[0:06]Connor Leahy is one of the world's leading minds in artificial intelligence. He is a hacker who sees the rise of AI as an existential threat to humanity.
[0:17]He dedicates his life to make sure its success doesn't spell our doom.
[0:24]There will be intelligent creatures on this planet that are not human. This is not normal. And there will be no going back.
[0:39]And if we don't control them, then the future will belong to them, not to us.
[0:46]Leahy is the CEO of Conjecture AI, a startup that tries to understand how AI systems think with the aim of aligning them to human values.
[0:59]He speaks to The Inner View about why he believes the end is near and explains how he's trying to stop it.
[1:20]And Connor Leahy joins us now on The Inner View. He's the CEO of Conjecture. He's in our London studio. Good to see you there. Good to have you on the program, Connor. You're something of an AI guru. And you're also one of those voices saying, we need to be very, very careful right now. And a lot of people don't quite have the knowledge or the they don't quite have the vocabulary or the deeper understanding as to why they should be worried. They just feel some sort of sense of doom, but they can't quite map it out. So, maybe you can help us along that path.
[2:00]Why should we be worried about AGI? And tell me the difference between AGI and what is widely perceived as AI right now?
[2:11]So, I'll answer the second question first, just to get some definitions out of the way. The truth is, is that there is really no true definition of the word AGI, and people use it to mean all kinds of different things.
[2:22]When I talk about the word AGI, usually what I mean by this is AI systems or computer systems that are more capable than humans at all tasks that they could do. So, this involves, you know, any scientific task, programming, remote work, uh, science, business, politics, anything.
[2:40]And these are systems that do not currently exist, but are actively attempting to be built. There are many people working on building these systems, and many experts believe these systems are close. And as for why these systems are going to be a problem, well, I actually think a lot of people have the right intuition here.
[3:00]The intuition here is just, well, if you build something that is more competent than you, it's smarter than you, and all the people you know and all the people in the world, it is better at business, politics, manipulation, deception, science, weapons development, everything, and you don't control those things, which we currently do not know how to do, well, why would you expect that to go well?
[3:23]Yeah. It reminds me a little bit about the debate about whether we should be looking for life in, in the universe beyond our solar system. Stephen Hawking said, be careful. Look at the history of the world. Anytime you sort of invite us a stronger power, more, more competent power, they might come and destroy you. But then the counter to that is that you're mapping human behavior, human desires, passions, needs, wants, onto this thing. Is this natural to do and fair to do because humans created it? Humans, humans created the parameters for it.
[3:59]So, it's actually worse than that. In that it's really important to understand that when we talk about AI, it's easy to imagine it to be software. And the way software generally works is it is written by a by a programmer. They write code, which tells the computer what to do, step by step. This is not how AI works.
[4:21]AI is more like organic. It is more like it is grown. You use these big supercomputers to take a bunch of data and grow a program that can solve the problems in the data. Now this program does not look like something written by humans. It's not code, it's not lines of instructions.
[4:41]It's more like a huge pile of billions and billions of numbers. And we know if we can run all these numbers, if we execute these numbers, they can do really amazing things. But no one knows why. So the question is less, you know, will humans impart emotions into these systems? We don't know how to do that. It's more, if you build systems, if you grow systems, if you grow bacteria, who are designed to solve problems, to, you know, solve games, to make money or whatever. What kind of things will you grow? And by default, you're going to grow things that are good at solving problems, at gaining power, at tricking people, at, you know, building things and so on, because this is what we want.
[5:41]You reverse engineered GPT2 at the age of 24, which was a few years ago. That's part of the legend. I mean, that's part of the, the credentialing of you before they say, well, this guy's saying we're in big trouble. They say, well, by the way, you know, he knows what he's talking about because technically he knows what he's doing. Tell me, tell me about the pivot point between being a believer or an enthusiastic about this to becoming a warner. What happened?
[6:12]So, uh, the story goes back even further than that. Um, reverse engineering is, um, a bit generous. It's more like I built a system and I found out that no one can reverse engineer it. And this is a big problem. Um, but it was even before that. So, I've been very into AI since I was a teenager, because I want to make the world a better place. And I think a lot of people who believe in AI, a lot of tech people who are doing the things that we think are dangerous, I think most of them, maybe not most, but most of them, probably, are good people. They're trying to build technology to make the world a better place. You know, I, when I grew up, uh, technology was great. You know, the Internet was making people more connected. We were getting access to better medicines, and there was, you know, solar power was improving, like all these great things that science was doing. So I was very excited about more science and about more technology. And, well, what is the, what is the best technology then intelligence? If we just had intelligence, well, wow, we could solve all the problems. We could do all the science. We could, you know, invent all the cancer medicines. We could, you know, develop all the cool stuff. So I was thinking when I was a teenager. And this is I think a common trajectory is that people when they're kind of like first exposed to some of these like techno utopian, AGI dreams, it sounds great, you know, it sounds like such a great, great solution. But then as you think about this problem more, you kind of realize that like the problem with AGI is not really how to build it. It's how to control it. That's much harder. Just because you can make something which is smart or that solves a problem, does not mean you can make something that will listen to you or that will do what you truly want. This is much, much harder, and this is and as I started looking into this problem more in my early twenties, I started realizing like, wow, we are really, really not making progress on this problem.
[7:58]So in that worst-case scenario, whether we have an apocalyptic ending for all of us, we get destroyed existential or we become enslaved in The Matrix, or whatever it might be. Tell me how it actually happens in your mind. How does this AGI, um, assume control? I mean, there there's these famous moments in Terminator and elsewhere. One of the Terminators, that final scene where the nuclear bombs are going off all over. I mean, there are lots of different ways people have imagined this. The way you see it, tell me how it happens and how if things continue to go in in the direction that you fear, how long will it take to get there?
[8:41]Well, of course, I don't personally know how exactly things will play out. I can't see the future. I can give you a feeling, though, of how I expect it to feel, how to expect it to feel like when it happens. The way I expect it to feel is kind of like if you play chess against a grandmaster. Now, I'm really bad at chess. I'm I'm not good at chess at all. But, I, you know, I can play, you know, a little bit of an amateur game. And then, but when you play against a grandmaster, or someone who's much, much better than you, the way it feels is not like you're having a heroic battle against the Terminator. You're having this incredible back and forth, and then you lose. No, it feels more like you think you're playing well, you think everything is okay, and then suddenly you lose in one move. And you don't know why. This is what it feels like to play chess against a grandmaster. And this is what is going to feel like for humanity to play against AGI. What's going to happen is not some dramatic battle that, you know, the Terminator will rise up and try to destroy humanity. No. It will be things get more and more confusing. More and more jobs get automated, faster and faster. More and more technology gets built, which no one even quite knows how the technology works. There will be mass media movements that don't really make any sense. Like, do we really know the truth of what's going on in the world right now, even now with social media? Do you or I really know what's going on? Well, how much of this is fake? How much of it is generated, you know, with AI or other methods? We don't know. And this will get much worse. Imagine if you have extremely intelligent systems, much smarter than humans, that can generate any image, any video, anything, trying to manipulate you. Well, and being able to develop new technologies, to interfere with politics. The way I expect it will go is that things will seem like mostly normal, just like weird, just like things are getting weirder and weirder. And then one day, we will just not be in control anymore. Just it won't be dramatic, there won't be a fight, there won't be a war. It will just be one day, the machines are in control, and not us. And even if there is a fight. So, sorry to interrupt you. Even if there is a fight or a war, they've handed us the gun and the bullets, and we've done it. I mean, it's us that might, might do all of this, precipitated by being controlled in some way.
[10:56]No, absolutely possible. I don't think an AI would need to use humans for that, because, you know, it could develop extremely advanced technology, but it's totally possible. Humans are not secure. It is absolutely possible to manipulate humans, like, you know, everyone knows this. You know, humans are not immune to propaganda, not immune to mass movements. Imagine if, you know, an an AGI gives Kim Jong-un a call and says, hey, I'm going to make your country run extremely well and tell you how to build super weapons. In return, do me this favor. I mean, Kim Jong-un's going to think that's great. And it's very easy to gain power. If you're extremely intelligent, if you're capable of manipulating people, of developing new technology or weapon, trading to on the stock market to make tons of money, well, yeah, you, you can do whatever you want.
[11:41]So, you're sounding the alarm. Jeffrey Hinton, seen as the founder or father or godfather of AI, he's sounding the alarm and has distanced himself from a lot of his previous statements. Others in the mainstream are coming out, heavily credentialed people who are the real deal when it comes to AI, are saying we need guardrails, we need regulation, we need to be careful, maybe we should stop everything. Yet, Open AI, Microsoft, Deep Mind, these are companies, but then you have governments investing in this. Everybody's still rushing forward, hurtling forward, towards a possible doom. Why are they still doing it despite these very legitimate and strong warnings? Is it only about the bottom line and money and competition, or is there more to it?
[12:34]This is a great question, and I really like how you phrase, you said, they were rushing towards. Because this is really the correct way of looking at this. It's not that it is not possible to do this well. It is not that it's not possible to build safe AI. I think this is possible. It's just really hard. It takes time. It's the same way that it's much easier to build a nuclear reactor that melts down than to build a nuclear reactor that is stable. Like, of course, this is just hard. So, you need time and you need resources to do this. But unfortunately, we're we're in the situation right now. We're currently in a situation right now where, at least here in the UK, there is currently more regulation on selling a sandwich to the public than to develop potentially lethal technology that could kill every human on Earth. This is true. This is this is the current case. And a lot of this is because of slowdown. It's just, you know, governments are slow. People don't want, and vested interests. You make a lot of money by pushing AI. Pushing AI further makes you a lot of money. It gets you famous on Twitter. You know, look how much these, like, these people are rockstars. You know, people like Sam Altman are rockstars on Twitter. You know, people love these people. They're like, oh, they're they're bringing the future. They're making big money, so they must be good. But, like, I mean, it's just not that simple. Unfortunately, we're in a territory where we all agree, somewhere in the future, there's a precipice. Which we will fall down if we continue. We don't know where it is. We don't, maybe it's far away. Maybe it's very close. And my opinion is, if you don't know where it is, you should stop. Well, other people, who, you know, gain money, power, or just ideological points, like, a lot of these people is very important to understand, do this because they truly believe, like a religion. They believe in transhumanism. In, in the glorious future where AI will love us and so on. Like, so there's many reasons. But I mean, yeah, I mean a cynical take is just I could be making a lot more money right now if I was just pushing AI. I could get a lot more money than I have right now.
[14:39]How do we do anything about this without just deciding to cut the under sea internet cables and blow up the satellites in space and just start again? How do you actually, because this is a technical problem and it's also a moral and ethical problem. So, where do you even begin right now or is it too late?
[15:01]So, the weirdest thing about the world to me right now, as someone who's deep into this, is that things are going very, very bad. We have, you know, crazy, you know, just corporations with zero oversight just plowing billions of dollars into going as fast as possible, with no oversight, with no accountability, which is about as bad as it could be. But somehow we haven't yet lost. It's not yet over. It could have been over. There's many things where it could be over tomorrow, but it's not yet. There is still hope. There is still hope. I don't know if there's going to be hope in a couple years, or even in one year, but there currently still is hope. Oh, wait, hold on. One year. I mean, that's, come on, man. I mean, we're probably going to put out this interview like a couple of weeks after we recorded it. A few months will pass. We could all be dead by the time you, there gets 10,000 views. I mean, just for, just for, like, explain this timeline. One year. Why one year? Why, why is it going so fast that even one year would be too far ahead? Explain that.
[16:05]I'm not saying one year is like guaranteed by any means. I think it's un un-unlikely. But it's not impossible. And this is important to understand. Is that AI and computer technology is an exponential. It's like COVID. This is like saying in February, you know, a million COVID infections, that's impossible. That can't happen in six months. And it absolutely did. This is kind of how AI is as well. Exponentials look slow. They look like you go one infected, two infected, four infected. It's not so bad. But then you have 10,000, 20,000, 40,000, you know, 100,000, you know, within a single week. And this is how this technology works as well. Is that as our computers get there's something called Moore's Law, which is just it's not really a law. It's more like an observation that every two years, our computers get about, you know, there's some details, but about twice as powerful. So, that's an exponential. And our technolo and it's not just our computers are getting more powerful. Our software is getting better. Our AIs are getting better. Our data is getting better. More money is coming into this field. We are on an exponential. This is why things can go so fast. So, while I'm not, like, I, you know, it would be weird if we would all be dead in one year, it is physically possible. You can't rule it out if we continue on this path. The powerful people who can do something about this, especially when it comes to regulation, when you saw those congressmen speaking to Sam Altman, they didn't seem to know what the hell they were talking about. So, how frustrating is it for you that the people who can make a difference have zero clue about what's really going on? And and more important than that, they didn't seem to want to actually know. They had weird questions that made no sense. And so, you're thinking, okay, these guys are in charge. I mean, no wonder the AI is going to come and wipe us all out. Maybe, maybe we deserve this.
[18:05]Well, I wouldn't go that far. But, um, this used to annoy me a lot. This used to be extremely frustrating. Um, but I've come to I've come to peace with it to a large degree. Because the thing that I've really found is that understanding the world is hard. Understanding complex topics and technology is hard. Not just because they're they're complicated. But also because people have lives, and this is okay. This is normal. People have families. They have responsibilities. They have There is a lot of things people have to deal with. And I don't shame people for this. You know, like, you know, I have turkey, you know, with my family over Thanksgiving or whatever. And, you know, my aunts and uncles, look, they have their own lives going on. They maybe don't really have time, you know, to listen to me give them a rant about it, so I don't. So, I have a lot of love and a lot of compassion for the things that are hard. This is, of course, doesn't mean that that solves the problem. But I'm just trying to say that, like, it is, of course, frustrating to some degree that there are no adults in the room. This is how I would see it. Is that there is sometimes a belief that somewhere, there is someone who knows what's going on. There's an adult who's got all of it under control. You know, someone in the government, they've got this under control. And as someone who's tried to find that person, I can tell you this person does not exist. The truth is, is the fact that anything works at all in the world is kind of a miracle. It's kind of amazing that anything works at all with how chaotic everything is. But the truth is, is that there are quite a lot of people who like who want the world to be good. You know, they might not have the right information. They might be confused. They might be getting lobbied by various people with bad intentions. But, like, most people, most people want their families to live and have a good life. Most people don't want bad things to happen. Most people want the other people to be happy and safe. And luckily for us, most normal people, so not elites, not necessarily politicians or technologists, most normal people do have the right intuition around AI. Yeah. Where they see, like, wow, that seems really scary. Let's be careful with this. And this is what gives me hope. So, when I think about politicians and them not being in charge, I think this is now our responsibility as citizens of the world, that we have to take this into our own hands. We can't wait for people to save us. We have to make them save us. We have to, you know, we have to make our voices heard. We have to say, hey, how the hell are you letting this happen? Like, we're, one of the beautiful things is that, you know, to a large degree, politicians can be moved. They can be reasoned with and they can be moved by the voters. You can vote them out of office. So, that's a good argument for democracy. That's that's a great argument for democracy. That's that's wonderful.
[20:44]You know, democracy is the worst system, except for all the other ones. Yeah. Um, so to the point of people's feeling and and I asked about this at the very beginning that intuitive feeling of like, something's up here. There's something ominous. There did seem to be a little bit of a plateau with something like ChatGPT. So, initially, people were very anxious, very surprised, but very wowed by what this thing could do. Could write your university thesis, whatever. It could, you know, do all these these fancy gimmicks. They seem like magic tricks. But then, once the hype died down a little bit, uh, people began to input new things, ask maybe better questions, and you could see some of the limitations of something like, you know, ChatGPT and its, uh, forerunners. And that led to, and this is anecdotally, a lot of friends of mine going, oh, maybe this thing, maybe we're okay for a while because this thing has severe limitations. Address that for me because a lot of people are still sort of like, well, I know there was the hype, but now I'm not so sure. Tell me about that.
[22:06]So, there is a story. I'm not sure if the story is actually true or not, but it's a good metaphor. Where if you take a frog and you put it into a pot of water, you know, cold pot of water. The frog will sit there happily. If you slowly then turn up the heat on your pot, the frog will sit there, there's no problem. And if you do it very slowly, if you very slowly, slowly increase the temperature, the frog will get used to the temperature and won't jump out until the water boils and the frog dies. I think this is what is happening with people. Is that, um, people are extremely good at making things which are crazy normal. Is that if it's a normal thing, if it's a thing all your friends do, then it just becomes normal. This is like during war, why people can slaughter other people. It's because if all your friends are doing it, well, it's normal. It's like, yeah, you slaughter people, it's normal. Yeah. Killing people's fine. This is how it can happen. And the same thing applies here. Is that, well, okay, you can talk to your computer now. Like, sure, we can argue about, oh, ChatGPT, it's not that smart. You could talk to your computer, like, slow down. If this was a sci-fi movie from 20 years ago, everyone would be yelling at the screen, like, what the hell are you doing? Like, this thing is obviously, like, crazy, like, what the hell is going on? But because it's, you know, available now, you know, cheaply online, it doesn't feel special. So, the way to address this is, I think, a lack of coordinated campaigning effort. And what I mean by this is, is that the general, when we think about our civilization, not just individual people. When we think about our civilization, how does our civilization deal with problems? How does it decide which problems to address? Because there's always so many problems you could be putting your effort on. How does it decide which one to pay attention to? And this is actually very complicated. And it can be because of a natural catastrophe, or a war or whatever. It can be because of some stupid fashion hype, just like some viral video on TikTok makes everyone freak out. Sometimes, yes. But usually, if you actually want your civilization to address a problem, a big problem, it takes long, hard, grinding effort from people trying to raise this to salience, to raise it to attention. Because again, people have lives. You know, like, most people don't have time to go go online and read huge books about AI safety and like, oh, how do we integrate ChatGPT or how do we deal with like, the safety threats? They don't have time for that. Of course, they don't. And I'm not trying to judge these people. I understand. It's not their job. In a good world, there should be a group of people that deals with this. The problem is they don't really exist. Right. Before we go, I'm glad you mentioned that people don't know where to look. If there was one resource that you could point people in the direction of, so that they can educate themselves about the reality of this situation and can bring themselves up to speed, that would be what?
[25:14]There's not one who I think has the whole thing, which is a big problem. Someone should make that resource. If someone made that research, please let me know. But what I would probably point people towards is control AI, which is a group of people who I'm also involved with, who are campaigning for exactly these issues, who are trying to bring humanity together to solve these problems. Because this is a problem that not you or me can solve. No human can solve these problems we're dealing with right now. This is a problem that humanity has to solve. That our civilization needs to solve. And I think our civilization can do this. But it won't do it without our help. It won't help, it won't happen without us working together. So, if there's one thing I can go on Twitter or Google or whatever, go to control AI and support them, you know, listen to what they have to say. And this is the campaign that I'm behind as well. I'm, I support them. Okay, we'll put the link also in the YouTube description, uh, if anybody wants to check it out. Connor, you have a brilliant mind and I'm really grateful that we got to talk. Thank you very much for joining us on the interview. Thank you so much. Take care.
