The Future of Synthetic Data

Hello and welcome to the Virgast, the flagship podcast of music that sounds early but not exactly like other music. I'm your friend David Pierce and today on the show we're talking about training data. It's the raw materials of everything that is AI . And we probably don't understand it or talk about it enough. I'm talking to Alex Reisner, who's a staff writer at the Atlantic. And over the last couple of years, he has spent a lot of time investigating training dat a and how all of these books and all of these articles and all of these YouTube videos and all of these ons are compiled into these gigantic data sets that AI companies then use to form the basis of their models. I think the way that these models are created and the sources of these data has a lot to do with the way that we feel about AI, particularly generative AI as a creative expression, and understanding how this data works, where it comes from and how it gets used, I think is really important. So we're going to have Alex on, we're going to dig into it. I'm very excited about it. But first, here's everything else happening on the verge today. This is ninety seconds on the Verge for Thursday, june twenty fifth, twenty twenty six. Apple just raised the prices of a huge number of its most important products. IPads and MacBooks in particular are now up anywhere from one hundred to five hundred dollars and even like the Apple TV is way up. It's now two hundred dollars , roughly the price of nine hundred Amazon Fire Six. We knew this was coming, but this is still a big moment. Apple is as well managed a supply chain company as anybody and exists at really high margins. If it can't keep prices down in this era of AI driven shortages of memory and storage , nobody can. And I suspect this is going to get even worse. My advice from yesterday still very much holds, which is Prime Day is not over. Go get your deals while you still can . And here's a way to help you pay for all of those more expensive gadgets. Go get some money from Disney. If you have YouTube TV or Direct TV stream, you might be eligible for a cash payout from a new fifty million dollars settlement. The case started in twenty twenty two, and the argument was essentially that because Disney owns ESPN and Hulu, which are both powerful streaming services in their own right, Disney was able to drive up the prices of all of its rivals to unnecessary heights. Live TV is lucrative and competitive and is going to keep being messy in this way. This will not end things. But hey, go get some of that fifty million dollars. And finally, a quick PSA to advertisers everywhere. Yes, I get it. The whole idea of using AI creative and AI targeting to put your AI ads in front of AI people is exciting, I guess. But maybe check your ads to make sure that you're not, I don't know, advertising a bike with two sets of handlebars like REI did this week, that may not go over so well with people. Just a little free advice for me to you. You can read more about all of this at thege. com that's ninety seconds on the Verge for Thursday, june twenty fifth. All right, let's talk training data. Alex Risner from the Atlantic is here. Hi, Alex. Hey, David, thank you for having me. Very excited to talk to you. I think we have talked a lot on this show about why people feel the way that they feel about AI and the sort of instinctive reactions to the whole idea of AI . And a theory that I have is that a lot of it is about training data. And I just want to talk about that. You've done a lot of work on this stuff and you've been investigating how AI models get trained for a long time now. But I want to just lay a little bit of groundwork here . And just to ask the very obvious question up front , why does it matter what data is used to train these products? What difference does it make what's inside of these models? I think it is potentially the most aspect of a model is what it's trained on. I mean, if you take a model and train you it on , let's say it generates music and you train it on nineteen fifties jazz. That model will be very good at generating music that sounds a lot like nineteen fifties jazz. If you train it on a recent hip hop, it's going to generate music that sounds like recent hip hop . You know, these models have names like Chat GPT and Claude, but I think you could make an argument that the right name for a model is actually a description of the data it was trained on because that is a description of its capabilities. That's what it can output. And so I think that the training data is really it's really fundamental to the model maybe more than the architecture to some degree. Interesting. And I think I mean that kind of answers to some extent my next question, which is why is training data such a closely guarded secret from these companies . And it seems to me that there is both a straightforward business answer and maybe a slightly more nefarious answer as to why these companies would be so secretive. But what is your read on why this is such a closely guarded secret, what data is being used to train these models? Yeah, I mean the companies have argued that they need to keep this secret because the data that they have selected to train on is their competitive advantage, right? Like Anthropic has done a better job at selecting data than Google and OpenAI. And if they were to let that come out in a court case or be public in some way, they would lose their competitive advantage . You know, there's another pretty obvious reason, which is that they have gone about acquiring a lot of this data in ways that the people who've created the data, the authors of the books, and the creators of the videos and the music would not be happy about it. And in a lot of cases, they just don't know that their work is being used. And when they find out they're not happy about it. And I think it's a conversation that the AI companies have tried to just avoid having. So a big part of what you've been up to recently has been sort of reverse engineering this process to like peel the model apart to figure out what data is inside of it. And it seems to me that you've had to do sort of the reverse of what all of these companies have done, which is go figure out where these mass sources of books or web articles or videos or songs are . So tell me a little bit about your process. How do you go about finding these things that are kind of otherwise closely held secrets? Yeah, my process is actually maybe not that different from the company's process . I am a programmer by I worked in tech for twenty years. I built websites and apps and did some statistics work and so I've been aware of these models for a long time . And at a certain point I started hanging out in the forums where AI developers were hanging out and talking about the work and they're talking about what data they're using to train these things . And so that was a useful source. And I've also been reading a lot of their research papers . You know, selecting the data is really challenging. There's a lot of effort that they put into it. They have to come up with some not ion of what is high quality data, like what do and don't we want in the model? And they write papers about that. I think they want to be involved in this conversation that's happening about training data within the industry. One thing that's been really helpful to me is that there's an open source AI development community and they believe that the work should be done more out in the open. And I think a lot of them are doing really good and important work and they have really interesting things to say about AI just philosophically and socially . And so they're pretty transparent. Places like Hallen, AI and Eluter about what they're using to train. And so that's been helpful. But then even at the companies, you know, the AI world's a little bit like academia in that it is good for your career to publish papers . And the companies don't like that, but they also acknowledge that they have to let the employees publish something. And so the lawyers will go over it and tell them what they can and can't say . And over time, they've clamped down more. And so the companies are revealing less through the research papers , but yeah, a lot of my research is just from reading a Google paper. You know, for example, for this last article where they said we trained on tens of millions of songs. Yeah, I remember not that long ago Apple going through a big cultural issue with this with its AI team because Apple is so secretive and so reluctant to let any of its work be public that all of its researchers were like, well, if you're not going to let us publish, we don't want to work here. And it did that push and pull seems like it has it has kind of morphed in a bunch of different directions over time, but interesting to hear that it is definitely it is everybody's retrenching a little bit as this space gets hotter and hotter. Yeah, it's not one anymore. The research papers read very differently now than they did a few years ago. That's really interesting. But tell me a little bit about these databases. One thing I think I had not given enough thought to until I started really reading your work is that there's a real business in making and maintaining these databases . Like one company you wrote about Common Crawl, I think is like a thoroughly fascinating player in this and they essentially as I can tell, just crawl the internet and make it available to whoever wants it, which is weird and complicated and we should probably come back and talk about crime and crawl sometimes. But it seems like if you look around a little, these gigantic databases of books and articles and videos exist. Like are people making these and selling them? Why do these databases exist? Mainly , well, I mean, for AI training, why else would they exist? It's a huge, you know, it's an extremely labor intensive process to find , you know, and once In one sense, they're just you just go and download all of library Genesis or all of Ana's archive . That's one sort of naive thing you can do, but the companies realized pretty early on that they need to filter the stuff pretty carefully. So the organization you just mentioned come and crawl. Yeah, they've been crawling the web since the late two thousands maybe two thousand nine or something like that And they just make the whole thing available . Every month, there's a new they've scraped a few more hundred million web pages. And it's available to anyone who wants to do any kind of research with it . In fact, it's mostly AI researchers who were using it . But all the early large language models were trained on common crawl. If you go back and read those early open AI papers , everyone was training on Comicra Cwl and at first and the models were terrible because if you train a model on the whole internet, it's just, you know, there's all the it says all the junk that people say on the internet along with along with the intelligent things. Yeah , mostly junk. Yeah. And so it's statistically mostly junk. Statistically, yeah, mostly junk. And I think the early large writers models were proof of that . But yeah, I think they're, you know, common crawl is a non profit. So they would argue it's not a big business. They do get a lot of money from AI companies and AI investors . But yeah, I think this the topic of training data selection, the challenge of selecting the right data for a model is still really hard. The AI companies, I would say, are still have a very primitive understanding of what data will make their model better . It's an area of research that I think even they at this stage are not very good at . They do it mainly by trial and error as far as as far as I can tell. So interesting. Again, so the reason you're asking why do these data sets exist? I think it's people trying to share what they've learned from, you know, curating data sets in different ways and training models with them. Yeah, I mean, part of the reason I ask is I think one the things I have come to believe about the AI industry is that this shift that went from AI research being fundamentally a research thing, like if you go, if you go way back , OpenAI was basically a research organization, right? And you talk about common crawl and I think its early users were largely researchers and it was these things were academic things for academic purposes . And from what I understand, the kind of rules of the road are different , right? That like if you want to make copies of a bunch of things for academic purposes these things are generally considered less problematic, right? But if you then do it and become a trillion dollar company on the back of it, people are going to rightly feel differently about the way you went about getting that information. And just the speed with which AI commercialized , all of these companies just moved so fast from, we are essentially an academic thing to oh my god, we're making so much money. Everyone's filthy r ich that it feels like they just hoped everybody would ignore the ways in which they got this information. And so that's why I'm particularly fascinated by where these data sets even come from because, it does seem like in many cases like you're saying , it's not that there is some gigantic business in being the one to sell the songs to somebody. It's that stuff is being this stuff is being compiled for other purposes , it's just that now the main purpose because of the sheer volume of work is AI research. All this stuff has been sort of co opted from every other purpose to AI . And it just all feels so concentrated now Does that feel right to you? Does that make sense? Yeah, that I think that I agree with most of that . I do think I'm not sure how much these data set weres really collected for other purp oses. Compancrawl likes to talk about I think they're a one case . They probably have the strongest argument that their data could be used for other purposes. But when you go back, you know, they've been cited by over ten thousand papers. I didn't read all ten thousand, but I read a lot of them and they're mostly AI, right? And it's early. A lot of it is stuff that people wouldn't wouldn't mind as much as with generative AI, right? Like Common Crawl , I think without common crawl , you know, AI translation tools might not be as good as they are. I think it was really a huge hell because they scraped web pages the same page in multiple languages and people were able to train translation models based on that . So that was helpful. But the thing that you know there is still what I would call a data laundering network where the AI companies are still relying on they'll do a collaboration with a university and they'll have university the university download millions of images to train a model or download millions of articles to train a model . And the company can say like, well, we didn't do it. This was like an academic . You know, the same goes, Common Crawl is not the only nonprofit that's like doing a lot of this scraping for the AI industry. One of the datasets I reported on in the music, the article about music training data is this organization based in Europe called Lyion. They have a data set of twelve million songs from YouTube . So anyway, this is like is it academic? Like not really. Like this is technically yeah,, there's a univers ities and nonprofits, but they're all receiving money from the AI industry. This is kind of a diversion, but this I'm so struck in reading all of your work by how often YouTube appears as just it is it is everybody's favorite source for everything. Like it's it's what, you know, it's what OpenAI used allegedly to create whisper. It's what a lot of the music stuff is using. It's what there's a lot of stuff based on videos. Like what is your sense of YouTube's role as an AI training force ? Because it seems to be everywhere. Yeah, that's accurate. YouTube is an extremely common source . I think one reason is there are tools for downloading from YouTube that are really that work really well. They're really easy to use . And it's pretty common for AI developers to just use those tools. And it's just kind of it's become a custom . But also, you know, and that's that includes stuff on YouTube is just less protected, I think is one way of saying it. Like there's music, you know, if you're a musician, your song might be on Spotify, but Spotify's website has digital rights management protections. It's really hard to download from Spotify . It's much easier to get the same song from YouTube and so many songs are also on YouTube. So I think it's just ease of downloading. YouTube obviously would say out loud that this is not allowed, right? That it violates s a of ser serumv ice. And yet it seems to have either either it can't or it just hasn't done anything to stop this, really . Yeah , that's that's a question that's been in the back of my head for a long time and which I've asked YouTube in which they don't really answer. They did, they have said that they considered a violation of their terms of service to be downloading their videos. But yeah, they have them, you know, years have passed and it's just as easy to download from YouTube now as it was a few years ago. Yeah, I downloaded a YouTube video this morning. It is just a thing you can do . It's really true . You mentioned the sort of data laundering stuff, but it also seems to me that more and more the people doing the AI training are just completely unapologetic about it . Like you quoted Rich Skrinter, the CEO of Common C rawl , who literally said to you, like, if you don't let AI robots crawl your data, you essentially don't exist. Like, you're gonna miss out on the future of the internet. And I think about, you know, Mark Andreesin even a couple of years ago being like, none of this would work if we couldn't just take the training data that we needed . There is this almost like manifest destiny sense of the AI industry that we can have this data because the thing that we're doing is so important that we must have it no matter what. What is your sense of the trend there? Because it occurs to me that that's happening even as the backlash from people hate the experience and feeling of AI in part because of the way that this stuff is trained just keeps getting worse. Like these things are just running away from each other at record speeds . Yeah , that's a huge question . I think it has something to do with the fact that we just have not done a great job in this country establishing the value of data and who should be able to have data, right? This is something that privacy advocates have talked about. Jerem Lanier, I think, was one of the earliest people to be talking about this . I think he wrote in two thousand seven about that you should get paid, like you were being surveilled basically, and companies are taking they're monitoring everything you do. They're generating data from your online activity , that data is extremely valuable to them. It might seem like nothing to you, but you should be getting paid for that. People thought that was crazy back then, and I think a lot of people still think that's crazy now, but taking people 's music to build models that generate songs that compete with them is just the next version of that. And it's just gonna keep going. If we don't acknowledge that this data is incredibly valuable and figure out a way to write laws around that or just have, you know, better business practices or . This is just it's just going to get worse. There's just going to be more exploitation and more mining . The next step I think a lot of people have perceived for a while to be synthetic data, right? That eventually we're going to get AI models that are so good at making new things that then we can use those things to train new models. And eventually they don't need existing recorded music. That synthetic data is the future, and that's how we get to everything. And especially now, I mean you look at Spotify and there are tons of AI generated songs on Spotify that some people are listening to. There are a billion AI generated podcasts out there. Like the content is being made . Is that the next phase of training data? Is synthetic data coming into its own in such a way that we're going to start to see the next generations of these models built on the things made by the last generations of these models? Absolutely not. I don't think that there's any evidence that that actually works. I think when AI companies talk about training on synthetic data, they are they choose their words very carefully and they always ex aggerate the extent to which it's happening. There's a lot of paper, there's a lot of research out there on the phenomenon called model collapse, which is what happens when you train a model on its own its own outputs. It's, it very quickly it doesn't get better. It very quickly degrades . And it's not hard to see why that could be the case, right? Like AI is kind of an averaging machine, right? It's defin stiteatistical average between different types of content and putting that into some new kind of more average type of content. And there's just not enough weirdness or interestingness or something like that. There's some quality in the work that humans do, that's not in the work that AI does. And I think that's actually proved by the model collapse phenomenon. So I'm amazed that AI companies are going around talking about synthetic data still when there's so much evidence that it doesn't it doesn't work. So is there a next untapped place filled with data? I mean, these models keep getting bigger, they keep needing more data. They got to go somewhere, right? Is there a next sort of unturned place to go for these companies? I think they just pay people to make it. I think there's already a gigantic industry of writers who are writing for AI , musicians who were making music for AI . You know , after I published this story, I got an email from a company that is doing this. They claim to have paid creators over ten million dollars just to make things for AI training . It's very strange, but this is the next AI as your audience is the next frontier for creators . That is a deeply strange thing to think about, but yeah, also like, you know, we're going to put this on YouTube and it's going in there anyway, so who knows? Maybe this is all of our destinies no matter what. I certainly hope not and I don't think so I think we can have a conversation and arrive at a more reasonable future for ourselves and the culture I'm with you on that. All right, well, Alex, thank you so much for being here. Really appreciate it. Thank you, David. That's great. All right, that's it for the show. Thank you to Alex for being here. And thank you as always for watching and listening. If you have thoughts, questions, feedback of any kind, if you have a favorite AI song you want to send me, just not the Puerto Rico song. I already know that one. Send me all your favorite AI songs. No judgment, I just want to hear all of them. You can always hit up the hotline eight six Virgin one. You can send us an email, Virgaster atge the V ot. com. d We love hearing from you. And as a reminder, the best thing you can do to support all of this is to subscribe to the Vergevs. com slash subscribe. It gets you all of our podcasts ad free, including this one. It gets you all of our exclusive newsletters, it gets you all of our coverage . Terrance O'Brien on our team has been covering SUNO a lot and doing a really terrific job. Lots to come on that. Subscribe to the Verge. I think it's a pretty good website. Vergecast is a Verge production and part of Vox Media Podcast network. The show is produced by Josh Kahas, Eric Gomez, Brandon Keefer, Travis Larshuk and Aron Lacio. We'll see you tomorrow, rock and roll.