastructure. And for the enterprises when they want to build those applications themselves, for instance, around open source models, we allow them to do that by way of giving them access to the software.

>> And transformation works on open source too because those software just models are going to be faster, less expensive, but also they have been vetted too, people are looking at those models. So you're having that security layer into transformation.

Eiman Ebrahimi

>> Absolutely.

>> There's a benefit.

Eiman Ebrahimi

>> There's a lot of interest around open source models when it comes to private information independent of Protopia because they give the flexibility of where the model is going to be deployed, right? They can deploy it on premises, they can deploy it in their own clouds, but even then the question is how do I protect the information from being exposed to the infrastructure even though I have control over the model? And many a time enterprises would really like their service providers, maybe their cloud providers to host the models for them and manage the models for that because some of the larger models, for example, a Llama 405B for instance, which is very, very capable, it is not trivial for an enterprise to maintain a deployment of that at scale for themselves. So they would like somebody else to be able to manage that in the backend. But then these same questions around data privacy come up.

>> And you bring up the complication of not just the application but the infrastructure itself. I mean, a lot of the hybrid deployments are cloud core, core cloud on-premise and edge, that's distributed computing. Everyone kind of knows that, but that's now running cloud operations. You got to be aware of what environment the infrastructure is performing at. So I might use some cloud and some higher level sources on say, AWS that I'll run a training data center or I might even outsource to a GPU provider. They all have their own systems.

Eiman Ebrahimi

>> Absolutely.

>> And this is where it becomes key that you're on the other side of that.

Eiman Ebrahimi

>> Absolutely. The fact of the matter is that the implementation details of how the model is implemented on any compute, again, it can be a GPU cloud provider, it can be your hyperscaler, it could be some of these more the newer compute paradigms that exist out there right now. The data owner does not want to be concerned with the implementation details of where did the data get decrypted, where did it get exposed? They just wants to use something in a seamless manner. So in order to do that, being able to give the data owner confidence that your plain text information did not actually get exposed in the implementation details, that's something that can speed up a lot of this usage.

>> What I really like about your approach is that we've been kind of talking about this in theCUBE, but you go into much more detail is that I remember when security became a big deal shift left and you had the application developers, "Hey, I just want a code in my CI-CD pipeline, 'Hey, security team get it set up.'" They would slow things down. But finally then security became with containers, all set up with all the policies, all were kind of the things that were guardrailed out for the developer. So they're making the calls on security in line. The same thing is happening now, and you're providing that layer of reliability so that the app owner knows it's trusted. It's delegated properly.

Eiman Ebrahimi

>> Yeah. It becomes a matter of can we provide the right tooling for the particularly usage models that are evolving? If you think about the usage model that LLMs have introduced, where we see value to creating the ability for many, many different users to access many, many different models that are ultimately going to in some form be on shared infrastructure, maybe it's either multi-tenant on an infrastructure that an application provider is providing the different customers or whether it's a shared cloud environment that someone wants to deploy on. In both cases, it becomes a matter of a new usage model that hasn't existed before. And that begets requiring new tools to manage the different risks that are involved there.

>> Yeah. I want to ask, every company has a core competency or cadence, culture, DNA, whatever you want to call it, intel, at Moore's law doubling every whatever few months. What do you guys, when you wake up every day, do you guys think about, we have a lot of encryption for breakfast, is it data? I mean, what is the vibe? What is the culture of the firm? What would you say is your core competency, encryption, data, infrastructure? How would you describe the DNA of the culture?

Eiman Ebrahimi

>> It's a very great question. Our core focus and our core drive is how do we enable AI use cases in a safe manner without requiring these high-cost private, physically separated infrastructure environments? That's our core drive, enablement of AI, independent of what the infrastructure deployment options are, meaning we want it to apply to any infrastructure deployment option. But from a culture perspective that involves getting comfortable with solving some problems that have traditionally been attempted but not necessarily solved at scale. And so we operate from a place of there are always trade-offs in systems to be made and not all trade-offs have always been explored. And the trade-offs in the world of AI continue to evolve because the technologies continue to evolve very rapidly. And so we are focused culturally on where are the right trade-offs to make in the systems to enable what the end goal is, which is enabling AI at scale with what is efficiently available instead of always looking to the future to say, "Oh, there's going to be some hardware in the future that's going to come solve this problem for us," for instance.

>> Yeah Eiman, it's great that you bring this up because I've been saying on theCUBE for a long time that we're in this systems' revolution. Every wave has some sort of cultural cliché, design, design thinking, iterate fast, all these things. But I think we're in a systems' mindset right now, and you guys are, I think really in this area. Explain in your mind and your view, what is a system thinker? And because I think you're starting to see this being the core mindset of this AI wave, it's thinking about the consequences of decisions and the trade-offs, every decision has a consequence. The trade-offs are super important. You got to evaluate those. What does systems thinking mean to you?

Eiman Ebrahimi

>> Yeah, I think for us, systems thinking is heavily focused on making sure we have some core role posts that we're focused on. And for us, when we approach this problem from a systems thinking perspective, we set those goal posts as three things. We don't want it to be dependent on a certain type of hardware. It should be something that people can use independent of the hardware they have access to. We shouldn't need it to change the frameworks people are using to build their AI, because those frameworks as it is, are evolving very rapidly and that's where the innovation is happening. We have to build what we're building in a manner that, again, the frameworks that people are building can adopt. You'll see we have contributors that will raise flags per PyTorch or Hugging Face around things that we have found could be established in those frameworks where for instance, PyTorch just as a framework and language could help the design of things like what we're building. And so making sure we fit within what the frameworks are has been another one of those goal posts. And third is to have very, very minimal impact on latency because we believe inference is where the majority of the scale use is going to be. And when you think of scale use, that's when, again, back to the usage model topic, that's where the most data exposure is going to be. And so if we're going to be a big portion of the inference pipeline, we can't have a big impact on latency. Given those three goal posts, every other decision is about what can we trade off without coming out of those three targets such that we can make it easy to use for whoever it is that's building the models and whoever it is that's deploying the models?

>> What's great is that you have... I love this market because you have the developer frenzy and it's going so fast and it's innovative and it's really moving the needle, and then you have the infrastructure getting faster, smaller, cheaper, so you don't have the whim of any one thing. You will roll with whatever comes out on the hardware side and then take advantage of these new distributed computing footprint opportunities, whether it's CPU or GPU, you take advantage of that too. So you're riding the wave on both the under the data layer, which is infrastructure, and then on top the developers.

Eiman Ebrahimi

>> Absolutely.

>> You ride on both sides. You ride that growth.

Eiman Ebrahimi

>> And we think that that's the only way to provide value across the stack, which is what our goal is. Our goal is to unlock private data at the top level of the stack, reduce costs for the application layer in terms of how they deploy, and improve the efficiency of utilizing the infrastructure, whatever that is. In order to do that, you have to work with what it is that those entities are trying to build. And so we've taken that approach and been very firm on these are our tenets of where we make trade-offs around, but ultimately it becomes a matter of how the entire industry... So you've spoken about the industry evolving. How the entire industry ends up delivering value such that these investments that enterprises are making in AI become self-sufficient in some way and actually start returning value. In order to get there, I think that some of these efforts are pretty critical. Meaning if companies like Protopia or other parts of the stack do not focus on things beyond just the next application, then those applications will have a tough time actually coming to market and the enterprise will have a really tough time trying to create that value that's necessary.

>> It's really cool because right now the mandate in most companies is unlock the data value. So what I see happening is the first progression is what data do we have? Let's use AI to figure out what data we have. And then they say, "Okay, now what is the low-hanging fruit? Okay, let's get some wins." I call it small ball, get a single, get on base, get some wins, and then go for the home run hits later. And then once they have that, then they come and say, "wow, we just did this cool project. Now we have other opportunities that are now identified." Then the next progression is, okay, this is a net new opportunity, not just cost recovery efficiency, it's potentially revenue. So then that brings us to the private AI when they go, "Wait a minute, this is our data. We're not really getting this from the large public or proprietary foundation models. It's coming from the combination of the developer app and embedded AI and what we've done with the private AI." And they go, "That's great." Then they go, "Okay, how do we protect it from prompt injections and context poisoning?" Okay great. Then they go, oh... Now they go, "Oh, how do we recover from this?" Then they go, "Okay stop." So this is where I think private AI is at app right now and I want to get your reaction to that because how would you describe the progress bar in AI? I mean, IBM hired McKinsey to say that only 1% of enterprise AI is truly being realized, which is huge TAM for you, which is great. But all the private AI, they just seem to, I won't say you've stalled but they're in this inspection mode of like, "Okay, we know the values there? We're going to figure it out." Assume that that's going to happen. Everyone kind of sees some obvious use cases. Then they go, "Okay, what is our long-term plan, medium and long-term plan?" And that's where it gets in this, "Wait a minute. What is the resilience? What do I do?" So how would you peg the progress bar of private AI right now? Because no doubt it's a hot trend that kind of seeing it. Where are we on the realization of deployments in POC mode? How would you peg the progress of this private AI area?

Eiman Ebrahimi

>> I think the private AI area is in its early years. It's something that requires unlock in multiple ways. And one of the interesting factors around that is that people have imagined the first set of use cases of where that new revenue can come from or where really major efficiencies can be created for their relative businesses, but it's in that imagine state. And maybe they've run a few POCs with data that isn't actually full-blown being injected into the systems, but they haven't gone much farther than that. The inspection phase is trying to answer a lot of these similar questions around data exposure. How can I actually set it up? Who's going to build the model with my data? How would I give them that data? All of those questions do need to be answered for this value to be unlocked. And I think there's both a lot of opportunity, but also there's always an amount of time that you have to get there. And I think that's where a lot of the boards of companies are getting a little angsty about of, "Well, where is it?" Because at some point we're going to stop just throwing budget at this from various different sources, and we're all going to be in a bit of a pickle at that point.

>> It reminds me of when I travel, I just want to get on the plane and go to New York. I got to go to TSA, I to get taking off my shoes. I'll take my laptop out. I mean, all this process soon will be rendered. Then there'll be TSA pre with clear, and they go right through. I think we're in that phase. I just want to get on the plane.

Eiman Ebrahimi

>> That's right.

>> That's AI. So I, we're going to get there pretty quickly, but we still have to figure that out. I wish could had to do more on this topic. We could go an hour on this because it's super important. This is where the value is in the enterprise, the data is and the workflow are the IP.

Eiman Ebrahimi

>> Absolutely.

>> It's very clear. I don't want to mess with that. I guess my final question is is that in a steady state, okay, how would you see... Go down the road a little bit in your mind's eye, things progress, things move along, the hardware moves along, it's better, the edge becomes robust and got nice processing power on even small cameras or whatever, computer vision. What is your view of... At a steady state, what does Protopia look like? What does that mean? What's the deployment? Take me through your mind's eye on how you envision steady state of your opportunity.

Eiman Ebrahimi

>> Yeah. Steady state for Protopia, we believe is the ability for anyone creating a model on any given model creation environment to also create these privacy layers, the stained-glass transforms for those models. We believe that our engine, the way that we have built it will be integratable by all of those platforms. And what that results in now is the ability, this is an ability that whoever wants to use those models that have created stained-glass transform for them, to either access the model with plain raw information if the information is nonsensitive or if they're okay with exposing that data to the infrastructure, and they also will have always the capability of accessing those models for data that they wouldn't have otherwise through stained-glass. And this we view as a completely new way of looking at what applications you can be deploying where, which is a game changer for value creation at the enterprise level. And that's why we focus a lot on our channel partnerships of delivering how stained-glass can get to where the data is on the data provider side. And on the compute provider side, how can they create the transforms where they are creating the models?

>> Yeah, and very intentional relationships too, because data's involved. At the end of the day, latency you mentioned earlier is critical because if you have latency that doesn't match the gen AI runtime, if you will, your host.

Eiman Ebrahimi

>> Yes. Absolutely.

>> Yeah. Anyway, thanks for coming on theCUBE. Really appreciate you.

Eiman Ebrahimi

>> Thanks for having me.

>> Love the location here. Great to see you last night at the founder dinner, reception. Thanks for coming on theCUBE. Appreciate it.

Eiman Ebrahimi

>> Awesome. Thank you.

>> Thank you so much. Okay. I'm John Furrier, your host of theCUBE. We are here in our East coast studio kicking off Media week, day two of three days of wall-to-wall coverage on Wall Street. I'm John Furrier. Thanks for watching.