Welcome to the Vergecast, the flagship podcast of hydraulic-powered cutting machines. A very cool phrase that is going to make sense in 10 minutes or so. I'm your friend, David Pierce, and I am currently phone shopping. So, I have this iPhone 16, which is fine. It's blue, which is why I bought it, if we're being honest with each other. And the problem with it now is that I miss having some camera power, but what I really miss is having a battery that isn't awful. This battery is awful. Like, I'm at like 3:00 PM every day and I'm having to charge my battery. So, there's a world in which I could just, you know, replace the battery or upgrade to an iPhone 17, but I figure I'm a tech journalist. So, what if I just go out and experience a bunch of phones? So, I'm going to try a bunch of stuff. I have a Pixel here. I think I need to get a foldable phone, but I actually want your help, which is: what phone do you think I should get? I'm in a phase of being sort of unusually willing to switch from iOS to Android. Switching operating systems has traditionally been very hard; people largely don't do it. Um, I'm very willing to do it. I don't know if the answer is like I should go get the Samsung Z trifold for $3,000 or buy one of the flip phones that everybody's excited about, including me, or if the answer is just shut up and go buy the orange iPhone 17 Pro, which I will like very much. I don't know if you have thoughts, especially like weird thoughts about what phone I should get. I want to hear them. The hotline is 866-11. The email is
[email protected]. Get at me. I'm going to do a bunch of weird phone experiments on this show over the next couple of months and then I don't know what I'm going to do. I'm going to break iMessage forever. I know that for sure. But that's not what we're here to talk about today. Today we're here to talk about two things. First, we're going to talk to Will, a reporter at the Washington Post, about a big story he and a couple of his colleagues wrote about the way Anthropic and other companies are training their AI models using books in particular. There's some really fascinating details and some really big questions about how we're supposed to feel about AI inherent in all of that. So, we're going to talk about it. Then, Julia Alexander, our old colleague at The Verge, who is now at Puck, is going to come on and we're going to talk about Netflix and movie theaters. We've been talking a lot about Netflix recently, but I think this company is important and fascinating and also kind of a way to talk about the whole entertainment industry all at once. After that, we have a hotline about the smart home. Gen 2 is back answering weird smart home questions. It's going to be awesome. All that is coming up in just a second, but I've just realized that I have to go charge like 12 phones in order to do this experiment. So, here I go. It's phone charging time. This is the Vergecast. We'll be right back.
All right, we're back. So, over the last couple of years, there has been this slew of lawsuits against AI companies all about the way that they've trained their models. These AI models, these large language models, require just a vast amount of data. And to get all of that data, these companies are going and getting whatever they can, right? There was lots of reporting a couple of years ago about OpenAI essentially transcribing every YouTube video on the planet and then feeding all of that into its models. There's been a lot of talk about books in particular, a lot of authors and publishers suing these companies over the way that they are acquiring and then actually using that data to train their own models. So Will Armus at the Washington Post wrote a story with a couple of his colleagues about this thing called Project Panama, which was an Anthropic project to digitize and use just an unbelievably staggering amount of books to train its models. And it's not the only company doing this, but we have a lot of interesting data thanks to some newly unsealed documents in these cases. Uh, Will and his colleagues went through all of it. And so I'm going to make Will explain to me how this works and why it's become such a big part of our conversation about AI. It's a really fun conversation. Let's get into it. Will, welcome to the Vergecast.
It's snowy. It's icy. You've made it. We're all... I feel like everyone's internet is just hanging on by a thread right now.
I think everyone is just hanging on by a thread, period. I'm joining you with a little ice rash on my face from a minor sledding incident yesterday. Uh, it's all good. I'm ready to go.
Yeah, my kids have been home from school all week and I am just very slowly coming outside of my brain and body as the days go on.
Mine is in the house as we speak. So if you hear any uh squeals of delight or cries of terror, it's him in the background.
It's just don't worry about it. Everything's fine. Exactly. So, uh, you were part of a team that wrote this story for the Washington Post about, uh, Anthropic and this big book digitization project they underwent that I think gets at a bunch of the most interesting things going on in AI. There's a piece of this that is like how AI models actually get made that I think people largely don't understand that this I think elucidates in a really interesting way. And it also gets to a big galaxy brain theory I have about AI that we're going to talk about at the end. I'm very excited about it. I'm looking forward to... Let's just start at the beginning here. Uh, what was Project Panama and why is this something that Anthropic a while back threw real resources into?
Project Panama was something that Anthropic started in late 2023 or early 2024 with the goal to quote, 'destructively scan all the books in the world.'
It's, yeah. It's really unfortunate. It sounds like something a Bond villain would set out to do. Yeah. Um, they didn't actually want to destroy all the physical books in the world. Um, just one copy of each. Um, sure. So what they were trying to do though is to scan the books and the destructive part comes in with the way that the books were scanned. It's more efficient to slice the spines off and scan them. Uh, so you've just got a stack of pages instead of, you know, you ever tried to spread out a book on a Xerox? It's like a little difficult. So they just cut the spines off. I think there was some indication in the court documents that they would, you know, a recycling truck would come and back up to the warehouse afterwards and get what was left of the books. But the point was that they wanted to digitize all the books. And the reason they wanted to do that is to feed the content of those books into their AI models which power the popular chatbot Claude.
So why books? I think a thing you discovered, and a lot of this information we should say came out in court documents based on, I mean there's just a million lawsuits about all things AI right now, but one of the big ones I think that you got a lot of these documents from is a sort of broad suit from a bunch of publishers and authors saying that their books and works were used illegally in the creation of these models. Is that right? That's where this data comes from.
Yeah, that's correct. And so you said why books, and I think the first answer to why books is because everything, right? It wasn't just books they wanted to, you know, the AI companies wanted to vacuum up everything that's published on the internet, everything that's published in government records, every video, every photo, every piece of art, everything, right? They wanted it all because they're trying to build general super intelligence and they wanted to know everything. But the second answer to why books is because books were specifically seen as a source of really high quality on average. I mean, there's crappy books out there as we all know, but like on average, books are high-quality content. They've been vetted by someone. Someone went to a lot of effort to write them. Someone deemed them worthy of publishing. Um, they might be fact-checked. They might contain information that you can't get anywhere else. You know, that's not on the open internet. So, Anthropic, there's some evidence that Anthropic's executives saw books as a way to catch up. And backing up for a second, I mean, Anthropic is an underdog in this massive AI race to control the future of technology with bigger dogs like OpenAI, like Microsoft, like Google, like Meta. And Anthropic saw books as a way to kind of bootstrap their way to the state-of-the-art in terms of the quality of their models.
Okay. So I do find the quality piece of this really interesting because it does seem like if you were to just train a model on, I don't know, YouTube videos, which is a thing that has been talked about a lot, right? If you were just to transcribe and feed every YouTube video in existence into an AI model, it would learn to talk in kind of a very specific way. And it would learn to talk in a very different, very specific way if you just fed it the content of social media posts or blogs, or I'm just thinking about the things that are sort of readily available on the internet. Like if you just train a model on Tumblr, it would talk like Tumblr. And books, it occurs to me, are, if you want something that speaks in coherent sentences, that maybe books are the right place to start with that. At the top of it, you have something that is designed to be read and consumed and understood by lots of people that has been through a whole process to make sure that that's okay. So, in a way, it seems like it is kind of the pantheon idea of how you want to teach these chatbots to communicate, right?
Yeah, I think so. And you know, to be clear, Anthropic wasn't the only company that thought books were important, but I do wonder if the emphasis they put on books, and this is pure speculation on my part, I do wonder if the emphasis they put on books is one reason why a lot of people swear by Claude as the best writer out of the chatbots. You know, people who want to use AI for writing often will say Claude's better than ChatGPT or the other options out there.
That's a good theory. Um, so walk me through the process here. I think one of the things I just kept thinking about reading your story is like if I'm at the beginning of this journey and I'm like, okay, I want to get a bunch of books into my model. I literally don't know how I would start. So where, how does one first of all, I think I assume step one is acquire many, many, many, many books and put them all somewhere. How do you do that if you're Anthropic or anyone else?
Well, I don't know if this is the best time to bring this in, but the way they actually started was not by acquiring real physical books. The way they started, which is the same way that according to court documents, um, OpenAI is alleged to have started, which is by pirating vast libraries of already digitized books.
These are called shadow libraries. They're available online in shady corners of the torrent world. Um, and there are allegations in a separate case against OpenAI that uh, one of the OpenAI executives named Ben Mann, back I think back in 2019, like way before the ChatGPT boom, um, downloaded the entirety of this famous shadow library called LibGen from a torrent site, right? Um, LibGen I should say is like if you imagine what the Pirate Bay is, it's that for books. Like it literally, there's a pirate ship on the website. Like it's not unclear what this library is trying to do.
Right. Right. And I mean the backstory of LibGen in particular is fascinating. We don't need to go too deep into it, but it has its roots in Russian samizdat, you know, when academics were trying to get access to verboten materials uh to evade censorship. But nowadays it's traded around online as this vast, like, probably illegal repository of millions of digitized books. And so OpenAI allegedly downloaded this. And then the guy who allegedly downloaded it at OpenAI was one of the people who left OpenAI to co-found Anthropic. And so then he gets to Anthropic and he does it again. He downloads all of LibGen again at Anthropic and there's like the court documents, you know, there's screenshots of his browser with this, you know, with the torrent site open and like LibGen partway downloaded. Um, and so uh, that's how they started is just straight up piracy. I don't know if I need a 'allegedly' in there. I think they've been like, they're not really contesting it. Um, it was basically straight up piracy.
It's all over like their emails to each other that this is a thing that they're doing. And that is a fairly accurately stipulated part of what happened here.
Yes. And again, our story is about Anthropic because that's the company where we found newly unsealed court documents that shine a light on this big Project Panama to acquire real books and scan them. But once again, I just want to emphasize that like as far as we know, all of the AI companies were trying to get their hands on books and there's evidence that more than one company was downloading these pirated shadow libraries. Um, and not just Anthropic.
Okay. But you got to tell me about the physical books, Will. This is the thing I still cannot wrap my head around. I understand downloading lots of things illegal on the internet. I've done that. We've all done that, right? Like I get how to download things on the internet. I've never done that. You were like, download. That's right. We should be clear. Uh, this is not legal advice. Um, but I literally, I just found myself being like, 'Okay, if you just came to me and were like, David, you have unlimited resources. Go get me 100,000 books as fast as you can,' I straight up would not know where to begin. Um, and one of the things I thought was so funny about this is um, I think where Anthropic began was by hiring the guy who had done it already, which leads me to believe there is like a how-to-buy-books-by-the-million expert out there in the world who is just running around doing this for people.
I love these venture-backed tech companies. They're like, 'Okay, we need to figure out how to scan a ton of books. Let's go hire the guy who literally oversaw the Google Books project, like the biggest books scanning project in the history of the world. Let's just go get that guy on the payroll and uh let him figure it out.' So, they hired this guy, Tom Turvey. This is why, David, this is why they did not hire David Pierce because you wouldn't know where to start. But Tom did.
I would have not done a good job. I will say I got to the last round of interviews for that and then they were like, 'How do you do it?' I was like, 'I don't know.' And then that was it, man. Next time. Next time. Um, so, uh, Google Books, by the way, non-destructively, as I understand it, non-destructively scan books. They use this much slower process involving robots to go through and scan all the pages without breaking the spines. It's kind of like fascinating. Um, I went down many rabbit holes in the course of reporting the story and that was one of them. Um, but in this case, they didn't feel like they had time for that. There's also a difference between Google Books was scanning books from libraries, right? And so, this is going to include rare and valuable books. You don't, I mean if you're a library you don't really want a company going through and slicing the spines off all your books. Um, in this case, uh, what Anthropic did, there is some evidence that Turvey did a bit of outreach to publishers and authors like, hey, could we, what would happen if we wanted to license your work, you know, use it to train our AIs, could we, would you be open to that? And pretty quickly he decides that's not viable. It's going to be too expensive. A lot of authors and publishers aren't open to it. Even if they are, like it's not going to work at the scale and at the speed that Anthropic needs, presumably the scale and speed they need to compete in this all-out race to build super intelligence first. And so instead he turns to these vast used book warehouses um like Better World Books which I believe has an agreement with the Internet Archive. Um, there's another one called World of Books um that I think is based in the UK. I apologize if I'm getting them mixed up. Um, but these are places that actually literally have hundreds of thousands if not millions of books. Mostly not rare or super valuable books, I should say. So, destroying them isn't, you know, isn't necessarily a tragedy for humanity. Um, and then they're super cheap. You know, he buys them in bulk um in giant amounts and then finds these scanning shops that can do it super fast and get uh, you know, I think they at the very least they got, there was one order in the court documents with a vendor who was supposed to supply them something like half a million used books at once.
Wow. That is nuts. So, okay, let me tell you the picture that's in my head of how this works and you tell me how correct I am. The way that I picture it is, do you know those money counting machines that you just take a stack of bills and sort of stick it into and it just like feeds it through really fast and sorts them all out for you? I'm imagining that essentially, but for a book. And so, you get a thing that slices off the binding. So, what you have is just a giant stack of papers. You feed that stack of papers into this machine and it just rapidly slots them through and scans them one by one and out you have a digitized OCR book.
That is my understanding as well. And I should say I didn't observe this happening. I'm, you know, we're, I think we're both extrapolating from what was in the court documents and the files but that sounds basically right to me.
Okay. They called it a hydraulic-powered cutting machine which is another of the extremely good phrases that come up in this. We have shadow libraries. We have destructive scanning and we have hydraulic-powered cutting machines. Like there's just, this all feels very low-stakes Bond villain. Like you said, there's a lot going on here, right? Okay. So then you end up with whatever, it seems like somewhere between hundreds of thousands and millions of books worth of stuff that goes into these models at the end of the process.
Yeah, that's right. The exact numbers are still redacted. Um, but we're extrapolating from the fact that some of these orders were for hundreds of thousands of books at a time.
Okay, got it. Um, and then you know, fast forward to these models start to ship, people start to use them, and people start to find their own work in them, and then you get to these lawsuits. And so my biggest question for you here is we've talked over time about a lot of the different ways that these companies have gone and gotten material. And I think your point about Anthropic's experience of saying we went out and we asked people if we could buy it or license it and found it to not be expedient, ordinarily what happens is then you don't do the thing. And what all of these companies, and like Anthropic is not uniquely guilty of this, all of these companies just did it anyway and it became the thing that happened. Um, and I want to come back to that because I think this gets to the galaxy brain theory I'm going to throw at you in a minute. But is there something different about what happened to and with books than what happened with all of these other sources? We've talked a lot about, you know, source material from all kinds of different stuff on the internet being fed into these models and there are a lot of questions about it, but it does feel like the way that books are being both discussed and actually litigated is different. Is it different?
Yeah, I think it is a little different. And one of the reasons is because so much of what went into these AI models, particularly at first, was stuff that was just available on the internet, right? That was the low-hanging fruit. And if your stuff was just available on the internet, it's not that you couldn't possibly have a copyright case. There have certainly been lawsuits about that stuff. But I think that there's a sense that you have a stronger case if you had something that was supposed to be bought. Like you weren't supposed to be able to just get it for free and they got it for free somehow anyway without your permission. That's seen as, I think, a stronger copyright case. And in fact, we can get into this a bit, but in some of the court rulings so far, that in fact has turned out to be the case. The judges in both the Anthropic and the Meta case, albeit for different reasons, both found that the actual training of the AI model on the book content was fair use or at least that the plaintiffs in Meta's case, at least the plaintiffs had failed to show that it wasn't fair use. So the part that everybody thought that these lawsuits would really revolve around was this unauthorized training. And that's the part that at least two judges so far have been like, actually that part was okay. The problem was how you acquired these books in the first place. So I think that is what makes books different is the fact that they're difficult to obtain and that you had to go to these wild lengths to get them in the first place.
Interesting. As opposed to, you know, there's a thing that someone wrote that is just at a publicly available link on the internet that the legal definition notwithstanding, that feels different from a book that you ordinarily would have to go to a store and buy.
Yeah. Exactly. And then there's in-between stuff. So like one of the most high-profile AI copyright cases is New York Times versus OpenAI. And in that case, there's been a lot of discussion about how did OpenAI get around the New York Times paywall, right? Like what did they do to circumvent that in order to get access to all the New York Times material?
Okay. Um, suss out the difference here for me because I could sort of see in the piece you spend a lot of time reckoning with where we are on the legal spectrum here. Is it illegal to train the model on this information? Full stop. No, because it is transformative and fair use. And my sense from the reporting I've done and from the story that you wrote is that there are a lot of people who are like, 'Okay, that is an early decision. That is kind of where the winds are blowing, but this is by no means settled law.' Is that right?
That's right. Yeah. It's very much still unsettled law. And the two big rulings that we looked at with the book authors are Anthropic and then this other case, Kadrey versus Meta. And I think that it was really interesting that the judges in the two cases, even though they both ruled that it was fair use to train the AI models on the book content, they had totally different reasoning. In fact, the judge in the Meta case, who ruled afterwards, criticized the ruling of the judge in the Anthropic case. So the judge in the Anthropic case said, 'When you use books to train an AI model, that is quote exceedingly transformative end quote.' And transformative is one of the key tests of whether something is fair use. So, like, you know, if you copied a million books and then you turned around and sold them for 10 cents a piece, like that's probably, I'm not a copyright lawyer, but that's probably a pretty clear-cut copyright violation. But he's saying if you take a million books and then you feed them into this AI model and you put out a chatbot that like answers people's questions and doesn't produce entire books verbatim, which by the way there's nothing in the record that indicates that Claude would spit out these books verbatim even if you asked it to. That's something different. It doesn't compete with the books and so that's fair use. In the Meta case, the judge was like, 'Actually, I think you could probably give me lots of good arguments that using books to train AI models is going to end up hurting the market for books, right? You're going to end up with AI models that can produce a book of their own, even if it's not exactly the same. Um, you're going to end up, and we already have this. I mean, there are AI rip-offs of books all over Amazon. I mean, anybody who's published a book lately has probably seen some crappy AI clone of it turn up on Amazon.' Um, and the judge in Meta was like, 'You guys just didn't, you failed to argue that, like you didn't bring me the evidence of that, so I can't rule on that.' But he left open the possibility that it might not be fair use. But in any case, what both companies are still, what Anthropic ended up having to settle over rather than go to trial and what Meta is still facing litigation over is how they acquired the books in the first place.
Right. And the thing I've never been able to figure out is why that fact doesn't kind of obviate and make irrelevant the other fact. Like I got this data illegally. Now I can legally use it. Just doesn't, I don't know. I have a hard time following from one to the other. Does that make sense?
Yeah. And I did too. I mean it's completely counterintuitive. I don't know how deep we want to go on this, but the Anthropic ruling as someone who's not a lawyer was really counterintuitive to me. So actually the books that Anthropic scanned, you know, acquired legally, scanned and then fed into commercially released models of Claude, that was ruled to be fair use. What the judge was not going to let them off the hook for and what they were going to face trial over and potentially tens of billions, maybe hundreds of billions in statutory damages had they lost um in the worst way possible, was the books that they did not use for training commercially released models. Because he's like, if your defense is that this was fair use, that the reason you acquired these books was to do something totally different with them, then the ones that you didn't use, that can't be fair use. It can't be fair use if there was no use. And so what they got in trouble for were the ones they didn't use, ironically, and had just stored and decided not to put into their commercially released models. There are other legal experts who think this decision was like totally weird and backwards, but that was the ruling. And so Anthropic ended up settling with the authors in this case for $1.5 billion, which if you're Anthropic is a lot of money. Um, you know, probably not so much if you're like, if you're Meta, right?
It is both a lot of money in raw terms and not a lot of money compared to I think the scale some of the people against this are talking about. Uh, I mean, we've been hearing, I think it was Marc Andreessen a while back who was like, 'If we had to pay a fair price for all of this data, the AI industry would go out of business.' And all these people were like, 'Yeah, dude, that's our whole point.' Like that is the thing we're trying to say. We agree on this fact. Um, and so I think again all of this is unsettled, but this brings me to the thing that I've been thinking about a lot. And a thing I have spent a lot of time reporting on and talking to people about and just trying to figure out is why the reflexive backlash to AI is so intense and where it originates from. And a theory that I've developed and I'm just curious what you think about this is I think there is a sort of original sin in the AI community and I think it lays at the feet of OpenAI for starting the race the way that it did. Um, because, and there's a bunch of this in your story, but OpenAI basically started as this sort of academic research team building things for academic research purposes essentially. And they did a bunch of things that in academia are generally done right, like take a bunch of research papers and put them into your AI model. Less problematic when you're doing something to extend research than when you're doing it to try to make billions and billions and billions of dollars. They did this thing and then as OpenAI became a commercial product, it didn't change the way that it acted to source and train its models. And I think, again this is just a theory, but I'm increasingly convinced that I'm right about this, that I think the way, the sort of cavalier way that OpenAI approached building and training its models and then the fact that it released it and started this war where everyone had to catch up as fast as they possibly could or risk being essentially left behind by this like train that was already out of the station. So you get Meta having to essentially run the same playbook and take the same shortcuts. You have Anthropic essentially having to run the same playbook and take the same shortcuts. Anthropic in many ways it seems like actually tried to do this in a more upstanding way by actually going and buying some books in the world, which is something, right? Like that adds to the fair use questions, but at least they bought the damn books. That counts for something. And to me it's just like if you back it all the way up, if they had just come to the end of, well we can't do this, this will take a while and cost us a lot of money, and they had decided to take the time and spend the money, I think as a society we might feel really different about AI.
I think that's a really compelling theory. I think a lot of people, I think some of the anti-AI sentiment does come from this feeling that there was a grand theft at the beginning of it, that nobody was asked for permission and they just took everything anyway and now they're making, you know, they're poised to make billions upon billions of dollars and transform, you know, replace, put people out of work with a product that was built on the back of those people's labor. I think that's, you know, that's at the root of some of the backlash against AI. I think there's probably even more to it than that. Um, I should, a couple quick notes. I think you're right. I mean, I think there is a way of looking at what Anthropic did where you're like, at least in the scheme of the different tactics that the AI companies took. I mean, Anthropic was doing it like almost relatively the right way, right? Like they started by pirating stuff and then once Claude was out and commercial they were like, let's try to find a more legal way to get these books and try to put ourselves on better legal footing. Yeah, there's a way in which what Anthropic was doing was like trying to do the relatively right thing. But I think there is a sense from the people who created all these works in the first place, the books, the movies, even the blog posts, that the actual right way would have been to come and ask permission and maybe pay people to license even if it's a little bit, you know, um, just pay the fair share. But um, so far that's not the way the judge's rulings are going. So, but I want to come back to this sort of what does it look like to be the good guy here? Because I think if I were to be um unexpectedly slightly sympathetic to the Anthropic team here, um, it does seem like there's a lot of stuff in the documents that have been out and there's been a lot of stuff in these trials that has essentially made it seem like the...
The people at these companies, including Anthropic and Meta and others, who have felt like they had no choice but to play this game this way, right? That they looked at it and they said, 'Okay, there is no good guy here. There is only winning and losing, and our opportunity is either to play this game the way that we have to play it or we lose.' But you've been in these documents a lot more than I have. Do you have a sense of if there were folks inside of these companies who identified kind of the white hat way to go about building these things that might have actually worked? Was there an alternative?
There is relatively little discussion in the documents I saw beyond what I've already described of better and more ethical ways to do this. I think, you know, again, this is not coming directly from the documents. This is an inference, but I think you're right that there's a sense as you read these court files that it was understood that, you know, we can't afford to lose this race, and, you know, let's try to get legal sign-off, let's try to get approval for what we can do, but losing isn't an option.
Right. I mean, on the Meta side, right, there's some implication that these debates went all the way up to Mark Zuckerberg, and he's the one — like, Mark Zuckerberg has made very clear that he feels it is an existential risk to lose the AI race. So it's not surprising that he would push, come to shove, say essentially by whatever means necessary.
Yeah. And so what we have there is we can sort of infer that from a document where there's an internal chat between Facebook employees, and one of them says, after prior escalation to MZ, being presumably Mark Zuckerberg, you know, GenAI has been approved to use LibGen, which is again this shadow library. And so, you know, we'd have to speculate about exactly what Zuckerberg knew. I mean, he said in an, I believe in a deposition, that he didn't know what LibGen — he said, 'I don't really know what that is.' But, you know,
You can make a meal out of what that sentence actually does and doesn't say. You know what I mean?
Yeah. And these are not direct quotes, by the way, but this is my recollection, my best recollection of what the documents said. And I think one more thing that's noteworthy here is that in any field, you know, under capitalism, there are companies racing to try to be the winner because, you know, if you're the loser, it doesn't matter what your ethics are if your company doesn't exist anymore, right? But in AI in particular, this industry is so shaped by this idea that we have to be first in order to save the world. Like, that's the founding principle of OpenAI — like, we have to be, let's found a company that's going to build AI in an open way to save the world from these hypothetical other companies that would build AI in a closed way. And then it turns out that in a lot of people's view, OpenAI did move to a more closed way. And so you get people splitting off and forming Anthropic, for instance, saying like, OpenAI has strayed from its principles. We can't trust them to save the world from runaway superintelligence anymore. We've got to do it. We've got to be the saviors. But again, if you don't win the race, you don't get to dictate how the superintelligence turns out. And so it's this really, it's this sort of moral paradox where everybody's like, the ends justify the means. And so, you know, we might have to break some eggs along the way to save, you know, because the ultimate goal here is to save the world.
Yeah. And if you believe the stakes are that high, and I think a lot of people in this industry earnestly do, then yeah, the breaking a few eggs along the way is a pretty worthy sacrifice.
And that's why some critics believe that maybe some regulations would be in order in order to constrain that race in some way.
Some critics, some reasonable people, some normal people, some objective journalists, David.
All right. We got to take a break, but Will, if you ever get your hands on one of these book scanners, I demand that you come back and show me how it works. This is the only thing — I feel like I have a pretty decent picture of how this thing works, but I just, I need to know how big the scanner is, is part of it. Like, is it the size of a room? Is it a thing I could put on my — I just, I need to know everything about how to scan books very quickly. I now have the Big Lebowski line in my head like, 'You need a scanner, David. I can get you a scanner.' We'll see. You can have me back on if I get you the scanner.
Deal. All right. Until then, Will, thank you for being here. This is great. Thanks for having me. We got to take a break. We'll be right back.
Support for the show comes from Built Rewards. Renters out there, listen up. It's about time you earn something back on your biggest monthly expense. I mean, you can earn loyalty points on pretty much anything else these days. So why not on your rent? Built is the loyalty program for renters that rewards you monthly with points and exclusive benefits in your neighborhood. With Built, every rent payment earns you points that can be used towards flights, hotels, Lyft rides, Amazon.com purchases, and so much more. And starting this month, Built members can earn points on mortgage payments for the first time. That means you'll be able to get rewarded wherever you live. Plus, using Built unlocks exclusive benefits on more than 45,000 restaurants, fitness studios, pharmacies, and other neighborhood partners. Join the loyalty program for renters at joinbuilt.com/verge. That's j-o-i-n-b-u-i-l-t.com/verge. Make sure to use our URL so they know we sent you.
All right, we're back. So now let's talk about Netflix. We talked last week about Netflix and podcasts, and now I want to talk about kind of the exact opposite end of the spectrum, which is big, huge, splashy movie theater movies from Netflix. Netflix has made a lot of noise about the way that it's going to think about movie theaters and movies as it tries to buy Warner Brothers. This has become a big part of this deal is what will happen to the movie studio and the movie theater movies that Warner Brothers makes. And Netflix says all the things you would expect it to. But the question of what does Netflix want from movie theaters and movies in general, I think is really fascinating and brings up a lot of questions about like, what are movie theaters even for anymore? What does it mean to make a movie theater movie in 2026 if you're Netflix or anybody else? So Julia Alexander, who was our former colleague here at The Verge, she was actually on Decoder last week also talking about Netflix. We're all very good at planning here at The Verge. She's now at Puck writing about all of this stuff there. I asked her to come on the show and just basically explain to me what Netflix is doing, whether we can fix movie theaters, and whether any of it even matters. So we had a lot of fun, and I think we did fix movie theaters. So let's get into it. Julia Alexander, welcome to The Vergecast.
So excited to be back after many years away from this.
No one ever gets away. This is one of the things Nilay always says is no one ever actually leaves, and I'm very happy to have you back.
Yeah, I'm excited to be here. It's one of the few podcasts I sincerely listen to. When people say I listen to your podcast, I do actually listen and watch Vergecast.
So, do you have a lot of thoughts about Brendan Carr? That's actually what we're going to do now.
Yeah. Yeah. That's when I'm not texting Nilay about the Green Bay Packers. It's just Brendan Carr.
Perfect. So I brought you here because I want to talk about movies, and I want to talk specifically about movie theaters. Because we're in this interesting moment. You spend a lot of time thinking about sort of the machinations of the whole media business, and I think a lot has been said about what's happening to movie theaters. Movie theaters have obviously been in kind of a bad way for a minute here, but now with this Netflix Warner Brothers Discovery deal potentially happening, movie theaters are being talked about again. And I think the more I think about movie theaters, the more it makes me think about the whole structure of what we consume and how. And so I just want to dig into some pieces of this with you. But my first question for you, as somebody who pays a lot of attention to the business of movies and TV, is: do you think Netflix is serious when it says that if it buys Warner Brothers Discovery and takes over one of the biggest studios in Hollywood and would suddenly become, if it wanted to, a huge player in the theater business, do you think it's serious about caring for movie theaters and being a real player in movie theaters?
This is one of those annoying, weaselly yes-but answers. There's a lot to the Netflixification of theatrical. And the issue is that a lot of what they're saying, as you know and as you have talked about before on the podcast, is they're just saying stuff to get this through the regulatory hurdle. So of course Ted Sarandos and Greg Peters, co-CEOs of Netflix, are going to come out and say like, 'We love theaters. Actually, we've been thinking about theaters forever.' Which is what they said on the most recent earnings call after years of publicly saying that they don't need to be in theaters, that their commitment is to their members, that they can build up just as much anticipation and excitement for film as a theatrical release can do. Now all of a sudden they're saying, 'Actually, we've always loved theaters, and we just didn't make it a priority because we were too busy making Squid Game or something for our streaming service.'
So the yes portion is that we have a lot of data that does suggest theatrical movies perform better on streamers, and there's a lot of reasons for this. They spend a ton of money on marketing, so the awareness of these films is higher. We associate — if you look at kind of the qualitative studies being conducted by theatrical exhibitors, of course, but also studios and kind of audience surveys — the sentiment around theatrical releases is that they are higher quality. People will typically go to theaters for IMAX films, so they associate it with this really intense experience that they like.
