AI Sovereignty & Cybersecurity with Dr. David Bray

Something happened. A funny thing happened.

Something happened. Are you able to share anything about—

Yeah, I'm keeping it at the high levels, but I think it would be safe to say I would say 2 events happen in the world that people can look at and presume there's other things happening too, which is one, we have seen an onrush of increasing cyberattacks and ransomware attacks impacting Japan. And it's unprecedented, the number. And we can actually then talk about what we've, what's happening this week as well. And then similarly, folks may be tracking open source report about France as well, that France has also had an onrush even to the level of refrigerators being compromised. Which causes the food in them to spoil and having to throw this out. And this is on top of the already existing problematic scenario that France is experiencing, which is they've got some riots and some protests over the budget and education and the like. So it seems like, and the reason why I got pulled in last week and why I apologize that I couldn't make the conversation we did have scheduled, it's very timely because it seems like now world governments around the world are having to figure out. Now that these AI tools are available and not just through the frontier models and the private companies, but increasingly open weight AI models can be used to do this as well. How do you deal with the onslaught of cybersecurity vulnerabilities and keep up the pace? Because the challenge is an attacker using these tools doesn't have to care about breaking things or if things don't work, whereas defenders, they have to both push back the attacker, But also keep the lights on, keeping things running. And it seems like at least for the next 2 to 3 years, things are going to get harder before they get better.

And then the other timely item, which most people here may have heard of, is the PHP vulnerability. There was, I believe, also a Node.js vulnerability yesterday that was announced yesterday that's been around for years, that's been exploited with a lot of people in IT and across the web scrambling for patches. Do you have any perspective, color on that outside of what we maybe are seeing in the news?

Yeah. So I think, so without saying the specific entity, I got pulled into some of these conversations in March when it was clear that the frontier labs that were doing AI had technologies that were essentially really good vulnerability scanners. And I got pulled into these conversations because I'm, again, I'm a nonpartisan. So Both says, yeah, I might ask my opinion some of the times and then shoot at me also. And so you just have to be okay with that. And the March conversations were what to do. And I think, again, like all things, some people were looking for some old playbooks that we know we either, we'd either make this limited access, possibly classified, which of course wouldn't work, especially for private companies, or somehow we thought we could contain it. And I was really puzzled by this because I was like, again, These tools are going to get out there, whether from the actual companies that are producing them and selling them. But we know also that there are increasingly numbers of what's called open weight AI, some of which that are legitimately produced, some of which are a result of what's called distilling. And so distilling is when an actor, either with permission or without permission, distills the key insights from a trained frontier model. But basically it is 95% as capable. But it needs much less compute. It could even run locally on a laptop or something like that as a result. And so distilled open weight models will be out there. And so anything that a government, whether it's the United States or others, attempted to do would miss the fact that one, we're not in a world right now where anything convinces me that the 193, 195 countries in our world will reach agreement on anything at the moment. But then 2, even if you somehow magically did get them to agree to everything, There's still the dark web. And we know already there's a thing called WormGPT. There's 3 different flavors of WormGPT, which I don't recommend you use, but it's basically the dark side cousin of ChatGPT, 3 different flavors with no filters. So the reality is that even back in March, it was clear we're going to have to prepare for a tsunami of which in the next 2 to 3 years, and you're seeing it right now in real time, we will know more about cyber vulnerabilities than ever before. And so on one hand, you might be, that's awful. But the other thing it just points out, those cyber vulnerabilities were already there and they might've been exploited by a few actors, possibly without visible awareness. Again, like you said, with the PHP vulnerability, it was known. It's just nobody knew how to construct a way to exploit it. And eventually we figured that out with the help of AI. But what this is really doing, in my opinion, is it's actually, and I'm using the word little d, democratizing or making available the access To discovering cyber vulnerabilities at machine speeds, essentially everybody, both good and bad actors. And so the consequence is really just make it so that we similarly can make it available for you to rapidly patch, rapidly fix, and also at the same time rapidly push back if and when a cyber attacker is using these vulnerabilities and you haven't fixed them yet. Now I'm actually seeing interesting enough, and this gets to questions of digital sovereignty. I was in Eastern Europe. In August, and I saw how a 6-person computer programming team using an open weight model basically every week are creating a whole new digital fabric. So it's not even just a product. They're actually saying, give me a tool that allows me to do geolocation, to do analyses, to do essentially defense functions. But they're rebuilding instead of building individual products, they're vibe coding an entire software fabric. And they're tailoring in features every week. So their iteration loop is they can customize it every week, but they're also changing the cybersecurity attack surface every week. So any known vulnerability or any hacker that knows a vulnerability only has a week window to actually use that exploit before the attack surface has changed. And so I would submit that over the next year or 2, we're going to see the very definition of what is software change, where it actually might be you're not buying individual products or services. If anything, you're getting a custom digital fabric unique to yourself. that's being updated every week or less.

It seems if these types of vulnerabilities become— we become aware of them more and more frequently, it seems like that may be a requirement. And it seems like there's maybe— there's some short-term pain here. Oh, yeah. But that maybe long-term, everyone is more secure, everyone's more aware, to your point, versus just a few bad actors. So maybe long-term, great, wonderful endpoint, but near-term, a lot of uncertainty and probably a lot of fear.

Yes. Again, if we look back at history, when automobiles first came out, one of the unforeseen things of automobiles coming out in the 1910s was you had interstate crime. Now people could commit a crime in a different state than they lived in. And so police did not know how to find the suspects. And the solution wasn't ban cars or prevent cars. It was actually, let's evolve better ways for local police, if they collect fingerprints, to be able to look for not just suspects in that local area, because again, interstate crime might be happening, Bonnie and Clyde, But they could actually go to a library, again, protecting privacy where possible, and be able to find the suspects. And so, like all things, the same thing with the radio. When the radio came out, actually, you can go back, 1930s, some people thought the radio was going to be the end of free societies and Western civilization as we knew it, because it was going to be the dictator's toolkit. Sure. So at least in the United States, that actually gave rise to a place where I did work, the FCC, which said, if you're going to be using radio, including amateur radio, You actually have to take a test, demonstrate you know how to use it expertly, but also promise not to use it for bad or exploitative means. And you have a license. And if anyone calls you on the radio and says, what's your call sign? You must identify yourself. Otherwise, the FCC Enforcement Bureau will come investigate. So oftentimes when new technologies come out, we have to, we have to advance broader social technical solutions to solve them. And that's part of what I find is currently missing a lot in this this AI debate is we have people, again, some very well-meaning, that come from AI companies or AI CEOs themselves, and they're reaching for technology solutions to the problems that are created, when in fact we know from history it's got to be a broader systemic whole-of-society solution, not just a technology fix.

Flowing naturally off of that, where are we with the FCC of AI and agents? I know there's a lot of scrambling, and I was originally going to talk about this later in the hour, but It's just such a relevant follow-up conversation to what you just said. Is what's being done right now from maybe a legal framework, or we hear about the heads of the big frontier models, you know, trying to get together or being frenemies, what have you. What's your take on where we're at with the FCC of AI?

Well, one, I'm glad you said an FCC of AI because I was actually at a Singularity University event. It was with Hod Lipson and others who also are with SU. This was, I think it was like 4 or 5 years ago even. We were talking about as AI comes, how are we going to deal with the fact that there's going to be, we didn't call them agents, but we called them different models that you could use. And at the time I said, remember, the beauty of what the FCC model was, was it allowed each nation to have licenses or to give operator permission to folks within their nation. It may very well be we need something similar for if an AI agent is on the web and it's doing something. First, it should be able to identify itself. I'm call sign whatever, and I have been given permission to operate by this human. I think the big challenge that we're really going to face with that is, at least in the United States, and I celebrate protecting privacy, we don't really have a good way to identify you consistently online. That isn't also at the same time eroding privacy. And so we're going to have to come up and maybe we will hopefully advance. I know there are solutions out there where, again, it can say that this is something that David has authorized or something that Cody has authorized, but it's not giving all your information about who Cody is, where Cody lives, and everything like that as well. But it does point to— again, I think there's some very real pragmatic solutions when it comes to the world in which there's going to be 10 times or more AI agents in the world. Which is really just 3 things. Every individual, every organization, every family, every country should be able to say, what are the agents that I give permission to operate to my digital persona, my data? And then similarly, what are the systems am I okay with them touching? I might sign up for an AI agent to help me with planning my retirement, or I want an AI agent to help me plan my calendar, but I'm not giving okay to either of those agents to access my bank account or my family photos or something like that. And then I should be able to stop an agent if I haven't given it permission to operate on my digital persona. So I should be able to block it. And there could be consequences. If I say no, then maybe I'm not getting in the fast line at the airport security desk. I've got to go through the, you know, the more detailed screening, but that's a choice that I've made. And, or if I see an AI agent that I said, okay, you're allowed to do something with my retirement planning, but not access my bank account. If it tries to access my bank account, I should be able to stop that too. And if we have those 3 things in place, and I think it's going to have to be separate from the companies, the companies can't do it themselves because it's just too much of a conflict of interest. Then individuals, companies, families, and countries alike will be able to still be free and be able to say explicitly, I gave permission to those agents and not these.

There's a comment in the chat, and my mind went to the same place. Is this a A use case for blockchain?

Yeah, it could be ledgers, distributed ledgers, definitely, because you want to be able to have provable proof that you did or did not give permission, which then easily solves the liability question. I know some people don't want to solve the liability question for certain reasons, but it does solve that if an AI agent does something to your person and you haven't authorized it to, then that would be a case for legal liability. Again, we have ways to hold whether it's governments or companies responsible if they do something to you that's not legal or something you did not give permission to.

That is, I think, a great framework for where we end up. What do you see as like, as where we're at now? This is an impossible question, but where we're at now, like, yeah, how do we get there? What's— and for the people listening, what do they do? You know, is it just be aware of what's happening? Is there actions people can take to help us get there? Yes, there's a lot of work to be done. Back it up to a high level around AI sovereignty, what this conversation is originally about. And we've been chatting essentially about sovereignty mixed with cybersecurity. But what's your definition of AI sovereignty?

Yes. So it's a good one because I like to remind people that the original definition of sovereignty was absolute law by absolute rule by king or queen. And I don't think we're talking about that here. But I think it has come to mean, how do I continue to do my work Whether I'm a company, a country, or an individual, even if I have possible dependencies on someone else. And what this points to is, again, if you go back to the events as just recent as this spring, um, when one or more AI companies decided to pick some fights with the U.S. government, that had the unintended consequence of making every nation around the world say, wait, wait, wait, if a U.S. AI company thinks that they can fight with the US government over what can and cannot be done with their product, what's to prevent them from thinking they're going to pick a fight with me? And so every country around the world was pausing and saying, how do I continue to do the rule of law in an AI era when I might be dependent on a foreign company or a foreign country to help do my AI? And the reality is, again, 193, 195 different countries around the world. Most of them will probably not produce their own frontier models, but hopefully they could form coalitions regionally based on language or shared interests to train open weight AI models that they then could use for their own internal purposes and not have that same external dependency. Now, clearly I don't work for any of the frontier AI model companies because me saying that, I'm actually undercutting their value proposition, which is open weight model is actually going to be much larger than paid frontier models. But again, just if you look at the realpolitik of the moment, no country wants to have so much dependency on a foreign company or a foreign country that they cannot do the rule of law. And that extends then further to any company that's not an AI company. You know, it's AI, it's not like cloud where it's just storage of your information. I mean, this is actually getting into your decision loops. This is getting into your intellectual property. This is getting into the secret sauce that makes your business your business. So again, if we define trust as the willingness to be vulnerable to the actions of an actor you cannot directly control, and that could be vulnerable to an actor that's a company or a country or an individual or even an AI, I think most companies and countries don't want to be that vulnerable to be so dependent on foreign companies and foreign countries. And so that's why digital sovereignty is coming up. The other thing I would say is it's also a watershed moment where it's been very clear that the speed at which free societies operate in terms of their deliberative processes to update their laws is woefully insufficient for the technology era we're going. And I'm not saying become an autocracy. Definitely not. I've spent too much. I mean, I believe in free society and free markets. That's why I do what I do. So we've got to figure out how do we operate at machine speed? And still do all the things we cherish in terms of civil societies, privacy, and the like. And at a minimum, when I'm here briefing folks in DC, I say, if you think you can spend 12 to 18 months crafting the perfect policy and then roll it out and have it be relevant for 10 years, I've got a bridge to sell you. Nowadays, I think you have at most 3 to 4 weeks to get a notional policy on the table, and it's probably going to only last about 6 to 9 months before it's got to be updated. So it's really about how do you do policy and programs in parallel, which most elected officials, not to mention their staffers, have no clue how to do. So again, it's going to be, it's going to be messier before it gets better. But this is again an opportunity for Singularity and others to step forward and say, it's not just about the tech. It's about everything else that goes around it in terms of companies, countries, and societies.

Uh, in a recent interview that you did with us for our alumni newsletter, You said that within 2 years there will be 10 times as many agents on the planet as people.

Yes.

Can you explain your reasoning behind that and then the implications of that and what that means for software people?

Sure. So in terms of, there are certain groups that track network traffic that we believe to be associated with AI agents. And in just the last 6 months of 2027 alone, that traffic has grown more than 80x. So, so if anything, a 10x guess in 2 years might be conservative. So network traffic is growing in terms of the the presence of agents. The reality was actually 2013 was already the year where the internet saw more machine-to-machine traffic than humans. So we humans have been the minority in terms of machine traffic since 2013. 2013, interestingly enough, by coincidence, was also the same year that there was about 7.1 billion human beings on the planet and 7.1 billion network devices. Now in 2026, we're up to 8.2 billion human beings. Wow. And 55 or more billion network devices on the planet, not all connected to the internet. Several are air-gapped. And so it is— there's enough evidence with the growing network traffic in just 6 months going 80x. Again, there's just so many network devices and the like. The— you think about how many people, how many apps people have on their phone. You know, it is, it is a very conservative guess, actually, even though I know it's shocking to people to say in 2 years or less, 10x the number of agents relative to people on the planet. But it's coming. And a lot of them will be behind the scenes. The question I have is then how do you give informed consent and permission to those agents that are behind the scenes to make sure that they're not disadvantaging you or using your data in ways that you don't really want to have happen?

So you've described sovereignty as having 3 separate layers. Can you explain those layers? And then I'd also like to follow that up after that with a, you know, what's the most relevant or important for people to pay attention to right now as far as which layer to look at?

Sure. So if you think about it, AI doesn't operate in a vacuum. So the first layer that we really have to think about is the hardware and power electricity level. And again, wherever possible, most countries want to at least reduce their external dependencies on electricity production from other nations. Several in Europe are learning this lesson fairly painfully because over the last 10 years, some of them got dependent on a country that is now being very aggressive towards Ukraine that will remain nameless. But they got dependent on that country for electricity. And then obviously, when the relationship went sour, that's a problem, but also on the hardware. And on the hardware, not to make people have anything more to worry about in our era of interesting change, but 4 years ago, I got asked to help out with an effort where the Department of Defense, the US Department of Defense, had 2 routers that had been operating for more than 10 years in underwater nuclear environments operated by the US government. And the routers occasionally ran hot. They were sent back to the vendors twice. The vendor said everything's okay, but Department of Defense was still concerned. So they had another company come in that had won a competition for what would be called non-destructive hardware verification. So making sure the hardware is only doing what it claims to be and nothing else. So they ran a scan and they came back, we're not getting a good score. Do we have permission to open up the box? Department of Defense said yes at the time. So they opened up the box and less than 4 hours later, they found foreign, I won't name the company, but it was foreign Daughterboards the size of my thumb, so very small, soldered onto the router that were attempting to phone home whenever the router booted up and basically get administrative access. And again, because it's a router, you're not going to see it because that's the very thing you trust to tell you everything's all good and nothing else is going wrong. So it was, it was fairly sophisticated. And you might say, well, that's just a one-off, or they were just targeting the military. Well, no, there was actually then a scan of a warehouse that was supposedly from US-provided equipment. The failure rate was 53%. Wow. So, and that was 4 years ago. More recently, you can see there's stories in the news, Open News and Press and the like, that says like the UK has some of their underwater drones apparently phoning home again to that large country that will remain nameless. I'm sure they were just backing up the data as a service and it was a benevolent, benign intent. But I put that out there because nations around the world, not only do they have to reduce their vulnerability to energy and electricity sources from other nations, But they also now have to worry about if you're not producing the chips yourself, and we know most countries around the world don't produce the chips themselves, including the United States, how do you reduce your vulnerabilities if those chips might actually, or that equipment might be attempting to phone home? So that's the first layer. And then once you build up on top of that, the next layer is data. And I think data, especially in our AI era, gets forgotten, or at least not talked about enough, in that I would love to unlearn the meme from several years ago that data's the new oil, because I never knew how that lasted more than 5 minutes. You know, oil, use it up, it's gone. Data, if you involve people in it, one, they make better data, they fix the data, they actually create more data, better data. And so I think what we really need to think about is data's a form of human voice, whether you're a customer or a citizen. And that data layer is important because if that data layer is compromised, or people don't trust what you're doing with that data layer, if they think you're doing surveillance, then the work of a country or company is undermined. And so that's the second layer. And then finally, the layer on top of that is the actual AI model itself, which again is who's providing it? How do you actually reduce your dependencies if they decide to pull it back or not allow you to use it? But I think it's important to think about that data as being the middle layer, AI being the top layer, but most importantly, how do you address even the hardware and the energy? Because I feel like right now we might in the United States be building a lot of data centers, And have we actually verified that they're not phoning home?

Right. So for people here who, you know, run companies, or maybe are in public sector, lead a department at any level of government, what's the most prescient thing for them to be aware of? And or what can, what can people actually do other than kind of just be like, oh no, this, this is a problem?

Yeah. So the first step is there are pragmatic things you can do in here and now. And so wherever you are, if you're in a government agency, if you are in a private sector capacity, if you're at the local government, first step is reach out to your stakeholders, constituents, or customers and say, look, this is— we know our old playbooks won't work. We want to at least have an ongoing conversation with you about your top concerns, top priorities. Again, we may have a limited budget, we may have fixed budget, but engage a conversation that will get you so much goodwill as opposed to saying there's nothing I can do. I'm going to hide myself underneath my desk and not surface. So in this case, a lack of interaction itself is losing. And if you at least say to people, look, we know our existing playbooks won't work, then invite them to share their thoughts and say also at the same time, we will pressure test our ideas that are unique to whatever we're doing too. And again, we may have a limited budget, we may have limited capacity and the like. And so what you're doing is you're— it's been statistically shown If you increase the collective intelligence of the problem solving, you will actually find solutions that are better off than if you're narrowly constrained. And it's actually worth knowing Herb Simon, who was a forefather of AI back in the 1950s and 1960s. He won a Nobel Prize actually, but for economic work, because Nobel Prize does not go out for AI. He actually got his PhD looking at decision-making in New York State. And actually what got him the Nobel Prize was he proved that generally people make decisions that are within their comfort zone. They don't step outside of their comfort zone. And so I would invite all the people that are listening to this call right now to like cross your arms. And then if you try and cross your arms the other way, I guarantee you that's gonna feel a little uncomfortable, but that's just a small thing. And so if you can show some empathy for your workforce, if you can show your empathy for your community and your citizens, if you can show some empathy for your customers that we're asking them to do uncomfortable things and then invite them to be part of the solution. So they're problem solvers as opposed to problem holders. That is the way forward when there's no clear textbook on how best to solve this.

Amazing. Well, we're getting a lot of great audience questions, so we're going to go ahead and switch over to those for a while. And I assume that this will take us the rest of the hour. If not, I have more questions myself, but—

I've seen a lot of threads coming in, so I look forward to the questions.

So the first one from Fernando, Fernando Vera. What competencies will people need to critically assess, challenge, and collaborate with autonomous AI agents?

Love it. So one of the best things I have seen in the competencies and education space is the idea that it's not no AI, because that would clearly not work, but it's the idea that you need to have the ability to wrestle with and engage these autonomous AI agents. And what do I mean by that? Well, when you go to school, your initial years of schooling is you're just learning facts after facts. But at a certain point in time, hopefully you get to the point where you can actually challenge the teacher. And we all know those classes where we actually are challenging the teachers being the best classes, because as you go into the workforce, hopefully you're not getting hired just to repeat facts back, because if that's all you're doing, done, then I guarantee you search engines, let alone AI, will beat humans at that game. But if you can instead push back and say, why are you giving me that answer? And you recognize the different AI tools you're using. If you're using generative AI, you need to be a little bit more circumspect, because guess what? Sometimes generative AI is spot on, and sometimes generative AI, as we all know, will go off the rails. And so knowing how to push back, knowing how to better interrogate the machine— I'm not saying prompt engineering is the silver bullet, it's not— but it's being able to say, why do you think that? And then be able to provide data that can actually guide it to a better place. And so seeing it more as an interactive colleague as opposed to a replacement, I think, is essential. I think it's then also the idea of how you do that collaboratively. So it's not just you and the machine, but it's the relationship you have across the organization. As to how you're interacting with the different agents. And again, if we're going to a world in 2 years or less where there's 10x agents, how do you pitch that agent with that agent? And then Cody, you and I might be working on a book or something like that. How do we both engage together with ourselves and our agents? So, so the good news is, if anything, I'm actually thinking while there may be a displacement of jobs that were more rote, repetitive, just simply fact-based work, if anything, I'm going to see there's going to be an increase in jobs that talk about machine integrity. Machine rhetoric, machine interrogation. We know when computers first came out, they thought that was going to actually reduce the workforce. And if anything, let's look at what happened with computers. There's more people in the workforce. So I would say if there's anything I recommend is learn the art of how do you push back and how do you wrestle with the machine? And then similarly, when do you know the strengths of different AI approaches, be they generative, neurosymbolic, predictive, computer vision, or the like?

This next one from Daniel. Do you think we'll start seeing only human or human-only commerce networks, things like that?

So just like how records have come back in our digital era, I think people may pay a premium for things that are attested to be only human-made, that this was knitted by a human or this was painted by a human. But at the same time, I can remember, well, I wasn't around at the time, but We know from when the printing press first came out, there were several scholars, including religious scholars, who said, who would ever want to read a book? We all know that hand-printed books are the only source of knowledge. Printed books are slovenly. They were actually calling them really weird, rude names or calling them, you know, books were equivalent to apparently prostitutes and the like. Like they were really saying books are low. No one's ever going to use a book. And we know what eventually happened. So while I do think there may be some premium paid for human-produced content, just like how we pay for art because art is in the eye of the beholder, I do think it's, it's going to be that we're also going to be okay with AI-produced content as well. Now, that's not saying embrace AI slop. I don't think anyone's saying that. And I also think we're going to have to have ways of helping artists, musicians, and the like find revenue. If anything, I've been saying for quite a while, I feel like we're actually repeating Napster. And in the sense that we're not respecting IP, we're not honoring what the artist or the musician produces. And I think there's 2 ways to look at that is that when Napster first came out, yeah, IP was disregarded. And then a few years later came iTunes, which does compensate the artists, but the artists still had to change their model to more live performances. And so something similar to that will probably happen where we need to actually have a legitimate model that does compensate the musician or artist for whatever they've contributed to the data commons or whatever was used. May point to more live shows or live interactions.

Kind of a bit of a probing question from Robert. Why do we assume countries can work together for cybersecurity unlike everything else? And then that doesn't even take into account disagreement within countries.

Oh, I agree with you 100%. I'm not assuming they're not gonna work on anything. If anything, I would say they're proof that over the last 30 years, countries have Possibly gotten together and nodded their heads and we'll work on cybersecurity and then gone back to their corners and continued to not work with each other. But we also know from 1899, all the countries of Europe got together and said chemical weapons, really bad things. We should all ban them. So they all agreed to ban or pause chemical weapons. And less than a year later, they were all pursuing chemical weapons in secret such that when World War I broke out in the first 6 months, both the Central and the Allied powers used chemical weapons against themselves, the very thing they supposedly had banned. back in 1899. So that's why I say I celebrate AI technologists, but don't let them try geopolitics because that might not be their forte. And so whatever we do, we need to recognize the realpolitik that at the moment, you know, there are, there are nations that will be friendly to themselves, to each other, but they're still going to pursue their own interests. Now, that does point to what is an interesting tension here, which is even before AI, we have the internet, which now has, you know, more than half the planet now has access to the internet. We have smartphones, which I would submit to you, your smartphone does things that President Reagan, US President Bush 43, President Clinton, and US President Bush 41 would have loved to have had in the Situation Room of the White House. I mean, you can call anybody at a moment's notice, hopefully with their permission. You can use an AirTag or RFID tag to track anybody in the world or their assets. Again, hopefully with their permission. And you can use commercial space apps to get satellite footage as recent as 15 minutes ago. at 0.25-meter resolution. And now when you bring in AI, we are giving you the power and capabilities of the CIA and the KGB in the early 1980s. And we're giving it to you for $100 a month or less. So this raises really interesting questions, which is organizing by geography, does that still make sense? Or is something else going to take its place? Is it networks, networks of shared affiliation, shared beliefs? And what does that look like? The fact that Europe in the last decade passed GDPR, their Privacy Act, and said you must treat someone from Europe, even if they're outside of Europe, consistent with European privacy laws. That's extraterritorialism. California has done the same thing. If you're a citizen of California and you're outside of California, for AI purposes, we must treat you as if you're in California, even if you're in New York or London or wherever. So what we may be seeing is internet, smartphones, and now AI It may be more bubbles of who do we decide to align ourselves with versus national borders. But the challenge is, is how do we navigate that messiness given that our government officials are used to operating with borders? And if anything, the technology might be saying, I see your borders and I don't care.

This next question from Andre is, has a good amount of context behind it. So how do non-frontier AI countries ever catch up or close that gap The context being more than 150 countries have 100% dependency on hardware and models in AI. And then, you know, in the next 5 years, advanced compute may be starting to get placed in outer space, humanoid robotics being manufactured in a couple of specific countries. How do you see the gaps of the EU, US, and China growing over the next decade in those capabilities?

Yep. You got to identify your friends. I would say to start with, and what do I mean by that? So I have been deeply involved over the last 9 to 10 months with Europe. I was recently named a distinguished fellow with the German Marshall Fund to the United States. And as a bit of a history, obviously after World War II, the US did the Marshall Plan to help rebuild Europe. And then as a thank you gift, Germany created the Marshall Fund to the United States as a thank you gift. And So Europe is obviously trying to figure out, but it's not just Europe, we can talk about the Global South. They all are trying to figure out how do they play the game. And the first thing I would say is don't play the game the same way the United States and China did. That's a recipe for— don't is all I can say. You cannot out-US the US and you cannot out-China to China. But what you can do is say, okay, in Europe's case, you do have this ethos of that, you know, Post-World War II, we'll see what happens in another 10 to 15 years, but post-World War II, Europe wants to work together. They believe in the commons. So again, if I was Europe and I was recommending the EU, I would say train an open weight model and you'll make money off of consulting off the model that knows your languages better than anybody else. And that's the open weight model you use for consulting. That's the model you use for your robots. And so sure, someone else might produce a robot, but you can say you now need to use our model because it's gonna speak our language, it's gonna know our culture, it's gonna know our context better than anybody else. And the same thing for the Global South. It's a little bit more difficult because of the geography of the Global South. And again, as we already said, you can't trust the chips. So you're going to have to do the, what I would call the reverse of what Reagan said, which is verify and then trust. There are tools and tactics that can come in to actually check the chips. And if anyone ships you a chip that's compromised, first time you send it back to them quietly and say, try again. Maybe the second or third time you get a little bit more concerned. And by the fourth time you actually call them out publicly and say, You're shipping me compromised chips and you use shame and blame. Same thing for actually trying to figure out how to reduce your energy dependencies as well. But again, it's, it's, it's, I would say for countries that are not China or United States, don't try to play the same game. If anything, figure out how you can work across borders and maybe that very pressure of we have this shared challenge because these 2 countries are racing ahead in ways that we can't will actually serve as a good forcing function to encourage your national and your local leaders to work across borders because they have that shared challenge together.

I think we have time for maybe 2 more questions here. One from Roman. How realistic do you see decentralized open-source AI infrastructure as an alternative to centralized frontier models, strict state-level chip regulation? Yeah, what's your, what's your take on decentralized versus centralized?

I am very much pro-decentralized. And this is again why I'm probably not popular with certain companies, but that's okay. I'm fine with that. I think the only way free societies and free markets remain free is if you have the option of being able to run things locally. Doesn't mean we all will. I mean, some people will say, look, for my mother or for, you know, my family, I'm going to actually trust this company to do it for me, but I should have the ability to run it locally if I want. And then I can go to a company if I want that, I can scale up if I want. And maybe sometimes I do want to reach for that really top-notch private frontier model. But I think if, you know, in some respects, I know people say the AI era we're in feels like the dot-coms, but I would submit it's the complete opposite. You know, when the early internet was rolling out, if you wanted to, you could launch your own Apache server or you could hire someone to do it for you and you could write code on it and you could set, I mean, for those who are around GeoCities or whatever you wanted to, you could send up your own things. And so it was much more decentralized. than what we experienced in the last 4 years, which was very capital intensive. You couldn't do it yourself. And so I hope we actually do find a way, whether it's, I mean, again, I think the US is kind of blinded at the moment because it's pursuing a certain model, but maybe it is Europe, maybe it is the Global South actually leapfrogging and saying, we will avoid that because that didn't really work out that well and showing a more decentralized model. And then the US will also join the party. After that.

The last question here from Stan, I think, is, is very, just a great way to wrap us up. So what are we not paying attention to that we should? And what questions should we be asking that keep falling through the cracks that you may be aware of? Because consensus doesn't always mean it's the best idea.

Yes. So the one thing we haven't talked about, and I would be happy at a later date to circle back, and I saw someone in chat say this, our e-encryption regimes are about 10 to 20 to 30 years old. Um, I actually once saw, this was just a week or so ago, I saw a demonstration that scanned the certificates you have on your local computer to give permission for things to happen on your computer. And there were certificates as old as 1999, if not the 2000s. And we're talking, you know, only 512 bits of key length, because that was good enough. And that's because the software manufacturers use old expired keys. And so that's something that impacts us locally, but it also impacts servers and the like. And so even before we get to quantum computing and whether or not we're going to have quantum decryption, I think we will, the existing key lengths that we have are insufficient. And now that we've unleashed AI in the world that can do vulnerability scanning, I would not be surprised if another wave of AI vulnerability scanning points out how bad some of the encryption keys that we are currently running on our software are. Hmm. and how exploitable they are. Not that, again, even before you get to quantum, if an AI knows you and knows, say, David Bray well and says, well, David Bray, he speaks English and like, I'm going to narrow my attack surface to find his password, they can actually reduce the time-space to find that password. So I think the thing we're not talking about enough is, um, we may discover, in addition to the vulnerability wave that's happening, uh, we may also be facing challenges with encryption. And even before Q-Day, before quantum computing comes along, there may need to be a way to upgrade using software, our encryption stack locally, globally, and the like. That would be the conversation I think we need to have.

Amazing. Any other final thoughts or tips or notes for anyone? And then also, is there a best way to follow you or learn from you on an ongoing basis?

Yep. So I would say lean into the moment, be a good Stoic, don't freak out. Imagine you were in London in 1939, you know, whatever it is. I mean, I know a lot of people, you know, fear is a mind killer. Instead, lean in and be bold, be brave, and be benevolent. And to find me, uh, best way is to find me on LinkedIn. Please, if you do ping me on LinkedIn, please say pineapple in whatever you message me. That means you at least watched this video and got to this point. Uh, 'cause even in '26, I get a lot of bots on LinkedIn. So if you use the word pineapple, I'll know you at least watched this.

Nice. Passwords. Love it. We will be sending out the recording of this, a recap, the transcript. I'll also include a link to David's LinkedIn and any other resources that we covered throughout the talk. So look out for that email within the next day or so. And just want to say thank you to everyone for being here. Thank you for all the amazing questions. I think we only got through about half of my questions that I had because the audience is on it. And obviously this is a very relevant topic. So Thank you, David, for being here. Thanks for the work that you're doing right now. I think it's very critical in the world. And for, yeah, for being willing to share your experience and expertise with us.

Thank you, Cody. And thank you, Venus.

All right. Thanks all. We'll see you next month. More to come on our next discussion series. And yeah, onward. Hope you have a great day or night.