N.Y.C. Says Airbn-bye + How Far Would You Go for a GPU? + The A.I. Songs of the Summer

23:36Okay, so we've talked about chips on this show. We have. Against my will, but we've talked about it. So a few months ago, we did a show about NVIDIA and how they kind of became a central part of the AI economy by making these things called GPUs, which are these high-end graphics processors that you need to build large language models. And this week, NVIDIA reported a 101 % jump in its second quarter revenue compared to last year, which they said was in part because of the huge demand for these chips. And the other day, my colleague Aaron Griffith published a great story titled The Desperate Hunt for the AI Boom's Most Indispensable Prize, which was sort of all about the crazy, insane lengths that companies are going to to procure these high-powered GPUs.

24:23It's like, you know the game show Supermarket Sweep where you have to put as many groceries in your cart as you can before time runs out? Yes. It's like in the AI industry, there's a giant supermarket sweep, and all anyone wants to put in the cart is GPUs. Exactly. Remember on Supermarket Sweep how they always went for the hams first? Always the hams, the big cheeses, the diapers. Yes. Yeah. So these are the hams of the AI world right now. Everyone wants them in their carts. I thought we were the hams of the AI world, but go on. So I wanted to have Aaron on because I think it's just a great story and a good way to understand how insane the market for these AI chips is right now and what lengths people are going to get them.

25:02So, Aaron Griffith, welcome back to Hard Fork. Hi, thanks for having me. Hi, Aaron. So I was hoping that we could just start by talking about kind of just the lengths that you describe in your story for how far some of these companies are going to procure GPUs. You talked to one executive who basically called GPUs a rare earth metal. So what made you want to do this story and what did you hear while you were reporting it? Well, I've been going to a bunch of AI events, hackathons, meetups, pitch events, just to get a sense of what's happening in this kind of boom time. I mean, there's tons of companies springing up, and I want to know which ones to pay attention to.

25:39And I noticed that one of them, that when founders were pitching their startups to a bunch of engineers that they're trying to recruit, one of the main selling points that they were including alongside, you know, all the regular perks like free lunch and stock options and all this stuff was like, Oh, and we have access to all these H100 GPUs, or we just raised$40 million and we're allocating$8 million of it to GPUs. And I just thought that was kind of a new thing that I'd never heard before. Like, this is a recruiting tactic. Like, we have access to compute. I was like, well, isn't that just kind of a given?

26:14We live in the cloud era. We live in an era of computing abundance. You know, like, why is this suddenly a hot commodity that you're bragging about having access to? Right. Startups used to brag about doing your laundry. And now it's like, get a load of this chip. Yeah, yeah. It seemed so bizarre. I was like, it almost reminded me of like the supply chain issues in the beginning of the pandemic. It's just like, what is happening here? Like, we live in a world where we can't get access to everything we want at the tap of a button. What? And so I started, you know, talking to more and more startups.

26:45And all these comparisons came out. The rare earth metal, people talked about it like it was drugs. Like, you know, you have to know somebody and have a hookup. People are just walking down Market Street in San Francisco in trench coats, like opening up and saying, hey, buddy. So that's actually, that's why I take issue with your supermarket sweep comparison, because, you know, that would be like if the supermarket shelves were empty. And it's more like a sort of Hunger Games situation. I see. It's a post-apocalyptic supermarket sweep where there's nothing left on the shelves. Right. Yes, exactly.

27:14So the more I started talking to people, the more I realized this is an enormous problem that's impacting startups. Erin, the most upsetting thing that I learned from your article was about the existence of something called the GPU song. Oh, I'm so sorry to have introduced that to a broader audience. Which is a song that was put out by someone calling themselves Weird AI Yanko Chip. And it is a version of We Didn't Start the Fire that is all about the quest to procure GPUs. And I think we should listen to it.

28:01Elon Musk, GCP, Sam Altman, AMD, Inflection, for an Alan Turing test. Oracle, Anthropic, Google Cloud, Symbiotic, NVIDIA, Microsoft, AWS.

28:20US. Hugging face, DGX, Lambda Labs, AlexNet, AMD, InfiniBand, A100, LLMs, Twitter, Tesla, SpaceX, Tiny Corp, HGX, Call of T, SMC, EMODM, Stability. GPUs are fired! Everybody's hunting for some H100s. GPUs are fired! Who is this song? Do they play this at parties in Silicon Valley? I don't understand. Did you interview weird AI Yanko Chip? I didn't. I didn't. I had limited space and I didn't want to oversell. Well, my understanding is it's very hard to get on the phone. It's sort of like, you know, getting an interview with Beyonce. So it's not, to my understanding, just that startups are using their GPUs to recruit engineers.

29:12they're actually being used by lots of people as different types of incentives. They're also using GPUs as collateral. This is something that I learned for the first time from your story, which is that at least one company has gone out and sort of raised money, like gotten a loan with their GPUs as collateral, the way that like you might use your house as collateral for a loan. Yeah, that was crazy. And so I guess my understanding is that that market is not new, but it has gotten a lot more intense, you know, in the last year. So I guess there's always been the ability to finance your GPU or your servers or anything like that with this special kind of debt that, you know, if you default, the debt provider just they just get the GPUs.

29:54And so, yeah, there are now, you know, there's all kinds of weird shadow markets that have sprung up to service this disconnect in the market right now. Erin, what is the craziest story that you heard about a startup or a company trying to get NVIDIA's GPUs? They're making hundreds of phone calls. They are calling in every favor. I mean, this one guy that I talked to that I quoted in the story, Eric Donas, he was calling his friends at hedge funds that like run quant funds, seeing if they might have extra capacity lying around outside of trading hours. I mean, he basically stopped short at what he called the equivalent of buying some GPUs from a guy in a van.

30:36There are like third party brokers that are, you know, obviously seeing an opportunity to get in here and maybe mark them up or something like that. Yeah, that's what I want to know is like, who is the stringer bell of GPUs? Like, who is the person who's just like driving around vans in San Francisco, just slinging GPUs out of the back? I mean, that could be these guys from SF Compute Group who were also in the story. It's basically two startup guys who had trouble finding GPUs for their own startups. And then they somehow got a line in somewhere after working the phones for a little while. And they're now banding together a bunch of startups and researchers to essentially like share an allotment of GPUs.

31:16And they wouldn't even tell me where they were getting them from because they were worried that if they publicize that, that somebody might swoop in and buy them out from under them. It also strikes me that it puts NVIDIA in a very unique position in the tech world right now, because not only are they essentially the bottleneck for all AI development, right? They sort of are controlling, as you said, Casey, like kind of the pace of overall progress by virtue of having these chips that everyone wants and there not being enough to go around. But they're also kind of the kingmakers in the sense that they can choose, okay, you get 1 ,000 H100s and you don't get any.

31:51And we like your project and we don't like your project. So, Erin, what has NVIDIA said about how they are deciding how to allocate these very in-demand GPUs? Yeah, do you have to write like an application essay like for college where you talk about the traumas that you had and how it informed your journey? I'm sure that would help. So they do have programs for startups that maybe need a smaller amount of GPUs, because that's part of the problem is that small startups, some of them don't even know how much they need. And they certainly don't need like a year's worth of reservation that would cost them, you know,$50 million or something like that.

32:22They only need a small amount. And so they work with this company called CoreWeave that is, you know, specifically providing them for smaller companies. And they're obviously moving as quickly as they can. And they, just like everybody else in the world, were kind of taken off guard a little bit by this huge AI wave. I mean, it just so happens that the chips that they make that are good for video games also work really, really well for AI. And so they have a huge head start on AMD or the other chip makers that are trying to catch up to them now. So, I don't know. I think a lot of people in the industry do think that the backlog will catch up in the next six months, hopefully.

33:00maybe not as long out as a year. But, you know, I think the shortage will probably die down soon enough. I mean, maybe as some of these companies start going out of business, if some of them weren't great ideas or they pivot to whatever the next thing after AI is. Yeah, you wrote in your story about the sort of challenge of being a smaller company or a smaller research project that maybe doesn't have connections at NVIDIA, doesn't have, you know, an in with Jensen Huang or someone else high up there who can say, okay, you can buy these H100s. You wrote about this company called Docugami, which was trying to, what do they do?

33:38They fold up documents into origami. They use AI to read and process documents. So like long PDFs, their AI can help companies turn it into useful data. Got it. So they were looking for these GPUs to do their projects and they ended up actually getting them from the government. Yeah. which is not a place that I expected to have access to a lot of cutting-edge GPUs. So explain that. Explain how startups are kind of trying to use the government to get access to these things that maybe they wouldn't be able to buy on their own. Yeah, that was kind of an interesting thing. I also spoke with some university researchers who, you know, they have a little bit of a closer connection with the government in some instances, too.

34:20I mean, there are programs, like the National Science Foundation has these grants that they give out for companies that are doing research. And then there's this program called Access, which is what DocuGami used. And any startup or researcher who wants to apply for it can get access that way. So that's an interesting avenue that I wasn't aware of either. I mean, most people in the tech industry like to complain about the government moving slow or being slow to adopt technology. But here's one instance where it might be very helpful. Well, I'm trying to think of what research project Kevin and I could propose to get the government to give us an H100.

34:58And then resell it really quickly for a QuickBooks. Exactly. And then just see what kind of price we could get for it. Yeah. You also wrote in your story about the companies that aren't getting these GPUs, that are not sort of part of the elected elite that NVIDIA has chosen to receive these H100s and how they are kind of making do without them. In In particular, you wrote about a company called Hebbia, which used to get GPUs through a cloud provider, but ended up doing something different because they didn't have enough of the chips that they wanted. So talk to us about that company and sort of how other startups are coping when they can't get the GPUs that they want or need.

35:34Yeah, so what Hebbia was doing, so they have online, some startups don't have a product yet, and they don't have anything to show for it. Hebbia has a product and they have customers. And so they basically asked their customers, could you please go to our cloud provider and beg them and tell them how important our product is to you and how much, you know, you need us to get those GPUs so you can keep working. And so they're basically pulling any string they can. And then they also did the equivalent of like digging around in the couch cushions. Some of their engineers, you know, found some old gaming computers that they had that had some GPUs that are, you know, less efficient, not necessarily designed for the kind of AI work they want, but they're using them to run smaller experiments in a server, in a closet in their office in Soho, which, you know, of course, my first question is, isn't that very loud and very hot?

36:25And they're like, yes, yes, it is. and they're like, we put it in a closet, we close the door, we have liquid cooling units there to keep it cool, but it is very noisy. It also seems like, I mean, NVIDIA seems like they're trying to do some things to kind of expand access to make it so that it's not just the biggest, richest tech companies that can get these things. You know, maybe some startups, maybe some university researchers. It seems like the government is also trying to expand access, but it does seem like we are coming back to this fundamental issue with the state of AI right now, which is that the market really favors incumbents, right?

37:01It really favors companies that have billions of dollars, that can order, you know, tens of thousands of these chips, which are like$40 ,000 a piece, and that can do so without it really like affecting their overall financial picture. So talk about that. Talk about the sort of balance between how the sort of companies that make these chips like NVIDIA would like them to be allocated. And just the reality that these things are very expensive, and you kind of need to have billions of dollars to even get really started with building one of these models. I mean, that's partly why you see so many of these startups raising money from what you would think is maybe their direct competitors, but they're the incumbent players, you know, like Google, Oracle, even Salesforce, Amazon, they have been investing in these smaller startups and NVIDIA also.

37:50And so it really pays to have some kind of relationship, potentially financial, with these large cloud providers who are kind of the ones who are doling out the access to this. I mean, OpenAI and its investment from Microsoft, its partnership with Microsoft Azure is like a great example of that. And in my story, I write about Inflection AI, which is one of the bigger AI startups that raised$1.3 billion from Microsoft and NVIDIA. and they are spending 95 % of that or more all on compute power, which, you know. Wait, we have to just spend a moment there because what you just said is actually insane.

38:29You know, usually if you're a startup and you raise a big round of funding, you're using that to go out and hire people, maybe to get a bigger office. You're using that to go into research and development. You're using that for operating expenses. This company, Inflection, raised a huge round, One of the biggest rounds I've ever heard of for a startup. And they're saying that 95 % of that money is going to go right back out the door to buy GPUs. That is extraordinary to me. It is. I think the quote from the CEO, Mustafa Suleiman, was, it's eye-watering. I mean, it's madness. This company only has like 30 employees or something.

39:06But if you are trying to build like a general LLM, the kind of thing that's trying to compete with open AI, you know, you need that amount of compute power. And, you know, he's like, we're already at capacity. They're going to be running 22 ,000 of NVIDIA's H100 chips. And, you know, he said that other startups have asked him, like, oh, do you have any extra capacity? Like, can you share a little? And he's, yeah, exactly. And he's like, no, we're already at full capacity. Sorry. And now, does NVIDIA really care who gets the chips? Like, isn't their job just to sell the most chips for the highest amount of money?

39:44Why is it to their advantage to try to spread these things around? In general, I think the tech industry understands that there is supposed to be a cycle of innovation coming from the bottom up, starting with the small startups. And if we do not have that in this wave, that's kind of a problem. We're going to be missing out on a lot of fresh thinking and innovative ideas that come from small, scrappy startups. And those small, scrappy startups often grow into big, powerful customers that spend lots of money on things like chips and cloud computing. And so it's smart to try to get those customers early.

40:19So the cloud providers and NVIDIA know that they do want to have those relationships with small startups, especially the most promising ones, and help them flourish because eventually they'll grow into big companies that buy lots of their products. How long do you think it will take for the supply of these chips to meet the demand for it? So there are kind of like amateur investigators out there who are writing these elaborate blog posts trying to game out, you know, when they think the backlog will lift. And some of them are saying, you know, maybe by the end of this year, maybe in six months, maybe in a year, which is a scary prospect for a lot of people in the AI world.

40:57In the meantime, you know, they're just trying to make do with whatever they can. Startups that I've talked to who have called their cloud provider and asked for more compute have said that they're put on wait lists that are, you know, four months was one. One said that it was a year-long wait list. You know, they're all trying to skip that line however they can. Aaron, if hypothetically I were interested in building a large AI language model to, you know, say, replace a podcast co-host with an AI, where would I go? Where can I get some of these sweet, sweet GPUs? I mean, I would start with the university researchers.

41:36They seem to have the most connections. They might be able to hook you up with a broker, maybe someone who knows about a random server that Google has left over, you know, in some European country, perhaps, or maybe, you know, some like French OEM who's selling, you know, computer parts, but their deal fell through. That was one example that someone told me about. What I hear you saying is that it might be easier for me to just keep my faulty human co-host. Now you're talking.

42:08NVIDIA, please help me. I'm on a desperate mission here. All right, Erin Griffith, thank you so much for showing us inside the wild and wacky world of GPU procurement. You're very welcome. Thank you, Erin. Thanks for having me.

42:26well Casey we've heard one terrible AI song on the podcast this week and when we come back I gather you're gonna play some more for me you ain't heard nothing yet because we're going deep into the world of AI music and YouTube's big plan to get it under control

43:02Well, Kevin, as great as he is, I couldn't let Weird AI Yanko Chip be the only musical artist we featured on today's show. Because believe it or not, there is a lot more going on in the world of AI music, and I think it is long since past time that we caught up about it. Let's do it. Okay, so you may remember in April, we discussed a somewhat controversial song called Heart on My Sleeve. Remember Heart on My Sleeve? Yes, this was by Fake Drake. That's right. Somebody used a synthetic version of Drake's voice created with AI and put him together with a synthetic weekend, the musical artist The Weeknd, and they made a song.

43:40And it was quite successful. And eventually, Drake's label stepped in to get it removed. And our question at the time was sort of, what happens now with AI music, right? Are we going to see a lot more of this thing? Right, because these AI synthetic voice tools, they're getting a lot better. You know, when we first started this show, I made like a version of your voice that was synthetic and it kind of sucked. It was not that good. It was terrible. And it was less than a year ago. This really was not very good unless you had access to a vast amount of speech. Right, but now these AI synthesizers, they're getting quite good and people are able to create very realistic versions of not just like people's voices, but they're singing voices as well.

44:20That's right. But I think the important thing that happened in April was that when that song went viral, I think we started to understand, oh, there is going to be a market for this sort of thing. And the fact that Drake didn't really sing it, that these aren't his lyrics, isn't going to stop a lot of people from not just seeking this stuff out, but also wanting to make more songs like it. And what we have seen, I think it's fair to say, has been a summer of debate between YouTube and the labels and artists and individual creators and DJs about how all of this should work. Should you be allowed to make a song with a synthetic version of Drake's voice?

44:58Can you post it on YouTube? And of course, what is most important to everyone involved is who gets to keep the money if the thing is successful. I confess I have not really been paying that close attention to this area since we did our story on Fake Drake. You're not scouring the billboard charts every week to see what's hot? I'm just listening to Rich Men North of Richmond on repeat, my friend. It's time for change in this country. All right, we'll get to that in a future episode. So, yeah, what's been going on? Here's the thing. I was browsing YouTube the other day, as I sometimes do, and I love to listen to music on YouTube.

45:32But almost always, it is music that has been written and recorded by a mainstream artist, right? What I saw for the first time the other day was YouTube's famous algorithm recommended to me an AI song. And I listened to it. What was it? Well, actually, I just want to hear it right now. Yeah. Okay, now keep in mind, I never search for any of these terms. This would just present it to me as something I might like. So why don't we go ahead and hear this first clip that the YouTube algorithm recommended to me.

46:02Do you recognize the opening words? Oh, yeah, this is Frozen. That's right, Let It Go. But maybe not with the original Elsa. The snow glows white on the mountain tonight Not a footprint to be seen Yeah, you're right. This is not Elsa. It's not Elsa. Yeah, this is the touring cast. No, who is this? Let's see if maybe when we get a little bit closer to the chorus, we're able to tell. Like this swirling storm inside Couldn't keep it in Heaven knows I could try Don't let them in Don't let them see Be the good girl who always have to be Conceal, don't fear, don't let them know This slaps, right? Yeah.

46:55Let it go, let it go Okay, I confess I do not know who this is. Okay, this is Freddie Mercury from Queen. No. Let it go, turn away and slap my guard Can you hear it now? Yes. Yeah, that's really good. Yeah. Freddie Mercury, famously not alive. No, he tragically died. Yes. But Queen is still one of the biggest bands in the world if you go look at their Spotify monthly listeners. But look, I saw this, and of course my first instinct was, well, that's probably going to be bad. But then my second instinct was, I should listen to it because maybe we'll want to talk about it on the show. And so I did, and by the end of the song, I thought, it only sounds to me like Freddie in some places, and yet it was still passable, and at the end of the day, it's a catchy song, and I enjoyed listening to it.

47:42Yeah, that's pretty good. Yeah. So after that, I kind of started to go down the rabbit hole, and I started to hear even more of these AI covers that have been popping up on YouTube. So maybe let's listen to another one. This is from a great channel. I first encountered it on TikTok. It's on YouTube as well. It's called There I Ruined It, and they do a lot of really fun and weird editing with songs, and increasingly, they've been using AI. And this one is from the legend of Folsom Prison himself, Mr. Johnny Cash. Hello, I'm not Johnny Cash.

48:20I'm a Barbie girl in a Barbie world. Life is blaster. It's fantastic. You can brush my hair. Undress me everywhere. Now, this sounds amazing to me. That's really good. Come on, Barbie, let's go party, because I'm in a Barbie world. Let's maybe just hear one more, and this one is also very beautiful, and it may unlock some emotions in you. You know, maybe if you feel the tears come, just let them come, because I think it has an important message to share. It's from a channel called AI Covers, which seems to be promoting one of these companies that is out there creating AI voice models. But let's give that one a listen, too.

49:09Here we go. This is sounding Rascal Flatts-ish. Yeah, very Rascal Flatts coded. Yeah. Yeah.

49:29Who is this? I would say this is an animated character.

49:38Is this, I genuinely do, is this, is this Pinky? No, it's not Pinky.

49:53It's Eric Cartman from South Park. Oh, it's Cartman. It's Eric Cartman saying life is a highway. Rascal Flatts version. That's really good. Yeah, and by really good, of course you mean it's not, It's something you want to turn off Yeah, I would like that out of my ears immediately But I do appreciate the creativity Sure, so look, on one hand Everything that we've just heard, I think Is an evolution of the mashup I used to love mashups In the 2000s in particular DJs would take two different tracks And they'd sort of get the stems and they'd get a vocal from another song And they'd slam them together It's very like girl talk, you remember girl talk?

50:30Yes, I love girls, I saw girl talk live On multiple occasions So look, at the same time there is the potential for so much more here than just what artists were doing with mashups in the early 2000s, because now an artist's voice can essentially become a new creative tool. You can inhabit the voice of an Eric Cartman or a Johnny Cash or a Freddie Mercury and just have them go nuts singing whatever song you like. Right, so it seems like this kind of thing is going to become ubiquitous, and some of these AI-generated songs will be kind of novelties, like, ha-ha, isn't it funny to make Eric Cartman sing Life is a Highway.

51:08But I imagine that as with Drake and his fake song and The Weeknd and the tussle between their music label and the platforms, I imagine that this is going to be very controversial because Freddie Mercury is not alive, but Queen's record label still presumably has a vested interest in making money when their voices are used. That's right. And so that brings us pretty naturally to what actually happened this week that I think has put this story back in the news. And that's that YouTube and Universal Music Group, which is one of the major record labels, they announced what they call the Music AI Incubator, which they have outlined for us in a couple of blog posts.

51:48And basically what YouTube is going to be doing is working with the UMG artists and songwriters and producers to use AI music tools and just kind of get a conversation going. They've got a bunch of artists who are participating so far, including the estate of Frank Sinatra. But the most important thing is that YouTube has also hinted at creating a system similar to Content ID. Now, you've written about YouTube for a long time, so you know what Content ID is. Yeah, Content ID is basically if you, like, use a song in one of your videos as a background track or you're singing along to it. Basically, YouTube has a way of kind of stamping that such that it gets, like, credited and royalties get paid to the artists who made that song.

52:32Now, what's interesting, though, is that if YouTube develops a version of this for AI-generated tracks, essentially something that would catch a fake Drake song the second it was uploaded so that royalties could be immediately distributed to Drake or his record label, then we start to get into a really interesting and tricky legal world. Yeah, so I have some questions about this because I have not been following the story super closely. So say that I am an internet creator and I want to make a video in which Frank Sinatra sings Vampire by Olivia Rodrigo, Song of the Summer. Sure, Song of the Summer.

53:11So can I do that? Or will YouTube take my video down? And what does this new program change about that? Great question. So currently, there is no law that says you can't train an AI on someone's voices to create a synthetic version. It's being litigated, but it's not clear. And I think what is scary, if you're somebody who just wants people to be able to be creative and not to have major labels telling them that they can't be creative, is that you can imagine a world where in the not-too-distant future, you're a young creator and you create a song, and maybe you're not even trying to sound like Drake, but you upload it to YouTube and YouTube's new AI detector says, this sounds enough like Drake that we're going to start giving royalties to the Universal Music Group.

53:59And who has the power in that situation? It is not going to be the teenager that uploads the fake Drake song. That person does not have legal representation. They are not going to fight this. They're just going to get their thing booted from YouTube. And then, you know, if it's a good song, they're going to have to go figure out a way to post it somewhere else. Now, obviously, all of these moves that YouTube and UMG are making are sort of defensive, right? They're saying, like, we don't want this technology to sort of become a free-for-all Wild West scenario where anyone can make any song with any person's voice.

54:31A Burning Man situation. Yeah. Right. We want to be paid for this. We want to sort of keep control of this. So let me say, a lot of my thinking on this comes from my friend, Nilay Patel, editor of The Verge, who wrote a really great piece about all of these issues for The Verge this week. And what Nilay points out is that when it comes to AI, YouTube and Google, which owns YouTube, are taking very different approaches. So let me explain this. So when it comes to music, YouTube is saying, well, the artist's voice has value. We want to protect that value. We might develop a whole system to determine when someone is using that voice so that the value can be apportioned and we can keep our record label partners happy.

55:16Eric Hartman must be paid fair wages. At all costs. At all costs. You do not want to upset Eric Hartman. Okay? And then there is the World Wide Web, which is populated by publishers who are publishing text of all subjects under the sun. And what is Google doing with that content? They are slurping it up into a large language model. They are not paying anyone for it. But they are going to use it to create the next generation of Google search, which they will then monetize in any way that they can think of. And the publishers might not see any compensation for that at all. Right. Okay. So at YouTube, they are saying that artists whose voices are synthesized should be entitled to some compensation for the act of their voice being sort of fed into these synthesizers and used to make new music.

56:08It sure seems like that is the direction that we are heading with YouTube and music. And then on the other side of the house at Google, you have the text-based AI models that are hoovering up information from the internet, using them to train AI models and using that to create synthetic text. But without that sort of financial and legal link between the outputs and the original things that were used to train it, right? So most publishers, whether they're newspapers or magazines or blogs or book publishers, they are not under the current model being compensated when, say, someone asks Bard to generate a news story that draws heavily on those sources.

56:53Am I getting that right? Yes, completely. So what do you think explains the very different treatment that Google slash YouTube is giving AI between musicians and publishers? This is a really simple business story about who has leverage, right? If you're YouTube and you lose access to even one of the major labels, you alienate a bunch of customers. And if they can find - Why? Tell me why that is. So, because let's say you're a kid, you can't afford a Spotify subscription. When a new Olivia Rodrigo song comes out, you go to YouTube to listen to it. All of a sudden, you can't find it there. But guess what?

57:30TikTok still has a deal with major labels because maybe they've taken a stronger stance on this than YouTube. All of a sudden, kids go spend their time on TikTok and YouTube has a deal. Well, and you're jogging my memory here about something that happened years ago at YouTube, which is that the music labels fought back against not just people listening to music videos on YouTube, but actually if you were doing a cooking video and you had a Bruce Springsteen song playing in the background of the video, even if the video had nothing to do with Bruce Springsteen, Bruce Springsteen's label would get very upset about that and wanted to be paid for that.

58:02And so YouTube actually struck deals with the music labels to allow that. And by the way, that's what Content ID does. It is scanning for Bruce Springsteen in the background of your other videos to ensure that the rights holder can get paid if they flag it. Right. So that is part of the sort of truce that YouTube formed with the music labels is like, we will pay you royalties for the usage of your songs on YouTube. And in return, you can't go like getting people's cooking videos pulled off YouTube because they used a Bruce Springsteen song in the background. Right, right. You know, YouTube, it's in their interest to make these things very easy on the people who are uploading videos, because if it's super hard and your stuff is getting yanked down every time you try to do anything, eventually you stop doing it and YouTube withers and dies.

58:45So if YouTube antagonizes the music labels, they risk not only having like these official music videos getting pulled off YouTube, but also like a much more annoying world for anyone who is trying to make a YouTube video that has copyrighted music in it. Absolutely. And on the upside, there are only a tiny handful of these major labels. So you only need to strike three or four, maybe five deals, and you'll be able to get access to the vast majority of all the music that people are going to use. I see. So then you go over to the tech side of the equation, and we live in a world where Google is maybe one of the main forces keeping what remains of the media industry afloat.

59:24And the reason is that when you search for things on Google, Google is directing a lot of traffic to news sites, which news sites are then able to convert into advertising revenue and into subscription dollars. And because so much of their traffic is derived from search, publishers will do essentially whatever Google says. So this is a case where Google has all of the leverage and publishers have very little. So this seems like what you're saying is we need music labels for writers. Right, but it's hard because if you own the rights to Olivia Rodrigo's Vampire, for example, Song of the Summer, that is a hugely valuable thing that, you know, thousands of creators might want to use in their videos.

1:00:11One single story at the New York Times, at the platformer, as great as we would probably argue that they are, it's not going to have the value of even one song in a major label's catalog. Well, speak for yourself. My stories are read out loud in the background of many thousands of mukbang videos, and they are a huge part of the fabric of the internet. So, but this just seems to me to be untenable, right? It's like, does a creator have rights or do they not? And why is it so different between music and tech? I understand the business reason why it's different, but as a principle, it seems to me that those things are very similar.

1:00:52What do you think should happen here? What is the fair resolution of all this AI-generated music stuff? Fortunately, Kevin, I know this whole time you've been thinking, Casey, this seems very complicated. I hope you have a simple solution to it. And the good news for you is that I do. What's your solution? Going forward, we have to sing all of our articles. If we can copyright all of it, if we can get a major label to license us and publish everything we write, we're golden, baby. Casey, I would like you to pull up yesterday's version of your newsletter and do a rendition and I'll find a karaoke track for it.

1:01:27Well, thank God I wrote it in rhyme. Okay, I'm going to use, what style do you want to use for this? Do you want like trap or indie folk? Oh, indie folk. Maybe something like the Lumineers, maybe. Okay. So I'd like you to sing over the karaoke version of the Lumineers, Ho-Hey. All right. Are you going to play that for me? Yes.

1:01:49Well, today let's talk about how an idea that we once took for granted on social networks is now under threat. Facebook, YouTube, and TikTok all grew on the strength of recommendation algorithms that learned to predict your interests better than even you could yourself. And then it would just go on from there. Yeah, they're calling this the real song of the summer, actually.

1:02:18Casey, I think you actually just got signed to Universal Music Group because of that. There's an A &R scout on the line, and they're interested in a multi-record deal. This guy's got what the younger generation needs. He's going to be packing shopping malls all over the country. I'm sorry. engineered by Alyssa Moxley. Original music by Dan Powell, Marion Lozano, and Casey Neude. Shoo-be-doo-bop. Special thanks to Paula Schumann, Pui-Wing Tam, Nel Gologli, Kate Lopresti, and Jeffrey Miranda. As always, you can email us at hardfork at nytimes.com. Let's keep those AI covers coming.

1:03:36Thank you.

1:04:03Thank you.