Human Progress

Chelsea Follett: Joining me on the Human Progress podcast today is Kristian Stout. He serves as Director of Innovation Policy at the International Center for Law and Economics. His work addresses competition, telecommunications, and artificial intelligence policy—very big topic these days—with a focus on how law and economics informs regulatory design. Alongside his ICLE role, Stout has been a partner and counsel at A & S Technologies. So he’s a lawyer, software developer. He has seen all sides of this industry. He joins the podcast today to talk about AI and some recent developments. Thank you for joining me, Kristian. How are you?

Kristian Stout: I’m very good. Thank you. Thank you for having me.

Chelsea Follett: Thank you for joining me. So let’s start with a big-picture overview. AI feels like it is a very new topic almost because we’re always seeing headlines about new advances. But how new is AI? Can you sketch out sort of the big picture of the history of AI and data centers for us before we dive into the current debates?

Kristian Stout: Sure. Yeah, it’s understandable why it feels like something very new because when the computer starts talking to you, that’s unusual. But this is part of a trend in computer science that goes back to the 1940s, really. So for many decades, researchers have been working on both sides of this coin. One, getting ubiquitous computing available throughout the United States. And then, on the other hand, how to make computers smarter, frankly. So as far back as the ’40s, there were people, researchers, trying to understand how you could model, for instance, human neurons inside of computers. Now, it was just one neuron, and they were just trying to understand what that could do for computers. But that’s what we’ve been thinking about since then. That’s what our computer scientists have been thinking about. So I think it’s important to understand that, you know, as far back as, say, that era, 1940s, 1950s, even though those computers were nowhere near as capable in so many dimensions as they are now, we’ve been framing the whole project of computing as trying to introduce more intelligence into it, more sophistication. And what we have today is the result of compounding developments over those many decades.

Kristian Stout: So none of it is really surprising from a research perspective. There have been surprises. The surprise is not that computers are capable of doing these things. It’s how we’ve managed to get them to the point where they can do these things. So, you know, and frankly, we’ve been living with the fruits of more intelligent computing over that time just sort of silently and under the covers for most consumer uses, even though researchers have been trying to push the boundaries of this stuff. So, you know, as far back as—well, I mean, you can really say in the ’70s you started getting some more of these sophisticated algorithms being deployed. But the ones that consumers are most familiar with, and they just don’t even think of as artificial intelligence, are things like autocomplete. So when you were able to have really advanced statistical methods for having very responsive searches for users, first in the desktops, say Google, Yahoo, those kinds of search engines, and then all throughout, like, your devices, like phones and so forth, where you start typing and it figures out what you might be wanting to type or what you might be wanting to search for, like, those are all forms of artificial intelligence.

Kristian Stout: And as recently as, say, 12 years ago, we started really focusing on machine learning. For a long time, the news was about this sort of big data, is what we were calling it before, again, the computer started talking back to us. Big data was all the rage. All of the investment money was going to big data, was going into these different machine learning techniques that would be useful for forecasting all kinds of behaviors and making the computers smarter at what they’re doing. That was artificial intelligence. Nobody was writing news articles about it as artificial intelligence. People weren’t afraid of it as artificial intelligence. There were still anxieties around it, for instance, like what that augmentation meant for labor, meant for the future of work, things like that. But we didn’t frame it in terms of, like, Skynet taking over. Now it becomes more popular to think about it that way because, again, you can interact with the computers in a way that feels very human-like. But none of it is, like, new in a kind sort of way. It’s new in a scale sort of way, is a better way to think about it.

Kristian Stout: So we’ve finally gotten to the point where we have them doing the things we’ve been wanting them to do, we’ve been trying to get them to do for 70 or 80 years at this point. But that’s not new. The same thing with the data centers. Data centers have been with us since, I believe, the 1950s is when you can really start to source that. Again, very small scale. But IBM was very—what made IBM so successful was the development of the mainframe model in which you would have—sometimes you’d have it on premises, networked inside of your research environment. But sometimes, you know, the early internet, for instance, is where this came from. You would have a mainframe that researchers would share over long distances with phone lines. It was very slow, but that was an early form of what you can think of as a data center. It was like—there was like the kernel of what became a data center. And then over that time, we’ve come up with more and more complicated ways to get really advanced compute into areas that then you could share, you do time-sharing, so forth.

Kristian Stout: And, you know, the first data centers that we think of now as the cloud really started, say, around 2005, 2006, when the scale of the internet really was taking off and we needed a lot more computing support. So what we’re dealing with right now is, again, a scale difference, not a kind difference. We’ve been trying to do this since the 1950s.

Chelsea Follett: And that scale difference has brought with it some interesting political debates. If you look at public opinion polling, the public’s view of data centers has seemingly soured even as people are using AI more and more. And so we have seen more and more political moves to ban or restrict the construction of data centers. What is your take on this development?

Kristian Stout: I try to have a nuanced and generous take on it because I think there are multiple things that are going into it. First, I think all of your viewers are going to be familiar with the idea that humans often do not express exactly what they’re saying, what they’re thinking when they respond to public opinion polls, right? So you will get a different result if you say something like, “What do you think about data centers?” versus “How do you feel about an industrial project in your area that might bring jobs and might have these costs and benefits?” You know, you’re going to get different answers to that. It’s the same thing with AI. You know, Dario goes out and says something about how it’s going to kill us in 10 years, he thinks. And then, so then you go and poll people, say, “What do you think about AI?” Like, “I would like it not to kill me in 10 years.” But then if you then had a poll that said something like, “How would you feel about a computer technology that enables you to do all of your week’s work in two days, and then you have like two days extra of leisure?”

Kristian Stout: And they’d be like, “Oh, well, that sounds kind of good. I like that.” So, you know, part of it is a categorical thinking problem. We try to put—you know, this is something that enables us to survive as human beings. We find ways to look at the natural environment, and we create groups of things that allow us to do very fast decision-making. That decision-making is often not complete enough for us to actually answer complicated modern questions. So my generous take is I think there’s anxiety that leads people, when they’re given generic questions, to sort of respond and say, like, “I don’t like that thing.” I think most people, if you actually drill down into the questions with them, they have a more nuanced response to it. So that’s just like a framing point that I always try to keep in mind, and I try to advocate that other people keep in mind. In terms of the actual anxieties that I think people are experiencing and how that feeds into what we’re doing with AI and data centers right now, I think that there is an understandable concern among the general population when they interact with these technologies to think, like, “What does this mean for the future of my job?

Kristian Stout: And what my kid—like, what are my kids going to do when they’re out of high school?” Right? So, like, you have kids now, you’re putting them in middle school or grade school, you’re like, “How do I even advise them on, like, what the future—how are they going to have kids of their own? How are they going to have families? What are they going to do?” Right? So, like, I understand that anxiety floating around there because the way that we’re thinking about what AI might be able to help us do in the future will significantly transform the economy. I personally, from the research I look at and just from interacting with these technologies, I don’t think it’s going to cause mass layoffs. I think, in fact, it will be a bit of a genie’s wish granted, where it gives us all these awesome new powers, but then it will have costs, meaning that we’re going to be expected to be even more productive. So I think employment will go up. What will happen is we’re going to invent new forms of consumption that then we need to satisfy.

Kristian Stout: So I don’t think anybody’s going to actually get a vacation, is the short answer of it. But this is not a new public response, right? Like the whole idea of sabotage, right? The word sabotage was from people taking their wooden shoes, the sabots, and throwing them into the machines that were replacing the weavers back in—I forget what year it was—like 1400 or something, right? So, like, people have been mad about machines taking their jobs for hundreds and hundreds of years, and it’s not unreasonable. Like, I don’t want to downplay the anxiety, right? So, like, you know, everybody likes to talk about the buggy whip manufacturers when Henry Ford came around and say, you know, to kind of make fun of the buggy whip manufacturers, right? But I don’t want to make fun of them. I want to understand, and I think we should be understanding of that anxiety. So if I went to a buggy whip manufacturer in 1900, 1910 or whatever, whenever the Model T was, like, coming off the assembly line, and they were, you know, and I said, “Look, these cars are going to transform the world.

Kristian Stout: These motors, like what we’re doing with transportation, are going to transform the world.” And these people said, “Well, what is the world going to look like?” Like, in 1910, there’s no way I could tell them what 1990 would look like. It just wouldn’t make any sense. Planes, rocket ships, cars, like all of these things, they build on each other. All I could say to them at that time is like, “Look, it’s going to be better. I want to, you know, personally, I want to help you figure out how to transform into this new economy. You’re not going to be completely unemployed. You’re not going to make buggy whips anymore, though.” But they want to know, like, “Well, what is the buggy whip of 1990?” There is no buggy whip in 1990. It’s a completely different world. It’s a better world. It gives more people more opportunities and more access to things. But there’s no way to communicate that adequately to people who have invested themselves in a certain thing. I don’t think there was mass starvation for the people who made buggies and whips and so forth.

Kristian Stout: They all figured out how to get involved in the new economy. And I think we’re sort of at a similar kind of precipice where AI has got a lot of promise to do a lot of interesting things, but it’s what researchers call a general-purpose technology. It’s like electricity. It’s a platform that people are going to develop new ways of doing things with. It’s intelligence. It’s commodified intelligence that we’re now going to deploy in the same way that people, when they started standing up the electric industry, were able to create a field in which people would now use electricity to do all kinds of things. If you went back to Thomas Edison and said, “Tell me what you think people are going to do with electricity,” that dude had no idea what the world would look like right now. There’s no way you could know that. What it did was it enabled really smart people to take this new source of power and this new source of driving innovation and actually develop things that people wanted. We don’t know what it is until the market discovers it, unfortunately.

Kristian Stout: So the anxiety is understandable. What I think we need to do when we’re looking at this is be solicitous of people who have that anxiety and figure out—you know, one of the nice things about this particular technological revolution is it helps us have cognitive adjuncts. So the technology itself can help us figure out how to onboard more people more quickly and more seamlessly into this new intelligence revolution in a way that previous technological revolutions couldn’t. But I think we do need to be patient, and we do need to understand that there is this anxiety, and it’s not going to go away overnight. And we also should not just brush it under the carpet and be like, “Oh, look, you’re just being a Luddite.” I don’t think that’s a fair way to treat the situation.

Chelsea Follett: That’s great historical context, and I love how you brought attention to the world-transforming, society-transforming potential of

Kristian Stout: Yeah.

Chelsea Follett: this technology, because there is so much focus on the risks, and you don’t hear as much talk about the benefits. Even as many people compare it to the Industrial Revolution, you don’t hear people sketching out the potential positives of a world post this AI revolution as

Kristian Stout: Yep.

Chelsea Follett: often. And you said something very interesting that I would love for you to elaborate on. We hear so many fears related to technological unemployment, but you said that you actually believe with AI, employment will go up. That is the opposite, the exact opposite of the typical narrative that we hear. Can you elaborate on that?

Kristian Stout: Yeah. So, I mean, this comes from—I, full disclosure, my predilection is to look at things through an economic analysis lens. So in economics, people think about it as—I was actually just having a conversation with a friend this morning about it, and I was talking about economics. Like, “Oh, well, that’s just about, like, money and profits.” Like, no, that is not what it is about. We use money and studying the economy as a proxy for understanding human motivations, human desires and human action, right? All economics is is the study of human action, really, at root, trying to understand why people make the decisions they do, what are the things that drive them to make those decisions, and then how to understand the trade-offs among those things. If you really take human nature seriously—this is both a good thing and a bad thing in a way—we have an unlimited amount of desire. The reason why we are where we are is because there was once upon a time a caveman who’s like, “You know what? I think I can make something that makes me kill that animal a little bit better.”

Kristian Stout: That process has not stopped for however many hundreds of thousands of years humans have been looking at their environment and being like, “I think I can do a little bit better.” Part of that “I think I can do a little bit better” is, “I would like to relax more, and I would like to eat more, and I would like to have more mating opportunities. I would like to have more of everything,” right? We have an unlimited amount of desires. Now, there are tons of spiritual traditions around the world that try to help us learn how to tame that and be at peace. So far, they have not been working very well. What we have seen instead is that when a new opportunity presents itself, our desire fills the volume that you provide it. So if you have a technology that comes along and says, “Hey, now you can get your workweek done in a day,” I don’t think people are going to stop working at one day. I think—because the other thing to keep in mind, too, is that we’re status creatures. We’re status-driven creatures.

Kristian Stout: The reason why we value money, the reason why we value resources is not, for most people, not for the resources themselves, the money in itself. It’s because of the status that it provides us relative to other people. All of our preferences in the world are based on comparing ourselves to the Joneses, right? Like, that’s the reason why this theme keeps coming up in literature for the millennia that we have literature. You know, I don’t think it’s as base as—I don’t want to impugn it as saying it’s something like envy, but, like, that inherits from this drive that we have to compare ourselves to others and judge our value based on others, which means that there’s never enough. If someone else is perceived as having more status, we’re always going to want to increase our own status. Now, if the world now suddenly expands the volume of status opportunities, our desire for status does not go away, right? Like, that’ll feed back down. Now, the economy might look different. It might be that just obtaining the best car is no longer sufficient. It might be—it’s the same way that…

Kristian Stout: So you can almost say that modern wealthy people in the United States compete on their investment in their children, right? When I was a kid in the ’80s and ’90s, no one went to anybody’s practices, right? Like, you would get dropped off. Now they’re almost like this social event where people vie to be the coaches on the teams and who’s the most involved in all the things at school. Like, you know, my parents did not do that when I was a kid, and most people’s parents did not do that, because the field about what counts and how we invest our time and our resources changed, and now we’re investing it differently. AI is going to give us the same thing. It’s going to expand our capability to get more work done more easily, but the things that we value are going to expand, and it’s going to give us more opportunity. So we’re going to fill that time with more value-seeking opportunity. And it might not look like sitting at a desk for 40 hours a week, but it will still exist, and there will still be employment and resource opportunities within that, because humans like to hire other humans to do things.

Kristian Stout: They like to buy things from each other. They like to trade with each other, and we’re going to fill that gap. The gap, it’s not like it goes away. The computer is not going to be consuming for us.

Chelsea Follett: Absolutely. And one of the jobs that people predicted would be among the first to be automated would be radiologists.

Kristian Stout: Radiologists, yeah. There are more of them. Yeah, yeah. Yeah.

Chelsea Follett: Yes. And what we’ve seen there, interestingly, is the employment actually has gone up for that field.

Kristian Stout: And the demand—they don’t—we don’t have enough radiologists, and the radiologists are one of the most augmented by artificial—early, really augmented by artificial intelligence resources. And if you think about it, it makes sense. Like, the human body is a complicated thing. Even with the volume of scanning that’s capable right now, we want more health. We’re never at a point where we’re like, “You know what? I’ve got enough health. The doctors did good enough in 2017. Let’s just stop there.” There’s always more margin that you can get at. Like, once they figure out how to solve baldness, I’m going to want not just you to solve baldness. I’m going to be like, “Well, you know what? Like, I naturally have curly hair. Can you make it so I can have it straight sometimes? And, oh, now can you solve my gray hair problems?” And then, like, it will never end, is what I’m saying.

Chelsea Follett: That makes sense. We’ve covered the unemployment concerns, concerns about technological unemployment. What about the other risks that you alluded to earlier, the existential risks, the safety risks of advanced artificial intelligence that you hear so much talk of recently? What do you make of this?

Kristian Stout: So this is another area where I don’t want to downplay the fears and anxieties that people have because, you know, I get it. I understand it. I grew up on a lot of science fiction. So, like, you know, I’ve been thinking about it for decades, even before we had computers talking back to us. But let me frame it with one example, and then I can talk about the more specific stuff. Think about nuclear. Nuclear energy comes from a technology that obviously has potential really existential risks, right? Like, we’ve seen what you can do with nuclear bombs. At the same time, nuclear energy also would almost entirely solve all of the environmental concerns that we have, right? So, like, on the one hand, you have global warming that’s, you know, potentially caused by human conduct with our different kinds of emissions. But then you have this technology that, if we had been adequately developing it for the last 70 years, basically would make all of our concerns about energy access go away. If we had been smart about it for 70 years, we would not be in the place we’re at right now what we’re thinking about.

Kristian Stout: Which means that we allowed our anxieties, because some people could use nuclear weapons in a way that existentially concerned us, to freeze another adjacent technology that, if we had been responsibly developing it, we would have, like, small cell—we’d probably have, like, small cell reactors in our houses right now, right? I’m guessing here. I don’t know. I don’t follow nuclear that closely, but I know that some of the ones that have been in process that we’re thinking about deploying on some of, like, the data centers or just to reinforce our grid are much smaller, much safer, much cleaner than what we grew up thinking about. But essentially, the nuclear industry was frozen in place, and, like, we had started—like, all the world was, like, celebrating when we shut down nuclear reactors about 10 years ago. Now everybody’s kind of reversing course, like, ‘Oh, maybe we need to stand these back up.’ So we need to be careful about understanding when there are existential things you want to keep in mind and you want to have responsible people looking at and governing correctly. But you don’t want to lose all of the benefits that you can get just because there are existential risks that you’re concerned about.

Kristian Stout: So that’s, like, a framing thing that I think is important to keep in mind. And I’m not equating AI with nuclear at all. I think AI is very different. So now what would be the existential risk that people are concerned about? It’s like something like a mass extinction event, if I’m trying to guess what the best case—so first off, everybody who says that there’s an existential risk, I try to look at what they’re saying, and it’s never well substantiated. It’s a little bit like—I don’t know, most people don’t understand what I’m saying, so I don’t know why I keep saying this—but there was a South Park episode that featured the underpants gnomes, where they had this plan for world domination. There are these gnomes that were walking around town. They had this plan for domination where they would go and they would collect everybody’s underwear, and then the kids were like, ‘Okay, now what do you do with it?’ Like, ‘Well, here’s our plan. We collect the underwear, then it’s a big question mark, and then it’s world domination.’ It’s like, ‘Well, what’s that question mark?

Kristian Stout: Like, how does that connect those two things?’ And I feel like a lot of the really doomerish talk about AI is a little bit of an underpants gnome thing, where you start from, like, ‘Well, first we have this really intelligent computer system, and then we have death.’ It’s like, ‘Okay, well, all right.’ So, like, if I’m connecting the dots, maybe it has something to do with capabilities and, like, developing a new kind of virus or something, right? And maybe it can get access to, like, a benchtop synthesis. But, like, we currently have—like, right now, if you and I want to start a company, we could place orders from DNA synthesis providers that we could, like, technically string together really bad stuff that we could release right now. Like, we currently have an existential risk around DNA synthesis. So, like, why is this one different if, like, right now you could get a benchtop device and figure out a way to get it to produce smallpox? Like, what’s the difference? I haven’t seen that yet. We’ve seen human labs without AI help release viruses in the world that gave us a pandemic and gave us a problem, right?

Kristian Stout: So, like, we’re dealing with that. So, like, what’s the difference in kind? And I haven’t seen that filled in there. Or it could be—another one could be these devices become so capable, or these machines become so capable, they can hack into systems and, like, bring down our power grid or launch the nuclear weapons or something. I mean, yeah, I guess theoretically you could say something like that. There’s kind of an interesting facet—feature of human society, though, is that, like, our systems are actually not that well designed, and it almost is like a firewall against that kind of takeover, where they kind of fail on their own, and, like—but they fail in almost like a firewall kind of way because our human institutions are so complicated and separate things out that we get frustrated. But part of our frustration in upgrading our society also provides us with almost roadblocks for these systems to be able to just go in and easily, like, turn switches and flip them on and off, right? Like, things would fail in one part but wouldn’t fail in other parts, and, like, you know.

Kristian Stout: So the path to actually, like, world domination through our existing human institutions does not seem very clear to me. I haven’t seen anybody spell out how it would actually overcome the roadblocks that we ourselves experience when we’re even trying to do good things. So, again, the harms that I think are more likely—like, because there are potential harms that can come here—are not existential, in my opinion. First off, the whole idea of, like, the self-replicating AI that’s just going to break out and install itself everywhere is kind of ludicrous on its face. These AIs have to run on a hardware rack that’s at least $100,000 and has to have constant power, constant cooling. There’s no way they’re going to break out of there and install themselves on your computer and, like, take over your house. Like, they just won’t run. Like, they have nowhere to go. They have to be in these data centers. So even if they figured out a way to be self-replicating, they’re going to self-replicate in the exact same data center they’re already in because that’s the only place they can run.

Kristian Stout: Now we could have more capable models that can do more stuff in the future, but right now that’s what we’re looking at. In terms of the actual harms that we should be concerned about, I think there are real harms. They’re just not existential. Bad guys, criminals, are now much more capable of ripping off your grandma, stealing all her life savings, or executing attacks that lock down your computer and can do ransomware against you. Those things are a lot more capable right now. Deepfakes. You know, there’s a footprint right now. You’re doing a podcast of me. There’s enough footage of me out there where you could easily voice-clone me and call my family and get them to drain my bank account to, like, save me from some, I don’t know, like some criminals that have allegedly kidnapped me, right? Like, those are real problems that we need to think about. But they’re a different kind of problem. They’re the kind of problem where, you know, far be it from me to say the government needs more money, but maybe law enforcement agencies need more support in order to have, like, their own AI and cybersecurity types of divisions where they’re capable of finding these scammers when they’re out there, finding these, like, AI agents that are going out and doing attacks on services, defrauding people, and they have more resources for enforcement and suing and putting the bad guys in jail, right?

Kristian Stout: Like, we might need more defensive-level support in that sense. At the same time, we might need more resources and more devotion toward thinking about, like, real agreements internationally. Like, so, like the DNA synthesis thing I mentioned, I know the companies that do that are really trying hard to do a good job there, but they might need more support from the government. So, like, our Congress, they might need to stop being so dysfunctional and actually look at, like, well, how do we have these technologies responsibly even apart from the AI question, right? Like, how do we make our infrastructure better apart from worrying about AI, right? That’s a much more complicated, difficult question, but it intersects with the AI question, I think, in a more reasonable way that we should actually be focusing on. And my concern with the doomer talk is that when we think about it as an existential risk, that, like, Skynet’s going to take over, we stop thinking about those, like, real problems that we need to come up with real difficult, annoying democratic solutions to.

Chelsea Follett: And that doomerist talk can also inspire policies that could have a real slowdown effect on

Kristian Stout: Right.

Chelsea Follett: the AI industry at this critical moment, right? If you think about Europe and their technology industry compared to the relatively less regulated U.S. technology industry, you see a big contrast. Could

Kristian Stout: Exactly.

Chelsea Follett: you talk a bit about how the policy environment affects the ability of this industry to continue to innovate at this critical point in time?

Kristian Stout: Yeah, no, that’s a great point. So, by way of example, the European area, before it was the European Union, was one of the leaders in, like, pharmaceutical production throughout the 20th century, one of the best places to get innovative drugs. Over the last 30 years, because of policy choices they’ve made, they essentially have lost—I mean, they’re still developing these drugs and they’re still innovative, but America has just, like, blown way past them because we have a regulatory environment that is—some people criticize the regulatory environment as being a little too safety-conscious, but nonetheless, we have a very innovative, very pro-innovation regulatory environment for developing these pharmaceuticals. So that’s why these keep coming out of the United States, and why we’re having all these discussions about drug pricing around the world is because, like, we’re, like, the only place that’s, like, actually developing this stuff, and everyone else is kind of just trying to, like, lower the price on it without doing their own innovation. We could risk something like that, and we actually could risk something worse, I think. First off, assuming China figures out a way around its demographic crisis and it keeps becoming—it’s a very innovative place.

Kristian Stout: Assuming it keeps enough people there working that they can keep being innovative, I don’t believe they’re going to slow down on trying to innovate on AI, on trying to—you know, they’re standing up their own chip producers to get around our export controls, and some of the early returns on their chips look very promising, like— I myself would not mind getting a hold of some Chinese memory and chips just because they’d be more affordable, and I could run some of my own AI a little bit more efficiently in a more cost-effective way. So they’re not slowing down on that side. They also have half of the world’s most brilliant AI researchers, so they’re not slowing down on the developing of the model side. I run, myself, I run a lot of the open-source models that come out of China. They’re fantastic. They do a great job. What the costs could be, apart from—so, like, on the European side, they’re largely a consumer of our technology products. They don’t need to be. There are brilliant people in Europe. There have been brilliant people there for centuries. They’re constrained by the regulatory environment.

Kristian Stout: They had a report a couple years ago called the Draghi Report, where they were trying to do some soul-searching about why it is they keep falling behind the United States and what they can do. And they sort of get it that they’re regulating themselves into a corner, but then their proposals were, like, a little incoherent with that. And they were saying, well, maybe we need to regulate a little more in this way to try to offset. I didn’t love the proposals. I love that they’re thinking about it, at least. The cost to Europe is that they’re six months behind us, say, right? Like, everybody who rolls out a product in the United States gets it first, and then they have a team of lawyers that they pay to figure out, like, well, how can we deploy this in Europe? And then Europeans eventually will get some form of it. Maybe not, but usually they’ll get something. That’s one cost. The real cost, I think, on the AI side is—it actually goes back to the real harms and real threats that we need to be concerned with.

Kristian Stout: We need to be thinking about a future. Now, take this with a grain of salt that I’ve read, again, a lot of science fiction as a kid, and one of the areas, one of the book series I really loved was the cyberpunk novels. And, like, the whole concept of the cyberpunk novels, or one of the concepts, was there’s almost, like, personalized AI defensive measures that everybody has. Like, everybody has their own counter-AI. There’s AIs and counter-AIs, and, like, your AI is essentially supposed to go out and be your agent. It’s like an agent. It’s supposed to go out and do the things you need it to do. Part of that is defending you from the bad AIs. The future is probably going to be something like—you know, you see these agents starting to roll out from these big providers right now. And then there’s also open-source ones. The future is going to be something like you have an agent whose job is to be, like, almost like your virtual twin and check your identity theft exposure, check your financial risk exposure, go out and make sure that, like, all the opportunities for your professional life are being maximized, or, you know, it’s like servicing those things.

Kristian Stout: It’s supposed to go out and, like, work on your behalf to both protect you and then also advance your interests. That is probably what the agents are supposed to be doing and they’re going to be doing. If we slow down innovation in the United States, we don’t get there. We become eventually dependent on something like a Chinese provider to give us that technology. So if we hamstring the open-source development in the United States and the proprietary as well, I think we don’t just fall behind and become consumers of technology. We lose access to the American-style defensive AI and pro-individual assertion of their own autonomy in this space. And we become dependent on the Chinese Communist Party for providing us its values embedded inside of its AI because we won’t be able to produce it here. So we have to be very careful that we don’t just, like, we’re not losing—again, economics is not about money. Economics is about the study of action and what humans can do using their own volition inside of markets. We want to make sure Americans are able to exercise American values broadly in the future.

Kristian Stout: And if we lose the race to AI because we have bad policy, the cost is going to be becoming sort of de facto citizens of Chinese AI infrastructure in the future.

Chelsea Follett: And it really is hard to fully, I think, as you mentioned, imagine that post-AI revolution future for us now, because it’s like being at the precipice of the Industrial Revolution. You have no idea how different the world’s going to be—

Kristian Stout: No idea.

Chelsea Follett: —in the near future. We don’t even know the potential benefits that we could be giving up. So I want to get your take on the recent AI Safety Accord. So this is a matter of self-regulation. Some have critiqued it because it’s non-binding. Of course, that self-regulation does not exist in a vacuum, right? There already are laws against fraud and many other potential harms that people want AI-specific regulation on. Some people want that. What’s your take on the AI Safety Accord and this idea of industry self-regulation?

Kristian Stout: So I think it—I answer every question you give me with, like, a backstep framing. I apologize for that. But, like, one backstep that I would make on that, pull back on, is there’s no such thing as self-regulation anymore anywhere in the world. It’s just, like, we do not live in a laissez-faire environment. We live in a highly regulated environment. In the United States, we’re very pro-innovation, but we are highly regulated. Regulation is not just someone at the FTC sitting down and designing, like, designing a code of conduct you have to apply to. The legal system itself is a regulatory apparatus. It’s just we do it through courts, right? The common law type of approach the United States has, where we have adjudication through different court levels, you know, it’s spread across the states and the federal jurisdictions. And then there’s—you can even say some of the adjudication at the regulatory—we have, like, sort of pseudo-courts in the regulatory system. All of that is meant to create the—make regular, to regulate, to make regular conduct inside of the economy and inside of the United States. That backdrop means that there are liability rules operating in the background of every firm that’s out there that will potentially destroy your company if you run afoul of them.

Kristian Stout: So this, you know, the labs—people call them labs; they’re actually giant corporations—they’re functioning right now without anybody having brought a serious suit against them. But I guarantee you they have counsel inside that says you need to be careful. Now, you wouldn’t know it by some of the rhetoric, but we need to be careful about how we’re doing X, Y, and Z because, you know, we haven’t had any major cases on this. But, you know, there’s a path to where you could see negligence liability, products liability emerging based on the way these products interact in the marketplace, the way consumers use them, the way you’re selling them to people, the way deployers are using them. There’s all kinds of exposure for liability. It’s only a matter of time before you start to see some kinds of settlements or cases emerge out of this. So when we say that they’re self-regulating, what we’re really saying is actually something pretty good. We’re saying we would like this industry to start trying to inform courts and administrative regulators how to understand how this technology works, how consumers are using it, what we consider to be safe and fit for purpose, so that when a court actually has a case come in front of it, they’re now going to look at what’s considered the industry practice as they’ve evolved it, and then the court will say whether that’s good enough or not good enough.

Kristian Stout: So they’re not going to—by saying we’re encouraging them to do self-regulation, we’re not saying we want you to just get away with whatever you want, don’t worry about it. We’re saying you need to take the first crack at helping us understand how to understand this technology and how it fits in society. And when something goes wrong, you can be guaranteed that a court or regulator is going to look at what your efforts were and whether they were good enough, and if they weren’t, you are going to face massive liability over those things. So this first pass at trying to stand up self-regulation, I think, is actually really great, and there are many areas where this actually functions and feeds into the legal system. So Underwriters Laboratory is known for taking a look at products and figuring out whether they actually work, actually are safe and useful for what they’re for, and giving it a stamp of approval. And that, like, carries a lot of weight. Like, there’s a lot of self-certification of these things, which carries a lot of weight in the economy and which will feed into legal processes.

Kristian Stout: It might take a little bit of time, but that’s the way we do things in the United States.

Chelsea Follett: What do you see as some of the roadblocks, even in our relatively less regulated environment compared to Europe, that could potentially cost us the AI race and cost our descendants who knows how many benefits of this technology? Is it permitting delays, zoning restrictions, densification limits? What are some of the regulations that are potentially the most damaging that we should be aware of?

Kristian Stout: Great question. We probably need a couple hours to dig through all that. I’ll try to be concise. Two broad classes of where I think there are problems right now. The first one is the state versus federal distinction. The second one would be this, like, zoning and classification, data center pushback thing. On the state versus federal, because the federal government has not been able over the last two years to enact something like a federal framework for how we want to think about AI regulation, where you would divide capabilities between what the feds would be good at and what the states would be good at—because we don’t have that operating right now, we have a lot of states rushing to fill the gap. California has actually been somewhat restrained so far, which is unusual for California, but more recent news has suggested that they’re going to explore more regulation over, like, the development of models and the deployment of models. Which, what I don’t think you want, what I think would get in the way of regulation is if we allowed all the states to come up with their own regulation of the development of AI itself.

Kristian Stout: That would create—it wouldn’t be 50 states. You’d probably see it break down to, like, 20 states that have, like, different kinds of requirements, maybe 10. But even 10 states, to try to figure out, like, how you square training a model, safety testing a model, and then how you offer that to consumers, that would create a major slowdown. It would not make it impossible to develop AI, but it would definitely put a lot of brakes on the process. And brakes on the process when right now the estimate—you know, this is not empirically quantifiable, I don’t think, based on benchmarks alone. But, you know, the received wisdom right now is that Chinese models are about six months behind American frontier models. Maybe it’s a little less in some areas. Maybe, you know, maybe they’re a little closer. Maybe they are at six months, something like that. If you were able to introduce enough regulatory friction in the United States where major labs felt like they needed to, like, severely slow down their deployment, not just, like, wait a couple months for testing or something like that, you could easily see China start to outpace us or at least be at the frontier level.

Kristian Stout: Now, that doesn’t mean that, again, America automatically loses the AI race, but it means that as China is going around the world shopping its technology stack through, you know—because it does this through its Belt and Road Initiative-style approach to AI—that more of the world adopts the Chinese stack, more of everything moves towards the Chinese stack, and the American stack becomes an also-ran. And that is not a world I think we want to live in. And I want to be careful here. I don’t think China is evil. I think they’re a fine country in many ways. I’m not a big fan of communism, so, like, maybe I don’t like their government. What China is developing, you know, great for them. We should see them as a trade adversary, and in a friendly way, but we still want America to win. We still want American values to be the ones that are respected, because I think, on the whole, that’s kind of better for people to have those sort of democratic values baked into their AI stack that they’re consuming versus the Chinese Communist Party ones. The other major roadblock is what you alluded to in this, like, permitting data center approach.

Kristian Stout: I see different numbers on the research about how hard it is to actually build a data center. I think SemiAnalysis had an article out a couple weeks ago where they suggested that maybe it’s, like, a little overblown exactly how hard it is to put the data centers out, that, like, the core hyperscalers are still able to get these ones out that they need. And it’s, like, some of the more marginal data centers that are having trouble getting deployed. Some other research I read suggests that it actually is a problem, that there’s a lot of roadblocks to getting the data centers out, and it happens on a number of fronts. One would be on this, like, local zoning issue where different localities are just saying, “No, I want a moratorium on data centers.” When I look at it, it looks like it’s more like temporary moratoria where the localities are just trying to figure out, like, whether they can handle it. Sometimes it’s a little more permanent. I live in New Jersey. Apparently we have a whole bunch of them all throughout the state. To give it its credit, when I look at the local data center issues—I was looking at Kansas’s recently—some of the concerns are totally understandable in that a hyperscaler or somebody who wants to build a big data center comes into a community, and it’s often a bespoke process where these, like, locals—this local council is like, “Okay, so, like, you know, what questions do I ask?

Kristian Stout: Like, what data can you give me?” And, like, different ones get different answers, and it’s not because anybody’s doing anything evil. It’s because the process—the public processes that the states have in place often can be kind of opaque, and it’s not clear how to get the exact answers. Like, how—when—at what point do the jobs promises that the data centers bring in materialize? You know, how do you judge those things? So frequently, a lot of these, like, local councils, they don’t know how to answer these questions. So I think the answer there is the state governments need to take seriously this idea of figuring out how to almost, like, normalize, create, like, a regular process that someone who wants to come in and build a data center has to go through in order to just be able to expose all of the information that the local communities need to understand exactly what they’re being asked. Because without that information, like, yeah, I get it. Like, some residents will be like, “Why would you do this? You’re just—you’re trying to do some shady backroom deal.” It’s like, it’s usually not, but without that information, it’s hard to justify.

Kristian Stout: The other side of it is the permitting side, where we have this long-standing problem in the telecommunications area where, like, wireless, cable, fiber providers, they will frequently run into not a 50-state problem. They will run into a 50-state-times-however-many-counties-are-in-those-states problem, where they need to go in and get permissions for rights of way in each of these localities in order to deploy their infrastructure. Sure. Energy suffers from a similar problem where, when you want to go in and say, “Hey, I need to bring online a 20-megawatt data center or something like that,” there are processes for that, but they can be onerous. It can be difficult to do self-provisioning, depending on where you’re at. Right now, I think the estimate is when you want to build a data center, it’s something like a seven-year lag before you’re allowed to even connect and have the energy running on the regular grid, which is why a lot of these places are starting to look at whether they can do self-provisioning, meaning putting their own on-site energy production. But even there, then you run into complications about whether you’re allowed to do it or what your grid interconnection needs to look like.

Kristian Stout: So we have a very fragmented and not efficient way to get the energy into these data centers. We have a fragmented and not efficient way to make sure that all the telecommunications networks are in the area, able to feed it the bandwidth that it needs. And then we have the legitimate problems at the local level of: what information do residents need to know in order to vote on it?

Chelsea Follett: Now, because this is the Human Progress podcast, we usually try to end on a positive, progress-oriented note. So my final question for you is: what are you most optimistic about regarding AI and/or AI policy?

Kristian Stout: So that’s another one that I could probably go on for two hours. So I don’t—I try not to be too much of a fanboy for technology, but, like, you know, again, I grew up reading a lot of science fiction. So, like, as the world keeps giving me all this cool stuff, I get really excited about all the things that are emerging. And, you know, all the problems that I flagged and all the anxieties and concerns, again, I take them very seriously. I don’t think that they will ultimately be showstoppers. I think as people start to learn about how this stuff integrates into our lives, they’re going to accept it. So by way of an example, this is not directly on AI. It kind of is. Actually, no, it is AI. So Meta has these Ray-Ban glasses with the cameras and the AI that you can talk to, right? People get upset about them because they don’t want to be, you know, picture taken in public, and we’re still working through the privacy issues there, right? But think about what this does for people who are vision impaired, to now have on their face something that helps them navigate the world that they did not have before.

Kristian Stout: They can, you know—assume they’re not completely blind—and they can kind of see something and say, “Oh, what am I looking at over here?” And the AI tells them right in their ears, like, “Oh, this is Janice. You met her two years ago.” And, “Oh, I remember Janice.” Now this person can now participate in society in a way that they couldn’t before. And as someone who has normal vision, I can’t even appreciate what it’s like to have to go through life and feel like I can’t fully participate in society because my vision keeps me from interacting with my environment and seeing people who I know until, like, I have some other cues that help me. Like, being able to get those people online, like, fully participating is just mind-blowing to me. It’s the same thing. Think about what AI could do. You know, we focus on, like, the jobs anxieties about, like, white-collar jobs. Like, you know, we’re white-collar jobs people. Like, you know, we type all day. That’s what we do for a living, right? And, like, “Oh, well, the AI can type, and it can take our job away.”

Kristian Stout: Okay, fine. But think about what it can do for people who previously were not quite able to even get into the typing jobs, right? Like, they were kind of below-average intelligence. They did okay in life. They’re able to scrape by. Right? Now they have a personal assistant who helps them cognitively navigate the world. For someone who has a below-average intelligence or, like, much below average, just imagine the utility of having someone that can help you understand the incredibly complicated modern world in a way that you couldn’t before, and that would freeze you out of society. That kind of cognitive adjunct for—you know, it’s obviously going to help us do even more impressive stuff. But to have something that helps you when previously the world was, like, kind of confusing at the level you were at, now you can participate in society. We have more people having more participation in society. That can only be for the benefit of society, in my opinion.

Chelsea Follett: Thank you so much for talking with me, Kristian.

Kristian Stout: It was a pleasure. Thank you. I’m happy to talk whenever you want.