This Is How AI Starts a Nuclear War

Conversations with Coleman

In this episode of Conversations with Coleman, Robert Wright discusses his new book, The God Test: Artificial Intelligence and Our Coming Cosmic Reckoning. He explores how AI represents an existential earthquake in human

Key takeaways

• AI's advancement in language and coding represents a paradigm shift beyond mere automation—these systems exhibit emergent cognitive abilities that mimic human thought processes.
• The rise of AI is not just technological but existential; it demands a global moral upgrade to prevent conflict and ensure humane outcomes.

Main topics

• AI as an existential earthquake
• Emergent intelligence in large language models

Notable quotes

"We may be building this, you know, global superintelligence that could wind up running the planet possibly to our great disadvantage. If that's the case, then we should think about how you build a good god."

Conclusion

Robert Wright concludes that AI is not merely a tool but a profound challenge to humanity's moral and social

Transcript preview

Speaker 1 (0:00) Welcome to another episode of Conversations with Coleman. My guest today is Robert Wright. Robert is a journalist and best-selling author whose work explores evolutionary psychology, philosophy, religion, and more. He's best known for his books The Moral Animal, Non-Zero, The Evolution of God, and my personal favorite, Why Buddhism is True. But he has a new book out, and that is the topic of today's conversation. It's called The God Test, subtitled Artificial Intelligence and Our Coming Cosmic Reckoning. This is a book about how AI is going to impact the future of everything we care about. It's also, by the way, a great book for people who want to understand how large language models actually work, but don't want to get lost in a bunch of comp sci jargon. Robert is among the best at making complicated scientific mechanisms easy to digest. I've been a fan of Robert's for a very long time, so it was great to get him on the podcast. So without further ado, Robert Wright. Speaker 1 (1:13) Okay, Robert Wright. Thanks so much for coming on the show. Hey, thanks for having me, Coleman. I forget. Have we both been on each other's podcasts? I've definitely been on yours. I think this is my Speaker 2 (1:24) debut on your podcast, unless I'm, you know, forgetting some major phase of my life, which is possible. Speaker 1 (1:32) Yeah, well, I've been following you for a long time. You were very early in the podcasting game. Is that... correct? Blogging Heads? In Speaker 2 (1:42) a sense, yeah. Blogging Heads was the first video dialogue platform. Certainly, there was devoted to discussion of politics, public affairs. A lot of the then semi-obscure bloggers who are now famous, like Matt Iglesias and Ezra Klein, were on the platform. And there was an audio download from the beginning. Now, it wasn't really a podcast per se, because the audio wasn't distributed via RSS, probably. And I think that's... in a technical sense, the definition of the origins of the podcast per se. I don't know. But yeah, 2005, you could listen to us, watch us. Mickey Kaus was then my, he was my co-founder and sparring partner frequently. Speaker 1 (2:28) Yeah. And you've written many books, Non-Zero, Why Buddhism is True, which I think I gave you a bunch of praise for last time, a book I really like. Today, we're talking about your new book, The God Test. This is a book about artificial intelligence, how it's going to change the landscape of our social lives, of everything, really. I'm curious, why did you choose the title The God Test? Yeah, for three reasons, actually. First Speaker 2 (3:02) of all... I think AI, I mean, I should say, first of all, I think AI is going to be an earthquake. I'm on that side of the argument. It's going to be huge. I try to explain in the book why the capabilities have advanced so far, so fast, and why, once you understand that, you've got to expect them to keep advancing. So I'm on, like, team earthquake, and I argue that we face, which is to say humankind faces, the kind of test... God would give for one thing, because I think we can only handle this technological challenge, navigate the revolution to a good end, if we approach it as a global community with much less in the way of division and war. And to do that, I think we're going to need to undergo something like a moral upgrade. I mean, individual human beings. Speaker 2 (4:00) are going to have to get better at. you know, understanding the perspective of adversaries and adversarial groups and so on and work their differences out. So there's that. And, you know, I guess that idea got just in that sense is kind of reinforced by my taking a somewhat cosmic perspective because I do frame the AI revolution ultimately. In the context of the entire, you know, 3.5 billion year history of life on Earth, you know, what does this represent in that context? In other words, as not just a threshold in technological history, but the history of organic life. God test also refers to this idea that we may, and I'm kind of agnostic on this. I mean, I don't rule it out. And I look at. the idea closely in the book and unfortunately kind of find myself unable to dismiss it. We may be building this, you know, global superintelligence that could wind up running the planet possibly to our great disadvantage. If that's the case, and, you know, people who have this view refer to this as a god. If that's the case, then we should think about how you build a good god, like one we can get along with, you know. Speaker 2 (5:26) And, you know, then finally, I do, as I've done before, fool around with the possibility that evolution and subsequent, you know, so-called cultural evolution, that is to say this process that human intelligence has launched that includes technological evolution, politics, religion, everything, that whole sweep of evolution. all the way from primordial ooze up to kind of the brink of, I would say, the global social organization of humankind and even, in a sense, the construction of a global brain. I fool around with the possibility that that represents the unfolding of some larger purpose, possibly divine purpose, possibly some other one, but it has a kind of directionality to it that I think makes that speculation non-crazy. And I should say that's compatible with a materialist view of the mechanics of the unfolding of these evolutionary processes. So anyway, is that enough meanings of the term God test for you? Or would you like me to go on for another hour? Speaker 1 (6:33) No, no, that's good. So let's back up to that very first premise before we kind of go through that bit by bit. First premise is that AI is going to be an earthquake. Speaker 1 (6:47) implicitly you are arguing against the people here that believe AI is an interesting technology that isn't going to fundamentally change everything. And so to give them their due and make their argument a little bit, it's like, okay, AI's got better than us at chess 20 years ago at some level. You know, chess was something we thought only humans could do. It turns out, you know, we can make machines that do it better. Big whoop changed the chess world for sure. Math is, you know, 100 years ago, 200 years ago, I think everyone would have thought math is the special province of human intelligence and one of the great tests separating human beings from other lower life forms. Speaker 1 (7:38) And then you invent the calculator and you invent things that can do math way more reliably than certainly the average human and even the best humans, even if they're not creative. It changes math. It doesn't change, you know, it changes certain elements of the world, but in the mundane way, in the way that any breakthrough technology does. What is it about language and coding, which are really the two... breakthrough technologies in AI of the past five years, language and coding. What is it about those specific skill sets that is going to make these AI breakthroughs an earthquake as opposed to just a really cool innovation? Speaker 2 (8:25) Okay, let's start with the chess thing. So as it happens, when the IBM computer beat the world chess champion, Garry Kasparov, in, I think, 1999, Time Magazine asked me to write their cover story about it. And I did, but I really didn't consider it interesting, what was happening then in AI. In fact, I steered the story away from... AI in a sense, and just made it a story about consciousness, not really about cognitive processes, because there was just so much kind of brute force in IBM's computer. I think it was named Deep Blue. And I didn't think what was going on inside the machine was very much like what's going on in the human mind. So I've been writing about technology, not only technology, for a very long time. In fact, you know, for people who are listening to this and haven't. assessed my age on the basis of my looks. Let me say I, in 1983, interviewed the guy who is now most commonly identified as the, quote, godfather of AI. He himself would, you know, emphasize the contributions of others, but he's as good a candidate as any, Jeffrey Hinton. I was writing a piece on AI for a journal called the Wilson Quarterly. And, you know, I. So I've been following this a long time and it was he he then represented this maverick view. And in retrospect, I did not understand it. His is the view that prevailed in those days. It was certainly not mainstream. It's referred to as deep learning, you know, neural networks, massive parallelism or some terms associated with it. And when the LLM revolution broke like. So in early 2023, I guess, I went back and read the piece I had written. And then I went back and listened to a lecture he had given recently. And I realized there's something I didn't understand in 1983 about his approach. And this gets at exactly your question of why this is different. And it's that. In that piece, I described a model that would be dealing with language. And it was a neural network in some sense of the word. In fact, it had been described by a guy who had co-authored something with Hinton. But he was a psychologist. He wasn't really an AI guy. So the model I was using was misleading. But in any event, the key thing about this model is that... you know, it did have a way of representing the meaning of words, but... In this case, humans would have to transplant into the model their prior understanding of the meaning of words. They'd have designed the system of representing meaning of words and feed in their knowledge that like this word has two different senses and just build this elaborate network architecture that reflected our understanding. And it turns out that. You don't have to do that. And this applies to everything, not just language, but it's the reason self-driving cars are getting so good. The same thing works with video, whatever. It works with robotics. And the key thing is this. You feed certain kinds of data into the machine that are... the kinds of data humans produce, or the kinds of data that humans get before producing the data. So in the case of an LLM, it's a human sentence, and the machine's job is to guess the next word. You know, in the case of a self-driving car, the machine's job is to, looking at this scene, do what a human would do, move the steering wheel to the right or something. But the key thing is this. All you have to do is feed it that data. You don't have to... tell what data is. And with these models, it will like reverse engineer the cognitive and perceptual mechanisms that are at least functionally comparable to the mechanisms humans use. So, and this is not emphasized. It's, it's, there are maybe a few AI researchers who might even kind of contest my argument that, uh, we should really think of the training in these machines, not just as a process of learning, but as a process of evolution, because some of the machinery in particular, the system for representing the meaning of words, or at least the specific kind of modes of representing the words are kind of, which in the case of humans is almost certainly a product of natural selection, right? That basic mechanism, this. This pre-preparation for ingesting and mastering language is probably a product of biological evolution. And I think that gets reverse engineered in a sense in the course of training. And I think a lot of things do. So the short answer to your question is, it's like, and this was not true of the computer that beat the chess champion, okay? They coded that line by line. And this, they don't even understand exactly what's going on in these machines. My guess is that it's often comparable to what's going on in our minds, which we also don't understand, by the way. And that's the key. We don't have to understand the human mind to build an artificial mind that functions like ours. That is the key. Speaker 1 (14:06) So can you tell me why those AI researchers might disagree with what you just said in as simple terms as possible? Speaker 2 (14:16) Some of them, for example, might have a totally blank slate conception of the human mind and think that we are born with a brain that's just a very general information processor. Most evolutionary psychologists would be on my side of this. They'd say that actually evolution developed a... a number of different kind of little mechanisms at different times in the brain. For example, the mechanism for mastering language. And so the brain is, you know, is not this highly general and blank system. But there are people, I think Rich Sutton is one who's famous and made great contributions to the field. And it's associated with the phrase bitter lesson, which people can Google if they want. I think he's like a pure Skinnerian. You know, B.F. Skinner thought the mind was very much more like a blank slate. And all the positive and negative reinforcement that's important to turn us into functioning cognitive and perceptual machines happens during our maturation. Whereas a lot of other people would say, actually. Evolution is a process of positive and negative reinforcement. You know, the things that get gene spread are positively reinforced, those mutations. And that in the training of a large language, you're doing both at once, really. You're on the one hand, the machine is mastering a specific language the way a child would. That's kind of learning. But it's also developing more generic mechanisms of linguistic. processing that are probably products of natural selection. But I do think a lot of the disagreement gets down to the question of whether you think the mind is a blank slate. I don't. Speaker 1 (16:05) Okay. So let me reframe the question slightly because of your answer. So I start out by saying, okay, big progress in AI and chess and machines can do math, but it doesn't change the world, right? Why does language and coding change the world? And your answer is, well, really the big difference is between AI that you program, where you're basically just programming a set of instructions, which is what Deep Blue was that beat Garry Kasparov, and something like AlphaZero, which ended up being much better than Deep Blue. This is the chess engine that was based on neural networks. But it still leaves the question for me. Okay, so now there are chess engines that are neural networks, that are this really cool breakthrough, and they turn out to be even better. They still don't change the world, really. They change the chess