Domyn and the future of AI in regulated environments

Welcome to this episode of Use From The Top, the podcast by Arma Partners. I have the pleasure today of being joined by Ulyan Sharka, the CEO and founder of Domin, an AI platform built for the most demanding and regulated environments. Hi Ulyan, it's so nice to see you today. Good morning, Sikha for having me. There is a lot of noise around AI companies disrupting established sectors. How does Domin fit into that market? Because very few of them actually have a proven value proposition. And I'd love to hear from you how you explain what you do to someone that hears it from the first time. What gets them the most excited about what you do. We focus on AI ownership. We enable our clients to be active participants in the AI revolution by taking control of the technology so that they can protect their data but also they have a chance to shape it. So we enable AI not just as a tool but also as a product for our clients. This leads me to why I believe most of these projects are failing nowadays. It's just the normal process of understanding the technology and the risks that come with it. Not just the opportunities that you see first hand. This is a true confirmation of how disruptive the technology is. I'd just like to ask another question on that. Do you see these limits expanding? Obviously the technology is bound to progress. Do you think in the future the limits will still be there and your answer will remain the same? I think we are leaving a reality distortion because some of the limits are not true limits of technology but of how people perceive it. We are dealing with a new form of artificial intelligence which we call generative AI. And generative AI, if you look at it, you can immediately understand that it cannot do certain things. For example, when we think and when we assume that these models can be autonomous and they can make decisions that are deterministic, that's impossible and we know that upfront from a science perspective. We're talking about models that are pre-trained. We're instructing them with a lot of data and teaching them what they can do. And then we're in addition to the pre-training phase we're also aligning them to certain values and to certain tasks that they can perform. So I think we know the limits upfront but somehow we're not we're failing as an industry to communicate this effectively to clients, to enable them, to remove the noise, to understand what these models cannot do even before what they can do. And then to frame it in the right way when it comes to translating the capabilities of these models into a use case that can provide value but that can also be safe. Because again, these models are not autonomous systems. Generative AI alone without software and without the application layer is not possible in production environments. For example, if we enable a trading system based on generative AI and if we ask a model if it's safe to execute a trade or not, we're just going to get hallucination because this is how the technology works not because there is a limit. This is a probabilistic brute force data processing system. So then you need additional components around the model, like knowledge graphs, for example, to reduce hallucination by better understanding data relationships. You need to query real-time databases that provide facts that will somehow police, you know, the reasoning of these models and enable you to make a decision that is safer. So what I'm trying to say is that it's not generative AI failing and generative AI having limits. It's the approach of implementation that organizations are taking, mostly influenced by this distortion of reality where we think that AI can do everything but in reality it cannot. And this is I think ashamed because for the first time we're enabling humans to talk with machines that is revolutionary, that can solve so many real world problems. So we should focus on those things that AI can do and not everything. Great response. You mentioned a really important word for AI alignment. Beyond the giving the AI the right playbook and the right rules for the game, is alignment to human values? Is that a topic that is very important for you at Domain? It's absolutely important because of the responsibility that comes with AI. Again, these models are instructed, are trained, and they perform based on such activities. So the bias that comes from the model provider is something that not only should be transparent to who is adopting so that they can have a safe approach in terms of where they're going to use it and why they're going to use it. But alignment also opens for customizing these models, tailoring these models to the needs of clients and use cases. And I think that post training is not enough because post training limits the way customers can shape these models. So this is why at Domain we have made a very bold decision to enable our customers to own a copy of the model. And we're also open the pre-training recipe. So we give them the keys so that they can continue to train the model like we do at the beginning when we train it at the base level. And by doing so they can fully control the outcomes to the maximum possible extent. And they can be an active participant in this revolution, which means that they can shape the model as any other company out there. So we want them to be the AI companies of their industry, not just companies that adopt AI in their industry. That's very interesting. And I guess being more in control also, it's very important for these environments. You mention responsibility and that usually comes with accountability. What's your view on accountability for AI? Like who's responsible if there is a mistake? I think there is only one short answer, which is the human. And I think every assumption that AI is this form of conscience of a new social entity that we need to somehow regulate is completely wrong. AI is just a tool. We could think of a future evolution of AI. We can imagine where this could potentially go, but this is not today. The current architecture is science that we understand. It's science that of course is so deep and so wide that sometimes it's hard to explain the very, very, very basic, I would say facts around, you know, what was the single sentence in a training data set of 20 trillion words that influence that specific answer. This is just, you know, a technology problem, but it's not an understanding problem. We understand how this model works. We understand what they're doing. We build them as humans, so we should be accountable and responsible. So I think it comes down to the ground truths. So I think this is the challenge that we're having now. We should be more transparent about how these smallest work and not confuse, you know, consumers and organizations that adopt them by showing, you know, things that do not exist. And this is mostly, you know, part of every technology revolution. So people like to dream and to imagine what technology can do for them. So we take this seriously. So for us, it's extremely important that our clients understand the capabilities of technology. And then we also have them to prioritize in a way that the implementation is not only responsible, but it's also going to produce value because you can have very amazing beautiful use cases with AI, but potentially they are not worth it sometimes. And you made a very specific choice or maybe you're going to tell me if it's a choice or not, you went fully integrated vertically from cheap to front ends. Was that actually a choice or was that a necessity? So it was part of the choice of addressing the needs of regulated industries by embracing regulation as a competitive advantage, which means that you need to understand zero trust architectures at zero back door architectures. You need to be able to deploy your technology in air gap environments and in order to achieve that full independence, you need to have a segregated product that you own and that you can take accountability for. Otherwise, if it has one single dependency, even that is convenient from a cost perspective. Sometimes you need to reinvent the wheel in order to enable this business model. And I think this is one of the core pillars of our value proposition. We really understand the needs of compliance and segregation regulated industries. And we have turned that into a strength and not a weakness that we address, you know, with I would say by finding alternative solutions or compromises, which is a typical approach. For example, of other vendors, they have their own foundational technology, their view of how this should work on efficiency level from a technology perspective and then they find ways to patch it so that it works for the client. We start from a different assumption, we start from the assumption that this is not about decisions made by executives in these regulated environments, but is about the rule of law and other aspects that we have investigated, have understood and have implemented in the technology. And this was something that we learned, you know, by working with clients at the beginning, we thought that sovereign AI and segregation could work. work only at the heart of our solution, which is the orchestration layer and the foundational layer of the LLMs that we built from scratch. But then we discovered this extends to authentication. It extends to the deployment model to data integration and so on and so forth. And we ended up building a true AI operating system, which took longer than expected. But this is what is helping clients to move very, very quickly from pilot to production today with Domin. And also putting us in a position where we have more demand that we can manage. When it comes to licensing, the open enterprise license breaks with traditional licensing models. How would you say this changes your conversation when it comes to speaking with stakeholders such as financial institutions, governments? Concept is beyond control. Because sometimes when we speak about on-premise solutions, we kind of address some of the needs of regulated environments. But while this was true for software 2.0, where we have a clear separation between the data ownership and the provider of the tools. So you sign with a ERP company. You have your data in their database. You don't like them anymore for whatever reason. You will terminate the agreement with them. You export the data and you go to another provider. If you need to customize the models, and if you need to have this compliance approach to AI, with generative AI with software 3.0, this is totally different. So when we put data into a model, it's like putting sugar into a cake. Once you bake it, you cannot take it back. So what this means is that a government in order to use a third-party model, they would need to actually export confidential information to the AI provider. So there is no other way for these clients to adopt AI in highly confidential and mission critical environments, other than owning a copy of the model. And it should be not just open weights. It should not be just open from the perspective of all the details and keys that are needed in order to pre-train it, but also from a perspective of legal ownership. They need to own it on a legal basis, so it becomes their artifact so that if they bake their data into the model, they still own their data. I think you've already answered that of the question. But when we talk about generalist LLM's going into highly regulated environments, like recently with defense, would you say that this is mostly what keeps them from penetrating the high-risk environments? Or is there anything else that prevents them from scaling in these environments? I think this is the core factor of slowing them down to actually enable these clients to use AI. On the other hand, is also the approach they have taken to governance. So they are pushing this idea that AI can eat software. Probably is going to happen in the future. We can all see that. But from a responsibility approach, we all know that is not the case today. So if I'm in a government today, a defense organization, leading asset management or bank in the world, or an advanced manufacturer that builds products that are worth billions. So if I make one mistake in engineering a design or whatever part of the process I'm automating with AI, that is going to result in billions, if not dozens or hundreds of billions of damage, I cannot afford to make a bet on this idea that generative AI can replace most of my software needs. So I think it's also a perspective of innovation. So some of the Gen AI providers, I think they live in the future. So somehow they are drunk from this potential future that we're going to enable with AI. And they are completely ignoring, on my opinion, the needs of the current market. And their responsibilities that regulated industries need to deal with in order to adopt AI in certain levels of the business. So I strongly believe that there is going to be a composite market. So this approach, actually, to innovation makes AI more efficient because you can centralize it, but then you cannot use it for highly mission-critical use cases. And that's fine, because if you're a bank and you're buying intelligence from the public web, paying by the token with one of these centralized models, it's OK. It's the best option, actually, to buy intelligence. It's cheaper. But then if you're processing your own IP, you have intellectual property, whether you're building pricing models or you're building a new product that has a competitive advantage from a perspective of process, you want that to stay with the company, because that is going to be the heart of the business in the future. And you just talked about regulations. Regulations are often seen as going against innovation in some ways, throwing it down. How do you see the EUAA act? How do you think it's going to reshape the competitive dynamics? And also how do you see the landscape evolving in other major regions like the US or China? I think the European AI Act is suffering the geopolitical tensions, because this framing of Europe versus China versus the US is somehow branding each region with one keyword. And unfortunately, Europe is known for regulation. And that is being somehow weaponized against Europe. While I think that some regulations in the past around data have been unreasonable, I think the AI Act in Europe is absolutely spot on, because it is starting with first principles. So the European AI Act is saying, you know, you cannot build weapons. You cannot violate human rights. And you cannot create systemic risk. Three concepts that are extremely reasonable. And I think they should be the playbook for everyone in the world, not just Europe. And when it comes to the competition between regions, I think we're missing a point here. This is not about governments. This is about the private sector. AI is currently being developed by the private sector. I think that the disruption that Europe is talking about, both at a social level and at a economic level, is true for the US itself as well. And China, of course. So the US is going to have the same challenges of Europe. So if we build AI as a monopoly, if we put the world's intelligence in the hands of a few private companies, it doesn't matter if we're talking about the US or Europe. That is unacceptable because how it works and because what it means for the future of freedom to operate in the private sector and the future of democracy in the public sector. Because it means that if every single transaction of software is going to be controlled by a handful of companies, of course they have a power and a monopoly of knowledge that could potentially control the world. And so that's a great bridge into sovereign AI because it's basically on the list of pretty much every country right now, but the definition of sovereign AI keeps evolving. What's your definition of it? And how do you see it changing in the future? So sovereign AI has been again weaponized against Europe because Europe is obsessed with sovereignty, but I think sovereign AI is just the desire and the ability of nations and organizations in the private sector to own their future. So when control is needed, it's not a choice. It's not that someone is obsessed with this idea of controlling the technology they use. We're talking about a shift in how the technology works and it forces these organizations, these nations, to think of sovereignty because they can see a scenario where they lose control. And unfortunately we have had cases where technology companies have been shutting down. It happened with a bank if I'm not wrong, a few months back. It has happened with another scenario. So I think we need to find the checks and balances by defining sovereign AI as a way for organizations to keep their freedom to operate. This is not about becoming local. This is not about destroying cross-regional partnerships. This is not about being European only or US only. This is about enabling everyone to be an active participant, to be free in this democratic world and open market, instead of just being a follower and becoming a slave of a system where at some point someone can shut you down. And again, this should not be framed at the geopolitical level. This should be framed at a technological monopoly level. So I think these monopolies, they happen to be, in one part of the world, they could very easily have been in another part of the world. So it's not about where they are. It's about what they mean to the world, to the free world. And this is why we should care about sovereignty. I'm not just because Europe is obsessed with it. And when it comes to your customers, obviously, innovation is great, but blue cheap partners they care about outcomes, right? So how do you balance innovation with ROI when it comes to large organizations? So it's about being honest and co-developing with them. This is a revolution where we all would love and want and dream AI to do everything for us, that amazing black box where we throw our data and it solves all of our problems. But this is not the case. So we need to start with that. honesty first. And we are very clear, you know, with the state of the art of the technology, we look for people that understand this and this open to having this honest conversation. And once we find those clients, those early adopters that, of course, have huge ambition in implementing AI, but also they are realistic about what can be done and they also have a long-term view in how this is going to shape their business. And they know that the cost of mistakes in the short term is going to be very high if they don't take a realistic and pragmatic approach. So we're helping them, first of all, by enabling them to understand the state of the art of technology and navigate the noise that unfortunately we are living with, you know, both at a media level and unfortunately some leading players in the market, they keep saying big lies, in my opinion. And this is not even about marketing anymore, you know, everyone somehow tells a good story to sell their product, but here we're talking about big lies, about where the technology stands in terms of development. So we help our clients navigate that environment first and then we sit down with them to make this profitable for them. You have a huge differentiator owning your compute is a major commitment. How does that differentiate you from the other AI companies that are cloud-dependent? We started this project not because we want to challenge cloud providers. These are among our key partners in the market, but because of the gap between the demand and supply in out-say-i compute more than traditional cloud services. We have a very strong partnership with Microsoft. We love them. We're doing great things together. We also work with other providers. The reason why we started building our own cluster is primarily because we need it as an AI factory. We are an AI software company. We're enabling our clients to become AI enterprises. This is what we do and this is who we are. In order to fulfill that mission, we need to be independent and autonomous in innovating, in training our models at scale. And by making just some, you know, basic calculations, we came to the conclusion that it is more convenient for us from a cost perspective to have a cluster to train our models and potentially to leverage that also to build new solutions like what we call "chip to front-end", which is a solution that segregates workloads from the chip level to the end-user interface and for national security agencies and for some mission-critical environments, especially in the public administration level. It helps because you can take ownership. I mean, we understand these clients. Even though they have a combination of services in the solutions that we are involved with, they want us to be responsible. How can we be responsible of things that we do not control? So this is why we're investing in this sense to better serve our clients and to mitigate risks in terms of strategic autonomy for the company as we aim to build a one trillion company in Europe with global ambitions. So this is our North Star and we understand the tools that are needed to get there. We adapt very quickly and as we understood the need for this, we just did it. And, Ulyan, how does that affect your unit economics? How does it make it different for you day to day? It's absolutely a way to improve our margins. It helps us to own the full chain. So initially we thought this was going to be mostly about improving the way we produce our software, which of course comes with higher margins because the cost to produce the models and to retrain them, keep them state of the art would be lower as a consequence would have higher margins. But this is actually enabling new solutions with even higher margins that we thought before. And this is going to result in making the company profitable ahead of schedule. So it's very strategic for us. And earlier you talked already about the US, China, other regions. When it comes to founding a company other than access to capital, what do you think makes it very different for founders in the US compared to Europe? I think it's about adoption. It's about adoption and leadership in the air perspective ecosystems. So if we talk about the US for example, once you have a working product and we've seen this because we're present in the US, we have clients in the US, they just move faster. And there's nothing we can do about this. This is the culture I think challenge. I fully respect the European culture in being risk adverse and having a different approach. I think it's just a matter of where the market is. I think this is going to change very quickly and the AI is going to help some regions like Europe that have a more conservative approach to adoption. As they understand, you know, the threats that this poses to their future and the needs for them to be more strategic. I'm happy to see some leaders move faster. And I think these early adopters will show their eternal investment and will be case studies and examples to unlock others to move faster. But it's just a matter of culture at the end of the day. And talking about change as enterprise environment moves towards fully AI native. Where do you see if any competitive edge for legacy SaaS providers still? I think it's about speed. SaaS providers have been in a comfort zone, I think, for too long. The cycles that SaaS companies are used to are not no longer possible anymore in this current environment accelerated by AI. So this is not about external factors. This is not about a new model coming out from one of the AI vendors and disrupting their features. It's about how fast do they move? Because if an AI company can build an AI product in two weeks, why a SaaS provider that has experienced customers and has all of the base components that would make an AI product successful, why they cannot do the same. They I would expect them to be even faster. So it's all about leadership in SaaS companies, how fast will they move? And also what is their vision for the business? If their vision is to consolidate, I would definitely encourage them if I were an investor to accelerate consolidating. If their vision is to build a market leader, they need to become an AI company and transform really, really fast. Or if they have a different view, you know, if they're comfortable in their niche, for example, because they're profitable and they're sending dividends to their shareholders, then I think they should think this twice because that niche is not going to exist for a long and this means that they're going to be this rapid. So these are the three categories I see in the market. And I think everyone has the possibility to make a choice because going back to the hard truths and the reality distortion and the lies that we mentioned before, there is time. Not enough time for some maybe, but there is more time than we think. This industry is at the very beginning of it. The models that we're using right now are going to be obsolete in 12 to 15 months. So this means that there are cycles and wide spaces where SaaS companies can actually monetize and potentially be third-movers, third-movers that can not only strengthen themselves, but potentially reinvent themselves and get them even bigger than before. And talking about vision, yours is global. It sounds easy on the paper, but I'm sure it comes with a lot of hurdles. What are the biggest challenges you encounter when scaling globally across multiple jurisdictions? We are doing this quite well by localizing the company. Of course, that is a standard playbook. It slows you down because then you have more complexity, multi-dimensionality to manage from a corporate perspective. So I think localizing the business in different regions is definitely the biggest challenge we have right now because you need to move fast and in order to move fast you need to be a global organization, one team, one goal. But at the same time, if you're building a product for the US and you expect it to be adopted the same in Europe and vice versa, that's not going to happen. So you need to plan for this complexity for this sophisticated environment. And that is a challenge because that requires some extra effort that regulated environments especially in post. And that, for example, other players in the B2C environment in the B2C market don't have. So they can move faster and sometimes it feels bad because it feels like you're moving slower, but actually you're moving with their right speed because trust is the most important currency in our industry. And for us, it's better to be slower in this case and to make sure that these challenges are properly addressed while building the most trusted AI company in the market rather than maybe moving fast, breaking things and risking to break the trust. But in the same time, you're growing really quickly. So what advice would you give to founders that grow quick to build the right team to make sure you have the right foundation? I think this innovation dilemma is something that every founder is dealing with. So I think you need to find the perfect balance between building trust, which is the most important currency. and see, I think, in AI, and it will increasingly become the only one and innovating. Because sometimes, founders need to take risks. And this is what is special about these companies that innovate. They move fast, but they also take risks. And they unlock new phases for the business and for the market itself. But I think with AI, we've gone too far. And we need to find the balance of preserving trust, building trust. Because this is like a slow down to speed up type of strategy. I think now everyone understands that AI is important. But we're not going to have a safe adoption. I'm not saying this is going to result in an AI win turn. But you can see the signs, the failure that we're seeing in the market. And that in the short midterm could actually be a problem for founders. If they move too fast and if they completely miss on trust, I think they're going to have true challenges to manage in the short midterm. So I would invite them to think trust first instead of AI first or innovation first. Finally, there is a question I'd love to ask. Do you see a dominant technology eventually reaching the consumers, those that are the most concerned, maybe about privacy, about using AI the right way? Is that something that's on your roadmap at all? The innovation that we define softer 3.0 is structural. So if it is true that a private company or a government cannot send their most confidential information to a model provider for model customization, which will be needed at some point to get most of the performance out of these systems, they are exporting themselves. They are literally getting their IP out for free. This is going to be true for individuals as well. So we see an opportunity in the future to enable citizens to own their AI. And we're already experimenting with edge models, models that can be isolated on our phones and our private devices. So we think that while we build the most trusted AI brand in regulated environments, this will be much needed in other industries and why not even in the consumer space. That sounds very promising. Thank you so much for joining me today, Julian. And thank you for listening to this episode of Use From the Top. I hope you've enjoyed it and you will tune in for the next one. [MUSIC]