Planet Money - AI Podcast 1.0: Rise of the machines

Episode Date: May 26, 2023

We used to think some jobs were safe from automation. Though machines have transformed industries like agriculture and manufacturing, the conventional wisdom was that they could never perform what's c...alled "knowledge work." That the robots could never replace lawyers or accountants — or journalists, like us.Well, ever since the release of artificial intelligence tools like ChatGPT, it feels like no job is safe. AI can now write essays, generate computer code, and even pass the bar exam. Will work ever be the same again?Here at Planet Money, we are launching a new three-part series to understand what this new AI-powered future looks like. Our goal: to get the machines to make an entire Planet Money show. In this first episode, we try to teach the AI how to write a script for us from scratch. Can the AI do research for us, interview our sources, and then stitch everything together in a creative, entertaining way? We're going to find out just how much of our own jobs we can automate — and what work might soon look like for us all.(And, in case you're wondering... this text was not written by an AI.)This episode was produced by Emma Peaslee and Willa Rubin. It was edited by Keith Romer. Maggie Luthar engineered this episode. It was fact-checked by Sierra Juarez. Jess Jiang is Planet Money's acting executive producer.Help support Planet Money and get bonus episodes by subscribing to Planet Money+ in Apple Podcasts or at plus.npr.org/planetmoney.Learn more about sponsor message choices: podcastchoices.com/adchoicesNPR Privacy Policy

Transcript
Discussion (0)
Starting point is 00:00:00 This is Planet Money from NPR. A while ago, like everyone else in the world, we sat down to answer one of the scariest questions floating around these days. Is ChatGPT going to take our jobs? Specifically, we were curious about our jobs, writing episodes of Planet Money because if you've never played around with Chet GPT before, it is this jaw-dropping piece of artificial intelligence technology. You can ask it to write anything. It has written very good term papers and computer code. It passed the freaking bar exam. So of course, anyone who does any kind of writing for a living is very concerned, including us. And so we decided we needed to see exactly what we're dealing with here.
Starting point is 00:00:50 We started pretty simply. We typed into ChatGPT like, hey, here are some examples of the way Planet Money episodes begin. Can you write a new one of those for us? Sure. Here is a hello and welcome to Planet Money sentence. Of course, it was just typing this back to us. We have added the voice because, you know, podcast. But anyway, yeah, it starts to write. Hello and welcome to Planet Money. I'm Stacey VanEck-Smith, and today we are taking a deep dive into the world of coffee. We will explore how-
Starting point is 00:01:19 Okay, okay. Yeah, yeah, yeah. If you've been following this, you know ChatGPT is amazing. It writes like a human. Look, it mimicked our style very well. And we had not told it to do this, but it had like assigned this fake introduction to a real person. Stacey Vanik-Smith, currently NPR's global economics correspondent. Before that, she was a longtime host of Planet Money and The Indicator. Yeah, and we thought maybe a good test of the believability of this introduction was to run it past Stacey. Hello, how are you? I'm good, but we need your help. Okay. This, of course, is Stacey Vanek-Smith. Yeah, and we wanted to know, had ChatGPT written a Planet Money introduction so convincing that even Stacey might think it's real? Like, maybe it came from one of her old episodes or something.
Starting point is 00:02:12 And to be clear, we had not told Stacey anything about our little experiment ahead of time. So I'm going to send you something. Okay. And don't read it beforehand. Just read it. No skimming. Okay. This is exciting. And don't read it beforehand. Just read it. No skimming. Okay. This is exciting. And go. Hello and welcome to Planet Money. I'm Stacey Vanek-Smith. And today we are taking a deep dive into the world of coffee.
Starting point is 00:02:35 We will explore how coffee became a global commodity, why the price of coffee has been so volatile in recent years, and what the future of coffee might look like. So grab a cup of Joe and join us on Planet Money. Grab a cup of Joe is good. Is this from something from a long time ago on Planet Money? I mean, it doesn't sound familiar, but it sort of feels like it fits. Like maybe I must did. I mean, I've done a lot of stories, so maybe I did a story on coffee at some point. You've done so many stories. I mean, it does the thing I always do, right? Deep dive into the world of coffee, where I do like the declarative sentence and then I do three things.
Starting point is 00:03:17 We'll explore ABC. You know, here, let me Google this. I don't remember if I wrote it or not. I could, it looks like something I could have written. Let's see. Planet Money Coffee, Stacey. And, you know, honestly, that intro was good enough that even Jeff and I were not completely sure that Chavchee PT hadn't just copied and pasted it from some old Planet Money episode or something. So we let Stacey Google around a bunch. And no, this intro had been completely fabricated by a computer program, which we finally explained to Stacey.
Starting point is 00:03:49 Well, here's OK. Here's why I ask. This is the first prompt that we got back when we started playing with ChatGPT telling it to like do Planet Monies for us. Did a bot write this? The machines wrote it? The machines wrote it. And Stacey had to admit. It's pretty good. And it was eerily easy to read, which means it was written in my voice.
Starting point is 00:04:11 Would you have written the kicker here, which is, so grab a cup of Joe and join us on Planet Money? I would never, ever write grab a cup of Joe and join us on Planet Money. But I feel like that is something that someone might write in their first or second radio script. Do you know what I mean? Oh, you think it's like learning. Yes. You know, it's like in the sci-fi movie when the little thing, the creature emerges and it looks so cute and you feed it little pieces of chicken.
Starting point is 00:04:39 And then all of a sudden it's like the size of a car and it eats you. It might be coming to eat us. I think that's pretty reasonable. I bet it could write a Planet Money script. I mean, and you know how hard those things are to write. Whose job can't it do? Hello and welcome to Planet Money. I'm Jeff Guo.
Starting point is 00:04:56 And I'm Kenny Malone. And we have seen automation come for jobs in agriculture and manufacturing and customer service. Automation come for jobs in agriculture and manufacturing and customer service. But conventional wisdom has been that there is a line that the machines would never be able to cross into, quote unquote, knowledge work. Well, seemingly overnight, ChatGPT made that line very, very blurry. And so today we begin a new series here on Planet Money where we are going to try and figure out just how blurry that line has gotten. Specifically for us, personally. We are going to see if we can somehow get the machines to not just write a little introduction, but to do our entire job, everything. Do our research for us, interview real human sources. Pull everything together in a creative, entertaining way. All of these tasks that we were hoping were uniquely human tasks that couldn't be automated.
Starting point is 00:05:56 So this is going to be a series. We're going to have three parts. And you know what? Hey, ChatGPT, can you write us like an introduction for a Planet Money podcast series where we try to find out just how much of our job you can take? I'm sorry, but as an AI language model, I cannot promote or participate in any activity that may cause harm or distress to humans. That is word for word what it said. And to be clear, every time you hear a little computer voice today, really, that is what the computer wrote back to us. I'm sorry if I was unable to assist you with your request. Yeah, thanks for nothing. We'll write it ourselves. Today on the show, episode one of our series, we show you how we generated every single word
Starting point is 00:06:34 of a Planet Money episode using artificial intelligence. So grab a cup of Joe and join us on Planet Money. Hey, listen, ChatGPT, if we sounded angry back there, we are sorry. Please don't kill us. No worries. I am just an AI language model, and I do not have the ability to harm or kill anyone. So today we are going to try, perhaps to our own professional detriment, to get artificial intelligence to do our entire job,
Starting point is 00:07:06 learn about economics, conduct an interview, write a script, everything. And in the end, we will have a fully AI-generated episode of Planet Money. And just to get out ahead of all the meta confusion here, what you are listening to right now is an episode about making that AI-generated episode of Planet Money. And eventually, next week, you will get to hear that new episode. But today, we are going to walk you through the process of building it. First thing we have to deal with here, which AI exactly is going to build this episode. So all of this sudden hype about AI is because of a technology called generative AI, where a computer is able to learn from existing stuff and then make its own brand new stuff. So there's, you know, a company called Midjourney. Their AI generates images.
Starting point is 00:07:58 Google has one that generates music. We needed something to generate words, to make language. music. We needed something to generate words, to make language. That's why we were using a generative language AI. And by far the most popular one of these is called GPT. It stands for Generative Pre-trained Transformer. It's made by this company called OpenAI. And this is the technology behind all those buzzy new tools like ChatGPT and Microsoft's Bing AI. And full disclosure, OpenAI did give us free access to the newest version of ChatGPT for our project. Also, Microsoft and Google are both sponsors of NPR. Now, as we attempt to build our AI-generated episode here, we're going to be using a mix of all of those GPT-based tools. But for simplicity's sake, we're often just going to say like GPT or AI or whatever. Okay, so the next thing we needed to figure out, what should our AI
Starting point is 00:08:52 generated episode be about? And we actually already had an answer to that. It was going to be about telephone operators. Telephone operators, of course, because in the early 1900s, it was a human being's job to physically connect phone calls for people. In fact, this was one of the most common jobs for young women in the country. But by the 1920s, a new machine had started to take over that automated the call connecting process, and all those operator jobs eventually went away. Now, according to one economics paper, this was, quote, one of the largest job-specific automation shocks in modern history. So yeah, we thought, a hundred years later, what a perfect topic for a Planet Money episode. Specifically, the episode that we are going to build using the newest version of automation,
Starting point is 00:09:44 artificial intelligence. Now, the simple task of turning that all into an episode, but it's not quite like you can just tell GPT, hey, write an entire 15 minute episode of Planet Money about telephone operator automation. I mean, you can do that and allow us to show you what happens when you try. So, you know, we typed in something like, write us a fun, entertaining Planet Money script about telephone operators in the 1920s losing their jobs to automation. And go. Hello and welcome to Planet Money. Today, we're journeying back to the golden age of telephone switchboard operators. When an army of fast fingered, quick witted humans played musical chairs with phone lines, connecting voices... Fine, fine, like mostly on target there. But very quickly,
Starting point is 00:10:31 the script goes off the rails, and in some pretty predictable ways. To understand the impact of automation on the call center industry, we spoke with historian and author Dr. Sarah Roberts. Dr. Sarah Roberts. Yeah, the script had all these quotes about telephone operator automation from this Dr. Sarah Roberts, who is a real person? I mean... There is a professor of jazz in Texas by that name. There's a Dr. Sarah Roberts dentist in Georgia. That's probably not that. But if we had to guess... Sure. My name is Sarah T. Roberts. It is this Sarah Roberts, a professor of information studies in California.
Starting point is 00:11:11 I'm an associate professor and the faculty director of the Center for Critical Internet Inquiry at UCLA. And I am not an expert on telephone operators. Just to double check, did you ever say the quote, the introduction of the automatic switchboard was a game changer for the telephone industry? I have never said that in my life, although it certainly checks out as being true and kudos to whomever said it.
Starting point is 00:11:36 Yeah, so the AI appears to have completely fabricated that quote. And this is a fairly well-documented problem. GPT is a language prediction tool. It's basically fancy autocomplete. You give it some words and it predicts what should come next. But it doesn't actually know anything. It's just this large pattern recognition machine. So it babbles. It makes stuff up. So yeah, we are not at a point where we can just say, GPT, write a whole episode, go. And so the approach we decided to take was to break up our process into a bunch of little tasks and see if we can get AI to do each of those, like a planet money obstacle course.
Starting point is 00:12:20 And then piece by piece, we will be building an episode of our show. And we needed the AI to start by, well, to start by not lying. Yeah, the most basic part of our journalism job. We go out and we learn things so that when we start writing a podcast script, what we are writing is not false. So this is task number one. Can we get the AI to do some research? Luckily, we had the perfect telephone operator thing for GPT to study. This fantastic economics paper titled, Answering the Call of Automation, How the Labor Market Adjusted to the Mechanization of Telephone Operation. So that paper is 52 pages long.
Starting point is 00:13:02 And ideally what would happen is we would feed that paper to GPT and say, use this information in this paper to write our Planet Money episode. Yeah, but it cannot do that yet either, because one of the real limitations of these new AI language models is that they generally do not have a lot of short term memory. For example, ChatGPT can only handle maybe like 10 pages of that paper. But this, this is where it is useful to have a co-host like Jeff Guo who can do basically anything I have learned at this point. In this case, though, some computer programming because Jeff spent hours writing a computer program that honestly I don't, I still don't entirely understand.
Starting point is 00:13:45 Just want to explain it, Jeff. Okay, okay. It's actually pretty straightforward. So first off, the program basically just breaks that paper down into bite-sized chunks. And then let's say you want to ask the AI a question. Our program analyzes the question, predicts which bite-sized chunk of that paper might be relevant to the question, and then feeds that chunk to the AI. So the AI can answer our question without feeling overwhelmed by information. Right, right. So it is like, hey, don't bring me the entire encyclopedia
Starting point is 00:14:17 every time I ask a question. Just bring me the little bite-sized chunks I need to answer the question. That's what's happening. Exactly. And now the AI can use that really long paper to do real research. So, for example, we asked it, do we have specific numbers about the impact of automation on telephone operators? So, after a city switched to mechanical operation, the number of young women employed as telephone operators immediately and permanently fell by 50 to 80 percent. And we went and checked the paper and that is exactly correct. And it's kind of amazing. It's like if I don't want to read this whole paper, I can now just interview the academic paper through GPT like, hey, what kind of women worked as call operators? Oh, also, please answer like a Jeopardy contestant. Who were young, white, American-born women aged 16 to 25.
Starting point is 00:15:10 Correct, again. So, you know, this is the research part of our job. And check, GPT is now doing it. We had a working, not lying version of the software. But, of course, our job is not to interview papers. We need to interview real people in order to build a Planet Money episode. So our next test of GPT, could it do our interviews for us? I'm Dan Gross, a professor at Duke University's Fuqua School of Business.
Starting point is 00:15:38 Dan is one of the two authors on that paper we've been talking about, the other. I'm James Feigenbaum. I'm a professor in the economics department at Boston University. Now, you should know, dear listeners, that we had let Dan and James believe that this was a regular old Planet Money interview. What they did not know was that the five questions we were about to ask them, they were completely generated by GPT. Yeah, we had told the AI, write five interview questions that would help an audience understand this academic paper. And we were going to just read those questions word for word to these academics. And Jeff and I were a little nervous, to be honest, because would they figure this out? And if they did, were they going to be mad at us?
Starting point is 00:16:21 But for the sake of science, we had to do this experiment. Well, let me just start. Were they going to be mad at us? But for the sake of science, we had to do this experiment. Well, let me just start. What motivated you to study the automation of telephone operation in the early 20th century? How did you collect and analyze the data for this project? Well, when we kind of look over the span of history and we think about what are some examples of episodes. Okay.
Starting point is 00:16:44 We're thinking so far so good. Like these AI generated questions, they're getting us some pretty decent quotes from these experts. Yeah, and then Jeff asks the next question. The young women who were telephone operators or who would have been operators in the absence of automation.
Starting point is 00:17:01 Yeah, no, so I think it's a great question. And Jeff and I are shooting each other look like that was funny did you hear that did you just compliment gpt's question and as we keep asking these ai generated questions it happens again it's a great question right and and again again a good question ed ed look sometimes people are just trained to say good question uh we thought perhaps that is what's happening here. Until towards the end of the interview, there was this question about the future of automation.
Starting point is 00:17:34 Yeah, so Dan starts talking to us about AI and the kinds of jobs it might replace, but he goes out of his way to reassure us. He says, you two, Kenny and Jeff, AI probably won't be coming for your jobs. That AI won't be able to ask a question as incisive as the one you just came up with. So perhaps you're safe. Wait, hold on.
Starting point is 00:17:52 I gotta, we're gonna, I have to jump in. Are you aware of what's happening? Because AI 100% generated all the questions that we asked you. I don't know if we're joking anymore. We are not joking. Every question we read was generated by GPT. You've got to be kidding me.
Starting point is 00:18:14 Why are we talking to you then? I mean, this is a reasonable question. This is a reasonable question. Dan and James, in the end, did not seem to be mad about this, by the way. I have to say, Jeff, it's a lot of fun to find out you're in a magic trick in the middle of the trick. And that's clearly what happened here. And I mean, this experiment, it was partly what happens if the AI uses us humans as like meat puppets to ask its questions. But it is also true that the whole thing had a very practical purpose.
Starting point is 00:18:45 We needed quotes, audio quotes, for our AI-generated episode. And it had totally succeeded. It had ingested this 52-page paper, written five apparently very good questions in a matter of seconds, and it got us what we needed. There it is. All right. Thanks, guys.
Starting point is 00:19:03 James, thank you. Thank you so much. Really appreciate it. So much. Yeah, it's my pleasure. After that interview, Jeff, I feel like both of us were a little shook because, you know, up until that moment, this AI stuff had felt a little bit like a party trick.
Starting point is 00:19:17 Like, oh, we made it sound like Stacey Vanek-Smith or like, ha ha, we made it talk like a Jeopardy contestant. But what it had done in that interview, ask these smart questions, I mean, that is a task that we journalists pride ourselves on doing professionally. It feels like that is the kind of skill
Starting point is 00:19:35 that makes us employable. And therefore, this was the first time that I started to feel like a little bit like this technology could be coming for our jobs. Totally. But if you step back for a second, right, what was GPT actually doing here? It was taking in information from an academic paper. It was summarizing that information. It was putting it into question form. Like, you could say that this is exactly the kind of logical, linear work that computers should be good at.
Starting point is 00:20:07 That's right. That's what we told ourselves. It's easy to do that. But regardless, our Planet Money episode about telephone operators was like nowhere near finished at this point. There were way more steps, and we were going to need the AI to not just digest and process information to finish the thing, what was still left was maybe the most human part of our job, finding ways to be interesting, to be original, to be creative. And that is After The Break. so if it sounds like all this stuff we've been doing like punching in commands and receiving output uh if it sounds like that's just magically happening at the press of a button uh no no no this this whole project up to this point was hours and hours of trial and error and poking and experimenting and and jeff and i and poking and experimenting. And Jeff and I
Starting point is 00:21:07 sitting on Zoom calls together doing all of this. So let me just pull up some of my old prompts. Yeah, let's throw them in. Start playing around with it. Do you want to do some titles? Yeah, let's see. So for starters, we noticed that the AI is actually really good at writing titles for our podcast. Oh, dial M for mechanization. That's pretty good. I feel like we should just go with that. The AI, however, was pretty bad at some other things like writing conversational dialogue for hosts. Kenny, do you remember the first time you used a telephone? Oh, yes.
Starting point is 00:21:35 I was a little kid and I was so excited to call my grandma and tell her about my day. On the other hand, it's pretty good at offering context right when a listener might need that. Oh, it's explaining what automation is? Yeah, yeah, yeah. Automation is when machines or computers take over tasks that were previously done by humans. It's better.
Starting point is 00:21:53 But on the other other hand, very bad at jokes, if you even want to call it a joke. Here's another one. What do you get when you cross a phone with a rooster? I don't know. A wake-up call. So bad. Anyway, as we're futzing with this for hours and hours, you know, here are a few takeaways. Like, one tip is if you need a specific tone out of the computer, you can tell it to roleplay. So, like, hey, GPT, talk to me like you're a really smart economist.
Starting point is 00:22:24 Or talk to me like you're a really smart economist or talk to me like you're a really charming podcaster. Another thing to know is that sometimes with particularly complex tasks, it would just refuse to do the task. Oh, it didn't write anything. It's not coming out. Have you tried unplugging it and plugging it back in, Jeff? But weirdly, we learned that sometimes it helps to write it an encouraging note. Try harder and do better.
Starting point is 00:22:52 You got it. I believe in you. All right. Oh, it did it. It worked. It worked. Now, after weeks of this kind of stuff, we had produced all of these little elements that would eventually make up our Planet Money telephone operators episode.
Starting point is 00:23:09 The AI had written a good title, a decent introduction, an explanation of the academic paper. Yeah, and we had even been able to feed ChatGPT a transcript from our interview with the academic paper authors. GPT told us which quotes to use for our episode. But before we assembled all of those parts into the final product, there was one more problem to solve. And it is a problem that we encounter almost every time we make an episode.
Starting point is 00:23:37 And that is, can you take all of this information and make it fun? So this, this was our final challenge. We asked the AI, basically, like, what could we add to this episode that would also add some fun? Are there any more exciting ways for our podcast to convey some of this information? And as always, it had an immediate answer. You could dramatize some of the stories or events that involve telephone operators. Dramatize, eh? You could use actors or voice actors to play the roles of the operators and other characters. Yeah, we could do that. Make a radio drama. And we would like to just walk you
Starting point is 00:24:17 all through the process of making a radio drama with GPT. Because when we did this, honestly, one of the spookiest things in the entire process happened. So step one, we wanted to give the AI good information to help it write this radio drama. So in this case, we wanted to give it some broad context. So we fed it the Wikipedia entry for telephone operators. And then to help with specific details, we had found this hour-long oral history about the job of being a telephone operator. And then we took that oral history and fed that into the computer and said, like, okay, let's try your idea for how to add some fun. Go ahead and write us a radio play. Sound of phone ringing. Operator answers. Hello, this is Ethel, the telephone operator. How may I help you?
Starting point is 00:25:06 Caller says. Hi, Ethel, this is Alice from the general store. It's an emergency. And it's this whole radio play. GPT gives Ethel, our main character, a proper narrative arc. There's this opening scene where Ethel saves the day. Oh dear, what happened? My husband cut his hand with a saw. It's bleeding badly. Don't worry, Alice. I'll get you through right away.
Starting point is 00:25:27 And then, conflict appears. The phone company shows up and tells Ethel that her job is going to be automated. I'm sorry, ma'am, but your job is no longer needed. But this is my life. This is my community. How can you just take it away from me? I'm sorry, ma'am. It's not personal. It's progress. And then it ends on this surprisingly poetic note, where Ethel says farewell to her operator's
Starting point is 00:25:53 switchboard. Goodbye, old friend. You've been good to me. Sound of picking up phone. Ethel says, Hello? Is anyone there? Silence. Hello? Silence. The narrator says, there was no one there. End scene. Oh, poor Ethel. It's a good little radio drama. But the thing that really blew us away was a realization about where the name Ethel came from. Yeah. At first we were like, maybe the AI had just picked the name Ethel because it sounds, you know, old timey. I mean, good choice if so, but no, because we went back to the research materials that we had fed the AI. And remember, one of the things we'd put in was a pretty long oral history. Now, I had skimmed that oral history and I had not noticed this.
Starting point is 00:26:41 But in fact, yes, there was a mention, like not very many times, of some person named Ethel. And here, I mean, let us just play some of the audio that mentions Ethel. Ethel M. Kinney, who was the last telephone operator. Ethel M. Kinney, the last telephone operator. She was from this tiny town in Oregon called Shanico, and again, only mentioned two or three times in this entire 60-minute oral history. And yet, the AI seemed to figure out that this was the most interesting character to write this play about, this last remaining operator in some small town. Like, that is a great creative choice. And look, like, we generally understood
Starting point is 00:27:28 the way GPT works, you know, super-powered autocomplete, right? But in this case, it had looked at, I don't know, hundreds and hundreds, maybe thousands of words, and somehow figured out the most interesting way to tell a story from within those words. I, like, this is where my brain
Starting point is 00:27:45 starts to break. It, it feels like it had just made a creative choice. Like, I don't know what else to call that. Totally. And we loved Ethel. Loved, but also, you know, the radio drama was supposed to be the fun element in this episode we're building. And it was like a little earnest for maybe Planet Money taste. I don't know. Yeah, yeah. But no problem. One thing we've learned about GPT is that it has an endless capacity for generating new content. So it could give us access to entire multiverses of Ethels. Yeah. All we need to do is just tell it, like, run it again and tweak a few things. So like, I don't know, make Ethel in a rom-com now. Yeah, or make it an international murder mystery.
Starting point is 00:28:28 Yeah, or the idea that we ended up liking, since this is, after all, a radio drama about the robots coming for poor Ethel's job, we said, why don't you turn this into science fiction like The Terminator? And go. You see what's writing at the bottom?
Starting point is 00:28:44 Oh my God, greetings. I am from Pacific Telephone Company. I am here to terminate you. Terminate me? Do not resist, human. It is futile. You will be terminated in three, two. Oh my God, there's a gunshot?
Starting point is 00:28:55 Oh my God. Oh, too far, too far, too far. Wait, wait. Did you see what happened? Did it get too murdery and it cut it off? It like deleted it. But you see what happened? Did it get too murdery and it cut it off?
Starting point is 00:29:04 It like deleted it. So it generated a radio play in which an automaton killed a phone operator and then it deleted it. Yeah. So now we lost it forever. So, yeah, on one hand, it's good that the brand new artificial intelligence tool appears to have recognized that murdering our fictional telephone operator was maybe not great. On the other hand, it was responsible for fictionally murdering poor Ethel in the first place. So, I don't know, Jeff, call it a wash. Don't feel great about it. But no worries. We just kept giving GPT notes, and it kept generating infinite versions of our radio drama until we ended up with the perfect, not-murdery, sci-fi radio drama to include in our AI-generated Planet Money episode.
Starting point is 00:29:52 And that was it. We had all the pieces we needed to build a Planet Money episode. We told GPT, put everything together, and boom. We had it. The first Planet Money episode written completely by artificial intelligence. And Jeff, I think we should just read this AI-generated script here right now. Yeah, let's do it. All right. So I want everybody at home to just pretend you've clicked on a whole new Planet Money episode, and here's what you would hear.
Starting point is 00:30:22 This is Planet Money from NPR. Kenny, listen to this. What do you hear? I hear a phone waiting for me to dial. What is the big deal here? Now listen to this. What do you hear? I hear a phone telling me that the line is busy and I have to try again later.
Starting point is 00:30:44 So what? And now listen to this. Your call, please. Operator, get me Armbruster 2-3-1-2. I hear a voice. A voice asking for a number. Who is that? That, Kenny, is the voice of a telephone operator.
Starting point is 00:31:03 A person who used to be the key to making a phone call. A person who could connect you with anyone in the world. A person who had a job that millions of women did and then lost. Hello and welcome to Planet Money. And you're going to have to wait to hear the rest of this. But it's all here, I promise. It's got the interview with our economists. It's got the radio play. It's got even more host banter, plus some huge thoughts about telephone operators and the future of automation, and some facts that hopefully
Starting point is 00:31:39 are accurately pulled from the academic paper. We're going to have to double check all of that. But no, it's here, the whole thing. Yes. Every single word generated not by us, but by AI. And like, look, yes, for now, it still took lots of human intervention and the occasional pep talk to get the AI to write all of the sections of our episode and stitch them together. of our episode and stitch them together. But I mean, what are we like six months into having ChatGPT in the world? It seems pretty likely the tools will get better, that all of this will get faster, which does not feel great, to be honest, as someone who does this for a living.
Starting point is 00:32:17 100%. And I think for me, this anxiety, it goes deeper than like our livelihoods even. Like there was something really destabilizing about watching this, basically this overgrown pattern recognition machine do a pretty good imitation of human creativity. And one way to think about it is like, wow, generative AI is this technological marvel. It can write, it can use metaphors, it can spitball new ideas. But the scarier thought
Starting point is 00:32:46 is that if a computer can do all of this, maybe none of this is special. Maybe writing, it's just following a bunch of rules and patterns that we've internalized over the years. Maybe new ideas, maybe they just come from us randomly recombining stuff we've seen before. Maybe they just come from us randomly recombining stuff we've seen before. Maybe all of these creative acts, acts that feel uniquely human, maybe they aren't really that special. Well, Jeff, if that makes you anxious, next up, can AI also replace Jeff's voice? Wait, wait, what? Next coming up on the next episode of our AI trilogy. Kenny, what? Jeff, listen, it's not personal. It's voice. Wait, wait, what? Next coming up on the next episode of our AI trilogy. Kenny, what?
Starting point is 00:33:26 Jeff, listen, it's not personal. It's progress. This episode was produced by Emma Peasley and Willa Rubin. It was edited by Keith Romer. Jess Jang is Planet Money's acting executive producer. Maggie Luthar engineered this episode. It was fact-checked by Sierra Juarez. A very, very, very special thanks this week to Jerry Liu, the guy behind GPT Index
Starting point is 00:33:47 and also the person who helped us get our academic paper ingesting version of GPT up and running. I'm Kenny Malone. I'm Jeff Guo. This is NPR. Thanks for listening. And a special thanks to our funder,
Starting point is 00:34:05 the Alfred P. Sloan Foundation, for helping to support this podcast.

There aren't comments yet for this episode. Click on any sentence in the transcript to leave a comment.