Speaker A: Hello and welcome again, everyone, to another episode of Waiting to Be Signed, a special interview episode. We have with us today Ivona Tau, who you probably know from her generative AI GAN photography, very multimedia mixed discipline release as an fx hash. And of course, we've got Trinity here and myself, Will. Ivona, welcome to the show.

Speaker B: Hi, really happy to be here.

Speaker A: Thank you for taking the time. We were so lucky to have met you at NFT NYC a few weeks back and having you come, you know, agree to come on the show and talk to us because really, You know, something we talk about on the show a lot is like people don't really understand even the code behind a lot of releases on fxhash and struggle to kind of understand what's going on just with those types of projects. But then you've taken this extra level of adding your personal photography, the GAN network training. So I think it's going to help a lot of people to understand kind of what you're doing and the technology behind these projects.

Speaker B: Yeah, of course. I'm, I'm also really happy to have met you there and yeah, looking forward to this discussion. Uh, I mean, AI, artificial intelligence, generative adversarial networks, there are just so many buzzwords that people start using nowadays. And especially with the rise of Midjourney and DALL-E, there is a lot of confusion what can be done, what is easy, what is difficult. And we have so many amazing artists as well who are just pushing the medium forward. So yeah, a lot, a lot is happening. Definitely.

Speaker C: Can't wait to hear what you have to say about all of this. But maybe before we jump into all of that conversation, because I think that's going to be a huge rabbit hole that we could probably talk about for hours, maybe it'd be great to get a sense of who you are, what your background is, how you came to this entire crazy world of both first and foremost art, And then, you know, digital art, AI art, and then also how you got your introduction to fx hash.

Speaker B: So, to start with, I feel that, well, artists usually say that they have been artists their whole life and it feels like, like true thing, but more so I was always just curious about the world. And in the beginning I was curious, you know, when I was trying to paint when I was little, but then I realized I'm not very good at it. So I actually lost interest in painting quite quick. But then I realized that camera, especially film camera, is a magnificent being, that it captures world in a very different way than we see that with our own eyes. So this curiosity of world and this curiosity of looking at things of different viewpoints, of subjectivity, is actually something that always led me forward. And I think that to some extent, this also was the main reason why I went to study undergrad mathematics. A lot of people, when I hear that I studied mathematics, they're like, oh, you must be good with numbers. But the thing about mathematics is that you don't even use numbers in undergrad and master's mathematics studies. You just have to imagine a lot of super complicated things. Things, infinite dimensions, infinite spaces. So this is something that definitely takes you out of this world. And yeah, this, this fascination is something that I have still with me, this fascination with the different understanding of what actually is the, the essence of the dimensions that we live in. So going forward, my life was kind of Dual in a sense that I was studying computer science, mathematics, but also I was learning photography. I went on to study at Academy of Photography in Warsaw, so also focused on the art basics, art background, to learn about the masters that were there before our times. And at some point, I think it was when generative adversarial networks came around, so it was in. 2014, I realized that the field of computer vision is something that was made perfect for me in a sense that I could combine coding with photography and also try to use code for creative goals and maybe even start recreating what I have seen, what I have collected with my photos during the years with the use of networks. And the first experiments were really fascinating, but the first GANs were quite limited. They only outputted images of like 16 by 16 pixels. So that was not really of an artistic quality, more of an experimentation. But as the time went by and the methods got better and better, and also I was very involved in AI research. I'm actually just finishing my PhD. Monday, I will be doctor officially. So that's— Congratulations.

Speaker C: Yeah.

Speaker B: Thanks so much. Something to happen very, very soon. So yeah, I've been very active in AI research in the time as well and published some publications with regards to how can we combine different modalities, images and texts. So this is something that is very important within the AI research. How can we understand photographs and images? And also connect them with the text representation with natural language. So yeah, all of those fields that I have been really interested in came together when I started doing GAN art and AI art and combining that with photography.

Speaker A: Amazing. So I think as we continue this conversation, we should talk very briefly, do some definitions for anyone who doesn't know What a GAN, or a generative adversarial network. Now, of course, we know, you know, Will and Trinity, we 100% know this, but it's not for our benefit, but for the benefit of the listeners who may not know. Perhaps you can give, as best you can, kind of an explanation of what those networks do and how they play a role in the work that you make.

Speaker B: Yeah, of course. That's something, as you mentioned, is a very complex issue, but the simplest way of understanding that is that GAN, or at least this type of model that I train, is essentially me showing inspiration photographs to an artificial network that then tries to replicate what it sees. So I'm showing it examples of my work, and then it takes noise and creates something that is as similar as possible to, to my examples, to my photographs. And in the beginning, it does not really do a good job. It just captures the general shapes, colors kind of aesthetics. But as the time goes by and when I run the optimization algorithms on this network, so some very complex mathematical stuff is happening on top of that, then the network gets better and better at reproducing. But it's not really reproducing on a one-to-one basis, more of capturing the essence of what is the theme, what is the subject, of the general collection of photographs that I'm showing. So in order to be able to generalize, actually this network needs to see a lot of examples. So this is why my networks are usually trained on a couple of thousand of photographs and then fine-tuned on some smaller datasets of a couple of hundred photographs. It's really not about network and AI remembering examples and finding different more things between them, but it's more about AI learning to mimic the overall distribution of everything that is possible within this realm of a collection of photographs or this realm of the world that it sees for this collection.

Speaker C: So over time, and this is maybe part of the adversarial part, my understanding is that there are kind of 2 networks in play, right? Or 2 different parts of it?

Speaker A: Yes.

Speaker C: Maybe this is— I would love to be learning more about this as well, even though I am the expert as well. How does it work? How does it get better? Is it just through that one optimization algorithm or is it something that it gets better and better every single time it sees new images?

Speaker B: Yeah. So actually the essence of GAN is, as you correctly mentioned, is that there are 2 networks underneath that. And one network is taking noise as an input and outputting something, as I mentioned, that is trying to mimic the real-world distribution. But how it knows whether something that it's creating is good or not is by taking answer from the other network. And this other network is trained to distinguish between artificial inputs, outputs, and actual real images. So those 2 are playing a game in a sense that One is trying to cheat. It's trying to create something that the other would think is real, even though it's created. So with this play that is happening in phases, one network gets better because it's able to distinguish better and better. But on the other hand, the network that is responsible for creating the artificial outputs is also getting better. In some sense, they could play on forever and create something that ultimately will be indistinguishable by humans. However, this is not essentially what I'm trying to achieve with my creative process. I'm trying to find those hidden representations of what is the most important part of my collection of photographs. So when I'm trying, when I'm starting a project, I'm actually starting usually with a question, with a question in mind of what is the essence of, let's say, of this representation of my memories. How do they get seen through the eyes of network? And through the limitations of neural network that it's not able to perfectly capture what I have seen, there is something new that gets created. This is usually my goal in my artistic practice, to find out this, this new part of the distribution that Never has seen the real world.

Speaker A: I'm curious to follow up on that. So the underlying networks that you're using to either create the new images through noise or to judge them, right, those are both coded processes that are made by humans somewhere. So to what degree do, like, the built-in biases of the people who design those systems then influence the outcomes? And, and do you play with a lot of different base neural network software because you've identified like, this one's really good. And I think we kind of see this in, say, DALL-E and Midjourney, which both kind of do to a base user like myself, they, they quote unquote do the same thing, but they produce very different like aesthetic outcomes. So is that kind of a challenge on your end? Or are you actually going in there sometimes and being like, I have to just design my own because there's nothing out there that is making the type of outputs that, that I'm desiring?

Speaker B: Yeah, that's, that's a great question. And especially bias is such an important part of neural networks. And this is why data is actually what is most important. And this is also the reason why I work with my own data and I don't train networks on some available open source collections, because in those collections, as you mentioned, there is usually so much bias that might get brought up by those networks. And it's even the case that if there is 20% bias in the training data, let's say, the trained data will have 40% bias because networks are just so good at capturing those features of the data and just making them bigger. So I play with bias in a way that I use my own data and I want to find those biases in my own data. I want to see what network suddenly realizes that I do more, or maybe there is something that is really an important part of the collection or the theme or the topic that I'm exploring. So for me, this is actually a tool of exploring the biases. Essentially, by curating the input data, you're able to very strongly influence the outputs. And this is why a lot of people are actually saying that, yeah, the outputs of Midjourney are looking the same as There is, of course, it's very, it's possible to make something, find something in this artificial distribution that is quite different. But the limitations and the biases that have been in the training data, even though it has been so vast, this is something that is one of the most important building blocks of the network. It is actually even more important than the architecture, at least nowadays. Those architectures might be different in terms of training speed, level of details, as well as some properties of those trained distributions. But the visual aesthetics is something that is captured with the training data itself to the most extent. So if you want to have something that is completely different, right now there are no shortcuts but to train your own models regardless of what architecture it will be.

Speaker C: And so every time you start a new piece of work, are you starting from scratch or are you using previously used datasets or models that have already kind of understand your own biases in a particular way? Or is it something that is like kind of new and unique each time? You just have like many children.

Speaker B: Yeah, that's, that's a great one. And it's, it's also a very important part of my process is that As you say, I have many of those children, but some of them are smarter than the others. And there are even some of the models that I have been training for more than a year now, as I make them forget the previous information they learned and also learned new information. As I like to believe that some of the features that got learned, they get reused. And this is something that I notice when I take, let's say, a neural network that I trained on Night cities, and then I show forests to that, it suddenly takes those features of night cities, let's say buildings or roads, and then it transforms them in a very interesting way into organic and nature forms. And this is also a very important part of my explorations, is how those features get transformed, as in our minds they usually work as absolutely separate concepts. But for a neural network, this is something that it has to rework and transform. So yeah, there have been some models that I've been training and retraining on new data, and then on third data, fourth data, fifth data, etc., etc. What is even more interesting is that some of those older features might emerge in a very unexpected manner. So I really like to see a reminiscence of my previous datasets in some of my newer models. Of course, they forget quite a lot of stuff. There is a lot of forgetting in neural networks. But for me, just working with those models feels romantic in a way that they're living entities, even though they're super artificial.

Speaker C: Yeah. And I think looking at some of the work that's more specifically AI art that isn't on fx hash, You do kind of see this, like, this romantic interplay between, like, looking at Under the Waves, for example, where you have some things that are incredibly organic, like top-down waveforms in the ocean and then buildings. But when you're looking at the interplay in the loops, it's something that is neither one. It's just like this kind of very liminal in-between as it goes from like one morphological shape to the other. How does that really work with like some of your visions? Because before you were saying like it's very much of like a dreamlike state and that's where you get a lot of that inspiration. So it'd be really interesting to talk about that, um, or to hear about some of that interplay between 2 different types of like trained images, I suppose.

Speaker B: Yeah, absolutely. This is something that I really liked in terms of exploration of what Or the limitations of AI and also how it is different, the world and the representation that AI learns. So in this case, I think it's the closest we can get to AI imagination because when we train AI on cities or on oceans or on forests, we just show examples of what is real and it's trying to mimic that. But in this project, uh, in particular, in Under the Waves, what I did is I showed 2 different distributions and nothing in between. For us humans, we would normally just learn those separate distributions and say, this is it. I can imagine 1,000 oceans and 1,000 cityscapes. But AI works in a very different way. And I really like to explore how continuous of an infinite space it learns. So in this continuity, it means that there are no blanks. There is just always a way from one representation to the other. And this is how I was able to actually discover those in-between states. And for me, it was just amazing to see them happen as there was nothing like that in the training data. So it is like going one step forward from showing AI what it has to learn to, to leaving it to its own imagination of the in-between states. This is also one of the reasons why I really like to explore my own training data, because then inherently you're able to give a task Give a goal that is not just as simply defined as you would normally have. This is where this kind of magic happens.

Speaker C: Just one quick follow-up on that, because, you know, you're saying like, here are 1,000 cityscapes and this is the reality, here are 1,000 waves and this is the reality. And then the kind of the outputs of that are like a new dream reality. What happens if you would then feed 1,000 inputs of this new weird dream reality Is this kind of like the AI having its own like dream state and new imagination?

Speaker B: Yes, something like that. This is, this is actually something I did in a couple of my experiments and having this repeating loop of AI training then on something that is already not real inputs, but something that has been already trained. And this is actually a technique that has been used by many of the successful and like pioneer AI artists who are just experimenting very broadly with it by the next stages of only showing artificial inputs to the network and not even having the real data anymore. This is one of the things that I also did in my Mem Discontinued project that is debuting in Hong Kong this month. So it actually has seen only artificial stuff and also destroyed stuff. So it was learning to forget instead of learning to learn. And I really like to explore in how can we change this task of AI by modifying the data. So it has definitely been one of my explorations of the field in AI.

Speaker A: I think hearing all of this. really helps to answer some of the questions that people might have in the back of their head, or, or address some of the misguided criticism of like, oh right, the computer just putting the images into the model, blah blah. But like, hearing you talk about all the processes and the amount of work and the amount of authorship that goes into these creations, and like you said, you're coming at it with a question, you have a rough idea at least of where you want to go. I'm curious to know, you know, this is an easy question, like, what is art? You know, like, um, we have so many abilities now with these semi-publicly available tools, right? Like DALL-E and Midjourney for anyone to just jump in, work with it with an already trained model. And, you know, built-in biases aside of maybe the types of words you might use for a prompt, the layer of software that interprets those words and feeds them into the algorithm, and then the training data that was used to develop that stuff. How can we talk about, or how should we be talking about the line Between what is art and what is— I don't even want to say what is not, but like, how should we be talking about this stuff more intelligently in terms of like these outputs and these practices and like being an artist in this space?

Speaker B: Yeah, this is an extremely relevant question right now, and having a background as a photographer, it is actually something that I'm not afraid at all to talk about, as the photography analogies is the perfect here, I, I believe. To some extent, when when photography exploded, right now everyone can do a photograph. But does it mean that everyone is a photographer? Of course not. And it also does not mean that photography as an art does not exist. It just means that it is not enough to take a picture to create art. And very similarly to the tools that we have with Journey and Dali provided. Software is that it's getting easier and easier to create. And this is an amazing thing about technology being so democratized, but it's also getting harder to create art in a way that it's no longer— and this is a very good thing— it's no longer a novelty that you're doing AI art. In the beginning, we have seen so much appreciation for the work that was first work created with AI, first work created with GAN. But now it is not enough. Now it's all about the story, about the motivation, and what is the message that we are trying to give with this work? What is the background of an artist? What are they trying to explore and what is their motivation? I'm also not saying that it's not possible to create art with Midjourney or DALL-E, but it is just more difficult to find your own very individual voice. And if you want to create art with that, you have to strive to maybe use a very different kind of prompting or maybe use something on top of the process, maybe use the outputs from Midjourney as a part of your process that you will then recombine in Photoshop or in some other tools. And it's actually very in line with what we see currently as top AI artists. So for example, artists like Claire Silver or Jenny Pasaña, they're actually using the tools that are text-to-image, but they're amazing and accomplished artists because they found their own voice within those tools as part of their process, or maybe as an individuality in the process. So I would say that there are 2 ways to have your voice in AI art. It's either by finding this individuality or by working with your own data and working with your own models. This is then, of course, A lot more flexible, but I'm not saying it's not possible in any of those methods.

Speaker C: I think that seems to relate with, you know, some of the conversation that was happening on fx hash and in the Discord there shortly after, you know, Mythic Latent Glitches came out, right? Where, you know, it was unfortunately tagged as image composition, which is just that, that layer of like, here's a PNG file on top of a PNG file. Maybe we did something to it, but that's it. And, you know, I guess there is a world in which that's quote unquote what you did from a platform perspective, but without all of the extra intentionality and the work and the training that went into like the actual images that were then processed through the code.

Speaker B: The thing with using GAN outputs and then combining them in a generative, more procedural kind of usual for fx hash process. Is the thing that unfortunately someone called image composition tools as non-pure code. And that's, that's not really true when you're using GANs because then you're just doing some of the calculations offline. But yeah, of course, I also understand the need to kind of have different classes and have the conversation about what is pure code and what is website code and what is, what is in-browser code. So.

Speaker A: Yeah.

Speaker B: There are many different ways to divide those.

Speaker A: I'm curious what, um, you know, a lot of the work that you did prior to fx hash seems like it's been distributed, you know, on other platforms in, in, in a non-long-form generative series, right? Like fx hash requires your work to be when you publish there. What inspired you to experiment with that long-form format and take the outputs? It's not just the output, right? You've actually, there is some code and some kind of stuff that's done on top of the image to further distort it and warp it and create probably different colorways and stuff. So what brought your attention to fx hash? What got you interested in pursuing—

Speaker C: is it—

Speaker A: was it the more democratic nature of it than having something, you know, be at auction or be like a one-of-one that's more exclusive? I'm just curious, like, what brought you over here?

Speaker B: Going to fx hash was going out of my comfort zone in a very strong sense. As curation element when working with AI art has been very important to me. First of all, there's this element of curation when you're preparing a dataset, when you're shooting photographs. So this is something that I still had with both Metaclet and Glitches and City of Atom. But then I decided to experiment with not having the final say and also with working more with code I call defined code, even though both GAN and something you would see created in Processing are code works. But it's a very different process in a way that when you're working with Processing or JavaScript, you define boundaries. But in GANs, you don't define those boundaries. You just put some examples into the model and you see what gets generated, and it can sometimes be extrapolation of something you've shown to it. So I was of course very interested in how can I work with something where I have more control over the final outputs, but also this control is very different because it is left up to the forces of some stochastic processes and some random variables. So yeah, there was a lot of curiosity. As I mentioned in the beginning, curiosity is my main driver— curiosity of the world, of the methods of learning new programming languages, et cetera, et cetera. So I was very scared. I was super scared with both of my drafts to see the outcomes that will get generated, whether there will be enough of the variations, whether the ones that I really, really like will get generated. It's also something you can, you can never be sure of. But yeah, it, this experience has actually shown to me different ways of creating. And it has been very enriching to work with long-form generative art. It's also something that I feel it's still a bit controversial. So for example, when I did a collection preview of Midjourney and glitches in a gallery, and I've shown some random outputs, there has been so many questions from, you know, regular gallerists and art enthusiasts onto. But how? I want to buy this work. So I cannot. I'm not sure what I'm getting. And it was a concept that was still so confusing to many of those people. But I was really happy to try to explain and also to show the excitement of all that, that you don't know what will get created. And it also requires you a very different kind of coding so that when I'm training GAN, I don't really have to have all of the outputs to be of an amazing quality. I will usually scout the latent space and find the ones that I really, really like or the ones that correspond to my vision the most. So it doesn't really matter if there is some noise in the training data and some of the outputs are not perfect to my vision. But when you're creating a long-form algorithm, you really have to have all of them perfect. And it's something that I've also heard other generative artists kind of to maybe not struggle, but have a very different relationship towards too, because this is a very different creative process. So there have been some lessons learned, there have been some new experiences, but mostly there has been so much enthusiasm and excitement of being able to combine the two. So I also feel that we don't have that many artists who are combining the curated kind of generation, which would be similar to GANs or their own custom AI models with the kind of long-form procedural code. So this is one of the areas that I'm really excited about. And yeah, it's also not that difficult to completely combine in code if we had a little bit more memory on fx hash or maybe with some mobile style GANs that are already being created, it will be even possible to have this long form also with GAN itself. So this is also something that I've heard a lot of people trying to implement is to be able to mint GAN model and just prompt for, for the output. So kind of long form with GANs.

Speaker A: I think one person at least has actually successfully done it on fx hash.

Speaker B: Yes. Yes. There's, there's the one, one project, uh, the Moon Crescents, I think.

Speaker A: Yeah. Cyril Diagne. Yeah.

Speaker B: Yes. Uh, yeah. So the, the only limitation right now is that the outputs of this Mobile StyleGAN that they used have to be like 20 by 20 pixels because they're bigger than the model is too big to, to fit in the, in the repo. So there is this limitation that it's not possible for me at least to use with my data. Or in my process, but there have been some really interesting attempts.

Speaker C: It would be the, uh, GAN x 8BitDo ultimate.

Speaker A: Oh yeah.

Speaker C: Ultimate showdown.

Speaker A: Get me to make pixel art.

Speaker C: Yes.

Speaker A: I wonder, Ivona, you know, you kind of spoke a little bit about exhibiting your work and people who might be familiar with you from the traditional art world and used to buying, you know, buying physical work and being able to see it in a gallery and own it and not be like, oh, actually, if you want that piece, you gotta go talk to this person with a .tezos address who owns the NFT, right? Like, I'm very curious to hear from you, maybe to talk a little bit more about some of the resistance that you've heard. You know, we ask this question to a lot of artists when they come on the show. Like, it's not typical that creative folks embrace blockchain technology. There's a lot of resistance to it all over. It seems like all over the place. It's not very popular outside of the people who do like it.

Speaker B: Yeah.

Speaker A: So what has it been like trying to explain to people from the traditional art world why some of your work is on the blockchain? And as a side note, I'd love to hear a little bit about what is the process of selling something at auction at one of the major auction houses? And to the extent that you might be allowed to say, I think it would be really interesting for some people to hear like how different that process is between just like putting it up on fxhash and then you get everything into your wallet immediately, right?

Speaker B: Yeah.

Speaker A: And you get your royalties after the fact. So I'd love to learn more about the traditional world and your perspective on it.

Speaker B: Yeah, so starting from the second question, it's just so much more stressful. I don't recommend that. I mean, when you have an auction, uh, virtually, you just sit with your coffee, with your blanket on your sofa, and it's just so cozy. But when you're out in the actual auction house and you see All those old folks, you know, bidding live, it just, time slows and you're like in a completely different world, your heart racing super fast. Well, I mean, the physicality of all that is still something that I feel humans experience a lot more. So for me, this experience of course has been amazing, but, but also quite stressful. But coming back to the first part about how the traditional art world embraces the crypto world and NFTs. I have been so far quite positively surprised, maybe because I had zero expectations and I was just treating every opportunity as something that I would not normally expect from them. So every traditional gallery that was like, oh hey, maybe instead of just selling this print, we will also sell NFTs Just as an experiment was something very positive to hear. And from what I felt with a lot of galleries and museums is that they're curious, but they are very, um, very safely curious. They don't want to have a revolution. They don't want to throw out the paintings and the prints. They don't want to go on full-on metaverse mode. They just want to try some things safely, have an experiment, have a little bit of NFTs, and then see if it sticks. So, so far I've seen a lot of the trial and errors, and every gallery is, is doing that in their own way. There has also been the cases where the digital work has been shown in a very great way. There has been cases where digital work has not really been given justice. So I feel that there is still a lot of experimentation and first steps that are not always made completely in the right way. So while it's— I'm very happy to see that they're trying, sometimes it's not the best possible way that it is done. Also, being part of Web3 means that you're able to show your work all around the world and also to connect with galleries that have been on a different continent, or perhaps you didn't have the networking abilities to connect with those folks. So that has been one of the democratizing aspects for me to be able to get connected to those people from all around the world. While my work has been exhibited more locally, at least only in Europe right now, there is just no boundaries for that to be shown on other continents as well. So for me, this experience has definitely been very positive and something that I'm super happy about. Just having your work being auctioned in the same evening auction as a painting by Salvador Dalí is one, like, lifetime achievement unlocked.

Speaker C: It's been proven you are a peer of Salvador Dalí. Let it be spoken and just put right out there.

Speaker B: Probably a lot of those surrealists and pop artists and Dadaists, they lived today, like, Andy Warhol would be like top NFT artist for sure.

Speaker C: And it's interesting, you know, to hear about your experiences because, you know, I think that we've seen such a digital art explosion over the last couple of years because obviously digital art has existed for decades and it didn't need to be tied to the NFT world. This is something that we've talked about quite a bit, but it seems like the The larger global conversation around blockchain has really invigorated some of these conversations and I think that we see a lot of more curated pieces or some of like the one-of-ones being especially embraced by museums and galleries, perhaps for the specific reason that it is just a one-of-one, what you see is what you get. There's just an NFT receipt to kind of make it different or special. But there's fundamentally nothing superbly different about the work that might be shown. Whereas going back to your previous point about when you're showcasing some of your long-form art to galleries and it's just some of the random hash outputs, it seems like a lot of the trepidation might be more around the, there are 400 of these, none of them actually exist yet. And conceptually speaking, that is going to be a huge Mind shift or paradigm shift. Sure, it's a paradigm shift. Yes, a huge paradigm shift for how these people think about it. Do you think that there's like a world where, you know, having like the long-form random hash generation being like another layer of output or like cur— not curated, I'm trying to think of like, but just more that randomization. How do you think that that would be interpreted by Yeah, that's, that's a big question.

Speaker B: I believe that NFT world is really embracing the long form quite enthusiastically due to the fact that also when you mint something, it's also, it feels that you're part of the creative process. And this is something that has been very interesting. And perhaps that could also be a paradigm shift for the art world. We are embracing different forms of creation. So the physical art, the digital art, form of creation where we no longer speak about the artist creating the work solely, but also having viewer, participant, gallery, somebody who is experiencing the work as part of this creative process is something that has not really existed in similar ways. So I see it as a huge change, but also a huge opportunity perhaps for different forms of even ownership of creation process. Of course, this is something that they might be reluctant to do, but in a way, all the art revolutions are something that the old folks are reluctant to do.

Speaker A: Here's one that we, we've kind of talked about and kind of haven't. And this is also stemming from the conversation we had at NFT NYC where I made an assumption about how Midjourney's training might work. And you were like, no, absolutely. That's not how, not how it works. So what are, what are some of the common misconceptions that you've encountered with the public, with the traditional art world, with private collectors even who might be like really interested in buying from you about AI-generated art, GAN-trained art, etc.? Like, what are some of the big things that if you could just dispel these rumors right now and put it to rest, like, just what are people often getting wrong?

Speaker B: Uh, so the first one, and I think it's the biggest one right now, is that with all the hype surrounding Midjourney and DALL-E, 80% of the people think that AI art is text-to-image prompt generation. And I have been so surprised to talk to people and actually explain to them that, well, this is just a small part of what AI actually is. And this is just the part that got picked up by the media right now. And this is what got so popular. But AI is just so much more than just text-to-image. We have, first of all, we have of course GANs where artists are using their own datasets and training their own models. Then we also, we don't have to only create images. There is a lot of stuff where people are creating sounds, text, poetry, a lot of different modalities with AI. And yeah, so first of all, AI is just so much more than that. And the other one to training, to training text-to-image methods. So the thing is that there is actually no training happening in those methods. They have been pre-trained by huge companies. So OpenAI, entity that has, I don't know how many, but a lot of GPUs and a lot of memory and a lot of computing power and a lot of money that normal people or even normal huge companies don't have. So those huge models have been trained and have been made public by them, but it means that you cannot really retrain them on your own dataset. So this is something that I get asked a lot. How can I retrain my DALL-E? You just can't. You need like a billion bucks and be an OpenAI and have whole internet and then you can do that. So this of course means that there are those limitations. And there are those biases that we talked about, then those biases are of course the reason why those models are not really made so public. There are of course artists that are testing them, but when you have something trained on internet and contains a lot of nasty stuff and a lot of the preconceptions, racism, et cetera. So it's a scary thing to work with. Yeah, this is the second, like, like the biggest misconception. And then the third one that I actually get a lot asked about when people hear that I am a photographer and also a Ghanaian artist is people ask me, okay, I really like what you do. How can I make my photographs morph from one to the other? And I'm like, oh no, no, no, this is not how it works. And, uh, the, all of the frames in my works, even when they look super realistic, they have been completely generated. And even when you're looking at realistic forests or realistic waves or oceans, every single pixel has been created with code and does not exist, has not been captured on camera, has not been painted by hand. And this is just something that network has learned. So in order to learn that, you don't just take 5 photographs and make them morph together from one to the other. But you create something from scratch. So yeah, something, when you get something that is surrealistic, people are usually wary of the fact that it's generative and, and don't realize that. So I would say those 3 are like probably the main misconceptions I've encountered so far. A lot of the people, yeah, also as you mentioned before, also a lot of the artists, AI artists, uh, regarded as just, you know, one, one click. I mean, yes, you can have one-click-to-art kind of thing, but it's, yeah, as I mentioned, this is also not a very bad thing. We have that with cameras and with photography. So I'm actually not, not a critic of this approach. If you can create something amazing with just one click, then, then wow, you're just great. And if you have a story and you have a background and you have a goal and motivation and artist background, then good for you. Yeah.

Speaker A: It's really amazing to hear how kind of enthusiastic you are about those publicly available models because, you know, like you've said, it does lead to misconception and assumption on behalf of the public between, you know, what you're doing where you're owning and creating your datasets and modifying them and really working to develop your own systems versus, right, just a, you know, not to diminish text-to-prompt work, but it's different, right? It's just a very different way of going about it. I'm curious, and if this is too sensitive of a question, we can cut it, but did you take a look at that Reddit link that I put into the notes?

Speaker B: I looked at it, but I just skimmed through that. So let me just go back to it.

Speaker A: To summarize, it's a discussion between some folks about the legality of the training sets used for Midjourney and DALL-E in particular, and whether or not The underlying copyrights of the art and images that they've used to train it. I mean, I guess one way to think about it would be like very akin to like sampling in music. When an artist wants to sample music from the past, they pay a licensing fee to the original artist to do that. When you train a system like this on potentially millions of images scraped from the internet, some of which probably are copyright protected, but you haven't paid fees to use those images, And they're being sampled in some of these compositions, it opens up kind of a new legal territory in terms of the outputs. And, you know, these services are selling licenses to people to then resell the images that they output from the text, you know, that they, the text prompts they compose.

Speaker C: Yeah.

Speaker A: So if this is too controversial, too hypothetical, I'm just curious, like, for you, it seems like it doesn't apply, right? Because you are taking your own photographs and you're working like from scratch. But, and maybe this is where you could shine some light on it, right? When, when it's something like DALL-E, it has its entire set of training data and I give it a prompt. If I say Barack Obama something something, does it just like literally take a Barack Obama face or does it compose individual pixels that then recheck against? Like, I'm really curious, like, is it really sampling and recomposing images or are the images that it creates completely original and That kind of might get around some of that concern about like the training data they're using.

Speaker B: Yeah.

Speaker C: Yeah.

Speaker B: That's, that's exactly a very problematic question as the thing is that, well, let's, let's imagine collage, like collage, like photographic collage. Can you do that? And how much of a copyright do you have to the images that you use? If you take like 50% of the picture that is someone else's, of course you have a problem, but What if you take 3 pixels from somebody else's photograph? Then, well, I mean, maybe no one will realize, but then again, you're still taking somebody else's work. So maybe that's not completely right. And the problem with training AI models on huge collections of data is that you're essentially, you would have to take your number of pixels that can be, I don't know, 1,000 by 1,000, but like divided by thousands of photographs, like Every single pixel just contains small fraction of what is, what was kind of there. And it's not even like technically even the case because not a single pixel again was real. Everything has been kind of reimagined with AI based on the terms it had seen. So if it had seen 100 photographs of Barack Obama, it has this kind of vision of how Barack Obama looks like. But it's not really taking a single photograph of Barack Obama. So this is where the problem becomes real because you don't know which photograph it kind of learned from. It learned from all of them, but then again, from some of them it learned more, from the others it learned less. And this concept of the authorship gets really, really fuzzy and there is actually no there's even not much you can do from technical perspective to kind of attribute which training data influenced the outputs most. You can kind of compare which images it is most similar to, but then when you go pixel by pixel, you realize that those are just 2 different images and they don't have overlapping fragments. And yeah, this is, this is usually the problem because It makes some of this stuff too fuzzy, too difficult to have a case in court as it would not really stand. But also it raises a lot of ethical questions because you cannot just take somebody else's body of work and train AI on that and sell that. And I mean, technically you can, but ethically you cannot. And then the question is who creates those ethical boundaries. And even in the broad field of AI, there has been some manifestos by AI researchers that have been completely voluntary. Some of the researchers said, we will not create something that is against— that will work against humanity or will be used for malicious purposes. But this is something that you can technically do, but you just have to agree not to do. Yeah, with AI, there is still so much of this gray area. And I feel that at this point, the one thing that we can do is to talk about it. So yeah, so the more people realize what's possible, what are the limitations, and also realize, yeah, how it is done and also what should be done and what should not be done.

Speaker C: It's just a really interesting thing to think about, you know, because, you know, The intentionality and the ethical decision-making is really in the creator, not necessarily the creator of somebody who's making something on Midjourney or somebody who's making something on DALL-E, but even at the higher level. And I know that we're seeing a ton of inspiration and work being sold out on Versum or OBJKT, one-of-ones being sold based off of Midjourney outputs. And it's just a really interesting conversation that seems like It could never really end. And I don't know where I'm going with this because my mind is just going in brand new directions. So thank you. Yeah. For that.

Speaker A: It seems so challenging, right? Because like, as, as you kind of said, these images are composed and you, uh, from potentially like 1,000 underlying images, but then the resulting image is not actually pixel-level identical to anything. But in order to create these systems, they couldn't exist without all the data. And so it does kind of feel Like, just because the end result doesn't share any identity with an individual underlying piece of the training data, it doesn't quite feel like that is enough of a, an answer. Right. And it does feel like—

Speaker B: Yeah, exactly. This is why, like, the safest you can do, just work with your own data. This is why I do it.

Speaker C: I think that that kind of transitions into like another question that had been at the back of our minds. There are all these ethical concerns. You're training on your own data, your own photography. You know, you have a background in photography and that's actually kind of how you got into this artistic space. And so much photography is used as like the baseline training material and datasets for everything that you put out there. And even though it's gone through a huge GAN and AI process, and even though the outputs are very liminal, very dreamlike, do you still consider yourself a photographer? Or your output's photography or within that same realm?

Speaker B: That's a very difficult question. I would normally use a workaround and say, I just consider myself an artist and I no longer care if I'm photographer, post-photographer, AI artist, GAN artists. Like there's just so many terms. And I actually sometimes struggle when I have to define myself for a bio or for, for someone. Who asks like who who who we are to call you? There is just so many terms that come into play, like mixed media. What I think that is the most exciting is to work across mediums. So definitely photography is one of those that is like a very essential element to my practice, and this is something that I don't really imagine my practice without. But in some sense, it's also like post photography as it's no longer captured by lenses and it's no longer a representation of the actual world, but it is more reimagined through the technology. It is a very different process that is also so much more random. When we are using even experimental photography, we can sometimes apply different physical random processes, destruction to the negatives, or some weird techniques But it's still programmed physically by us. And when we use neural networks, there is this additional level of incompleteness and indefinitedness. So just some way it's like a very sophisticated contemporary watercolor technique that started with camera.

Speaker C: I think suffice to say, like, you may not find it like maybe ethically or morally correct to define yourself as a photographer, and maybe you don't want to because it's so much more than that. But on the other hand, some of your work would fit well within a gallery of photography. It's kind of neither that nor it itself, but it is really beautiful work.

Speaker B: Yeah, but that is a very accurate comment. And this is also something that I definitely don't run away from. I really like to think of myself as being a lot of things. Instead of, instead of like defining only for the very exact definition. And I also try to be very active on the photography scene as well, and I have not really completely abandoned that. Just to give you an example, next week I'll be going to Barcelona to Experimental Photography Festival to give workshops on how can we use AI in photography. So I'm still very proud to work with photographers and to collaborate photographers. And, and yeah, to kind of think about boundaries of photography that are also being pushed with AI.

Speaker A: Well, you know, we're over— we're just a little bit over an hour here. I'm wondering if maybe I, I came up with a, an idea for a question that might be a fun one to wrap the episode. But Trinity, do, do you have anything else you want to cover before I give my wild card question?

Speaker C: Yeah, I mean, maybe what we can do is do your wild card question Because I would like to end up with understanding what you're working on and anything that might be coming out for fx hash. I think that's probably the best way to wrap. So go for your wildcard question.

Speaker A: Okay. So since you're getting a PhD, I assume, in AI or something AI adjacent, like you said before, you work in the field. I'm sure you saw recently in the last month or so, this person who worked at Google who was making claims about the sentience of their AI and whether or not— I'm not You know, obviously I'm not going to ask you to judge if that actually happened. It seems increasingly like that this is not a sentient AI that was created, but I'm wondering if you're familiar with this thought experiment called Roko's Basilisk, and if you might give your more professional opinion. It's kind of the type of thing that a classic 4chan or Reddit person might philosophize on. Are you familiar with this?

Speaker C: Yeah.

Speaker B: So as you mentioned, there's just So much conversation and so much fantasizing about it and so many theories and not just the ones you mentioned. There is just, you know, I also get asked about what if my model suddenly decided it doesn't want to learn on my forest, but it wants to learn something else. What, what would it do? But, uh, But the sentiency and the question of, yeah, even the question if when we work with AI, is it a collaborator or is it a tool? Is there autonomy that is inside it? Is something that, well, it's very fascinating to think about. But on the other hand, unfortunately or fortunately, we are at this stage of technology where those questions are still so far from reality, I believe.

Speaker C: Yeah.

Speaker B: There is a lot of things that we don't completely understand, but still there is, there's no autonomy. This is like the simplest thing that we lack in the AI systems right now. And the definition of the goal, the definition of motivation, the tasks, this is something that cannot happen without the human element. So we are still very far from implementing that.

Speaker A: Yeah.

Speaker B: from the general artificial intelligence. However, how far we are, 5 years or 10 years or 100 years, this is the question that like no one has answers for. And I have no wild guesses. It's fascinating to work with something that is mysterious. And I think that we also, we humans, we like to anthropomorphize, well, to find those anthropomorphic features in the things.

Speaker C: Yeah.

Speaker B: that are mysterious to us and things we don't understand. And the forces of randomness and stochastic processes is something that we cannot fully grasp. We cannot fully grasp the concept of chaos, and we try to find the force behind it.

Speaker A: Does working in the field make you question or kind of reconsider the way like your own brain or human brain's work in general? Like, just do you ever find yourself drawing parallels between your thought processes or the things that you're, you're imagining and the way that your AI partners sometimes produce images? Like, how has it, how has it affected your personal philosophy for thought and existence?

Speaker B: Oh yeah, absolutely. So I'm super fascinated by the thing— what defines our consciousness And our personality. And I've been recently reading this book that actually I saw on Refik Anadol's Instagram, and the book was called Being You, and it's about the perception of oneself. And it's, it's really fascinating when we think about it that we just have those brains hidden inside our skulls without no light, no sound, and just know those neurons. coming in with a lot of signals that have a lot of noise, a lot of information, and we just make our own Bayesian guesses about the reality. So it has been very fascinating to also think how it is very similar to what neural networks are doing, and it kind of makes you think how our perception works. And yeah, I like to think about those things both ways, to sometimes think about my own perception, but then Also trying to incorporate those ideas and this philosophy back into training my models. Of course, it's a very simplified version of how this perception works, but the notion of forgetting of perception of oneself is something that I really like to explore with my creative practice. And it's also very often a theme of my work.

Speaker A: I wanted to ask that on behalf of my brother because he has a PhD in neuroscience and he studies this stuff. And I thought that would be a good crossover question for him to hear if he listens to it. Thank you for that. Wait, should we? Yeah, Trinity, do you wanna go with the last question here and we'll wrap it up?

Speaker C: Yeah, this has been such an amazing conversation and I feel like I've learned a lot and I hope that everybody who's listening has learned at least this much. Um, and thank you so much for taking the time to, you know, talk through everything. We really, really appreciate it. But maybe just one last thing before you go, cuz I know that this will be at the top of everyone's mind is, what are you working on? And is there anything that we should be looking forward to in the short-term or midterm or long-term future for what might be coming to fxhash?

Speaker B: Well, so I cannot give you like many details, but what I can say for sure is that I've started working on my 3rd collection on fxhash and This is something where I'm, well, combining again, my neural networks with the generative code and procedures and again, working with Longform. But this is something that, well, it's again, a bit out of my control zone. So it's even more scary than the previous 2 drops. And I'm learning a lot of new things and making a lot of mistakes, but also rethinking the idea of what is mistake. And what is the thing that we as humans learn within the process of learning? Very abstract, but definitely stay tuned for more.

Speaker A: That's not, not close. It sounds not close.

Speaker B: No. So it's like in, yeah, to, I think like autumn, September or so. Yeah. I'll be, I'll be taking some vacations in August, like 2 weeks off social media. No work to completely go off-grid. So yeah, then I'll be working a bit less.

Speaker A: Thank you so much, Ivona. This is amazing. I, I do wanna restate that I did actually know all of this before the interview. So this was, you know, I think really informative for everyone else, but not for me. I knew all of it. So, um, no, it's, it's, it's been amazing to have you on. I mean, I feel like we've packed so much into this hour. Like, I think it's gonna be super beneficial for everyone who's a fan of your work or curious about the space in general. And, uh, yeah, I don't know what else to say. Thank you. I hope you, I hope you enjoyed the process here and, and coming on the show with us and talking. It's really been a pleasure.

Speaker B: Yeah, thank you so much. I'm really happy to have been on your show and definitely a fan and looking forward to the next episodes.

Speaker A: Well, that's it for everyone. That was Ivona Tau, uh, future PhD. You'll probably be a PhD by the time this one drops. So Dr. Tao, congrats on that. And, uh, thank you. Thanks everyone for listening. We'll talk to you all soon. Later.