Poison Pixels

Speaker 1: Are you a plants guy, Scott?

Speaker 2: Plants as in, like, green things that grow in my house?

Speaker 1: Yeah. You got it.

Speaker 2: I am my wife is a plant person. I used

Speaker 3: to be. I used to keep tons of plants in

Speaker 2: my house, and I actually grew up in an acreage. And we had, like, a two acre garden.

Speaker 4: Sure.

Speaker 3: So I

Speaker 2: I have I have lived I have lived among the plants, but I would say currently, at this present moment, I am not a plant person.

Speaker 1: Okay. It's good to know.

Speaker 2: Are you a plant guy?

Speaker 1: I like a good fern.

Speaker 3: Okay.

Speaker 1: I got a little cactus on my desk. I like a plant.

Speaker 3: Nice. Nice. Nice.

Speaker 1: So there are a lot of plants that will kill you. Mhmm. But few are as famous for doing so as nightshade.

Speaker 2: Are we talking in, like, Dungeons and Dragons here? Because it sounds like it.

Speaker 1: I mean, it's probably in Dungeons and Dragons. That's the thing about nightshade, AKA Atropa belladonna, AKA the deadly nightshade. It's very famous. I don't

Speaker 5: know if you

Speaker 1: know what it looks like. It's this little green leafed plant with, like, pretty purple flowers, and these, frankly, like, quite delicious looking little shiny black berries. But don't eat them because it is famously very poisonous. Mhmm. There are a lot of other poisonous plants. You got your hemlocks. You got your foxgloves. But nightshade has the reputation. It's in the Odyssey. It's in Shakespeare. It's in Roman myths. It's part of Salem era stories about witches' potions. Call it whatever you want, Belladonna Nightshade. It's like a shorthand for poison and death across centuries of literature and stories right up to, as you said, Dungeons and Dragons. But deadly nightshade has this other association. The first part of the name, Atropa Belladonna, Atropa refers to one of the three, like, Greek fates, the people that would, like, cut the strings to end a person's life. The first part of the name refers to death, the poison part of the plant. But the second part of the name is Belladonna, which in Italian means beautiful woman. Mhmm. And that is because this plant, this shorthand for death, also had associations with aesthetic beauty. It was used as a cosmetic. This is kind of wild, but they would make eye drops out of it Interesting. Because those eye drops would make your pupils dilate, and that was apparently considered very hot at the time. Also, don't try this. There's a reason they stopped. But but you have this one plant with these two deep, long associations, poison and aesthetic beauty, death and art, which makes it a really cool name for the subject of this episode, which is a piece of software attempting to poison a system built on a foundation of and arguably producing visual art.

Speaker 3: I'm just interested in protecting human creativity in some sense. Right? This is, I feel like the companies where you and Garmin are taking a, a fairly for sighted view on AI. Right? Of course, these are awesome. They are able to copy style, generate images. But what I see is if we keep going down this route, say, you know, they get better and the artist getting replaced, And that more or less kinda marked the end of human creativity, at least in some aspects. And it's unclear where are we gonna go from there.

Speaker 1: That was Sean Shan, a researcher at the University of Chicago. And we had a conversation with him recently, didn't we?

Speaker 2: Lovely guy.

Speaker 1: Lovely guy. Lovely guy. Great chat. He is part of a team working on a suite of tools for artists to be able to turn their art into a kind of poison, a nightshade, which is why they named it that, for generative AI models. You apply these tools to a piece of digital art, and the art looks normal to a human. But if it's ever ingested by an AI model, it will hurt it like poison. Hence the name, nightshade. Not only will the AI model not recognize what's in the image and fail to, like, scrape the necessary information to imitate it, it's a poison in the sense that it will get it actively wrong. The artist creates an image of an elephant. The system doesn't just fail to see, oh, that's an elephant. It goes, oh, that's a little kitty cat. And so the next time it makes an image of a cat, there's a slightly better chance that it will have a trunk. That's the basic idea. Does that make sense?

Speaker 2: I love this. I love this idea. I I love that they did this.

Speaker 1: So interesting. Mhmm. This was a fun one. I really enjoyed this conversation. So here's our chat with Sean Shane about Nightshade, their earlier tool Glaze, which you'll hear us talking a bit about in the intro, poisoning AI systems, and David versus Goliath tech projects. Here on Hacked. Sean, thank you so much for sitting down with us to, to chat about this project.

Speaker 3: Of course.

Speaker 1: So I'm a layperson. Feel free to correct me as we go. I wanna get to Nightshade. I wanna get to Glaze and your team. But before we do, there was this sort of overarching question that snuck up on me as I was reading about your projects. When you train an AI model like Midjourney or DALL E on an image, you feed an image into it, what kind of information is it getting out of that image? How should we understand what it's learning about an image when you feed one into it?

Speaker 3: I see. Yeah. So at a very kinda high level, so this model takes in not only images, but they also take in corresponding pegs that kinda describe these images. Right? If you have, I don't know, like, a Van Gogh painting, you will say this is a painting from Van Gogh landscapes. So what it does do, the training process is basically associate alright. So Van Gogh is the name, and next time I get a question for generating a Go image, I should produce something similar to the image that I've been seeing. Right? So specifically how that work, they call diffusion model. I'm not getting into the details, but, it is very powerful. However, that each individual training point, they don't have a huge impact of the entire models. Right? So if you're training data only appear only once, it probably doesn't have too much impact. But there are many images that has to be shared many freak very frequently online. So also, again, memorize the model really to memorize pieces of that image and try to perhaps reproduce it where you prompt a image out. So that's kinda out of our how how this, models works.

Speaker 1: And then specifically, because I saw this term come up quite a bit in your research, what is style mimicry in generative AI? Is it the same thing as just feeding one image in? What is style mimicry for anyone that doesn't understand?

Speaker 3: Style mimicry is something more, targeted perhaps is, you know, the cases where I go online, I see this artist I really like. Right? I just wanna get her or his his painting, but I don't wanna pay him. Right? So what I can do is I can use this AM model to mimic their artwork, and I can do this with just downloading some images from their Instagram or their website. I just need pay images, And then I they call it fine tune this model. Basically, just means you have a base model. You just train a little bit more on that additional 10 images. And now now the new model will be able to basically output arbitrary content for from the same artist. Right? Very much the same style as how the artist would paint them. But maybe, like, the quality is not as good, but oftentimes, we see this as good enough to to replace a lot of the artists for many types of commissions. Sure. So that's kind of the style we were curtailed. Yeah.

Speaker 1: Interesting. So, hypothetically, an artist who might have had to get a commission from someone wouldn't need to get that commission if that person was able to use one of these systems to create a a piece of art that was to a a layperson pretty close to indistinguishable.

Speaker 3: Exactly. And and it perhaps not no longer hypothetical. There are cases, I think, artists got replaced by these models. So, like, it gets really bad. I feel like some artists today, they search their own name on on Google, for example. Like, the first thing that pop up is not their website anymore. It's like their model that mimic their style. And and, you know, if if I'm a customer, there's absolutely no reason for me to go

Speaker 6: to the artist, wait a couple months, spend

Speaker 3: a couple $100 to do that. Right? I can just use a model to do that right away. So

Speaker 1: Yeah. It's, I haven't I have more stuff I wanna get to, but it it reminds me of one of the first things I think most people do. You know, a year ago when a lot of people got access to these models for the first time, the first thing you do is you have it generate something totally abstract or just a concept. But, inevitably, you come to that question of, oh, I'd like this artist. Could you do that same prompt in the style of that artist? And I think that's the moment most people realize the full implications of, you know, this this tech we're just figuring out for the first time.

Speaker 3: Exactly.

Speaker 1: So if a person wanted to, let's say, deceive that model, to create an image that the model would misunderstand in some kind of way, to create data that would, to borrow the phrase, poison it to prevent such mimicry, How would they go about doing that? And I guess to get to your project, how did you go about doing that? Mhmm.

Speaker 3: Yeah. So I think for us, so we the product we call it Glaze, disrupting kind of, style mimicry step. So the idea is fairly straightforward. It's okay. The constraint we have is we can't change the piece of art too much. We can add some small changes. Hopefully, that's not too disruptive. But what we can do is we can carefully craft these small changes to, you know, to confuse these AM models. Right? Those are very smart, but these AM model has, a very different way to see images compared to how us humans see images. Right? And we can leverage that, basically, that gap to add in some small changes that are very small to humanize, but are very disruptive to how these models see images. And so this kinda how we deal with CLAY is we add some small changes to us humans, the same image. But the model see that image was saying, okay. This is actually a completely different style from a completely different artist. So, of course, it would not be able to steal or do the style maybe as you normally would.

Speaker 1: You used the phrase how they see the images. Yeah. Earlier on in the conversation, I asked, you know, what did these models see? And we we talked a bit about the text associated with it, the way that they're tagged and the sort of human reinforcement element. But I I guess when you're changing what the image sees without changing what a human sees, this might be too abstract a question, but what is it seeing? What is it seeing that's different from what I'm seeing? Or is there a way that I, as a human, will even really understand that?

Speaker 3: So I think, one analogy I I I tend to give is you can think of this as a UV light. Right? So, like, machine learning can be a UV light system. And, of course, they turn a mass number of pixel values. So they see certain things we don't really see. Right? So it's just some in some sense, open UV lights, see a lot of hidden things there. Right? That means we can snuck into a lot of changes on the normal kind of light frequency. But to us it doesn't really change much. But once you open the UV light, like these models, you will see so much different changes, it was super disruptive to how this model can understand, images. But to go a little bit more technical, these models are basically functions, right, they map raw pixel vectors, you know, basically functions where they map raw pixel vectors into a bunch of high level, not very like, black box features that they use to reason to generate different art. And the feature space can some sense be interpret interpretable, like, you know, Van Gogh images will be similar to other similar artist images. But beyond that, the models that are setting space is very hard for us human to understand. But because of that, we can also add in some small changes to really disrupt that space because we know the exact function that's being used to, you know, process images.

Speaker 2: Given that you're changing essentially imperceivable values to the human eye, but there are still values being changed kind of in the in the image binary. How hard would it be to then train the AIs to identify those manipulations and bypass them?

Speaker 3: That's a great question. So, yes. So since we released Glaze, there are quite a few kind of these type of attempts, to train the AI to recognize them. So what I'm gonna say is it is generally hard to do that without sacrificing its normal performance. Mhmm. Right? So, kind of without going into too much detail of this, so this is so these type of small changes is a vulnerability that, kind of researchers had identified for a very long time where this model had these problem that if you ask some small changes, you'll very easily to confuse this model. So there are quite a few research kind of, you know, how do we make model robust against these changes? And, in general, kinda after, like, five years of that line of research, it kinda agreed, the fact this is very hard to do. It seems that this is some fundamental property of this model. And, so so in order to be robust to these changes, you basically have to sacrifice a little bit of how your model perform. So we did some testing in our research, and in the case of generative models, the sacrifice is quite significant. Right? The reason really is just because you really have to be super precise to get a very high quality artwork. Right? And if you change a little bit, you start outputting some real artifacts that that's not very usable.

Speaker 2: So there's not like a there's not like a global fingerprint to Glaze that I can start to identify and just kind of remove or extract that that, malicious data from?

Speaker 3: Yeah. So I think there are people trying that. What we see has been felt or only working very specific cases. And the real reason is that we kind of proactively thought about this, so we added quite a bit of randomness into this whole glazing process. Mhmm.

Speaker 1: Starting some new isn't just hard. It can be downright terrifying. You put a lot of work into a thing. You're not entirely sure it's gonna work out. You're taking a huge leap of faith. I've started a few things. Now I know I was right for believing in, you know, the idea, the product, despite all of those fears and hesitations. But boy, does it sure help when you have a partner like Shopify on your side. Shopify is the commerce platform behind millions of businesses around the world and 10% of all e commerce in The US. From household names like, well, hacked podcasts merch, to brands just getting started, you can get started with your own design studio with hundreds of ready to use templates. Shopify helps you build a beautiful online store that matches your brand style. Did I mention that that iconic purple shop pay button that's used by millions of businesses around the world? I don't know why I wouldn't. I should. It's why Shopify has the best converting checkout on the planet. It also helps boost conversions, meaning less carts, sort of getting abandoned in the parking lot, and more sales for you. It's time to turn those what ifs into sign up for your $1 per month trial at shopify.com/hacked. Go to shopify.com/hacked. One more time, that's shopify.com/hacked.

Speaker 4: When you finally find your thing, you want the whole world to know about that thing. So you use a thing called Canva to make it an even bigger and better thing. Whether you want to create flyers for that thing, make presentations for that thing, or design merch for that thing, You can do anything. So people can see your thing, feel your thing, love your thing. The next thing you know, it's a thing. Canva, the thing that makes anything a thing.

Speaker 1: Study

Speaker 6: And play. Come together on a Windows 11 PC. And for a limited time, college students get The best

Speaker 3: of both worlds.

Speaker 6: Get the Unreal College Seal, everything you need to study and play with select Windows 11 PCs. Eligible students get a year of Microsoft three sixty five premium and a year of Xbox Game Pass Ultimate with a custom color Xbox wireless controller. Learn more at windows.com/studentoffer. While supplies last, ends June 30, terms at aka.ms/collegepc.

Speaker 7: When you need to build up your team to handle the growing chaos at work, use Indeed sponsored jobs. It gives your job post the boost it needs to be seen and helps reach people with the right skills, certifications, and more. Spend less time searching and more time actually interviewing candidates who check all your boxes. Listeners of this show will get a $75 sponsored job credit at indeed.com/podcast. That's indeed.com/podcast. Terms and conditions apply. Need a hiring hero? This is a job for Indeed sponsored jobs.

Speaker 1: I'm curious to learn more. So you brought up Glaze, which, as I understand it, is sort of the artist facing tool. It's the thing that you can put an image into to give it these qualities that are going to, cause it to be misinterpreted by an AI. I also understand there's Nightshade. Can you talk a little bit about these tools, where one begins and one ends, how they work? Take me through that.

Speaker 3: Yeah. Absolutely. So, Nightshade is a direct kind of follow-up to Glaze. You try to improve, do something more. So, you know, if we take to AI company perspective. Right? So if we look at this and the worst case with glazes, okay, there's some data I just cannot learn from. That's not too big a problem because there are just so many other art out there are not protected by glazes. There's so many, you know, historic artists that can throw those. So it's not too big problem. So for nightshade, what we did is okay. We can take this one step further is if you train on these nightshade data, what happens you will not only not be able to learn anything from these data, but you will also corrupt the base model that you already have. And so corruption can come in many forms, but we're showing the paper that you can basically have the model to output a cat when you ask for a dog or, you know, all sorts of weird stuff you can have the model to do. And this will have a pretty big impact on, you know, how trustworthy your model is. If you put out there, you can't even generate a basic, concept. Right? So that's kind of what nightshade does.

Speaker 1: Oh, cool. So Glaze is just about making it so it can't read. It it's turning it into text in a language that it can't read anymore. Nightshade is like, no. This is actually gonna deceive it. You're looking at an image of a cat, but what you're seeing is an elephant. So next time we ask you to produce a cat, it might have a trunk.

Speaker 3: Exactly. Exactly.

Speaker 1: Oh, that's very fascinating.

Speaker 2: So you're poisoning the well, essentially.

Speaker 1: Basically. Yep. Interesting.

Speaker 3: Interesting.

Speaker 1: What would it take either of these models for them to be really useful at scale in the real world? Where do these models and, say, large social media platforms where a ton of the data that's being scraped to train these, models are coming from? What role do they have to play? How does this go from being a tool that an individual artist has to upload an image into, get it back before they upload it somewhere else? How does it get more useful than that current state?

Speaker 3: Yeah. Absolutely. So we are well, we're fairly already kinda talking to all of this. But, so currently, the model is basically, as you said, artists will you know, download a tool and generate take quite a long time, so it's not very scalable. But we have been talking to our platforms. These are shared platforms. Some are very pro AI saying AI is great, but there are quite a few of them saying you they should protect the copyright of the artists who share the image online. So we're right now working with one of the leading platform in the space, and I think they already start integrating, Blaze into the the platform. So every time they share a image online, you basically have the option of Blaze it or or not. So that's kind of, where we're going. And we also talked to quite a few entertainment, gaming company. They're in a little bit sort of weird position. They're like, they wanna use AI because you'll it'll save their cost, but on the other hand, they have a huge artist, you know, base that they don't wanna piss them off. So I think I think they are a little bit tricky. So we're talking to some of them, we're talking to Riot game a little bit, but, like, they're also interested in integrating Glaze inside. Also, the entertainment company has basically the exact same problem. They're saying, you know, Disney recently sent a letter to Microsoft saying that you can't train our Disney characters, for example. So, like, they're obviously very interested in protecting their IP, but, it's unclear they're too sensitive of what's what's gonna be the future for for them. So

Speaker 1: Yeah. There's a really interesting, kinda tension that you bring up there where presumably some of these companies would love the idea of not having to pay the labor cost of a bunch of artists to generate some of these assets. But I would bet more than that, they would be really angry that you can type Mickey Mouse into a box and have it produce pretty good images of their very valuable intellectual property. And here you come with a product that goes right down the middle of that. Exactly. Do you feel like you're wading into the middle of a kind of a a a pretty long term fight that's gonna be playing out in that intersection of tech and intellectual property? Like, you're sort of you're setting yourself up to be right in the middle of that for the foreseeable future. Yeah. Yeah. Absolutely. I think when we start

Speaker 3: a project, we absolutely does not expect this at all. But the you know, once we released, I was like, okay. This is a real problem people are facing and and, you know, everybody is talking the space. But, yes, I think, right now, we're we're we're more kinda committed to just be in the space to I guess we frame ourselves as, you know, provide technical solutions when there's no regulation, really, the space at all. Right? Like, there are people actively hurting, and we're all waiting on the, like, the court cases. We're all waiting for, you know, the the legal action to take place, but, they're gonna take some time. So the technical solution really should be there to help some of the creators to to at least get by it for the second. Mhmm.

Speaker 1: You brought up regulation. I'm not I'm definitely not asking you to, draft up some regulation on the fly on this podcast, But, like, broad principles, maybe. Some some sort of, like, nice to haves, vague stuff. Again, not asking you to write it. What what kind of things maybe where do you think it's gonna go in terms of regulation? Let's start there.

Speaker 3: I think it's gonna be a little bit hard. I think there are just so many misaligned incentives in the space. Right? There are, you know, of course, we wanna protect labor. We wanna protect graders, but there's also the aspect of, we have to build the next big AI system before some other country Right? So the the discussion around this has been very slow. We'll see how the court cases goes, but, you know, even the court cases went well, there will potentially be new laws from the senate, regulating AI for better or for worse. And this is only The US and there's, you know, the whole Europe. Europe was going pretty well with AI act until recently. They kinda stopped their bunch of, tension around that as well. And then there is Japan and China, which is much more pro AI. I think Beijing just had a, passed a law yesterday. Say you can copyright, AI generated images and things like that. So, so I, my take is even in the longer term, I feel it's never clear. It's gonna be a clear cut. Right? It's not gonna just benefit artists completely. So I feel, that's why I feel like technical solution, some sense, is useful. But also, in some sense, technical research can also push for certain regulations, really. If we feel, okay, these are how many artists are impacted. Our our carrier want to protect their tools, protect their art, that may have some implication of the ongoing legal discussion. So Mhmm.

Speaker 1: Yeah. It's it's hard to say we need to pass some sort of law or regulation that platforms need to do something to images hosted there to prevent them from being scraped without the consent of the creator if that tech doesn't exist.

Speaker 3: Yeah.

Speaker 1: Like Exactly. You can't even start that discussion. Interesting.

Speaker 3: Or, like, there's, similarly, I think there are, like, a lot of platform trying to say, let let me compensate these creators. But, like, compensating through a model is gonna be very hard. How does each training data point contribute to a given generation and things like that? So technically, I think, the space is also very important.

Speaker 2: I can't I I can't help but just see the knock on effects too of, any kind of precedent set in court or legislation because, you know, we're kind of at AI, you know, one point o per se. And and it's only gonna get bigger and bigger. Like, one of the things that I look at when I see stylistic mimicry excuse me.

Speaker 3: One of

Speaker 2: the things I look at when I see stylistic mimicry is, like, you know, we're doing images now. We've got some text generating. You know, I can have an AI write me lyrics to a song. I can have AI modify an audio track that I sing to sound like any other artist. We're only a few steps away before I can just be like, hey. Write me a Drake song about, you know, dancing in the flowers, and it'll be like, boom. Here it is produced, and now out it comes. And it's like, you know, any kind of decisions that they make in the court cases today will impact all the future AI generated content. So it's just I I hope that they're weighing, you know, in the balances how severe some of these decisions will be

Speaker 3: Yeah.

Speaker 2: To the creative class anyway.

Speaker 3: Yeah. Yeah. Exactly. Yep.

Speaker 1: So you're not working on this alone. I'd be curious to know about, like kinda kinda tell me a little bit about the whole team. And it's specifically what kind of because this seems like such a new thing, what sort of backgrounds are people bringing into this project?

Speaker 3: Yeah. Absolutely. So I think, it's a little bit weird for us. Like, I say, we are a traditional research team, so I'm I'm doing my PhD to do my master here. And then we start out just as a research project. Right? So, traditionally, we kinda research in the space of security and privacy of machine learning system. So a lot of looking at how AI works and when does AI fail, where I look at things, you know, how do we make sure AI self driving car is secure? So we do a lot of that sorts of stuff. But then I think when generative AI really start picking up, we we start to talk to quite a few artists and then we see how so where the problem is, like, last year. So, okay, we should kinda steer our whole project, our whole focus into, you know, protecting artists. And and the reason really is that we're kinda in a very fortunate position because we study these AM model. We'll study their variability for quite a long time. Right? So we're like, okay. We know how to exploit this variability as a protection tool for these artists. So that's kinda how we started. So the team is fairly small. It has my advisors, my two advisors, in the space. They has been doing kind of CS or commerce science research for the longest time. And then there is, us, me, and my two other coauthors. We just do with our PhD, but our background is more on privacy and security of AI systems, and we cannot share, to do more of the generative AI stuff these days. So

Speaker 1: I'd I'd seen two different numbers floating around for this. In in regard to Glaze specifically, I've seen a 1,500,000. Roughly speaking, how many artists have used these tools you're working on to date to protect images?

Speaker 3: Yeah. So we don't have the exact number of how many people are actively using it every day, so we only have the number of how many people downloaded the app from our websites. So I think what in July was 1,000,000. I say we got to 1,600,000 as of, I say, last week when I checked. But this is just a number of doll just a number of full downloads. They download the whole package, download all the resources, and perhaps start using it. But, yeah, we did not keep track of anything after the download just for for privacy reasons. But we also have, Webglaze. It is kind of a service we put up for us who doesn't really have a laptop or doesn't have GPUs to run, Blaze. For that, I said we have 3,000 active users, but we have a huge wait list. We haven't really get around to put people on the wait list just because we don't have enough GPUs for the, for the moment. But yeah.

Speaker 2: If somebody wants to support the project, are you guys taking public support, or is it just an internal project for the University of Chicago?

Speaker 3: I think we very recently, worked through the university to have a donation platform. So we will go through the university, but a portion will come to us to continue research in the space. Great.

Speaker 1: You, so you've mentioned you've mentioned protecting artists. You've mentioned technical solutions being really important for figuring out the legal side of things as we move deeper into, you know, the AI era. Beyond those things, I guess just for you personally, why why do this work? Why why kind of take on this really long scale battle with these very well resourced companies that have a huge financial incentive to keep these models ticking? Why does this matter to you personally?

Speaker 3: I see it just okay. I think there are a couple answers and there are typical answers. Okay. I want to help people. Of course, I want to there's so much, like, very rewarding feedbacks from artists and working with them just you know, very enjoyed that process. But also I think, the the kind of the bigger reason really is that I'm just interested in protecting human creativity in some sense. Right? This is basically it. I feel like the companies or even government are taking a a fairly four sided view on AI. Right? Of course, these are awesome. They are able to copy style, generate images. But what I see is if we keep going down this route, say, you know, they get better and the artist getting replaced, And that more or less kind of mark the end of human creativity at least in some aspects. And it's unclear where are we gonna go from there because as we see today, these models are not really able to evolve on themselves. Right? They're mostly still feeding or mimicking existing art. So so a sad future will just be we're stuck with the same type of art for, you know, hundreds of years that we can't go back because there's

Speaker 6: no more artists. There's no more art

Speaker 3: school left. So So I think we really wanna push for that. Just, you know, give artists some leverage in this negotiation to to protect the human creativity at this point.

Speaker 1: Yeah. You used the phrase protecting human creativity, and it's it is funny how within about four seconds of using one of these things, you're sort of like left with these two competing feelings. One is kind of sort of like the technical awe and I'm sure as a very technical person, you really are like, wow, this isn't remarkable what you've managed to achieve. This is catastrophic potentially for human creativity, at least in terms of it being economically viable. There will probably always be someone wanting to pluck on a guitar in their living room. There will always be someone wanting to to draw on a pad of paper. But whether or not it's a job, a career, something that you can make a living doing, you you're immediately struck by how big a threat this incredibly cool tech could be to that as just a thing people can do for a living.

Speaker 3: Yeah. I think so. When we started off, when I saw these two, I was like, I was more on the other side. I was, oh my god. This is amazing. I can't generate. I don't know. Like, Darth Vader eating sushi or whatever. Like, whatever I want. And, and and but I think once we start talking to one more artist, specifically, like, I would have, like, give some of these talks to, like, the, you know, group of people. Always, you know, our parents come ask us, okay. Should I still pay my son's art school tuition? Like, these are the question we'll start getting. And so we're okay. This is very much the the the question that defines a human creativity in the future. So yep.

Speaker 1: Okay. My last question for you, for you, Sean, is where where do you think this goes next? We're having this conversation again in five years. Mhmm. What do you what do you what are the big moments? What where do you think this goes?

Speaker 3: That's a great question. I think it's hard to predict where AI is gonna be in five years. But Sure. But I think I don't know. I feel like there will be at least some regulation in the space. If we say there's zero right now, you know, it may not be great, but but there will be some. And I think a lot of our work just wanna fill the gap that regulation cannot catch. Right? Maybe they're able to take care of open AI and stability, but really not the random register online. So so these are the cases technology or understanding can really kinda, take in place. But also I say we are in this kind of a cultural shift with generative AI. Everybody using it, and just to understand how this impacts humanity, maybe for better or maybe for worse, to understand this space a little bit more and build tools to, help people to help steer AI to the place that we want it to be.

Speaker 1: Sean, thank you so much for your time. Really appreciate you, sitting down with us and chatting about all this. It's a very cool project, and, I look forward to seeing where it goes next.

Speaker 3: Alright. Thank you so much for talking.

Speaker 2: Yeah. Take care.

Speaker 3: Alright. Take care for joining us, Scott.

Speaker 8: You can't reason with the sun. Trust us. We've tried. This summer, it's time to put that angry ball of fire on mute. Columbia's Omni Shade technology is engineered to protect you from the sun's harsh rays that can burn and damage your skin. The sun is relentless, but so is our gear. Level up your summer at columbia.com to spend more time outside and less time slathering on aloe lotion. You're welcome. Columbia, engineered for whatever.

Speaker 5: Spotify, it's Jay Shetty. Are you one of those media strategy people? Scrolling through spreadsheets, searching for an audience that pays twice as much attention to your ads than they do on social? Let me introduce you to fans and they're here with me on Spotify. Trust me, I know fans. They don't skip, they stay for hours. They don't move on, they manifest. They're not a demographic group, they're fans.

Speaker 3: Spotify advertising. You're among fans.

Speaker 9: The most memorable gifts aren't found, they're made. Zazzle is a custom marketplace where you pick any product, a mug, a card, a tote, a phone case, and make it personal, a photo, a name, an inside joke, the kind of gift that actually fits the person. That's what 30,000,000 customers have been coming back to Zazzle for over twenty years to find. Right now, save 25% on your first order at zazzle.com. That's zazzle.com. Make it z amazing.