Skip to content

Podcast

‘Something Big Is Happening’ + A.I. Rocks the Romance Novel Industry + One Good Thing

Hard Fork

Source ↗ ← All highlights
  • AI Takeoff Is Reaching A Tipping Point
    • AI agentic coding and relentless agents are creating an inflection point in software development.
    • This momentum is shifting political and economic attention toward AI’s societal impact. Transcript: Casey Newton Well, Kevin, welcome back from our nation’s capital. Kevin Roose Yes, I was in D.C. Very briefly this week there for some book meetings, and it was very cold. But the bigger observation is that Washington, D.C. Is like freaking out about AI. Is that right? Yes. So everywhere I went, every meeting I had, people were sort of asking me, is this stuff real? Is it happening? Are we in the takeoff? Is the singularity approaching? And it does feel like the sort of political salience of AI has gotten much, much higher just in the past couple of weeks. Well, why do you think that is? So there are a lot of reasons for that. I think one of them is that I think there’s been a lot of people waking up to the new agentic coding of these models. We’ve obviously talked about that on the show, Cloud Code, et cetera. I think that is starting to kind of make its way out into the world. There’s also the stock market stuff that’s been going on with a lot of the software stocks that are falling because of the threat of AI. And then I think there’s just sort of this ambient cultural vibe shift happening that has led to a lot of people in my life who are not like AI bubble people texting me and saying, hey, is This really something I should be worried about? Is my job at risk here? And so today I think we should talk about this because, among other reasons, there is this viral essay that I’ve been sent now no fewer than three times just in the past day that is by a man Named Matt Schumer called Something Big is Happening. And it’s basically a distilled version of something you and I have been talking about on this show for a while now, which is like these tools are getting really good. They’re changing the way that programmers work. They’re approaching some sort of inflection point and everyone needs to be worried about this. Yeah. So all of this is starting to make me think that there is something big happening here, and I’m not sure it’s exactly what Matt thinks is happening, but I do think we are reaching an inflection Point in people’s feelings and senses about AI and where it’s going, and I think we should spend some time today exploring that. (Time 0:02:00)
  • Business Models, Not Just Code, Are Vulnerable
    • AI is likely to change software business models even if it doesn’t erase demand for software itself.
    • Outcome-based pricing and service-focused startups may replace per-seat licensing. Transcript: Kevin Roose Aren’t worth as much anymore. So I think in this specific instance, it’s not clear to me that there was like one particular trigger. People pointed to this set of plugins that Anthropic released, which included some tools for law firms trying to use AI. And some people think that was sort of behind a lot of the sell-off. I don’t think it was that in particular. I think maybe that was sort of the straw that broke the camel’s back. But I think there is sort of a mounting sense that, as we’ve talked about in this show, now you and I and anyone can theoretically at least build your own version of that, and these companies May not be as valuable. Casey Newton Yeah, I mean, there have always been people who would look at a product like Salesforce, which offers customer relationship management. You know, if you’re a salesperson, it helps you keep track of all of your different leads. And they’ve said, well, you know, that’s basically just a fancy spreadsheet. You know, I could make my own spreadsheet. And there are companies that are born almost every year that kind of take a direct run at Salesforce, and they say, we can build a better version of this. And I do think that when Claude Cowork came out and all of a sudden you could just take a bunch of files on your computer and throw them into Claude and get something useful back, there were People who said, huh, maybe we can actually just make a version of it ourselves. And I will say, as a small business owner, I used a plugin that Claude has for finances, and I just took all of my financial data that my bookkeeper has kept for me over the past five plus Years, and I threw it into Claude, and we had a nice long conversation about what was going on in my business. Now, it so happens that I don’t pay a startup to do this for me. This was just kind of a little bonus that I got. But could I imagine some people saying, hey, why am I using this kind of like bespoke financial startup to do business analysis when now I can just put these files into something that Lives on my computer? So I actually do feel like I have a sense of why the market freaked out a little bit. Kevin Roose Yeah. And do you feel like that fear is justified? Because there have been a couple reactions to this from the business community. One reaction is the sort of investor reaction, which is, oh my God, all of the SaaS companies are overvalued. Salesforce doesn’t have a moat anymore. Workday doesn’t have a motor anymore. Everyone’s going to be vibe coding their own versions of these tools and using them at their businesses. And the other reaction, which is sort of the reaction to that reaction is like, calm down. Don’t freak out. No one is going to be vibe coding their payroll software. That’s not how this works. So there are people saying, eh, this market reaction is overblown. And despite the fact that these vibe coding tools are very cool, they are not going to lead to the death of the software industry. Casey Newton So this gets at a really interesting question that I think is still unanswered, Kevin, and that is what makes these companies truly vulnerable? Is it the technology itself or is it that the technology will just enable different kinds of business models, right? So the argument that it’s the technology itself is the person who says, look, we’re just going to be able to vibe code all our own software now, or we’ll have like just very smart agents That can do all of the different things that every piece of software we used to buy did for us, right? That’s the sort of technology demolishes everything argument. And I would say that that’s like a minority view. Most of the people that I read and talk to do not actually think that that’s going to happen. But there is this other view, which is that the technology enables a change in the business model. And I think law is a really good place to think about that, right? Because lawyers are expensive. They charge by the hour. The most expensive ones charge $1,500 for an hour of a lawyer’s time. Well, what happens when you don’t need an hour of a lawyer’s time anymore? What happens when there’s a legal startup that does all your contract review essentially instantly? There’s going to be a different business model around that. And so if you’re a big white shoe law firm and you’re charging $1,500 an hour, all of a sudden it does seem possible that one of these little startups is going to eat your lunch. So that’s the distinction I would draw. Kevin Roose Yeah, so you see this more as like a case of the startup with 10 employees being able to use AI to do the work that would have required a thousand people a year or two ago. Yes, absolutely. Let’s talk about another really common business model in tech. Casey Newton A lot of these business software companies have you pay by what they call the seat. So if you want 10 of your employees to be able to use software, like let’s say Notion, you buy 10 seats. Another threat of anxiety rippling through Silicon Valley right now is maybe paying by the seat isn’t really going to make sense anymore because we’re actually just going to have one Agent that does that whole thing. We’re not going to expose that to employees. Employees don’t need to worry about that anymore. And so we’re just not going to buy seats. So again, if I had like one prediction to make here, it’s that you’re going to see more companies experiment with outcome-based pricing. You’re already starting to see this with companies like Sierra, which is a customer service startup, and they will sort of handle customer service inquiries that your business may Get. And you pay Sierra based on how many calls it resolves for you. So that’s the kind of thing that I think we’re going to start seeing more and more of. And to be clear, if you’re like an incumbent SaaS company and you have a seat-based business model, that is eventually going to be a problem for you. Yeah. So (Time 0:05:30)
  • Small Teams Can Replace Big Incumbents
    • Small startups can use AI to perform tasks that once needed large teams, threatening incumbents.
    • Economies of scale in labor-intensive services may evaporate when agents automate repeatable work. Transcript: Kevin Roose Yeah, so you see this more as like a case of the startup with 10 employees being able to use AI to do the work that would have required a thousand people a year or two ago. Yes, absolutely. Let’s talk about another really common business model in tech. Casey Newton A lot of these business software companies have you pay by what they call the seat. So if you want 10 of your employees to be able to use software, like let’s say Notion, you buy 10 seats. Another threat of anxiety rippling through Silicon Valley right now is maybe paying by the seat isn’t really going to make sense anymore because we’re actually just going to have one Agent that does that whole thing. We’re not going to expose that to employees. Employees don’t need to worry about that anymore. And so we’re just not going to buy seats. So again, if I had like one prediction to make here, it’s that you’re going to see more companies experiment with outcome-based pricing. You’re already starting to see this with companies like Sierra, which is a customer service startup, and they will sort of handle customer service inquiries that your business may Get. And you pay Sierra based on how many calls it resolves for you. So that’s the kind of thing that I think we’re going to start seeing more and more of. And to be clear, if you’re like an incumbent SaaS company and you have a seat-based business model, that is eventually going to be a problem for you. Yeah. Kevin Roose So do you think the sell-off is justified? Do you think people are panicking for the right reasons inside these companies and their investors? Casey Newton Here’s a funny thing. As a journalist, we’re not allowed to buy individual stocks. So I actually never have any idea of what the investors are supposed to be doing. So I’m not going to comment on whether I think, you know, the stock market is justified here or not. But I’m happy to comment on the overall picture, which is, do I think that AI is about to change a whole lot of business models and that a whole lot of businesses are probably going to have To either change dramatically or go out of business as a result? Absolutely. Kevin Roose Yeah. And I think this notion that, like, no one is going to vibe code their own payroll software, I think that is, like, I am not as convinced of that as some people. I think that if you are a business and you have, you know, 10,000, 20,000 employees and you’re paying by the seat for some piece of software, whether it’s payroll or workday or your HR Compliance software, those are real expenses. And so do I think that the, like the CEO of the company is going to vibe code the thing, you know, with one cloud prompt that is going to replace the software tool? No, but I can totally imagine a world in which you have sort of one or two full-time developers who are managing and overseeing and repairing your own internal software and you don’t Have to pay for a bunch of seats for someone else’s thing, I think that’s a very plausible outcome. So I am not as dismissive of the sort of the fear around these SaaS businesses as some people. And yet, if I were any of these large enterprise businesses with (Time 0:09:33)
  • Compliance Creates A Hands-On AI Market
    • Security and compliance remain major barriers, but enterprise customers accept hand‑holding.
    • AI vendors often embed engineers on-site to meet sensitive workflows and regulatory needs. Transcript: Kevin Roose And yet, if I were any of these large enterprise businesses with tens of thousands of employees charging, you know, per seat for software, I’m very worried. Yeah. Now, another pushback that we’re seeing to this narrative that AI is going to eat all the software companies is, well, it’s all going to be so insecure and sloppy and buggy that no one Is going to rely on it. What do you make of that argument? Casey Newton I think you’re absolutely right that it will be buggy and insecure, and you’re wrong that people won’t rely on it. If there’s one thing we’ve seen with Maltbook Mania, it’s that security is basically the last priority, at least for the, you know, bleeding edge maniacs who just want to try everything First. Kevin Roose Right. But if you’re a law firm or a bank or something like that, you do care about things like that. Casey Newton I think that that is true. And, you know, there are whole startups, and I know this because I see the billboards around San Francisco that specialize in various compliance functions. You know, it’s like, well, if you’re going to offer this kind of service, you have to be this kind of compliant. And so you’re going to pay us to make sure that all goes very well. That to me, like, because that is a repeatable automated process where you’re just trying to get your business to like match a bunch of checkboxes on a form, that just kind of seems like Something that you’re going to be able to train an agent to do. So like there are going to be some categories of things that I think are just going to be very risky for a long time. And then there are just these kind of like automated compliance functions that I just look at and I think, well, I don’t, I can’t think of a reason why a good agent wouldn’t be able to do That. Yeah. Kevin Roose And I think in these cases where it’s in a more sensitive industry or something that has more regulatory or compliance needs, like, I just think it will take a little bit more effort to Automate some of these functions, but I think it’s totally doable. So one thing that we’ve heard about just in the last week was this story about Anthropic developing tools with Goldman Sachs. So together, these companies have been deploying AI agents inside Goldman. And Anthropic actually has forward deployed engineers that will like go to if you’re a big customer, you know, like every other AI company, they will send people to your office to like Work to put agents into your workflows. And so that is also something that we’re seeing is like for the really sensitive things, it is not impossible to design AI tools that comply with all your various requirements, but you Might need a little more handholding. (Time 0:12:15)
  • Prepare Practically For AI Disruption
    • If you work in a white-collar role, familiarize yourself with modern AI tools now.
    • Get your financial house in order and avoid taking large new risks as disruptions accelerate. Transcript: Kevin Roose And this is by a guy named Matt Schumer. He runs an AI company. So there’s a little, you know, conflict of interest there. But he’s basically saying, look, the technical parts of my job are automated, not they will be automated or they might be automated, but I am no longer needed for the actual technical Work of my job. He talks about the advances in recent coding models, these agentic coding systems. He talks about how a lot of people have not tried AI since the original sort of LLM boom, and their impressions of AI are falling behind. Again, none of this is news to you if you are a listener of this podcast. But he’s sort of talking about this idea that these new models are contributing to their own development, this idea of recursive self-improvement. And he says that, you know, GPT 5.3 Codex, which came out just last week, OpenAI says that this is their first model that was instrumental in creating itself. So the AI models are now, at least if you believe the labs, contributing to their own development, starting to accelerate the development of these AI systems. And so what might have taken, you know, if there might have been six months between one model and the next a year ago, now that might be a month or two or even a couple weeks. Yeah. Casey Newton And look, I’m going to throw a hype flag on the play, Kevin, because I think that all labs have a vested interest in you believing that if you use their software, you could just tell anything To improve itself and it will become amazing. I’m not saying it’s not true. I’m just saying we should approach such claims with a degree of skepticism. Now, at the same time, I’ve talked to enough software engineers who work on this stuff that I do think that it is true in some ways, like that if you squint, that it is true. It is 100% true that they use these models in the production now of everything that they do, right? Like Claude and GPT Codex are deeply integrated into the workflows of both companies. And, you know, you look over the past three months, does it feel to you like there’s been an acceleration in the pace of releases? It kind of feels that way to me. We’ll see if that pace feels like it continues to accelerate, but it feels like things are moving faster now than they did, say, in February 2023. Yes. Right? So there’s evidence for it. I just always want us to be a little cautious with these claims. Kevin Roose I think that’s right. At the same time, I think that the timelines here are shorter than many people would imagine. I talked to an executive at one of the big AI companies this week who said that basically right now software engineering is kind of 90% automated. You still need a human to check in on the code that’s being written, to make sure it works, to fix things when they break. But, you know, that basically within the year, this person’s prediction was that software engineering will be fully automated. Now, that could take a little longer. It could happen sooner than the end of the year. But I think that is sort of the moment that a lot of people in the tech industry are looking at as sort of the beginning of what they call the takeoff. Casey Newton And one specific detail I would add here is that I’ve been talking to engineers who’ve been telling me about the specific ways that these agents work. And something that comes up again and again, I’ve heard Sam Altman say a version of this, is that these models just never get tired. They’re relentless. And so you can say to them, I want you to meet this objective, and they will just keep trying things until it works. And of course, there are some people who work this way, but those people don’t work 24 hours a day. They don’t typically work, you know, overnight. Sometimes they get tired, their morale drops. That’s not true of the agents. And so engineers I’ve talked to have started to see this behavior. And this is another reason why they think, aha, we’re starting to enter this takeoff phase is the relentlessness of the agents. Yes. Kevin Roose And so that’s basically the point of this viral essay is things are happening. Things are accelerating. It is not just coming for programming work. It is going to be a force in all kinds of white-collar fields. This is sort of the worker side of the AI panic. And Matt’s recommendations are, you know, basically figure out how to use these tools, try them out if you haven’t tried them in a while. He says, get your financial house in order. Basically, this, you know, the next few years could be very disruptive to your career if you’re a white-collar worker. So, like, don’t, you know, take on a bunch of new debt or anything like that. And then he… What am I supposed to do with this yacht I just bought this week? Casey Newton Oh my God, now you tell me. Kevin Roose So yes, I don’t feel like this is the best explanation I’ve heard of for what people should do. But I do think that this was something that has been rocketing around the internet. I think that people who have not been paying to what’s happening in AI are starting to wake up and maybe in ways that are panicky and maybe just in ways that are sensible. (Time 0:15:11)
  • Romance Authors Churn Books With AI
    • Alexandra Alter found romance authors using varied tools like PseudoWrite and NovelAI to speed production.
    • Coral Hart published 200+ books under 21 pen names and made six figures using AI workflows. Transcript: Alexandra Alter So the question that led to this story came up last year when OpenAI said that they would allow erotic content. They were going to make this change and start with age verification and then allow users to generate erotica, which is something that users had apparently been clamoring for. And so then I started asking around to figure out whether romance authors and publishers were feeling threatened by this. Was this something that they felt like might erode the market for traditionally published romance novels, you know, if readers could instantly generate their own love stories? And I was expecting to find a lot of hand-wringing and anxiety, which I did find. But I was also surprised to find a few writers who were willing to speak to me about how much they love AI and how they’ve been using it to churn out dozens of romance novels, and they feel Like it could revolutionize the genre. So that was a surprise to me because it’s a very contentious issue in the literary world. Most people, if they’re using it, are not open about it. And then the next question was, how good or bad is AI at writing sex and love stories? And the answer was, it’s pretty bad, and it requires a lot of help. Casey Newton So we’ll get into the controversy around it, but I want to hear first about kind of how this is actually working. Tell us about what is the workflow for a romance author who has decided, you know what, I’m tired of writing all these bodice rippers. It’s time to just hand that over to the large language model. Alexandra Alter It’s so interesting because there’s different platforms that writers are using. There’s places like PseudoWrite, then there are these sites that will generate customized erotica like Red Quill or My Spicy Vanilla. And then there’s just sort of your general bots, you know, your Claude, your ChatGBT, Gemini. And so writers have different tools that they’re using. But what I learned from talking to a couple of people was that if you learn how to prompt the bot correctly, it will write a pretty compelling sex scene. You have to give it kind of an outline. It helps if you give it a ton of information and tell it what subgenre you want to do. Because, of course, they’ve ingested all of these books from different subgenres. So you can tell it, I would like a reverse harem, mafia, enemies to lovers, slow burn romance. And it will deliver all those beats. It will require editing. It will require prompting. But I think, you know, what I heard from people who have played around with it a lot is, you know, some of them say they can write a book in a day and have it edited, ready to publish. Casey Newton Wow. Now, Kevin, have you ever tried pseudo write? I ask because you’re a pseudo writer. Kevin Roose Yes, I actually played around with pseudo write before ChatGPT. This was like one of the first AI programs that I ever played around with. Interesting. And it was somewhat helpful for me, but it was more oriented toward fiction writing. So I totally get what the appeal of the tool is. Tell us the story of Coral Hart, who’s a longtime romance novelist who’s been experimenting with some AI stuff. Alexandra Alter Yeah, she was one of the people that I found who was actually teaching other writers how to use AI tools to produce novels. And she has only been doing this writing with AI for about a year. She’s been writing novels, romance novels, for a really long time and was quite prolific, but realized she could absolutely supercharge her output if she started using AI. So last year, she created 21 different pen names and published more than 200 romance novels in all kinds of genres, super spicy erotica, tame, sweet teen stories, rom-coms. So it sort of had her foot in every corner of this market to see what would pop. And so just through a sheer volume kind of game, she ended up making six figures, she told me, selling these books. And in the process, she really learned a lot about which models, which chatbots would do what for her. She would combine them. Now she’s created her own proprietary AI writing system. (Time 0:29:22)
  • Prompt Precisely For Better Romance Scenes
    • Provide detailed outlines, subgenre cues, and kink inventories when prompting models for erotica.
    • Tell models to “slow down” to avoid rushed scenes and control repetitive phrasing. Transcript: Casey Newton It’s time to just hand that over to the large language model. Alexandra Alter It’s so interesting because there’s different platforms that writers are using. There’s places like PseudoWrite, then there are these sites that will generate customized erotica like Red Quill or My Spicy Vanilla. And then there’s just sort of your general bots, you know, your Claude, your ChatGBT, Gemini. And so writers have different tools that they’re using. But what I learned from talking to a couple of people was that if you learn how to prompt the bot correctly, it will write a pretty compelling sex scene. You have to give it kind of an outline. It helps if you give it a ton of information and tell it what subgenre you want to do. Because, of course, they’ve ingested all of these books from different subgenres. So you can tell it, I would like a reverse harem, mafia, enemies to lovers, slow burn romance. And it will deliver all those beats. It will require editing. It will require prompting. But I think, you know, what I heard from people who have played around with it a lot is, you know, some of them say they can write a book in a day and have it edited, ready to publish. Casey Newton Wow. Now, Kevin, have you ever tried pseudo write? I ask because you’re a pseudo writer. Kevin Roose Yes, I actually played around with pseudo write before ChatGPT. This was like one of the first AI programs that I ever played around with. Interesting. And it was somewhat helpful for me, but it was more oriented toward fiction writing. So I totally get what the appeal of the tool is. Tell us the story of Coral Hart, who’s a longtime romance novelist who’s been experimenting with some AI stuff. Alexandra Alter Yeah, she was one of the people that I found who was actually teaching other writers how to use AI tools to produce novels. And she has only been doing this writing with AI for about a year. She’s been writing novels, romance novels, for a really long time and was quite prolific, but realized she could absolutely supercharge her output if she started using AI. So last year, she created 21 different pen names and published more than 200 romance novels in all kinds of genres, super spicy erotica, tame, sweet teen stories, rom-coms. So it sort of had her foot in every corner of this market to see what would pop. And so just through a sheer volume kind of game, she ended up making six figures, she told me, selling these books. And in the process, she really learned a lot about which models, which chatbots would do what for her. She would combine them. Now she’s created her own proprietary AI writing system. But she has this whole kind of spreadsheet that she shared with me, which was sort of like, Claude writes beautiful sentences, but is terrible at sexy banter. Chat GPT will block you every time. Grok will do whatever you want. Goes for the filthiest option every time. Novel AI was literally trained on erotica and it’s out of control. So some of the writers said they actually had to prompt their bots to sort of calm down a bit. Kevin Roose I was so interested in the bits of your piece about how there’s a lot of steering needed to keep these things from sort of veering onto this very set of romantic tropes, right? If you don’t give the chatbot any guidance, it’s going to suggest, you know, that the character should be having sex in the bedroom or the shower. Boring! Yes. And so Coral, this romance writer who’s running these classes, is advising her students to, like give them a list of settings that are weird, like a winery fermentation tank or a stalled Ski lift or a horse stable. And she’s also recommending that her students give the AI a detailed inventory of sexual kinks that are not just the old, you know, typical ones. So this is actually more involved than just typing into a chatbot, like give me a romance novel about a big city lawyer who moves to the country and falls in love with the stable keeper. Casey Newton Oh, wait, that sounds interesting. What happens next? Alexandra Alter Yeah. Yeah, no, it was super fascinating, too. I sat in on one of Coral Hart’s classes, which was specifically about getting AI to write decent or even great. She said sex scenes. And she said, you are not going to get a good sex scene if you don’t carefully prompt it. You’re going to get weird euphemistic stuff. She mentioned that Claude had written in one of her recent drafts, his turgid manhood. And that was the kind of language she was getting. And so then she started writing lists of words that the AI loved, like shiver, unravel, manhood, moan, and blocking them, saying you cannot use these words. And I think the funniest thing that she said to students, she said it’s very important that you tell the chatbot to slow down because otherwise they just jump to the end of the scene. Everyone’s tangled in the sheets. (Time 0:30:42)
  • Authorship Shifts Toward Direction
    • Some AI-assisted authors view themselves as directors rather than sole authors.
    • This shifts authorship toward idea curation and away from word-level composition. Transcript: Alexandra Alter And I said, do you still think of yourself as a writer? And she said, I mean, not really. I’m more of a director. I’m a creator. She feels like she comes up with the plots and the characters, but she doesn’t necessarily think of herself as the quote-unquote author anymore, which is a different kind of species Of writer than we’ve seen before, I think. And a lot of people are very uncomfortable with that. Casey Newton One of the reasons that your story was so interesting to me is that romance as a genre relies on these templates, right? Like enemies to lovers or the slow burn, or I think there’s another one in your story, forced proximity, which I had never thought of as a romance template, but I guess it’s not dissimilar From our podcast. Kevin Roose Yeah. Yeah. That was actually the backup name for our podcast. Alexandra Alter Well, within forced proximity, you have only one bed, which is a great sub-subgenre. Casey Newton Very interesting. Well, we are getting a new studio, so we’ll talk about that. But because the romance writers use these templates, I think some might look at that work as maybe less creative than somebody who is writing literary fiction and is just sort of writing Whatever scenes come to mind. And I’m curious if you thought about that tension when you were writing this piece. And I want to be careful how I say this, because I think if you’re, you know, anyone who is working in genre fiction, if you’re coming up with all the plots yourself, you are bringing your Own humanity to that process. And yet I understand why some people are trying to automate it because they think to themselves, look, all of these stories hit the same eight or nine beats. And why bother writing them myself if the reader already knows where it’s going to go before they’ve started reading it? (Time 0:36:10)
  • Publishers Face Copyright And Contract Risks
    • Publishers worry about originality and copyright for AI-assisted books.
    • Traditional contracts require authors to warrant originality, creating legal uncertainty for AI content. Transcript: Kevin Roose Yeah. Alexandra, I’m wondering how the publishers are responding to this. The publishing houses that you mentioned in your piece all seem to be smaller or people that are self-publishing their books. But the big major publishers, I imagine, are starting to grapple with this too. And I’m curious what conversations you’ve had or are hearing about among the big publishers. Casey Newton Have they become turgid with rage? Alexandra Alter A lot of turgidity. It’s interesting because I think they see this happening in self-publishing, and self-publishing has become such a critical pipeline into traditional publishing. That’s where they’re finding these huge bestsellers, not just in romance, but thriller writers, even some self-help. So this has become kind of a feeding ground for traditional publishing. So they’re very aware that they are probably at some point going to acquire a book that has some AI in it. And most of them have policies that they’ve always had, which is their authors have to assure them in their contracts that the work is original. And what does that mean? If AI wrote it, does that mean it’s original? Some of them say that’s not original work, but if someone prompted it and they fed in their own ideas, then is it original? So it gets into these really sort of thorny areas. And I think there’s also a copyright issue for publishers because stuff produced with AI cannot be copyrighted. And publishers do not want to put out a book that they can’t hold the copyright for. (Time 0:42:32)
  • The ‘Ragged Prayer’ AI Quirk
    • Alexandra noticed the repetitive phrase “like a ragged prayer” appearing across AI-generated romances.
    • Writers attributed the phrase to model training data and began blocking it from prompts. Transcript: Alexandra Alter All right. So as I was reading through several of these AI-generated romance novels, I was commuting to work and looking at one on my phone, and I read this phrase, the hero and the heroine are in The throes of passion, and he whispers her name like a ragged prayer. And I was like, wait, I must have gone back to the other book I was just reading. I just read that exact thing. And I flipped back and forth and I realized that phrase was in several of the books and repeatedly within the same books. Whenever the hero says her name, he says it like a ragged prayer or sometimes like a jagged prayer or sometimes like a rough prayer. And so then I asked a couple of writers who I knew were using AI, like, what is this ragged prayer? And one of them said, I’ve actually had to block that phrase. It loves to say ragged prayer. Another one said, like, yeah, that’s an AI-ism. Like, one of them pinned it on Claude. And I couldn’t really, I tried to trace the origins of it. I did find it in a very popular romantic book by Sarah J. Maas, which was one of the many books that was ingested by Anthropic, according to the lawsuit that authors brought against Anthropic. And the phrase, set her name like a prayer, was in a sex scene in that book. But it’s just hard to know where it invented that. (Time 0:47:00)
  • Use Prompted Playlists For Better Discovery
    • Use Spotify’s Prompted Playlists to create tailored, data-driven playlists and avoid algorithmic drift.
    • Ask for constraints like play-count thresholds, sequencing, and regular auto‑updates to keep lists fresh. Transcript: Casey Newton So this is one that has been flying a little bit under the radar. It is a new feature in Spotify, and it is only available in a few countries, the US, Canada, New Zealand, I believe, and it’s only available to premium subscribers, okay? So there’s a little bit of a bar if you want to try this one, but in my experience, it has been well worth it. The feature is called Prompted Playlists. Have you seen these? I read about them in your newsletter, but tell me more. So from time immemorial, Kevin, we have all wanted the perfect playlist and we have devoted countless hours to crafting them manually inside iTunes, Spotify, whatever music system We were using at the time. But there are so many things that this system misses. And what I have found is that prompted playlists are a way to solve this problem. Here’s how it works. You open your Spotify app. There’s a tab somewhere there called create. You go to create. And if you have it, you will see prompted playlists in there. And it will show you a text box that looks exactly like chat GPT. And this is just basically a genie that you can throw a wish into and say, I would like a playlist like this. And then Spotify will go forth and do its best to create that playlist. Now, I’m looking at you, and I feel like you might be a little bored so far. Kevin Roose No, I’m not bored. I’m thinking about my own version of this and what I would use it for, but keep going. Casey Newton Well, I’ll tell you how I used it. I am a music nerd, and back in the 00s, I used iTunes to create these playlists that did something essential for me, which was they would keep track of my favorite songs and how recently I listened to them. And when it had been too long since I last heard my favorite song, it would just put those together in an automated playlist. Kevin Roose So like throw it back into the rotation, throw it back in the rotation. Casey Newton And while I vastly prefer the streaming era to the iTunes era in many ways, this is just something we lost. We took a very simple rules-based technology, we threw it out the window, and now we’re having to recreate it using something vastly more complicated. But as soon as I saw that these playlists came out, I wanted to see if I could get it to recreate the iTunes playlist of my dreams. And I wrote a very enthusiastic newsletter about this the other day. I will say my experience has been tempered somewhat in recent days, but maybe I should just tell you how I use this in case anybody else might want to try something similar. Please. So what I did was said, show me songs I’ve listened to at least 20 times, but not in the past two months. Please don’t repeat albums and create a well-sequenced playlist drawing from this set rather than just ranking by play count. And this was able to give me the playlist that I was looking for. It was songs that it knows I like because I played them, you know, a bunch and I haven’t listened to them in a while. And, you know, sometimes when you use an LLM, Kevin, they will lie to you. And so I actually called up Spotify and I said, put me on the phone with somebody who can explain to me whether this is actually working. Like, is it actually using my user, my listening data? And I wound up talking to the VP of personalization over there, a woman named Molly Holder. And she confirmed that, yes, when you type into Spotify, if you want it to use your listening data, and in my case, I have listening data on Spotify going back more than a decade, it actually Will do that. And this just enables all sorts of fun things. So she was telling me some people have used this playlist to say, make a playlist that is the opposite of my taste, and it will try to get you as far outside of your filter bubble as it can Kevin Roose Take you and if you’re listening to the hard fork podcast because you gave your spotify a same thing that said make a playlist with the opposite of my taste and it led you here welcome welcome Casey Newton You’re you’re safe now we don’t know what you were listening to before wait what is the opposite podcast to hard fork uh the megan kelly show the meg yeah the meg kelly So anyways, I encourage You to have fun with this. Some other things you might want to try. Molly was telling me if you’re traveling to a new country, you might say, make me a playlist of the top hits in this country or make me a playlist of some very popular songs in this country. You know, over the past couple of decades, that tends to work. You can also go really abstract. You know, you can say, make me the perfect playlist for eating fish tacos on the beach and just see, see what happens. So the thing I like about this is it’s a real kind of canvas for creativity and it is anti-slop. You know, I wrote about this in my column, you know, you have this great term machine drift, which is the process where an algorithm kind of grabs ahold of you and it leads you somewhere That you might never have wanted to go. This is anti-machine drip. This is you saying, hey, you already know a lot about me based on what I chose to listen to. I want to use that as the foundation to find more cool stuff that I might like. So this has been one of those AI tools that I actually found quite empowering in my life. (Time 0:51:43)
  • Birdsong Models Help Decode Whale Sounds
    • Transfer learning can let models trained on birdsong classify whale and underwater sounds.
    • Bioacoustic foundation models like Perch 2.0 help scientists categorize unknown marine audio efficiently. Transcript: Kevin Roose So Casey, I’ve got a whale of a story for us today. Love that. This comes to us from the great minds over at Google, who recently put out a paper that was about the use of AI to understand and interpret whale song and other underwater noises. So this is a new, what they’re calling a bioacoustics foundation model called Perch 2.0. And, you know, if I had to summarize this, I would say this is not a fluke. We’re not spouting nonsense. Free willy nilly. This is a new Kriller app for AI. Very good. No, no. Thanks. Very good. That was released that is about this new foundation model that they have built over there that allows them to categorize underwater audio, underwater audio samples from whales, dolphins, Orcas, you name it. And what’s really interesting about this is that it is not trained on whale song or any other underwater noises. It’s actually trained on bird song. And so they have found that these same sort of embeddings and techniques that had been trained on birdsong were able to also consistently label and categorize underwater noises. So it’s a kind of transfer learning. When you make these models big and general, if you give them one task, sometimes they also learn how to do related tasks. And so in this situation with Perch 2.0, they were sort of impressed and surprised at how good this model was at interpreting the underwater sounds that they were giving it. Casey Newton Well, so what are the whales saying exactly? Kevin Roose So this paper is really just describing a method of classifying sounds rather than sort of understanding what they map to in terms of like speech. We’re still not quite there with that, although there are a bunch of projects, including the Cetacean Translation Initiative, that are trying to understand the songs and noises produced By sperm whales. But this is basically giving scientists, marine biologists, a new set of tools to be able to, you know, if they hear something that they don’t recognize, to sort of classify that, to Maybe say this comes from this specific kind of creature, and that that will sort of help them detect and classify sounds going forward. Wow. Another thing I appreciate about these scientists is that they also have a sense of humor. Their paper is titled Perch 2.0 Transfers Whale to Underwater Tasks. They also make a reference to something called the Bitterne Method, which is sort of a play on an AI, a famous AI paper called The Bitter Lesson. So they are basically saying that some of the principles that we are finding for large language models with humans, where like if you make a model better at coding, it also gets better At math or some other related task. They’re also finding that that applies to things in animal studies. So if you make a model better at classifying bird sounds, it also gets better at classifying underwater sounds. So I don’t know what’s next for this particular project. This seems like an active area that a lot of companies, organizations are involved in. We may be getting close to understanding more about bird songs or whale speech, but I just thought it was interesting the idea that there is something generalizable about understanding Animal sounds as a whole, where you can take a model trained on bird sounds, use it to analyze underwater acoustic data, and find that it actually outperforms a lot of the more specific Models that are just trained on whale noises. (Time 0:59:24)