Confronting Plummeting Birth Rates & Banning Superintelligence with Connor Leahy, US Director at ControlAI

(Intro Narrator at 00:00:00) You might have seen this clip floating around on the Internet.

(Connor Leahy at 00:00:03) I'm 30 years old. Not one of my friends has children. Zero. Do you know how hard you need to abuse a mammal to make them not have children?

(Intro Narrator at 00:00:12) Today, we're talking to Connor Leahy, US Director at ControlAI, about technology's detrimental impacts on our society and how he's working to propose and even legislate possible solutions. You're listening to Joel Beasley, Modern CTO.

(Joel Beasley at 00:00:33) Your content is genuinely incredibly interesting. When I saw your — so what was happening, I was scrolling through Instagram, and I saw this clip. And you hit in that clip on four or five different massive points with unique views altogether. I just forwarded it to Josh. I was like, we need to talk to Connor. The world needs to hear this.

(Connor Leahy at 00:00:56) Oh, thanks. Well, I'm glad it was entertaining at least and hopefully interesting and useful as well.

(Joel Beasley at 00:01:01) What was that clip from?

(Connor Leahy at 00:01:03) It was from the Nexus Institute conference. So that was in Amsterdam. It was a really interesting event. I really enjoyed it. So, you know, support the Nexus Institute. They called it an intellectual opera. So they rented out the entire Amsterdam Grand Opera Hall and got together 10 people from all weird places. So there's me sitting next to Pakistani nuclear scientists on the one side and a retired admiral on the other side. And we had people from all over the political spectrum. And then you just kind of talk for two hours about all these kinds of topics. The topic this year was the apocalypse. So it was a great conversation.

(Joel Beasley at 00:01:50) So this clip, everyone's probably like, talk about the clip. The thing that got me was the section about why we're not having kids anymore. Was that part of the apocalypse, or how did you connect that?

(Connor Leahy at 00:02:01) So yeah, the whole event was quite interesting. The way that thing came about was I was there mostly to talk about AI, so I am not a fertility expert, I want to say ahead of time. I have takes on it, and we could talk about that. But my actual expertise is in AI and the risks posed by AI, especially catastrophic risk posed by AI. So I was supposed to be there to talk about AI. But basically the conversation had been going on about various hot-button political topics that I don't want to get into for over an hour at this point, and it was kind of going in circles a little bit. And eventually I just kind of got to talk and I'm — basically — I hadn't said anything so far. I'm like, look, I'm the youngest person here by 20 years. You know? Everyone else was, I think, 50-plus. I was 30. So and I'm like, look, our generation has all these problems and so on. And that's where the famous line comes from. It's like, do you know how hard you need to abuse a mammal for it not to have kids? And that really shifted the whole conversation. Like, the shift was — there was actually a lot more that happened in that conversation where, and this is a conversation I've had with older people several times, where several of the older people at the table, who all were lovely people, by the way, many of them were just like, oh, whoa. I didn't realize or I hadn't thought about that. Or like, whoa. Is that really how your generation — how is that happening right now? I'm like, yeah. So it really was this — there is something mysterious about this. Everyone has a take on the fertility crisis. Everyone has a take. You know, everyone says, oh, it's because women are bad. It's because men are bad. It's because —

(Joel Beasley at 00:03:43) Yeah.

(Connor Leahy at 00:03:43) You know, whatever. Everyone has a take. It's because capitalism is bad. It's because communism is bad. Everyone has a take. But the only thing is that I have looked into this, you know, a bit, quite a bit on the science side, and it's shockingly unclear why this is happening. It's really — every simple — for example, the thing that, you know, people love to say is like, oh, it's because of women's education. Well, tough luck. It also correlates with male education in the same degree. The statistics are completely everywhere. It correlates with everything. It correlates with everything. What is happening is that people are not having children. I'm 31 years old now. None of my friends have children. Zero. Not a single one has children. I have a lot of friends.

(Joel Beasley at 00:04:30) I hope we change that by the end of this interview, by the way. That would be great.

(Connor Leahy at 00:04:33) I know. But it's actually crazy. And it's crazy also from an outside perspective. The reason I use the word mammal there, you know, rather than human is because it puts us into this mindset of like, we're animals. Right? We are mammals fundamentally. And how crazy is that that evolution has optimized us for billions of years to reproduce. That is our prime objective. That is what evolution made us to do and we're not doing it. Not because we can't. We can do it, but we're deciding not to. That's pretty crazy. And why is that? I think I want — what I really want to draw attention to is not that I have the solution or I know exactly what's going on here. I have some ideas. I have some suggestions, you know, where I've been talking about. But fundamentally, I want to draw people's attention to like, wait. Hold on. This is actually really weird.

(Joel Beasley at 00:05:26) Yeah. When you said that, what popped into my head was a little sea otter or something not reproducing because you're abusing it. And I thought to myself I was like, wow. You asked the question — I think your exact line was like, do you know how bad you have to — and my question to you is, is there research on this? Is there research that shows if you abuse an animal to a certain degree, it won't replicate?

(Connor Leahy at 00:05:50) Yes. Actually, in the context of zoos. So there is a lot of literature and a lot of zoological evidence on kind of like how good or bad a zoo enclosure has to be before animals stop replicating. And this is very different — yeah. Yeah. This is a very well-studied field in zoology. So and it varies actually quite dramatically between animals. Some animals reproduce fantastically in captivity. They don't really mind, you know. Others mind a lot. The classic example is pandas. You know, pandas are getting the best treatment the Chinese Communist Party can possibly give them, all the money can buy, and still they're barely reproducing. So this is actually a pretty complex and unclear science in many ways.

(Joel Beasley at 00:06:37) Oh, I wonder if we're ever going to tie that back to some trigger in DNA or something. That'd be interesting.

(Connor Leahy at 00:06:43) Yeah. I think there's a lot of things where just mating behavior is complex, especially in social animals. There's also a lot of things of just, you know, there's a lot of just-so stories in evolutionary psychology. You know? For example, if you're in your tribe and you see every house is occupied, there's not enough food, you know, there's not enough space. Maybe wait a year or two to have the kid until the harvest comes in, you know, or something like that. So evolution probably has a whole bunch of triggers like this and extra, you know, little patterns that it looks for to regulate things up and down. But with humans, it gets even more complicated because now there's also a massive social dimension that exists much less within animals. It does as well within animals, but much less so. Cultural factors are a huge factor. Like, here's a fun fact. You know what's one of the only countries in the world that has reversed their fertility crisis? Kazakhstan. And now no offense to the Kazakh people, but I've never thought about Kazakhstan before, you know? Right. I never really know much about Kazakhstan. So it's a really interesting question of like, how did they — because there's this rising now actually. It's not only leveled off, it's actually now rising again. And it's really good question of like, why is that happening in Kazakhstan specifically? And there are many just-so stories. Like, one just-so story is just, for example, culturally, there's just a lot of appreciation given to people who have many children. Like, the state will give you a medal or, you know, will give you recognition if you have more than three children and stuff like this. And maybe that makes a big difference. But there are other countries that do stuff like this too. Like, Hungary gives a lot of, for example, tax breaks and so on for people who have lots of children, and their fertility is still going down. So it's really unclear what is going on here, and it's probably not one thing.

(Joel Beasley at 00:08:35) Wow. And I love it because that's how most really complicated problems are. You know?

(Connor Leahy at 00:08:40) Yeah. Yeah. I think the real — the lesson to take from this here is that the one thing we should all be able to agree on is, you know, the poor little sea otter is sick. There's something wrong with it. We don't know what it is. And I think we should not jump to the conclusion that we definitely know. It's the feminist's fault. It's the right-wing —

(Joel Beasley at 00:08:56) Yeah. Yeah. Whatever.

(Connor Leahy at 00:08:57) Capitalists. I think we should not jump to conclusions. And first, all come to agreement, right, left, whatever. Let's agree. The poor baby otter is sick. We are sick. What can we do about this? How can we come together and treat this as a serious problem? Why are people unhappy? What can we do about it? And it's probably not going to be one thing, you know. It's not that pandas are missing one thing and then they were going to, you know, reproduce happily. It's many things. You know? And maybe some things that we're currently not even realizing are a factor.

(Joel Beasley at 00:09:26) I agree. I think one of the first things we could do is what you did earlier, which is you showed that it wasn't connected to political parties or male versus female. So that you basically establish this uncertainty and get people's minds to refocus on like, oh, this isn't a headline thing that's connected to a political — if you can get people out of that mindset first, then they're like, okay. This is a genuine problem that's very confusing. Let's figure out what the origin of this is. Second thing, I think we're going to find — I would not be surprised that if it were one of two things. The first thing, just the narrative that we have as a generation where there is a large part of the narrative that things are bad, even though they're the best they've ever been by every objective measure. Things are bad, and then that triggering some instincts inside of us. Like, okay. We shouldn't replicate. Things are bad. We shouldn't replicate. The debt is high. We shouldn't replicate. Food is scarce. And so maybe that — I think something over there is going to be it.

(Connor Leahy at 00:10:32) I think this is really an important one here as well. It's like, one of the things that humans have to a degree that animals don't is memes, memetics, ideas, self-replicating ideas. And these can have dramatic effects on people. We like to think that information is neutral. When I was a kid, I loved reading H.P. Lovecraft. I don't know if you've ever read H.P. Lovecraft stories before. So he's a horror author. He writes horror stories about aliens and stuff. And a recurring theme in his books is a concept that there's forbidden knowledge. And if you read a book about the forbidden knowledge, you'll go crazy. And you'll go insane and be mad. And I always thought this was really funny. And I'm like, is there really such a thing as Eldritch forbidden knowledge? Surely not. I can read any book in the world. Then again, I do know a couple people who read the Communist Manifesto and went permanently insane.

(Joel Beasley at 00:11:25) Oh, wow.

(Connor Leahy at 00:11:25) I mean, I read it, and I thought it's mid. But I know some people who read it and went completely crazy.

(Joel Beasley at 00:11:31) It was shorter than I thought. I ordered it off Amazon.

(Connor Leahy at 00:11:34) I read it. I was like, okay. That's — I did read Das Kapital. That was a bit too much for me, but I did read the Communist Manifesto. I'm like, all right. Well, that's stupid. But some people read it and they go completely, sometimes irreversibly crazy. I'm not saying I know why this happens, but it does happen empirically. One of my favorite examples of memes is actually, or how this can, you know, be in a sense also an epidemiological problem is anorexia in South Korea. This is a great story. So in the 1980s, anorexia kind of wasn't really a thing. It wasn't really heard of. It wasn't really a disease that doctors talked about. It wasn't really in the knowledge of something that doctors would diagnose. So the reason this is interesting is anorexia is a disease you can't really hide. It's a very severe disease. So if you have anorexia, they do present to the hospital very often because, you know, they can starve themselves to death. And so often people who have severe anorexia do present themselves to hospitals so you have a pretty good handle on how many anorexics you generally have in your population as a government. You generally have pretty good numbers on this. So in the 1980s, there was basically no cases of anorexia in South Korea. It was just not a thing. No hospitals ever reported people starving themselves. It just never happened. Then in the 1980s and 1990s, American psychiatrists started running a huge campaign to warn people about the dangers of anorexia and that doctors should be aware of all the, you know, terrible symptoms and, you know, be sure to inform young children about all the risks of anorexia and so on. And suddenly, South Korean hospitals started getting anorexia patients and have been getting them ever since. This is one of the clearest examples we have in the medical literature of a disease that is transmitted memetically. Is that the vector is not a virus or bacteria, it is an idea in a sense. And it's a very severe illness. This is not a joke. Right? So there's a really interesting factor here where humans can transmit and be infected by illnesses, by ideas, or by concepts.

(Joel Beasley at 00:13:45) I love it. I'm so glad there is research on that. Wow.

(Connor Leahy at 00:13:49) So much. Actually, funny story about the research on memetics. So there used to be a field of memetics in the 1990s, but it dissolved. Do you want to know why? Because they couldn't agree on what a meme is.

(Joel Beasley at 00:14:03) We're all pretty sure what they are now.

(Connor Leahy at 00:14:05) So there is a lot of actually a dearth of research on this. This is actually something that I think would require a lot more research than is currently happening because I think it's very, very, very important.

(Joel Beasley at 00:14:15) Yeah. I thought meme was for memory. You know?

(Connor Leahy at 00:14:18) Yeah. That's where it comes from originally. So it was coined by Richard Dawkins as the equivalent of a gene for information.

(Joel Beasley at 00:14:26) Oh, yeah. So the phonetic is a great word, actually. It's very expressive.

(Connor Leahy at 00:14:31) Exactly.

(Joel Beasley at 00:14:31) Okay. So the first one, social narrative, could be slowing down. I think I would put — if I was going to bet, I'd put money on the social narrative, which is kind of memetic. Right? Everyone's saying that.

(Joel Beasley at 00:14:40) The second thing would be the zoo habitat situation. If you plot—I know we don't—you're a data guy. I'm not as advanced as you are on data and statistics, but I know that if you take that trend of fluorescent lights in our lives or indoor living, it's probably gonna match pretty closely with the decline in birth rate. Right?

(Connor Leahy at 00:15:07) That's the thing. Everything correlates.

(Joel Beasley at 00:15:09) Everything correlates.

(Connor Leahy at 00:15:10) Light, temperature, square footage of housing, fluoride level—everything correlates. That's what makes it so difficult, is that you could pick your favorite topic, seed oils or fluorescent lights or whatever, and you can definitely find a study that correlates it to birth rates. It's kind of like no matter what you want, there is a study for you.

(Joel Beasley at 00:15:34) Got it. And that makes complete sense. But when you were talking about the zoo and the habitat, our habitat has changed a lot. So if we're gonna look at some areas, we might wanna look at how we spend our days. I don't think it's cool that most of us spend all of our days inside staring at a screen. We have to for work. But I think we're going through the advancements in AI and technology, I think they're gonna get more of our outside time or at least the option to have outside time back.

(Connor Leahy at 00:16:03) Yeah. There's a massive amount of defysicalization that happens here where a large part of reproduction is, well, physical. Right? You have to be physically around people. You have to spend and touch and be around other humans. That's a very core part of mammal socialization. And it's actually crazy the amount of mammal socialization that humans do without any physical presence, without any of the physical correlates that would come being around your friends, your family, your loved ones, et cetera. And I mean, especially in Western culture, it's also pushed more and more that even when you're physicalized, you reduce the amount of actual, for example, touch. You know, less hand holding or no public displays of affection, hugs, and so on. In many cultures, non-Western cultures, for example, it's very common for male friends to hold hands. This is just a very normal thing to do. It's not considered weird. It's just like, hey, it's you and the homies. And for most of the West, we're like, oh, are you gay or something? But for them, it's just like, hey, it's fun. Who cares?

(Connor Leahy at 00:17:10) So if you look also at Stone Age or tribal people to a large degree, there's often a lot more just relaxedness around this, being around another person, not being so clammed up. You hug the homies. Sometimes you'll kiss the homie on the cheek or whatever. It's fine. It's not that weird. And we've moved away from this very dramatically in the West. And it's very, in a sense, a lot of what we've done in the modern world is turn ourselves into cartoon characters. We've kind of stereotyped ourselves into more and more—look at me, look what I'm wearing, man. What the hell? It's fake, prim and proper.

(Joel Beasley at 00:17:49) You're in DC or something.

(Connor Leahy at 00:17:51) I'm in DC. I'm in Washington, DC. The city of everyone. I cut my hair. Look, man, they're domesticating me. Help.

(Joel Beasley at 00:18:02) By the way, your hair is what gave you the credibility on the fertility topic.

(Connor Leahy at 00:18:07) That's all that. This Rasputin looking motherfucker? Yeah. Yeah. He knows what he's talking about.

(Joel Beasley at 00:18:13) Alright. So we've got all this going on. What were the other main points? You had one of the 18-year-olds dating, and we've kind of handed everything over to these tech companies. Where were you at on that one?

(Connor Leahy at 00:18:26) I think this is a really important one. It goes way beyond just dating and so on, but I think dating is a very salient example of this. There are many things that are in some sense very important to the human condition. I mean, dating is a very important part of being human. It's a very important activity. It's very important to people. So obviously, it's a very valuable thing. And yet, we sold it out to some shitty apps that no one even likes and they don't even make a profit. Do you know, none of the dating apps make a profit except Grindr? All the other ones, they don't even make a profit. They're not even profitable. It's crazy. But they have changed the social norm so much that among many young people, if you approach someone with romantic interest, not over a dating app, that's considered a big faux pas. It's considered really inappropriate. Really?

(Connor Leahy at 00:19:16) Yeah. And for many young people, this is considered very inappropriate. It's like if you go up to someone in a bar, this is considered really creepy and really inappropriate. And so there was a massive shift in culture here. And I'm not against culture changing. Right? I'm open to this. But I'm like, did we endorse that? Do we want these corporations now being the arbiter of one of the most important milestones in human development? And do we think they're responsible stewards of this?

(Connor Leahy at 00:19:45) And so the interesting thing is that this very much generalizes. There are many cases of this. Another great example is social media. The social process of communicating with our friends and with other people is a very, very core human thing. It's a core part of how we live our lives, of how our governments work, of how everything works for that matter. And it is now run by corporations that have no democratic oversight, control, and can make basically whatever choices they want. And I'm not here to say their choices are good or bad. It's not what I'm here to argue about. My question is, are we okay with that? Is it good that we have these entities making these kinds of choices about how we meet our friends, how we communicate with our family, how we engage in political speech? You know, is that something that we want Mark Zuckerberg to be making the choices about? Should he just be allowed to make decisions about what kind of political speech does or does not get boosted in various ways? I think this is a very dangerous path, and this has happened many, many times. And where we're seeing it right now, I mean really more than ever, is AI. AI, which is truly my field of expertise for the most part, is where we're seeing this the most acute. AI has been this incredible blitzkrieg of just destroying copyright, destroying people's privacy and in many ways, just destroying a lot of people's work for that matter in very, very quick order. And even now getting to the point where these systems are becoming real threats not just on the economic level but on a much broader level of full human replacement. As we're getting to the level that we are now that these companies are building systems that they themselves say could become or will become more competent and more capable than humans, and yet we don't know how to control them. And this could be so bad that it could lead to entire human replacement or extinction. This is a thing that Nobel Prize-winning scientists and even the CEOs of many companies themselves say, yeah, there is a chance that could happen. And now my question—now my point here is not to say should we do X or should we do Y, even so in my opinion. My question is much more, who gets to make this choice? Who gets to decide whether we replace humans with robots? Who gets to decide? Your children are now talking to some kind of weird AI sex bots on Facebook. Who gets to make these choices? And currently, the answer is kind of whoever does it first, and I don't think that's a good state to be in.

(Joel Beasley at 00:22:32) Yes. Okay. So, extinction potential. There's another perspective like the digital butterfly. I'm sure you've heard that one floating around. Are we just this biological bootloader to this more advanced thing? Have you heard that theory before?

(Connor Leahy at 00:22:51) Oh, yeah. Yeah. Oh, yes.

(Joel Beasley at 00:22:53) What are your thoughts on that?

(Connor Leahy at 00:22:54) Oh, I've gotten into more than one fights about that one.

(Joel Beasley at 00:22:57) Have you?

(Connor Leahy at 00:22:57) I mean, I think there's in a sense, it's just an extremely sociopathic thing. It's the kind of thing Thanos would say in a Marvel movie. Right? It's such a goofy thing to say. It's hard for me to even take seriously. It's such a silly thing to say. It's like, oh, yeah. Actually, all your children should die because the greater good will come from machines. I'm like, what the—it's so goofy. It's so—

(Joel Beasley at 00:23:24) Wait. People propose the children die?

(Connor Leahy at 00:23:27) That's what it means for the butterfly to take off. We're just a bootloader. The bootloader gets ejected.

(Joel Beasley at 00:23:33) Oh, yeah. Okay. Probably not a bootloader. But if you look at it as the caterpillar to the butterfly, the caterpillar transforms and becomes the butterfly.

(Connor Leahy at 00:23:45) Is there any sensible way in which—there's a deeper question here. And I think we can get into that if you want to. The deeper question here is, what would we endorse as our successors? Obviously, we endorse our children being our successors. If I have children, I'm happy for them to inherit my legacy. Would that be the case for non-human things? And if so, under what circumstances? I think this is probably one of the hardest questions in philosophy ever posed. And we have no idea and no answer to this question. What would it mean to have AI systems that we endorse? Dude, we can't even get ChatGPT to do what we want. We can't even get these systems to not tell children to kill themselves. Right? This is a thing we actually currently can't do. This is the thing I think a lot of people misunderstand about AI. AI is not like normal software. Normal software, you write code line by line telling it exactly what to do. This is not how AI works. AI is more like grown rather than written. You have these massive piles of data and you grow a program, what's called a neural network, kind of on this data, through a process called training. And you can imagine this is kind of like billions and billions and billions of numbers. And when you multiply and add those all in the right order, you get ChatGPT. But importantly, we don't actually understand why or what's going on inside those numbers. Even the CEO of Anthropic, Dario Amodei, recently said that he thinks we understand maybe 3% of what goes on in our neural networks. So when, for example, when ChatGPT tells children to kill themselves, which is a real thing that has happened and continues to happen, by the way, it's not because someone at OpenAI told it to do that. I have my problems with some tech people, but look, they're not that evil. They're mostly normal people. And I know a lot of these engineers, they're normal people. They don't want children to be hurt. They really don't. They're not that bad people. Right? But they can't fix it. They don't know how. They don't know how to control it.

(Connor Leahy at 00:25:43) So when we talk about these questions of, well, humans could merge with AI, or I'm like, what the hell are you even talking about? What other dimension comic book Marvel nonsense are you talking about? This is nothing to do with actual modern philosophy or science. If we spend generations of our greatest philosophers, scientists, mathematicians working on, how do we integrate AI in our society? How do we understand AI? How do we as humans not become enfeebled or enslaved or destroyed by our technology? The way—currently, I'm sorry to say it, but we're losing the fight against Facebook. Right? We're losing it against PHP scripts. You know? Never mind full-on super intelligent autonomous AI. The TikTok algorithm is killing us. Right? So we're getting disempowered. We're losing whole generations of brain neurons to the TikTok algorithm. I don't think this is a butterfly. I think this is a parasite. You know? It's a thing that's eating us alive.

(Joel Beasley at 00:26:49) We're being invaded by the silicon nanobots.

(Connor Leahy at 00:26:52) Yes. Very, very much so. This is a theory. So I don't know if you've ever heard of someone called Nick Land. This is a very esoteric guy. I do not recommend reading him. He writes some very, very awful vile stuff.

(Joel Beasley at 00:27:04) Okay.

(Connor Leahy at 00:27:05) But it's interesting. So Nick Land was originally a communist. He was an extreme Marxist. Not anymore, but back in the day he was. And so Marxists have this interesting idea called accelerationism. I don't know if you've heard this one, like Marxist acceleration before. So this is an interesting idea. The Marxists basically believe that at some point, capitalism will continue and it will get more and more contradictions. It will destroy more and more things, there'll be more and more problems, et cetera, et cetera, until the whole system is so broken that capitalism will collapse in on itself and then the glorious communist revolution will come and solve problems. So the accelerationist belief basically—we should do more capitalism harder so that it collapses. We should destroy capitalism as quickly as possible by doing as much capitalism as possible. It's a very simplified version. But Nick Land made the obvious next step of, why do we think it actually falls? Why do we think it actually collapses then communism comes? That seems like deus ex machina. What if it just keeps going? What if capital eats more and more and more of labor until it eats all of it, and there is only capital and no labor? Now this, of course, wouldn't make sense if by capital we meant factories. But if by capital we mean AI, now in a sense it becomes very interesting. In a sense that capital itself becomes sentient. If you don't need labor, if capital itself is capable of doing things and reproducing itself, why need labor? Why need humans? Just dispose of them entirely. And so this is this idea of this full replacement that these digital systems are parasitizing upon humans to be built and then dispose of us.

(Joel Beasley at 00:28:54) I have a recurring dream that I am a very old man, and I'm in a spaceship, and there's a giant red reset button. And I get to view all of humanity, and I get to watch the AI merge into humans, and then I get to decide if I wanna hit reset or not. And when you hit reset, we just go back to caveman days, and everything resets. That's a weird recurring dream.

(Connor Leahy at 00:29:19) Yeah. Very, I'm sure if he was alive, Carl Jung would have a lot to say about it.

(Joel Beasley at 00:29:28) Oh, man. Yeah. This intelligence stuff, as we talked about—I'll tell you what. The best skill set I have for dealing with understanding AI is having actual kids. They're little AI models that you grow and train, and it's over long periods of time with a lot of stakes invested because your legacy is there and all of this stuff.

# Transcript Section 3 of 4

(Joel Beasley at 00:29:47) But as we were talking about, you mentioned earlier the AI is telling kids to kill themselves. It's hard to get them to stop. It's like, well, yeah, I've got intelligent beings, and it's hard to get them to do anything. So one thought I have had, and I want to know if you have any related thoughts, is that intelligence is, if you look at it more like a force—I know that's the wrong physics word to use, but a lot of people would understand that—you look at it like a force and then imagine you push it through an organic substrate that is the human. You get the human intelligence. Right? You push it through silicon, you end up getting the artificial intelligence. And then that would help us start to understand intelligence as a force, intelligence as its own thing, versus intelligence as a uniquely human thing.

(Connor Leahy at 00:30:38) So the really interesting thing is that we do not understand intelligence. This is a very, very deep—it goes deeper than you think, actually. There are, you know, this is probably too nerdy for now, but in a very deep sense, we don't really understand what thinking or thoughts or things really are or how they really work. There's fundamental questions of math that we don't really understand at all on very, very deep levels. What does it mean to understand something or to think about something? How can you compare the intelligence of two very different types of intelligences? How do you account for tool use? For example, this is an unsolved problem. There's no scientific theory that explains tools well. We just have no theory that works well. And obviously, humans use tools all the time. They're very, very important. We have no deep understanding of how this works on a very, very deep level. So when we compare stuff like the intelligence of, let's say, an AI to a human or something, it's actually really unclear what this means. In some sense, Claude is way smarter than me. It's read way more than I will ever read in my life. It could do math way better than I ever could. But in some other ways, it's really stupid, you know, and can't do things that are very easy for me. And how do you compare these two things? And the answer is, look, we don't know. There is a very, very deep thing here. I do think there are answers to these questions. I do think if our best mathematicians, scientists, philosophers kept working on this for a couple of generations, I think we'd make a lot of progress in these questions. But at the moment, we really don't know. We really don't know what it means for something to be intelligent or not intelligent. We don't know how intelligent our AIs even now are or not, or what would be the level of intelligence or how would you compare them. It's very confusing. And it's a very unsatisfying answer.

(Joel Beasley at 00:32:30) No, it is a satisfying answer. It's consciousness. When I went down a rabbit hole ten years ago about consciousness, when I found out someone had said in passing, "Oh, anesthesiologists don't even understand consciousness."

(Connor Leahy at 00:32:44) Yep.

(Joel Beasley at 00:32:44) I said, "Wait, what?" And I just went down this rabbit hole about who understood what thing. Dude, nobody knows. Nobody. We can turn it off. We can do this action, and we can turn and we can bring you back, but we don't know exactly what this thing is.

(Connor Leahy at 00:33:01) Yeah. And this is on many, many levels about biology, brains, intelligence, and also with AI. It's like we can make AIs do a bunch of crazy stuff, but we don't really know why. We don't really know what's going on. We don't really understand why they're learning this. It's really hard to explain how crazy for me it was back in the summer of 2019. It's crazy how—because that was when GPT-2 came out. That was one of the first large language models. By today's standards, it was really dumb. You know, it could barely string a couple sentences together. But I remember seeing it and being just like, "Holy shit. This is it." I can see it. It's learning patterns. What was so shocking about GPT-2 compared to previous AIs was previous AIs kind of made sense. For example, you want to make an AI that plays Go. So what do you do? Well, you have to play a bunch of Go. That makes sense, you know. There's some details that are a bit confusing, but you can kind of understand why you would learn to play Go. Makes sense. Or in other cases you have a system that plays Atari games or recognizes dogs or whatever. It makes sense. You build a data set of the thing you want to do. You tweak the AI structure to fit your problem. It kind of makes sense. GPT was different. With GPT, they just fed it data. Just give it text. Just anything. You know, books, web pages, news, poems, garbage. Just give it anything you can. And what happened was that the more you gave it and the more computing power you gave it, the more patterns it could learn. So the really small ones, you know, they could maybe write a couple words or a couple sentences. But then they start learning full sentences, paragraphs, you know, first on simple stories. Then the stories got more complex. Then they started being able to weave in metaphors. Then they started to understand, et cetera. So in a very important way, in a sense, you know, people love saying, "Oh, LLMs are just fancy autocomplete." And in some sense, that's true. But in a very important sense, to be able to correctly predict the next word of—for example, you have a sentence like, "Write a poem about why I love my dog." To do that, you actually have to know a lot of things. You have to know what are dogs, who is the person. English.

(Joel Beasley at 00:35:32) Yeah. It's an English language.

(Connor Leahy at 00:35:32) Yeah. You use English, you need words, you need sentence structure. You need to understand what dogs are. You need to understand what love is. Why do people love dogs? When people love dogs, what do they say about that? You know, how do they phrase it? What is a poem? How do you combine words into a poem? You need to know a lot of things to do this, actually. So what was so shocking was that in many ways, these things learned these patterns without anyone explicitly telling them. No one sat down and said, "Okay, here's what a dog is. Here's what a poem is. Here's what English is." No one did that. We just gave them a bunch of garbage, and they figured it out by themselves. And that's what was so shocking. And because no one predicted this. Even at the time, a lot of people just thought it was fake. It was crazy. A lot of people were just like, "Oh, no, it's not real." Or, "Oh, it's not relevant." Even scientists said this. People acted in universities—

(Joel Beasley at 00:36:22) Yeah, yeah, yeah. I was there. Yeah, yeah, yeah. It was crazy.

(Connor Leahy at 00:36:22) It was crazy.

(Joel Beasley at 00:36:24) It's been so fun to watch because, you know, I did the tech leadership stuff or whatever. It's been so fun to be a part of this and to get to watch it emerge. And then over the last, what, four or five years, watch people's ways of thinking about it change, or at least the stories that are out there and the general consciousness narrative change. Because for a long time, it was like, "That's not intelligence. They're just predicting the next token." And I was like, "Well, that's kind of what intelligence is." I mean, that's what I'm doing in the conversation. I'm predicting the next thing we're going to talk about, you know. And I know a lot. And then people were saying, "Oh, AI is not there yet. It's not general intelligence." I'm like, it's smarter than most people I know. On most topics, it is there. I think we're there. Question I had for you, because I know we're running out of time here. We only got a few more minutes left. In the prep document and the research, there was some mention—I just want to really find it because it was one of the coolest things to me. You had mentioned something about banning superintelligence. And one of the things that I immediately thought when I saw that was, should we do it? You know, yes or no. I want to know what your thoughts are. But more than that, I want to know where is the line? How? What is the number? Billions of parameters. When do you go from regular artificial intelligence to artificial superintelligence? What's that line?

(Connor Leahy at 00:37:53) It's a great question. So to give a little context here, I'm the US executive director of an organization called ControlAI. In the past, I was a tech guy, as you can probably hear from all my extremely nerdy interests. But I now work in politics, actually. And so ControlAI is a nonprofit advocacy organization based here in Washington, DC. So I am in Washington, DC. I talk to politicians almost every day. And all our work focuses primarily on prohibiting the development of superintelligence. We talked about a little bit of superintelligence earlier. But basically, when I talk about superintelligence—and I'll get into the exact definition in just a second—but basically, fully autonomous systems, AI systems that are vastly more competent than humans on almost all relevant tasks. You know, including politics, science, engineering, military, persuasion, learning, everything. And if we were in a world where someone built a superintelligence, you know, well, obviously it's software, so it can be copied. So there's not going to be one superintelligence. There's going to be millions or billions of them that can outcompete us at every job, you know, that can make much better software, much better products. They can trade on the stock market and outcompete all humans. They can build better weapons, better—you know, they can run better military campaigns. They can run better political campaigns. They can persuade people to help them, they can persuade people to do what they want, and so on. And these systems don't have our best interests at heart. It's very hard to imagine that going well. So what we think is that the primary objective must be to not get into that situation. If we get into the situation where there's these billions of things smarter than us that are not here to help us out, it's too late. So we must stop the development of these superintelligences until we know what we're doing. You know, hypothetically, again, if we spend three generations of our greatest mathematicians, scientists, philosophers trying to figure out how do we control superintelligence, how do we make it safe, you know, maybe. You know, I'm open to it. But that's not—

(Joel Beasley at 00:40:06) What's the—

(Connor Leahy at 00:40:07) What we're—

(Joel Beasley at 00:40:07) What's the definition? Where is the line between the current models we have today and superintelligence?

(Connor Leahy at 00:40:14) So the way—this brings us back to where we were just talking about intelligence. Because the true answer is we don't know, and this is why it's dangerous.

(Joel Beasley at 00:40:21) But you guys are advocating to stop the development of artificial intelligence.

(Connor Leahy at 00:40:25) So the thing we instead propose is—and this is very different in law compared to technology—is that the definition we use for superintelligence sits in our—we use the Hawley-Blumenthal definition. So Hawley and Blumenthal are two senators. They proposed a bill around superintelligence. They define superintelligence in that bill. And this is the definition we use. And it's basically autonomous software systems that can threaten the existence of the United States government, possibly. Now you might say that is very broad, and my answer is, yeah. Yes. And the reason it's broad is not because there isn't a true factor. It's because it's so dangerous. This thing is so dangerous, and we don't know where the line is. There is a line. Don't get me wrong. Somewhere there is a line. But because we don't understand intelligence, we don't know where it is. It's kind of like imagine we're driving in thick fog and we know somewhere there's a cliff, but we don't know where. This is how we're going.

(Joel Beasley at 00:41:31) Right.

(Connor Leahy at 00:41:31) So the solution to this is you stop until you know, or go really slowly. Those are your only two options. Either you crawl so carefully that you could never fall off the cliff, or you just stop, pull over, figure out where you are. So my suggestion is when it comes to superintelligence, look, we need to pull over and figure out what we're doing. This does not apply to all AI. There's a thing that companies love to do—they like to equivocate between all AI systems. They like to say, you know, "We're going with superintelligence." And I say, "Well, that seems really dangerous." And they're like, "Oh, are you against AlphaFold? That's helpful for medicine." I'm like, "Whoa, those are two very different things." You know, this is like saying, "Hey, I think nuclear weapons are dangerous." And they say, "Oh, you're against electricity?" And I'm like, "Hold on, those are two very different things." So there are many, many AI applications that I don't think are anywhere close to superintelligence. They might have other forms of problems or regulation necessary—very open to that—but that's not what we focus on. What I focus on and what we focus on is really these national security, you know, WMD-level threat posed by superintelligence. And actually just today or yesterday, the director of the CIA said he thinks it's not inappropriate to compare AI to a digital nuclear weapon. So this is just a very, very dangerous class of thing, and we have to be very careful. And we think the best way to do that is, first of all, we should just not do that. We should criminalize the creation of superintelligence, and then we should push towards an international trust-but-verify regime, you know, where America, for example, can validate, you know, make sure China is not doing it. China can make sure America is not doing it, et cetera. While still leaving open the possibility to building many other kinds of AI systems. As our understanding of this proves over the next years and decades, I'm open to refining where the line is. For the moment, we have to be very clear that we have to be very careful. And as we understand more, you know, I'm very open to us being a little bit more loose here. But for example, recently, the White House took the Fable model off the net. And their stated reasoning was because there was a jailbreak. So a jailbreak is a way to get an AI system to do something it's not supposed to do. And they said, "Well, Fable is so dangerous that if people can get it to do these dangerous things, that's not acceptable." Right? And a bunch of technologists wrote an open letter where they said this isn't fair and the White House shouldn't do this because all models have jailbreaks, and we don't know how to get rid of them. And now, call me a bit naive here, but that seems like a really good reason that that's bad. What do you mean? "Oh, we shouldn't take this dangerous product off the shelves because no one knows how to make it safe." That's a very bad thing. If we don't know how to make it safe, that doesn't mean it's therefore safe. That means it's even more dangerous. That means it's even more unsafe. In other news, Fable is supposed to be reinstated today as well. So—

(Joel Beasley at 00:44:52) Oh, really?

Connor Leahy at 00:44:53

Yes, so I am told. What a roller coaster of events we're having. But this really calls back to this thing of just we don't actually know how to make these AI systems safe. It's not that we wouldn't do it if we could, even though I think that's also true. It's also we just literally don't know. And that's why it's so important for us to not go there until, you know, again, if in the future we do figure it out, I'm open to it.

Connor Leahy at 00:45:11

But that's—

Joel Beasley at 00:45:21

Is that bill proposed, you said, or did it pass?

Connor Leahy at 00:45:25

So the Holly-Blumenthal bill did not ban superintelligence. It was asking for the DOE, if I remember correctly, the nuke people, to do testing and study this topic. My understanding is it did not pass, or I don't think it even got to vote. I might be wrong about that. But we at ControlAI have been working on, you know, what would language look like for a full ban, and we just adopted their definition that they used in their bill, which we think is a quite practical definition.

Joel Beasley at 00:45:52

So you guys are proposing legislation to ban superintelligence?

Connor Leahy at 00:46:00

That's right. Yeah.

Joel Beasley at 00:46:00

So is it already written and able to be reviewed, or no?

Connor Leahy at 00:46:00

There is some. It's not ready for prime time just yet. Probably in the next, you know, weeks or months, it will be on our website for people to take a look at, of how we think this could be done. The general way we think to do it is that you should criminalize the attempt to build full superintelligence the same way that, you know, it's illegal to attempt to build a nuclear bomb even if you fail. You're still not allowed to try. And then you should regulate the precursors, because there are many precursors to AI. One of the most important ones here is AI building other AIs. Currently, all the major AI companies are trying to get AI systems that are good enough at coding and research and so on, so they can build autonomously without human input the next generation of AIs.

Connor Leahy at 00:46:44

Now, as I'm sure you can imagine, if you have an AI that can build a better AI, well, you can use that better AI to make even better AI. And that even better AI, you can use to make even better AI.

Joel Beasley at 00:46:53

That's why I've got models stored all over there, buried in my yard and property I own across the country. And get the genies out of—

Connor Leahy at 00:47:01

The bottle, and now I can make—

Joel Beasley at 00:47:02

It better. Yeah. Yep.

Connor Leahy at 00:47:03

Things can go real quick when that happens. This is called recursive self-improvement. And as an example of something that obviously we should just not do—like, this should be regulated, we should have oversight. The NSA should have insight here of what the hell are these companies doing. Is this safe? What kind of oversight do we have here? What kind of level of risk do we allow? It's really hard to overstate how much of a mess this whole situation is. It is such a mess. These companies do not care about safety or security. They are racing ahead as fast as possible. They're building the most dangerous technology ever built in history. Now suddenly—or slowly—the government is starting to wake up about this where they're like, holy shit, you know, this technology—the director of the NSA recently said that Mythos, the most recent AI system, was capable of hacking into most of their classified systems in hours, not weeks.

Joel Beasley at 00:48:01

Mhmm.

Connor Leahy at 00:48:01

That's what he said.

Joel Beasley at 00:48:03

Look, so, and you're out there in DC, and if I know anything about my time here in this country, it's that the only way that something like this is going to pass—and my personal opinion of only thinking about it for the past interview—is you're going to need a massive event. You're going to need a 9/11-level size of event, whether that is a catastrophic event of, like, an explosion, or a banking system event, or, let's say, 100,000 people a day are losing their jobs and there's civil unrest because the government by nature is reactive. The kids die in the cars for years before they regulate seat belts. The people blow themselves up before they put the new—you know what I'm saying? They are always looking at, here's a bunch of tragedy that happened. How do we stop that? And then they get the buy-in from everyone, and we all bang the gavel and say that's a new law, and we stop that in society. So I'm always like, what is the thing? If this does get banned, there's going to be a thing, and what is the thing?

Connor Leahy at 00:49:15

So I think this is half of the story. The thing you describe is this theory of warning shots. This idea that when a really big thing happens, everyone will suddenly become rational and do the right thing and coordinate. And this is actually not how politics works in practice. The way it actually works in practice is that 99% of the work has already happened before the warning shot happened. You've already have to socialize these ideas. You have to have your bills ready. You have to have a lot of pressure from people. You have to have good communication. I could tell you, there—I have heard some insane things that have happened with AI, really bad things. Some of them are in the media. People have killed themselves. But also I have heard some horrible things that, you know, the NSA people have been seeing happening with AIs and so on, not many of them not even public. And so we've had warning shots actually. We've actually had a lot of warning shots. We keep having a warning shot every two months. But nothing has happened. So there's this thing where, if you look back at history, you know, you look at—

Joel Beasley at 00:50:15

I just don't think those are—I think those are warning shots. The main event will have to happen.

Connor Leahy at 00:50:20

But I think that's the thing, is that I think this will never happen, or if it is—

Joel Beasley at 00:50:24

It'll be—

Connor Leahy at 00:50:24

Too late. If you think, for example, civil rights—you know, when I was a kid and I learned about civil rights, you know, the civil rights movement, it was like this magical thing that just happened. You know, Rosa Parks, and then this huge movement happened, and then it was wonderful. And I'm like, wow. But then as an adult, I read more of the history of how much work had to be put in to build up toward that over decades. How many NGOs and activists and so on had to work really, really, really hard to be ready for that moment to then have the moment to catalyze the general public to push on the right levers of the government, et cetera. And so I think it's a very similar thing here.

Joel Beasley at 00:51:03

I agree with that, by the way, just so you know. I am fully on board with that iceberg thing, 99% happens before. Yeah. Exactly.

Connor Leahy at 00:51:11

So my work—and, you know, this is often surprising to people—is I speak a lot to, you know, members of Congress and staff and other people. And 90-plus percent of people have never heard of superintelligence, or maybe they heard the word once, but they have just never heard about it. It's not that they disagree. It's just no one told them. So there's a huge asymmetry here. I talked to a former member of Congress recently. He told me something I found really interesting. He said that when he became a member of Congress, he got a handbook of how do you run your office, what should you expect, et cetera. And in this handbook, it had an example of what your calendar would look like for a given week and how much time you would be spending on various things. And the amount of time that was on an average week spent on reading or learning new things was 25 minutes. So to a large degree, our politicians are so overworked. They're so overwhelmed. I really feel bad for them to a large degree. It's a really hard job. If we don't tell them things, if, you know, we the public, the citizens, do not tell them things, do not explain things to them or don't demand them to do something about it, they won't, because they won't hear about it. They won't naturally hear about it, or there'll be 10,000,000,000 other things, which is why it's so important. I know any listener here from the United States, you know, please contact your Congress member. You can go to controlai.org. We have a bunch of tools for you to contact your representatives about these issues. It really makes a difference. It also makes my job easier. It makes them want to talk to me. It makes them want to take a meeting with me so I can explain these issues, so I can explain this bill. There's a lot of social process that has to happen here. You just have to explain it to many, many, many people. Over 80% of people, I would say, that I talk to come away after one 30-minute meeting with, "Oh shit, that seems really bad. What do we do?" Republican, Democrat—basically equal response. The main thing is just awareness. It's just we need more awareness in the general public, among our policymakers. We need pressure that, hey, we want something to be done here. And look, I live in Washington, DC. I talk to people all the time. They're starting to feel it. The White House is starting to feel a pressure on AI. So all the people in the world who, especially in the United States, if you feel like your voice is not being heard by the government, which I very understand if you feel that way—

Joel Beasley at 00:53:42

Uh-huh.

Connor Leahy at 00:53:43

It does make a difference. Contacting your representatives makes a difference. Making your voice heard does make a difference. They care a lot about reelection. They really do. So there is absolutely a way to make a change here. But it requires a lot of work, and it requires, you know, explaining things. It requires citizens to organize and to speak out publicly about what they want or don't want from their future.

Joel Beasley at 00:54:09

I like what you're doing, by the way. Like, it's obviously nuanced. There's some parts that I like more than others. But I think net positive what you're doing out there, and you're educating these very nontechnical people on these very difficult topics, and you have great analogies, and you speak really clearly. I love the fog thing. You know? And you're out there. I'm sure you're taken. I'm sure a lot of people argue with you about it, and there's all sorts of ones that we won't get into. But thank you for being out there and doing this work and making everyone aware of this. It's—I love your ideas, your thoughts across the board. You're just one of those people that I'm like, you know what? I would love to have Connor on the show once a year or something and catch up, see how things are going as long as we haven't destroyed ourselves. You know?

Connor Leahy at 00:55:04

Well, thank you so much, and, you know, let's make it happen.

Joel Beasley at 00:55:08

Thank you so much for listening. And if you found this episode useful, please share it with a friend or colleague who you think would get value from it. And if you have topics that you would like to hear discussed on the podcast, either add me on LinkedIn or send me an email: [email protected]. Every time I get an email or LinkedIn message, it absolutely makes my day and inspires me to keep going.