#336 From City Sewers to Sovereign AI with Russ Wilcox, CEO at ArtifexAI

Generative AI is transforming industries at an unprecedented pace. But as AI changes how you or your team work, one thing is clear. Your skills also need to evolve. At datacamp, we offer everything you or your team need to adapt and thrive with AI. Whether it's business users looking to get the Most out of ChatGPT and Copilot, or developers and data scientists looking to fine tune models. You can learn the entire AI skills spectrum on Datacamp. Power your AI transformation today. Start learning@data camp.com.

In a race with the east right now, specifically China, and it's two philosophical differences in regards to artificial intelligence. First off, here in the United States, we refer to AI as artificial intelligence. Meanwhile, the Chinese refer to it as intelligence, period. So that tells you a lot about the philosophical nature approaching these technologies. Sovereign AI essentially is the capacity of a given nation to be able to own and operate its own artificial intelligence systems based on the ethics and the views of that nation. And this often has multi levels to this, not only of how you're building the AI systems, how are you building the data centers that drive them, Are you building them sustainably? Are you trying to build them sustainably? Who are you employing? Are you importing new technology, Are you developing your own technology? And more importantly, what is this concept of data sovereignty as a nation? As I build more technical infrastructure, as I start to piece new sets of information together, who owns that?

Welcome to Data Framed. This is Richie. So far, I've mostly been enjoying the US China AI race because the competition has encouraged a lot of innovation. However, as AI becomes a core part of daily life, any problems caused by AI service disruption get larger. That means that relying on technology from other countries introduces geopolitical risk. And so both nations and companies need to think carefully about their AI infrastructure. This risk mitigation idea is called sovereign AI, and it's on track to become one of the hottest topics for executives in 2026. Today's guest is Russ Wilcox, the CEO at Artifacts AI. He's also the author of the Pacific Divide Substack, which chronicles the US China AI race. He has a fascinating story of how a broken sewer led him from a career as a data scientist to being an AI executive and an expert on international AI relations. Let's hear his tale. Hi, Russ. Welcome to the show.

Glad to be here. Thanks for having me.

Yeah, I want to start with the big one. So who's going to win the US China AI race?

That's a really good question. And first off, I want to say that I firmly believe in American Innovation. I think we got to where we are today due to the great researchers, the great institutional capitalism that allows us to really surpass the world, not just only in terms of artificial intelligence, but most technologies that we've developed. However, it is a deeper question than that because we are in a race with the east right now, specifically China. And it's two philosophical differences in regards to artificial intelligence. First off, here in the United States, we refer to AI as artificial intelligence. Meanwhile the Chinese refer to it as intelligence, period. So that tells you a lot about the philosophical nature approaching these technologies. In the United States, we have a bit of reticence, especially in terms of education, of how to apply it to our society. We worry about kids cheating, we worry about deep fakes, we worry about ethical situations that are very important. On the other side of the aisle, they've had AI education standards at least since 2022, if not going back to 2017 in terms how to use data and build a data driven society. So that kind of sets the backdrop of where we're at right now. Now who's winning? I would make the argument that in terms of innovation, America. Hands down, America has shown that with our novel architectures, Transformers, we have new state based models coming out, we've shown that we have the ingenuity to succeed. However, on the flip side of the coin, deep seek, that came out right, but that only takes one part of the market. We have to look at a deeper strategy. Where China is winning is on the integration side. Within the United States, we look at fragmented data sets, we look at data silos, and we rely on companies to try to break down those silos, unify that data, inform a prediction. On the other side, they already have that. They have what are called data reservoirs, where they have every single piece of available government data, public data being flown in into a single unified system that they call a unified compute architecture. Now that allows for faster prediction, more real time monitoring, and smarter decision making at scale. So we have to remember the AI race is not just about generative models like large language models like ChatGPT or Anthropic, but it's also reliant on predictive machine learning models that are driven on data and how you unlock data. And so if you take a look at the integration of artificial intelligence at scale and the aggregation of data, China is winning hands out. If you look at the novel architectures and algorithms that are coming out of the United States, America is winning. So the more detailed question is, what does the race look like in the next five years? As America seeks to export its AI technology, whereas China is applying it at scale, not only nationally, within their own borders, but also through what's called their digital Silk Road or a more modern Belt and Road initiative. So what we have to be cognizant of during the AI race, it's not this monopole of who's winning, who's losing. It's rather a deeper understanding of what does artificial intelligence mean at scale. How does that inform diff. Decision making across geol. Geopolitical borders? And how is it adopted, adapted not just by companies, but governments small and large and individuals both within the border and outside the border. And I would have to say at this point, China is winning that race.

Interesting. Okay, so I love the idea of different approaches and it does sound a lot like, I guess, recurring themes of, like, what happens when you have like a very much a free market economy like the US versus a more planned centralized economy like China. And so it's interesting that you say in some sense the borders don't matter and maybe it doesn't matter who wins the race, like as long as everyone benefits from the new technology. So, yeah, I guess. Do you want to talk me through, like, what happens with the rest of the world then? Like, should anyone care about this race or is it just like a purely political thing?

It's going to be geopolitical. There is going to be massive geopolitical ramifications. And in fact, one of the concerns coming out of China is this idea of an Oppenheimer moment, like what happens when you have these systems that learn so well that can aggregate new sets of data at scale and democratized throughout the world? How does that disrupt geopolitical events, policies, et cetera. So when we take a look at a foundational nationalistic strategy around AI, we have two competing models that have been competing for the last thousand years, or maybe thousand years a little too much. At least since the Enlightenment, we have this idea of a free market in democracy which is going to win the race, period. We started the question is, who's winning the race right now? And I made the argument that China is. I don't think that's the case. I think America and the west will win in the long run, but we have some hurdles. Democracies are slow and we have this pesky concept of federalism and commercialism, which is very important for states, rights for municipalities, but that creates data silos. And it's hard to create strong decision making around critical infrastructure such as the grid, housing, energy, because you have layers of bureaucracy and jurisdiction that have been historically siloed. And the government can't just go in and say I want all of Massachusetts's data. I know, I promise, I promise it's going to be for good. Doesn't matter. You can't do that. It has to be a private company or a private entity that enables that. On the other side you have the authoritarian model where it's data aggregation at scale. So one of the interesting positions that is coming out of the east right now is an extension of their Belt and Road initiative. And I think a lot of the users are probably familiar with that concept where China will go into a developing nation and say, we can help you, we can help you build infrastructure, we can help you build your ports and your bridges. We're going to be good friends. Yet if you read the loan terms, they're very non friendly. So there's, there's this tension here at the infrastructure level. Now as you start to extend this, you take a look at how does that play out in terms of artificial intelligence. Because China is also starting to scale its intelligence models and its intelligence infrastructure to developing countries that need this infrastructure. For example, Egypt, Indonesia, all of these areas are reliant on getting the technology they need to make modern informed decisions, improve healthcare, improve agriculture, all of these things that we want in a good world. So as an extension of that, China is starting to expand that out. So they've taken their data model and now are looking at strategic areas for implementation of AI tech. Now you have to ask, why would they do that? One, it's strategic from a soft power standpoint. A two, what drives artificial intelligence data? So they are looking to extract new novel sets of data the west currently doesn't have access to. And that presents a challenge around data sovereignty, AI sovereignty that we have to figure out. And I'm not saying that putting infrastructure in developing areas is something we shouldn't do. We have to be cognizant of that. That. For example, I believe there was a news story coming out of Denmark where the Danish bought a bunch of buses and they found out that, you know, they were purchased from China, there was actually tracking devices and pinged a Chinese server in Amsterdam. Right. You have to beg the question, right? And I'm not being adversarial because I believe that there is opportunity for collaboration to learn about how infrastructure is built, how we make these data distributions, data driven decisions, how you unify data sources. I am not a hawk. I am not saying we should separate these two. I'm saying we have to first fundamentally understand what the two differences are between these races. So China is expanding out from a data aggregation standpoint so that they can have the most real time up to date information. However, that's not historical data, that's a thin layer of modern data that they're going to build up over time. So how does the west compete? The west is empowered by democratic principles, debate diversification, historical records. Problem is that data is not easily accessible. It's locked in basements of city halls, it's an old PDF and thousand pages of reports that no one reads, but has an entire history or community, but also contain structured data. So while China is winning the race right now, America and the west has the opportunity to innovate yet again and supercharge this by unlocking massive amounts of historical data about what makes Boston different than Framingham, different from Chicago, in using that diversity of training data that has been transformed from unstructured information to structured information to win the race. So it's a little bit slower right now. And of course we have economic tensions that are exacerbating these, these challenges. But at the end of the day, those are the two fundamental differences between the AI race. And so you have to look at it as a more nuanced geopolitical perspective. Perspective rather than who's going to get chips, right. Chips are a big deal. That's important. But more importantly, what's going to happen to the underlying data infrastructure that's starting to scale and who controls that? That's going to be what's going to be interesting in the next couple years to see.

Absolutely. And so I think the idea that China's investing in lots of different countries is really interesting. I think it's currently the biggest investor of infrastructure throughout, throughout Africa and particularly, yeah, a lot of developing countries, like you say. So before we go forward, I should probably say I've got to be neutral on this. I'm a Brit, so being neutral between the US and China seems a sensible take. You mentioned the idea of sovereign AI here and data sovereignty. So do you want to talk me through, like, what are these concepts and who needs to care about them?

That is a good question. And I want to say, before we answer that, just to comment on your, your comments about being a middleman. I am no way trying to position an adversarial relationship from either side. I think that there are learnings that we can, we can, we can deeply value from both sides, but that requires communication despite differences and this idea of finding common ground and that will supercharge innovation so the basic idea of sovereign AI, you can make it similar to the idea of a nation and the idea of nationalism. So I'm in Massachusetts, I'm in the United States, I have a set of laws that govern how I interact. I can't go and murder you, I can't go and steal your car, right? All of these things that we all value and say, this is good, right? This is what makes us a function, system, society. Sovereign AI essentially is the capacity of a given nation to be able to own and operate its own artificial intelligence systems based on the ethics and the views of that nation. And this often has multi levels to this, not only of how you're building the AI systems, how are you building the data centers that drive them, are you building them sustainably? Are you trying to build them sustainably? Who are you employing? Are you importing new technology? Are you developing your own technology? More importantly, what is this concept of data sovereignty as a nation, As I build more technical infrastructure, as I start to piece new sets of information together, who owns that? So a good example is, let's say that I'm a nation, okay? And I'm. That nation's gonna go nameless. And I really believe in helping agricultural outcomes in Egypt, okay? So I'm gonna export AI technology in that AI technology might be in the form of smart sensors, it might be in the form of dual use technology like drones to capture data. And I'm going to build the pipelines to, to help this country develop a more digital ecosystem to improve agricultural outcomes. All of which is good. We want that, who owns that data. That data is highly valuable. Being able to predict crop yields, being able to understand where, when there might be droughts, who owns that data? That is the concept of AI sovereignty, but at a base level. So who owns the data that drives this? Who owns the infrastructure? Who owns the technology? And it's not monolithic. There could be aspects where we want to import technology. For example, Anthropic is one of the best large language models out there. There's not a competitor to that in the market right now. I'm sure all the people using GPT will say, oh, GPT 5 is, is good. No, no, no, sorry, Claude. 4.5 sonnet, hands down is the best coding assistant out there. I will state my, my, my flag on that. No, I was not endorsed by Anthropic SATA, but I'm just putting that into perspective, right? So right now we have the kind of two monoliths with Google coming up with Gemini, right? You have These two or three monoliths that are dominating the market, those are external API calls that a foreign country, an ally or an adversary has to send data and get data back. That's not an ideal outcome. So these, the countries are now starting to evaluate how we build our own internal AI systems and how do we build our own AI governance systems, how do we build our own data governance systems? And that's what sovereign AI is, is how does a nation have control over its digital infrastructure?

Okay, so it sounds like, I mean there's lots of different layers here. So you're going I guess right down from like data centers and all the kind of the low level infrastructure right up through the technology stack and then issues around like who owns the data, like the intellectual property around that. Okay, so many different areas. Then I can certainly see how if you are a government, this is something you really need to care about. Like does can my country support its own data and AI needs? If you're not a member of the government, is it something you need to be involved with or you need to worry about yourself a hundred percent?

So let's talk about data. That's the first thing, right? Because if I'm a, if I'm a private company, I might be reliant on an AI model like anthropic, right? And I'm sending data back and forth. First question that I have is what happens if there are sanctions across anthropic? Very unlikely. Very, very unlikely. What do I do? Does my model go down? Does my business die? What happens if I am using Deep Seq? What if I found that the way that I'm using Deep SEQ is not adversarially I'm using to structure data or something like that, but tariffs come in and sanctions come and say I can't use technology from another nation state. That presents a problem because not only are you impacting the company itself, imagine if you're a healthcare company and your services are interrupted to the constituent members of that healthcare community. So we're talking about AI models, but let's go back again to drive what drives AI models. It's data. So what happens if there's a disruption with data centers? What happens if our grid goes down? What happens if there are geopolitical tensions and I have a server in Norway, but Norway doesn't want to deal with US Tech anymore. What happens to my data? What happens if I'm building a data center in a very interesting regulatory environment? I have data going back and forth and that might be in a free some Sort of free trade zone. There's free trade zones popping up in a variety of areas, from indigenous areas, from isolated regions. Who owns and who's liable if I just decide to shut my data center down? These are the questions around AI sovereignty you have to consider both as a government but also as an entrepreneur. Because it's not just the data you're going to have to worry about, it's going to be the downstream regulatory effects you're going to be concerned with. What happens if your services are interrupted, what happens if compliance fails? And these are the questions that are still evolving that go into one of the questions that you asked a little bit earlier, which is what are the differences between America and China? But we forgot a big player, which is the eu. We know the EU had their AI act and I am not a big fan of that because I believe, I don't think really it inhibits innovation. I think they have good intentions. Just like my Senator Markey's bill for AI has good intentions, but good intentions paved the way to hell. If you read the actual policies, you could interpret Senators Markey's bill as requiring high school students to disclose the numbers they put into linear regression because it's traditional AI. So we don't even have a common linguistic framework at a national level to understand this, let alone at an international level to understand disputes around this data. So sovereign AI comes into the idea of a company saying where are my risk points and how do I mitigate that? Do I want, do I want all of my eggs in one basket? Do I want everything stored on an Amazon US east server that could be disrupted in a, in a, in a, in a traditional data security set? Right, but imagine that at scale with AI. And so this is going to be evolving as countries start to develop their own models, as small language models become more ubiquitous, as state based models that require less compute evolved and start to become that parity with transformers. These are the questions we have to start answering now because in two to five years it's going to be too late. We're going to have some sort of problem that we could have fixed if we thought about what does this mean to have some sort of sovereign control on information flow between my company, other companies and governments.

Okay, well, so actually I don't know about this bill from Center Markey. Do you want to talk us through that? And then I'd love to chat a bit more about the EUA act and the role.

See, you see there is a, a variety of AI acts getting put out from a variety of legislations, you start to see the states becoming the playgrounds for this to start. And I'm going to, I'm going to set the stage before I answer your question directly because again, a lot of this requires context around the generalized understanding of, of artificial intelligence policy. So just like any type of policy, you can have a local policy at the states and a federal policy at the national level. Right now the national AI space is equivalent to the Wild West. It's looking to create innovation and inhibit regulation. Right. And there's a trade off. There's a trade off between these two. And I'm not saying that we should have no regulation, but I'm also saying that we shouldn't have overbearing regulation such as GDPR right now as we're developing innovation technologies. But we also have the roles of states so the states can come in. And we've seen this with the California Data Protection act, we've seen this with the Texas Data Protection Bill. Right? You are going, if I'm a company operated in Texas and a company operating in California, I have two different frameworks of data governance that I have to deal with. And that already gives me a headache. Could you imagine having 50 different state policies you have to deal around with AI? So that sets the stage. So now when you go to the federal level, you have two kind of camps coming out of the respective party. One to drive innovation. And I'm certainly don't want to get into politics here. And the other side, which is saying, hey, we recognize that there is this nature of artificial intelligence do massive amounts of good, but also there's a substantive risk to this. So how to mitigate this? So in 2024, Senator Markey introduced what he called the AI Civil Rights Bill. And its goal within its intent was admirable to eliminate AI. Best to put guardrails on algorithms to help protect civil liberty, people's rights and livelihoods. Okay? The challenge is the policymakers themselves and those individuals that help. Policymakers are not technical in nature. They don't have substantive understanding of deep technical architectures of where this bias comes from. So you might label something called traditional AI as a technician. What does that mean? Does that mean K means clustering model? Because that could be considered machine learning to be traditional AI. Do you consider it generative models like ChatGPT or again, you go to the farthest originalist interpretation of a policy and you say, well, what's the foundation of all of these models, which are probability and statistics? So if I'm a high school student doing A linear regression. And I would never say that this was actually going to happen, but based on my interpretation and reading of this, if I'm put a bill out requiring the auditing of traditional AI systems and their disclosing of their training data, there is a world where, as a high school teacher, I have to give the data sets that I'm going to give my high school students to do linear regression to be audited. And as far fetched as that sounds, I'm not saying that was the intent or that would ever be applied. I'm saying there is sophisticated technical nuance that our legislators are not equipped to handle right now. And as we start to create informed, as we start to create policy, we're creating emotionally driven policy, which is never good. We want to create informed policy. And that's where I think there has to be a blend between think tanks, nonprofit enterprises, government, to come together and understand what is a common language, what is a common framework we can talk about so that we don't run into problems two, three, five years along the line because of vague interpretation of a policy that then has to go through the courts. That's how we lose China. Okay.

Yeah, certainly, I can see how if high school students have got to start auditing the work they're doing, it's like, well, that's kind of the job of the teachers.

Right.

Or whoever's creating their work. It doesn't need to go.

Go in and say, okay, Mr. Cotton, what, what data are you using to teach? Matrix? No, I'm just, again, being a physicist, I'm trained as a physicist. I always was learning to look at edge cases to start. Right. How do you interpret it? Interpret those as your kind of guardrails as you start to develop policy.

You mentioned at the start there are lots of fairly unambiguously great laws. Like the idea, like, don't murder people. It's one that I think most people think is a good idea. But then at the other extreme, okay, with AI policies, like, you have something where everyone's got to have every single model they've ever created reviewed that's completely unworkable. So do you want to talk, like, how do you get to having better policies around AI? Like, what's going to make this work?

Yes. And it's going to be something that we don't like to do right now. It's going to be sitting down at a table and saying, richie, I hate everything you say and everything you stand for. And you do the same thing for me. Right. We live in a tribalistic political system right now. There's no kind of denying that fact. I don't want to turn this into a political discussion. Right. But we, again, we have to observe the system we're sitting in. What it's going to take is when you talk to people coming together, and there are great folks on each side of the aisle doing some phenomenal work around AI regulation. Senator Todd Young, for example, is doing some great work to spur innovation, while also understanding that this is a very complex, adaptive system. To go back to your question, you have to bring people together and find common ground. And I'm not talking about regulations yet. Regulations should come at the very end. I am saying first, can we agree that maybe using automated systems to kill people with drones is probably not a good idea? Maybe that's a good place to start. Right. Again, not saying we're doing that now. Not saying that we should be doing that. I'm saying let's start with that conversation. Okay, let's first find these areas of common ground. And then you go deeper and you start to say, okay, what are areas of societal risk that go into this? And one of the first things I think of is the criminal justice system. Do we live in a. Do we want to live in a world where sentencing based on a criminal act is driven by a model? No. But does that mean that the final outcome can't be informed by some things on that that has been proven to be unbiased? No. So we have to think about, okay, what does a human and AI system look like together? Now, the Chinese have a very interesting framework around that that they call human AI. Human AI, which is humans and AI formed to be a blended model of humans plus AI to improve decision making. That's where we want to start adopting to realize that artificial intelligence should never be used to make final decisions. I will, I will defend that to the death. AI should always be used to inform someone to help make a better decision. So if we're talking about the criminal justice case, we say, okay, let's examine systems that can help make that decision making more fair, more just things that we all agree on. Okay, how do we do that? Well, we have this model. Okay, where was that model trained? Because we know that there has been systematic biases in the criminal justice system. So is that model inherently biased? Did you try to balance your training data to mitigate those biases? Is it explainable? So I think where we start is a conversation of where are there high risk fields around this? Another good example is around what I call cognitive sovereignty. As AI becomes more and more developed, it becomes more sentient. Like it starts to talk to you, it starts to inform you, you start to develop a certain relationship with it where you feel comfortable asking questions. Now, because we have had some horrible instances of suicides, there are safeguards that models start to put in there to say, okay, we shouldn't talk about this. And I agree with that. But where does that stop? Does that stop if I'm a researcher doing deep research and on, on the rise of fascism in the 1940s and I go too deep into a model and I start asking about Joseph Goebbels or things like that, how they did something right? Now that's an AI safety risk. And I think I agree from a certain context, but from another context you're inhibiting someone's ability to do deeper research. So that's the next step. So you define first a framework where you say there are absolutely areas where we should not be using artificial intelligence. It is too risky. The next layer is there's a higher risk tier of criminal justice policing. We found it a big two maybe things to do with genetics. Then you go to a deeper layer and you say, well, okay, healthcare is important, but we can still be a little less risk adverse because doctors are overwhelmed. We need better decision making. Right. So we have those three layer that we have those layers of the onion start peeling back to start making informed policy that doesn't inhibit us while also looking at other regulations that might inhibit artificial intelligence development. But simultaneously we have to have a deeper discussion around what is the underlying cognitive architecture of these. And I'm not making the argument that we are at AGI or using AI is going to kill your brain. I mean there's been studies out of MIT chat GPT on your brain where it says, okay, whatever, 30% of, of your front frontal lobe or whatever. I'm not making the, I'm not getting the statistics right. Right. But if you're using passive use of AI, you have metacognitive decline. But if you're using it in the right way, actively engaging with it, then you can use AI to be your superpower. So it's this duality that we haven't figured out. And as a society and as policymakers, we're not very good at dealing with dualities. We want things black and white. We don't like things gray. Well, you can't get any grayer than the proliferation of artificial intelligence systems globally. So we need to bring the voices together that don't often talk with technicians in the room with ethicists and philosophers in the room to really understand what is a framework that we can all agree on, but a framework that's not limiting. Because this goes back to the idea of sovereign AI. My American values are going to be different than your values in the uk, which are going to be different than my friends in Norway. Right. So we have to allow the flexibility for sovereign AI systems to adapt based on that culture, but we have to also have a common playing field of rules that we haven't started substantive discussions around yet. And then you're able to get at deep policy. That's my viewpoint on how we tackle this. Okay.

So I'm saying I really love the idea of lots of different groups of people just figuring out what are the real issues here and having conversations and seeing where there is common ground, where there are disagreements. It seems essential as a first step in order to get to good policy. The one thing that struck me, though, is when you talk about the different things that, okay, yeah, using drones or whatever to automatically kill people, terrible idea. Should probably ban that. And then you've got things like policing very high risk and whatever. I was like, I've heard this before. I feel like you've just reinvented the EUAA act risk classification system. It's very similar to that.

So, yes, but the risk classification system within the EUA act, again, is built on good intentions, but it creates hurdles based on its deeper language that stymies innovation. So we agree on the risk frameworks. I don't think anyone's doubting about that. And I am certainly not disparaging the EU AI act on that. I'm now asking the question is what does that disclosure look like? To, for example, if I'm a startup and I have to disclose my training data for the world to see, I lose my competitive advantage. That's going to inhibit innovation even downstream. I might be doing something in the drone space to do something. Which I'm not. Which I'm not. So there has to be a deeper balance in large and more substantive frameworks around the disclosure of training data, how these systems are audited, what are the ethical systems that you're dealing with. Right? Because even in a perfect world, if you're having these different risk systems, right, at the end of the day, everything comes down to an embedding. Right. How these models work is you take words and you put them into very long sets of numbers and you do some sophisticated math and then you try to predict the next token from a threat standpoint, what happens if I just change one of those numbers where 99.9% of the time I'm right. And a good example I might say is you have a classifier to look for BOMs in the TSA. Okay. And it goes through. This is not a bomb. This is not a bomb. This is not a bomb. But an agent gets in there and changes one of these little embeddings and now, oh, a bomb is now a toothbrush. Right. And again, this is not a real life situation. I'm just trying to have used it as an illustrated fact that there are not strong quantitative mathematical frameworks on how these systems operate well enough to create a highly regulated disclosure of training data, because we don't understand the axiomatic principles mathematically of these models. There's a great guy out of the uk, Anders Hansen, who was trying to actually create mathematical frameworks out of deep, deep neural networks and really try to under explain, understand explainability and those types of things. That's valuable. I don't necessarily agree with showcasing what the data was trained on because that's a, that's a defensive mode. I don't necessarily agree on how hard it might be to implement a risky AI system because you need to test it, but you need to test it in the right ways so you could have a live field test, for lack of a better word. So you have. I'm a startup, I'm working in the education space. I'm worried about that. My AI system, which is high risk, it deals with children who are, who are communicating back and forth, what is going on, what data is aggregating, all of this type of stuff. Right. I should have a set of frameworks, not necessarily defined by the government, but defined by a system of who watches the watchers, what are the ethical outcomes? Not just going into the regulatory outcomes, what are the ethical standards that go into this? That's where I'm getting at. And it's not getting at the heart of the deep implementations of risk factors, because we can agree with them, it's how they are applied, how are they audited, how are they implemented. And also regionally, you have to think about, the EU is going to be very different societally from America. Our justice systems are very different, our viewpoints on healthcare are vastly different. Right. So there's gonna be different sets of regulations you wanna go through. And the last thing we wanna do is necessarily just borrow legislation from the EU and plop it into a state bill, which is what I've seen in some of these proposals. We have to take A more nuanced, personalized approach to how we actually develop those regulations. Okay.

All right. So just to make sure I've understood this, so perhaps going back to the school kids examples, you got some AI helping school kids chat about their homework or whatever. You're going to want some guardrails in there, I guess. So it's not like swearing at children or whatever, but if.

Let's actually go down this road. Interesting, right. One of the frameworks that has been developing within education and artificial intelligence is this idea of a Socratic tutor. So this is the idea is it's a tutor where you can ask questions and it will ask a question back to you. It's that annoying ta. You had an undergrad that when you ask a question it would say, have you thought about it this way? Right. Why? Because it's. There are studies shown that that promotes metacognition or how we think about the way we think it improves learning outcomes. So what's happening is in that education chat, you're having an at risk individual chat with a system. That data is going someplace else. Right. Imagine what can be used by the data constructed from those chats. Can it be used to infer learning disabilities at scale? Can it be. If you have this system for education that follows this student throughout, say, grade three, all the way through college, can it usually predict what they should go into for work a la Gattaca? Can you have enough contextual information to understand the psychological profile in a student? And this is what I mean by the new, more nuanced applications of AI everyone's talking about, like the training data, but every time you interact with an AI system, you're creating new data that can be transformed into structured data that can create risky situations. And those are the types of conversations that are being lost right now in the regulatory sense because we're so focused on the deepfakes on the app, the high risk on the copyright. And I'm not saying that that's not valid, and I'm not saying that we shouldn't be cognizant of it, but I'm saying that that should not dominate the conversation. Because what happens when we're interacting enough with these chatbots where we can create a psychological profile on person? That's where we get into some really challenging ethical quandaries and we don't have the frameworks to assess that. So that's where my dream drive is, is how do we have a deeper conversation on those principles to promote more intelligible regulations rather than just thinking about you Know at risk systems and bias around that. So sorry to interrupt your, your question, but I thought it was a perfect kind of transition to give you a use case of this idea of cognizant sovereignty and the ability for AI to actually create new sets of structured data to build new models. On top of that we don't even think of. Absolutely. Yeah.

I mean it's a fascinating area and certainly in that sort of education setting where you're dealing with children, so many things that go wrong. Because if that information then feeds into like your, your criminal justice system records, it's like, oh well, yeah, we're going to give this person a longer sentence because when they were six, they said this and the AI flagged up whatever.

Yeah. And it goes into the data sovereignty. Right. So what happens to that data going on a server? What happens to that server? Right. So all of this system is intertwined. It's a com, what we call a complex adaptive system. You turn one little knob, you poke one little button. Right. And it has downstream effects that we can't predict right now. It's a chaotic dynamical system. And that's what we have to think about because exactly what you said. Right. What happens if we do create these unified data sets? There was a lot of good that could be coming from it. Right. We could understand. Are there systematic gaps in, in education? We know there are, but we haven't quantified them. Are there opportunities for hyper personalized learning? I think everyone agrees. Yes. But what's happening at that data layer, what's happening at that infrastructure layer is what scares me.

Absolutely. I mean so many ethical problems to think about, like many issues that need to be worked through and decide whether they independent legislation or not. I would like to take a step back though and talk about how you got into thinking about sovereign AI to begin with. And I believe it all starts with a broken city.

Sewer infrastructure is a challenge in the United States. The way we build power plants, housing critical infrastructure such as sewers goes back to your local government. I'm sure everyone loves going to their local government. I'm sure it's the best experience that you look forward to every chance you get to stand in a long line and wait for a permit and not know what happens with this regulation. So started it was a few years, it was a year and a half after Covid and I was in, in my city where, where I was living at the time. And the city had a budget about honestly $65 million. Okay. Most of that budget went to the school system. It also has a water quality problem. What that means is that the city didn't have a sewer system. It had what call what are a septic system. So what happens is when you go the bathroom, it goes into a tank and the material, the biomaterial degrades into phosphates and nitrates that then leach out into the environment. Okay, and now I know I'm getting fairly technical, but this is, this is important. What happens is then that goes into the water stream as you build and build more and it causes algae blooms when those algae blooms occur, and then when they die, they give off what's called nitrous oxide, which is 300 times more potent than a CO, than CO2. So it's a climate impact. Secondly, it's a hyper local issue, meaning that the state government can't tell me what to do. I have control over my little town jurisdiction. And this is a problem that scales globally, that impacts global markets by $10 billion each year. So that sets up the scale of the problem. So the problem is, is this city wanted to put in a sewer system. That sewer system, I think phase one costed $250 million. Phase one on a town that has a budget of $65 million, where the vast majority of it goes to the school system. How do you deal with that? So I went to a local zoning board meeting and I hear the presentation and I say, okay. I raised my hand, said, how did you learn to understand that this is the most important part of the Tao Da sewer? We listened to a consultant. Great. Where did a consultant get that data from? I said, can't you go through your permit history and see all of the areas where there might have been damage systems when they were built? Were there variances granted or exceptions granted because of one reason or another? And they said no. I said, well, why? They said, all that's in paper. It would take way too long to figure that out. I didn't like that answer. So I called my college roommate, his name's Birdie. I said, bertie, I have a great idea. What if we could create a system that changes the world from the ground up? Says, what do you mean? I say, what if we could create a system that actually could digitize in document? Digitization is nothing new. There are companies out there. Do that. We'll take it and put a PDF. I'm saying, what if we used AI at scale to extract structured data from these documents to map out where these problematic areas were. And in a weekend we did that and we showed a proof of concept saying that we could digitize all of this record. And I said, well, this is now very interesting because this only impacts the town I was living in. But they all share the same water supplies, but the town next door doesn't have information on this. So how are they talking to each other in data? Where. How's the state getting this information? And that's where I saw the opportunity. I said, what if we could create a system to build this sovereign data layer of public information for the United States for good so that everyone can have access to it? And that was apparently a good idea. I launched Artifacts AI as a TED Talk. A year later, I was offered to give a talk on this at the World Economic Forum. And then that scaled my career, where now I chair the policy committee for the American Society for Artificial Intelligence. I've had the opportunity to brief various departments, work with various members of Congress on AI. It all comes back down to that same problem, that in order to compete with China, we need to drastically reevaluate how we operate up in the United States to empower democratic fence principles. And that's where we come in. We use AI to digitize all of these records, connect them. So you understand, what is the risk if I were to put a sewer system here, or if I'm going to build a wind farm in Iowa, what regulations are going to hold me up? Why are they going to hold me up? How are they built the first place? And you could see a world where this, in a way, kind of goes down to a 1984 area of, oh, there's a surveillance state here, blah, blah, blah, blah. And I'm not saying that that's not a risk. I am saying we take the strongest security and ethical safeguards around that because we want to use AI to empower democracy. So one of our customers was dealing with is a city couldn't match their septic records with their building permits with their zoning records. And you connect all of that, and now you're able to quickly identify compliance issues, you're able to create informed policy, you're able to report back to the state on this. And my dream is that America creates what we call a digital utility that creates structured information to drive infrastructure decisions, to drive policy decisions decisions. It all started because when I went to go swimming in my local pond during COVID I couldn't because of algae.

I feel like that's as good a reason to try and save the world as anybody. It's like, I wanted to go swimming, I couldn't because of an algae bloom. Yeah. No, I love the idea. I mean, because we think about how all these, like, the large language models, they've got all the data on the Internet, therefore everything but all the stuff on paper, it's not there. And I feel like there's so many bits of data that are locked away. So it's cool that you're managing to unlock it. So I just told me to, like, what. What. What do you need to do to build all this stuff? Is it just like a lot of, like, grabbing paper records and then just scanning them in? Like, what's the take?

There's. And. And I'm not going to give away my secret sauce, of course, but they're. There's one fact that we have to admit. Governments historically have been underserved by technology services. I challenge you to go to your local assessor's website and just guess what year it was built. I'm guessing mine was in 2003 and never changed. So there's technical debt here. That's silo. That's one aspect, right? There's technical debt that's siloed within a municipality. There is. There are documents that you can scan and digitize. And we do do that, but not just to create the PDFs, to create that structured data. And then you take a look at all of the different other sets of data that makes a city unique. You look at their town council meetings, you look at public comments, you look at old state databases that no one has touched. And then you say, okay, can I develop a system driven by machine learning to aggregate that? The answer is yes. And, in fact, we can do this globally. We can understand development platter patterns in China, in Tanzania, all across the globe, because bureaucracy spans the borders. Now, I'm not going to go into, like, turn this into an intelligence pitch or anything like that, right? But what I'm saying is that government data is siloed. So the first thing you have to do is you have to break down those silos. And AI enables that. And then what you do is you categorize it and extract data out of that. And you can do that however you want. Whatever flavors of models you want, we have our own thing. And then we connect it. We connect that using machine learning. So you know that Russ had a meeting with Richie, and it's the same Russ that couldn't swim in his local pond because of an algal bloom. But you might hear that from a podcast, but you might see this from a zoning board meeting where Russ raised his hand complaining about this. That would historically not Be seen. And then is where the fun part comes. Then you get the entire histories and you look at deed records, you see where has there been red line why is things happen. And you start creating this ecosystem of structured data that is explainable, that is self reference, referentiable to create new novel unified data sets. And then you can build predictive models on top of that you can bring in geospatial data, you can bring in in social media data. And at its heart is an ecosystem of connected knowledge that can empower C development. So what it looks like is yes, there are times where I've had to go pick up boxes and pass them through a scanner and validate that that is correct. But at the end of the day you get novel sets of data. There are other times where we have what's called an omniscraper that could go out and scrape all of this different types of information and pick that up and link it together, whether you're at a state, local, federal level. And then you desilo it and you tear it down and you create a sovereign data layer of structured information that has competence in how you relate to these documents. And then the fun comes of building models and that's how cities should be built in the future.

It's very cool stuff. Um, it reminds me a little bit of like, I guess the, the original goals for like from like Google and Meta, where it was like we want to know absolutely everything about you, but that was like in order to sell you better adverts here it's like we want to have like make better cities. So a much, I guess a more noble, a more noble goal there is like let's go on a city level. Like this is all the stuff that's happening, this is how we make decisions better.

Yeah, 100%. Because if you look at the smart, smart city movement right now, and I've been fundamentally opposed to this movement for the last five years and years, is let's put in another sensor to measure traffic or air quality or this. And that's great, it's smart. Where does that data go? How does it reference the history of the town? How was it built? Why did we zone this one building this way? It's so funny. I think one of my co, my co founder Bertie I think went to a zoning board meeting or a planning board meeting when we first started in this city and they were debating why this one building was zoned this one way because they wanted to convert it to mixed use or something like that and they couldn't figure it out. We went back to our system and we said oh well, because in 1962 this person did this and this is why they did it. Right? And that's important to understand in context. Right. Or if you take a look at this at the federal side, large scale regulatory behaviors. You look at large scale acts of legislation, they are not new, you just append, you amend them. Right? Look at the Constitution, right? We don't create a new constitution. Historically we added amendments. Sometimes those amendments aren't contradictory in large scale policy documents coming out of FERC or LIPA or NERC or any of these agencies that require strong regulatory oversight, which I agree with. They are inconsistent with each other because humans are inconsistent. The person who sponsored this purse this one piece of legislation didn't understand the full context because they couldn't read a 20,000 page bill and understand why something happened. Right? So that's what I'm saying. AI enables democratic processes to function better. Another example of this is think about what's happening in local news. Back when I was a kid in the mid-90s, I could go to my newspaper and actually get a piece of unbiased news of saying what was going on in 2000s and growing, growing going forward. You had Sinclair Media just starting to buy up local newspapers and you don't have that level of report reporting anymore. So you have information absence from your local populace. That's why boater turnout, the municipal level is so low. I'm talking about between 9 and 13%. Is that democracy? No, it's not. Why? It's not because people don't care. They don't have the time, they don't have the resources to go to these six hour long planning boards meetings to get news from. But AI can democratize that, AI can enable that. Right. And it's this reframing from using AI as an efficiency tool from the corporate standpoint to an enabler in the public interest technology standpoint to improve how government functions.

That's a great dream. And I agree that it is quite difficult to get local news like in New York, some decent sort of cable news channels. I don't think there's a local newspaper. And if you want to go digging for like or what's going on then yeah, it's tricky to find the information.

And you'd be shocked what you find. You would be shocked. What you'd find is public record.

Okay, yeah, well hopefully I guess with your project it's going to be a lot easier for people to do that sort of thing.

Yes, that's the goal. That's the goal. Because again, what we're trying to do is we're trying to create transparency in this data. Any model we build is always going to be transparent. We talked about my viewpoints on regulation and things like that, right. They're explainable. Anytime you use AI in government, you need that level of confidence. Right? And that level of confidence starts at the data layer.

I want to pick up on one thing you said. You said quite often that a lot of the decisions go back to ABOL. Something happened with a permit in 1962. It sounds like to make this work, this has to be a project that works for decades. So how do you build for long term like that?

So that's a really good question and you'd be surprised. So some of our customers have, have wondered, you know, I'm building in Kalamazoo, Tennessee, right. That might have a population of 5,000 people, am I going to get data there? The answer is I've gotten better data from Huddo, Texas that had their records digitized going back to 1850. You could see when the first telephone pole was built, astounded. And then you go to Waltham, Massachusetts where you can barely get any public record. Right. So that's the first step is to realize is that there's not a correlation between population density and transparency of rights. And that tells you something about society. It's very interesting. But more importantly, this is one of the things that I'm talking with legislatures at both the state and local and federal levels is this idea of a data utility. God forbid someone thinks that I want something, some state owned enterprise like Amtrak or something like that for AI and data. That's not what I'm saying. I'm saying we need to have a data layer that functions like the highway system for information flow. Why did we build the highways? Why did we get that infrastructure we'll see more easily transport troops and supplies and supply supply chains. Right? Data is the new oil. Now you need a way to pipe it. And right now it's just getting piped to corporations. I'm saying there has to be some frameworks in a national investment around a joint public private partnership model or a data utility that allows folks at ferc, which is one of the large scale electricity federal regulatory environments that allow the state department of Public utilities to access all of this data. Why? Because inside these documents are the histories of rate hike changes, how the grid has developed. We know the grid needs massive monitor modernization, right? And people are talking about virtual power plants and all These great technologies, but it's missing the context of them. And so it's not just going to be the private side that invests in this. There has to be a concerted effort on the public side as well to say that this is a critical infrastructure problem that can improve decision making.

No, I love that. And I suppose we've had similar discussions over the last few decades around, like, the Internet, like it went from, oh, this is a cool new technology to actually, this is a utility, something everyone needs. And so, yeah, it should be available to everyone at a reasonable price in order to help the rest of the economy. I'm right there with you on. We should have the same thing for a data layer as well.

And I think there's a way of monetization this. Right. And I don't want to take this conversation too far, but you take a look at one of the challenges around housing affordability, right, is taxes. Towns make their money on taxes, mostly property taxes. So if you need more money, what happens? You either artificially inflate your property values or you look at the market and say, well, my $200,000 house is now $300,000. So the state gets more money or the town gets more money, or you do different funding modalities like Prop 2 and a half overrides here in Massachusetts, right? But modern small and midtown American towns don't have the infrastructure to compete financially. That's why this town only has a budget of $65 million. It can't afford its sewer system. So it has to go to the state, has to go to the federal government or grant. What if there's a way to monetize this internal data? So now we have to think about data as an asset, right? So historically, we go back to that first conversation where we say, why do we have laws, right? If I killed you 2,000 years ago, right, your brother or your son or your child could come and kill me and Hammer Robbie's code would say, yes, it's done, but if I put a fence on your property, I couldn't kill you. That's that. That's not equal. So you had this concept of an arbit, right, that said, okay, well, there's this set of common law that says, yeah, you're over the line. Here's what you have to do to mitigate that. But I'm not going to kill Richie, okay? And that's how our legal system developed over 2,3000 years around this concept of property as an asset. And we've built our economy around that. And I'm not saying we, we shouldn't. I'm saying that data, which is a far more esoteric asset, a harder to comprehend asset and something that has come into the world maybe 20 years ago at scale. We have to reframe the way we think about monetization and assets around this because we can't just raise our property taxes for the next 30, 40 years funding these projects. We have to find new ways of monetization and data and insights are the answers and that's how you can have this public private partnership model.

Okay, yeah, it seems like a useful thing to pursue hopefully over the next few decades. Maybe, in fact maybe we can go even soon that we've got plenty of listeners here, particularly I guess if you're in Waltham, Massachusetts where apparently there's not good public record systems. Maybe go and speak to your local government representative and. Yeah, see if we can get some things moving. All right, wonderful. So just to wrap up, I always want more people to follow. Is there anyone whose work you think is worth looking into in this area?

Yes, I think there are a couple places you can take a look. One of the first locations I would take a look at is a great non profit out of the Boston area called Just Horizons. They are doing some work on ethical AI standards. They're run by a work world leading ethicist and philosopher with technology, Wesley Wildman. They're doing some, some really great work on, on ethics and AI. So that is I think one of the first places to take a look. The second is there is a group called Gaia X out of the EU that is doing a very similar type of thing of data sovereignty of public data that I'm trying to do here with artifacts. So I think that's a, a good primer on what does it mean to be able to create these types of democratized AI systems. And then the third, which I'd highly recommend it can be a little dense, is the center for AI and Digital Policy. This allows for kind of a global academic network of folks can come together, really debate and figure out these frameworks around AI security, AI ethics, et cetera. Those are kind of the two or three external sources I would take a look. And then the last one is a personal plug for me which is if you're interested in actually learning more about my views on the AI China race, check out my substack blog called the Pacific Divide. I take an unbiased look. I actually read and translate the actual documents from China, compare them to regulatory frameworks and those kind of the tools that I would suggest, but I want to leave the listeners with one more thing is that we are in one of the greatest moments of human history. The only point in history that is similar that I've seen is the Agricultural revolution. The Agricultural revolution forced human society to change on how we work together. We evolved from hunters to gatherers to say no. Now we need to have a set of communities for common good. AI can do that. Right now we're developing these nationalistic AI systems and we talked about data sovereignty and all of that, which I agree with. But AI can also find common ground between people and create modern collaborative frameworks. So if you're scared about AI, I don't blame you there our risks around that. But I would challenge you to say what can it enable and for you to test it. And when you test it, don't test it like Google, don't ask it for information, don't ask it to search for something. Play with an idea that you have. Play a game with the AI. Try to explore how to create a collaborative environment and I think you'll be surprised.

Okay, wonderful. Yeah, certainly we lost the research around AI policy. I think there's so many surrendering things going on that area. But I also love your point about. Well, yeah, okay, AI, I mean, I guess it can be divisive, but it also bring people together as well. So yeah, let's hope we find some common ground in the future as well as like a good message.

You look at the trajectory of humanity, we have these like little ups and downs and big bumps and downs, right. But the general quality of living from someone from, from from now to even 40 years ago has grass. Right. We're in this diagonal sotal movement up and down, left and right. But we're an upward trajectory. And I have hope and I believe in humanity. I believe in democratic principles and I believe in collaboration. And AI is going to enable that might not be five years, but from now, it might be 10, might be 20 years from now. But AI is going to be central on that and don't lose hope on that and don't forget your humanity.

Wonderful. All right, great words. Thank you so much for your time, Russ.

Perfect. Thank you so much for having me.