E341 | Leonid Feinberg, Verax & Carlos Espinal, Seedcamp: Building Trust in Enterprise AI

0:05Welcome, everyone. On today's show, we have a good friend and Seedcamp founder Leonid Feinberg, who is the founder of Varax. Varax is the enterprise trust solution for generative AI. So he focuses on building trust within enterprises for the new line of tools that we're using more and more, whether it be through ChatGPT or any of the other LLMs in the market. Now, Loni's background is great. I'll ask him to share a little bit more detail about it. But having had quite a bit of experience in the space of both security and cloud. He had built a company prior called Cloudinger, which was acquired by Amazon, and then worked at Amazon for a good while thereafter before starting Varex.

0:52So with that very quick intro, Leonard, welcome to the podcast. Thank you very much. Pleasure being here. I wanted to see if you want to share a little bit more about your background, what led you to build Varax, and what is it? Yeah, absolutely. So as Carlos mentioned, my background is in starting enterprise B2B companies and helping enterprise customers with their challenges. Different types of challenges in my previous companies than it is for Varax, but everything else is similar. And that's Varax is my third company. and we're trying to help enterprises right now in navigating the difficulties and challenges of using Genitive AI in production.

1:37We saw with our customers that everything is working amazingly while you're running a Genitive AI product in a sandbox, but then when you move that product to actual production, then suddenly it becomes less sunny. Customers are losing visibility for what's happening in their product, both in terms of how the product is being used by real-life customer and user, and also visibility regarding how the product behaves, whether it behaves correctly, whether it behaves inoffensively, whether it behaves as intended or not. And so Varex, the product that we're developing, the Varex Control Center, gives that visibility back to our customers.

2:20And then it also does even more, and it allows customers to mitigate the risk of problematic behavior of their GNI product by intercepting the responses from the customer's product, analyzing them in real time. And then if the response is problematic in a way, then the various controls then could modify it and make sure that the problematic aspects are fixed before the response gets all the way back to the end user. Excellent. Well, let's unpack that from an architectural point of view first, because whilst you're sharing what it does, it begs the question about how your potential customers and people who are rolling this out are actually rolling this out.

3:04One simplistic view could be that you're literally just having one centralized LLM that has access to the entire organization's data and is used for just about everything. Whereas the reality is, it's probably something quite different. Do you want to walk us through a little bit about how enterprises are thinking about architecturally rolling out LMs? Are they using the same one across multiple different divisions? Are they ring fencing things? Walk us through. Yeah, absolutely. So interestingly enough, LLM is not an enterprise product by itself. LLM is just a building block, just like cloud services and others.

3:41So LLM by itself, just using it as in, is almost useless for enterprises, even enterprises that are using products like ChatGPT or Gemini. LLM is hidden somewhere in those products, and there are a lot of additional layers that those companies added to help the customers. So when enterprises look at LLMs, they are looking at LLM-based products. And then there's a multitude of LLM-based products. Every product contributes to a specific business case. And then this is how they analyze that. So every product has been analyzed separately by both the stakeholder within the organization. If it's a sales-related product, then sales stakeholders should look at it.

4:26And if it's engineering, then it's engineering. And then normally, at this day, the person responsible for AI in the entire organization, like a chief AI officer for the companies that have such a function or head of data science that most enterprises have instead. They also kind of look at the product from the AI perspective. But at the end of the day, there is no one tool to solve them all. There are multiple tools and Gen AI models are hidden somewhere in those tools very deep. Yeah, so I agree that that's a good starting framework to think about it. It's a tool. It's one of many tools. But as with most tools, right, you start getting into best practices, right?

5:13Think about hybrid cloud deployments. Think about where you keep your data and where the compute instances are. And think about all the things that we figured out in the last 15 years of cloud computing that makes sense to have a high availability organization. and deal with sort of all the issues that certain kinds of sensitivities and regulatory impositions have on companies. How is that evolving for your customer? Where is the data have to be stored? Which kinds of models are being used? What tools are integrating these AI tools that they're allowed to use? What is the first departments that are being used and that therefore you would be helping solve this issue of trust?

5:55Right. So that's a great question. I think that AI has two faces. It is still a software product. So everything that we know about software enterprise products is applicable. So architecture, data governance, privacy, it's all still as applicable to AI as it is to traditional software. But then the new aspects of AI that cause it to behave in a way which is unpredictable, in a way that is constantly changing. If it behaves in response to a certain input in a certain way today, that doesn't mean that it's going to respond exactly the same tomorrow. So these are a lot of new aspects that AI products have.

6:43So customers and the entire industry actually needs to develop new best practices that needs to cover these new challenges. So it's a combination. All best practices still applicable, but not enough. And then new best practices are still a work in progress. And today, every organization is pretty much developing its own. There are almost no best practices that are already industry standards when thinking about Genitv. Except maybe one. In today's world, Genitive AI and SaaS and enterprise don't really mix well. Most enterprises, when they use Genitive AI on their data, they prefer to run the model and everything else in their own environment rather than letting some other vendor run it for them and then send the data to them.

7:33So this is currently almost, from what we're seeing, almost a universal truth with our customers, especially when we talk about data, which is a bit more sensitive. Yeah, and that makes sense to me intuitively, because we have so much fear, anxiety about what the models are ingesting and retaining. And if it's corporate and sensitive data, the last thing you want is to have that be part of the cloud provider or the application provider's tool set to sell to somebody else, right? What do you think the future of that architecture is going to look like? Because right now you've just described it as it is today, but in a world where Varex is successful and you might not really care about this layer of things, you might be a layer above it.

8:17But what do you think that that's going to look like? Is it going to be a world where there's a lot of digital twins running around and cloud providers? And then every time you sign up for a new SaaS product, there is a synchronization period to sort of map out your data to come up with an equivalent? or how do you see this playing out? Right, so from my previous experience, for example, with how public cloud started, which had similar concerns of data privacy and what the public cloud provider would do with the data, I think a lot of that is about psychology and a lot of that is about trust. And generally speaking, companies would be willing to do more with their data or to kind of take a bit more risks on their data.

9:01if they trust the vendors and if they trust the processes. So if they have kind of assurance that the vendor is compliant using modern compliance certifications, et cetera. So I think that at the end of the day, assuming that AI is going to be as successful as people predict it to be, then I think that we will see AI falling into the current patterns. So yes, that providers using AI on the backend, And companies sending some of their data, are willing to send some of their data to those vendors that they trust them, are not willing to send other types of data externally to anyone. For example, still, you know, banks, even before AI, still are very reluctant using public cloud for the sensitive data.

9:49So I think this is going to stay pretty much the same right now. It's still, we're still at times where trust is challenging. But if companies like Varex and others would be successful establishing trust, I think that it's not going to be too different from what we know today outside of the AI. I'm going to challenge you on that a little bit, Leonid, because if you look at what made that transition to cloud trustworthy was the fact that you could spool and destroy cryptographically a lot of instances that were predetermined as binaries and that were stable and that were updated. But the premise behind a lot of the functionality of AI is learning, and that means retention.

10:35And retention means that you can't destroy as easily as you could a microservices or a virtual machine. So there's an element there of permanency that is required for that utility that is sold to a customer, which negates the traditional cloud metaphor of ephemeral computing and storage as needed. And so is that possible? So I think that it's a temporary limitation. It's my personal opinion. It's hard for me to support that. But seeing many companies realizing those challenges and trying to solve them, I completely agree with you. Today, this is different than the traditional public cloud. But when public cloud just started, it didn't have all these capabilities either.

11:19I remember the days when you could only encrypt your data on the cloud using a single key that was managed by the cloud provider. And that was a big deal for enterprises. And then all the cloud providers added different ways of managing encryption using keys provided by the customers in a way that the cloud provider could not really access the data themselves. So I think that these similar capabilities with time will be developed for Jantive AI. There are methods being developed as we speak for how you train a model without actually feeding it with the raw data. So there are different methods that are in development still, you know, still early days.

12:04But I think that this challenge is very significant. But there are some promising technologies to solve that. So I think this is going to be solved just like it was for the public cloud and other technologies. So part of the reason, and this is a good transition for bringing it back to Verax, because one of the things that I cannot tell is where you inject this trust factor, right? Are you injecting it at that creation layer where you're managing the retention and data and how it's managed? Or is it one layer above that, which is understanding how all these services add up and you've added effectively a semantic firewall that effectively limits what the output of all aggregate models in the enterprise spit out?

12:52or is it one layer above that? And so at the sort of asset level, where the assets consume the stuff and what they're allowed to see. And I'm trying to visualize where you see the future of injecting this enterprise grade level trust. So first of all, just a disclaimer, I think that we are just a part of the bigger enterprise AI trust story. I think we're solving very challenging problems but it's not where enterprise trust starts and ends with AI. There are other things, not all of them are about technology. Some things are about regulation or things like that. But I think the trust is a big umbrella here that needs to be addressed from any different angles.

13:37But for our specific one, which we think is one of the angles that are as relevant today as we believe they're going to be relevant in 10 years, we look at AI products as kind of non-human employees of an organization. Companies have humans working for them and humans make mistakes and sometimes humans are malicious and sometimes it's a combination of mistakes, small mistakes being done by multiple people that accumulate to one huge one. So companies are used to operating in a world which is not perfect. Risks are everywhere. And there are a lot of processes and technologies developed throughout the last millennia to accommodate that.

14:25We think that the AI world is somewhat similar to that. Yeah, software. But still, there are a lot of similar risks that that software is not perfect. It could do things that software couldn't do before. And the implications of that may be severe. So we are seeing ourselves as this layer kind of of the non-human managers slash supervisors of the non-human employees. So we look at everything from outside at the top of the stack. We don't know which models are being used. We don't know which data is being sent. We just look at the behavior of the customer's products from outside. We analyze every single piece of data going in and out in real time.

15:10And then we reach a conclusion again in real time about that piece of data. And we act upon that conclusion immediately. So it's like a WAF of sorts. It is a WAF of sorts, but I think that one big difference between like the analogies of WAF or a firewall, usually with this kind of solutions, you have kind of predefined set of rules, and then this solution applies that. With AI, I think that this strategy simply won't work. Redefining rules is not something that is very effective for AI. So we need to create the rules dynamically as we go. And the same rules could be modified automatically by the Berk solution tomorrow.

15:53Yeah. Yeah. Interesting times we live in, huh, Leonid? Interesting times. Lots of productivity gains and lots of challenges. And I guess maybe if we move away a little bit from the whole idea of what you're building, and I look at it more about the bigger picture of how different companies are behaving around this. Do you think that a lot of people are more worried about the cybersecurity implications of AI or more worried about the loss of productivity relative to their competitors? What do you think people worry about most? so i think that there are kind of two answers to that the current answer and the future answer because right now ai in enterprises ai is still almost not used it's still very early days and so a lot of the stakeholders like the sisal for example are still not seeing the risks of ai on a daily basis manifesting themselves.

16:58They're still very theoretical. So at this point, the AI risks are mostly addressed by people like the head of data science or the chief AI officer or the CTO sometimes because they are the ones pushing forward some of these products. And then, and because of their function, how they think, there is the, they are concerned about are very specific. They care about privacy, they care about damage to the organization, and they care that if that initial use of AI would not be a success, it would not lead to further use and to expansion. and I think that the once AI is going to be used more widely in production, then we'll see other stakeholders who would suddenly be exposed to the actual use of AI to start applying their risks and their concerns to real-life situations.

17:59And again, CISO is a great example. CISO are all very concerned, but their concern is very theoretical because in most organizations, most enterprise organizations, There is nothing really running in production yet, except maybe, you know, the chat GPTs of the world. And so their concern, practically speaking, is still small. Yeah. Okay. Yeah. It sounds to me like if I just listen to the tag cloud of your answer, on the high level, senior management, it's fear. And probably at the operator level, at the employee level, it's productivity. and that's probably going to be the future for the next five years as those two positions reconcile, right?

18:39If we talk a little bit about geopolitics and the role of AI in sort of shaping our world, first of all, what battles are you seeing right now in terms of how AI is being used by different nation states and what are the tensions that are brewing there? Yeah, so it's interesting because right now, again, just like with enterprises, the actual use of AI by nations for specific use cases is still very, very limited. A lot of those nations are looking into the future and saying, hey, when AI is going to be used across the board, these are the potential use cases for AI, we want to make sure that when these use cases become reality, we want some kind of advantage with them.

19:30But funnily enough, most of that is just get-to. So it's like it's trying to get the future. And right now, it's almost universal consensus that it is going to be very, very significant. And parts of the world with better access to AI, with better AI, better hardware, etc., would have advantage over other parts of the world. But I think that how exactly that advantage is going to manifest and what exactly is going and what it's going to look like, I think that no one really knows right now. And the difference is between people who admit they don't know it or people who claim they do. Yeah. Do you feel like the rest of the world, the world that, you know, we look at where a lot of these models are popping up, all the open source and closed source models are popping up.

20:22Do you think there's a risk globally that, and you can answer this from a nation state point of view or from an enterprise point of view, it doesn't really matter. But do you feel like there's going to be this huge divide that's further created between companies and countries that have models and have access to them and are okay with the data management of them and those that don't? I personally think so, yes. I am a big believer in the potential of AI. I think that AI could be used to improve many things on almost any level, starting from individual levels of consumer and user all the way to the state level.

21:06So I think that AI has huge potential. And I think that if we fulfill that potential, then yes. companies and states who would embrace that potential who would have the resources to use that and would have the state of mind and not turn away from it because we all know that AI is a controversial topic in general even before the ability to use it but I think the states that would have all of that they would definitely have advantage over states and companies again It's similar on different levels that don't. But do you think that it'll be possible to reconcile that without effectively having to sort of submit yourself to the powers that control the AI?

21:55In other words, I mean, think about right now when we think of the world as a dollar denominated world, there's a certain element of it that you buy into as simply being part of that. And so if that's like a sort of a metaphor, do you think that there is a risk for companies that are based, let's say, outside of the countries that own these models to adopt all the governance and all the data management and all the things that those data models come with as a sort of baggage? Do you think that this will really cripple the possibility for countries that aren't part of where these models were born from being able to ever have a chance of not getting a second priority on not only the tech, but also second sort of a risk of national security because of where their data is stored and how it's managed and how it's learned?

22:50And you're effectively at a disadvantage because if you choose not to engage, then you're way behind. So it is a possibility, but I think, again, purely guessing, I think the future is going to be more complex than that. And I think that kind of the models or the hardware is just one part of the equation. And then I think the regulation, et cetera, it's never one or zero. There are always gray areas and there is a lot of play around those gray areas. So I think that yes, even before AI, companies that have access to resources and access to compute power and access to talent, and they use that access to increase the gap between them and other countries.

23:34and maybe it's going to be a factor now. The gap is going to increase faster with the use of AI, but I don't think that something is going to change fundamentally there, that AI is something like magical new technology that would change the world from that perspective. I think it's going to change the world in other ways, but from that perspective, I think it's just going to take what's happening today and make it bigger and faster. them. If we go to a very detailed point of that, which is around data management, you know, there are many topics in AI. Data management is one of them, and it's probably one that you could spend hours on.

24:16And that includes everything from where the data comes from, what kind of data it is, is it labeled and classified? And then how is it ingested? How is it stored? What derivative data is created from it. What do you think is going to be the architecture of data management in the future? Once you're successful and AI is more broadly distributed, what is the future of data management going to look like? So I think that a lot of those challenges that you just mentioned are going to go away as the world evolves. So for example, data classification, etc which today is still a big challenge and i expect that to become a smaller and smaller challenges in the future and become commodity and then you know we could take any data unstructured messy whatever and everyone would have access to that data in in a very effective way so i think that all these data processing and challenges we have right now i think they're temporary i think the main two things that I think that are not going to go away is A, where to store that data, and I think that B is access to data.

25:29Again, data is king. The more data you have access to, the better your functionalities, your models are, etc. So these two, I don't see how they're going to change. Access to the data is a universal challenge. Just the implications are now bigger than before and storing the data. Also, I think that, again, before the AI was born, it was still a challenge. And before the cloud was born, it was a challenge. And it's going to be an even bigger challenge right now. But same thing, same access to private data, to others, et cetera, but in a different wrapping. Well, to wrap things up, and thank you for all your thoughts.

26:15This is one of those areas that we could go on for hours, and I love hearing your thoughts on them. But to wrap things up, I want you to think about your customers, not the ones that you have at the moment, not the ones that you have possibly in the moment. But imagine three years from now, you're talking to your customers in your pipeline, not even the ones that you've sold to. What are the top three pieces of advice you'd give to them as they think about engaging with you or one of your competitors? Sure. I think that the one big advice that I would give them is try to think about AI more flexibly.

26:51Because people are today in many of the organizations that we talk to are thinking about AI in one of two ways. Either it's just another piece of software, let's apply what we know to it and it's going to work just the same, which is incorrect. there are new challenges with AI that need to be solved. And then the second category says, well, AI is magic. We don't know what it does or how it does it. We don't want to have anything to do with it. It's like a comeback when all the challenges have been solved, maybe we'll consider using AI then. So I think that both approaches are problematic. And my biggest advice is to, on the one hand, to understand that AI needs new approaches, new procedures, new products, new technologies.

27:41But on the other hand, that with the addition of those, AI is manageable. As a matter of the risks, they're manageable and they could be mitigated. It's never going to be risk-free. I think that we established that. But I think that, you know, we should take the risks, embrace them, and think how you understand and mitigate them. and then it leads the way to a worldwide reaction. Sounds like the advice you're giving is really around not letting risk and fear paralyze progress and you can actually control more of those fears than you think you can. Yeah, exactly. Excellent. All right, well, with that, Leonid, thank you so much for joining us in today's episode.

28:25And for those of you that want to get in touch with you, what's the best way for them to reach out? So either to reach out to me directly on my email, leo at varaxai, or just to go to our website and use the Waze there. Excellent. All right, guys. Until next time. Bye. Thanks, Carlos.