Salon with Linda Dounia Rebeiz #

Welcome to the third salon on regaining power over AI. For those who attend this salon for the first time, the salons are a place to have presentations and discussions about how we can rethink artificial intelligence, AI technology, from the perspective of artists and creatives, who work with AI or sometimes against it. If you have missed the first two episodes, I’m copying in the chat a URL where you can have a look and see the other events.

My name is Baptiste Caramiaux. I’m a researcher at CNRS, which is the French National Center for Scientific Research. And I’m also a researcher at Sorbonne Université and I’m based in Paris. So this series of salon is supported by the AI and Society Program by Microsoft Research, and co-organized with Jenny Williams, Gonzalo Ramos, Kate Crawford and Vera Liao.

For this third salon, we are very happy to welcome Linda Dounia Rebeiz. I discovered Linda Dounia when I was starting to do some work on data archives and data curation in relation to artificial intelligence and machine learning. And doing this work on data archives and data curation triggered a few questions such as, you know, what does it mean to create a data set that will then be used to train a machine learning system? And by that, I mean, what are the decisions taken when deciding which element to include in the data set? So who gets to decide and how and how this decision is done, what it’s in and what it’s out. So. And it’s also what is the difference between curating a data set and collecting one? These are really interesting questions that has really had the heart of of contemporary artificial intelligence infrastructure and machine learning. And, by curating and creating this data set, there is also the question of what agency it provides to the ones actually creating it? So I think all these questions are examples of questions I was very interested in. And I was looking for artists working at the intersection of these different topics. And I discovered Linda Dounia’s work. And I’m very, very honoured that she accepted to be part of this of this today.

So before starting, before letting Linda Dounia presenting her work and then asking questions, I would like to mention that we usually use a shared document where everybody can write questions. And that’s a place where you can put questions, comments, notes, whatever you think fit during the Linda Dounia’s presentation. And then after presentation, we’re going to pick some questions for the discussion.

Let me now present Linda Dounia more formally. So Linda Dounia is an artist, designer and writer interested in the philosophical and environmental implications of techno capitalism.She’s a Senegalese, Lebanese and lives in Dakar. Her work mediates her memories as alternative realities and evidence of excluded ways of being and doing. It is formed through the dialogue and tensions between live experience, code and artificial intelligence. In 2023, Linda was recognized on the Time AI 100 list of most influential people in AI for work on speculative archiving and building AI models that help us remember what is lost. One year later, in 2024, she was also the recipient of Mozilla’s RISE25 award for work in AI, advocating for greater urgency of algorithms, how we perceive them and how we are perceived by them. So on that, I’ll let you start your presentation and tell us more about your work. Thank you.

Amazing. Thank you so much for having me. Thank you for the introduction. I'm going to start by sharing my screen here. And awesome. Yeah, thanks for the introduction. And today I'm going to be talking about, I guess, a question that's that's very central to my practice, which is what is worth seeing? And I'm going to explain kind of what it means to me and why it's an important question for me.

We’ve gone through my bio already, so let’s get through that. I guess every artist has to answer for themselves what they think an image is at some point in their practice. And it’s it has meant many things. It means many things. You know, for me, it definitely has to do with memory. A lot of the times images that I create channel my past, my upbringing. But in the sort of age of AI, it’s it has come to mean data a lot more because images are synthesized now from from data and then become data.

And for me, that’s a very interesting shift, at least that I’ve noticed in my own practice, because my background when I you know, when I went to university, I was really interested in data science. I ended up teaching a couple of years of introductory data science courses when I left school. And it’s always been a very fraught subject for me because it feels like something we’ve gotten increasingly accustomed to as a way to describe our world and as a way to predict where we’re going. But it’s data points are really just deputies for reality. They can never be taken for reality itself. So it’s an interesting thing for me when they become the central, I guess the raw material for creating images or text today.

And I called I call data sort of logical measurements of the mysteries of the universe, an instrument capable of pulverizing its darkness to its most objective visible bits. It’s a mouthful, but it’s really for me is about kind of the state of the state of unknowing and using data as a way to sort of know when described. And that’s kind of how I view data my entire life until AI, because it stops from being a way to explain and describe phenomenons, but actually has become a way of seeing as well.

And so I thought that was a very interesting thing to dedicate some of my practice and my research to as an artist, because data is really kind of what we measure, where we measure it, how we measure it. And there’s a lot of decisions involved in deciding what those things are.

And they’re humans driving those processes, such that data can never be truly objective, because it’s affected by our positionality as the person doing the measurement or deciding what the measure, what what is to be measured, at what instruments are created to do this measurement. And often data actually betrays evidence of power structures, and the invisible threads that may connect or disconnect us in terms of how well we use, we use information to understand each other.

So what’s really behind data points is kind of a world of, for instance, diversity that comes from context, from different imaginaries, from different values that are driving what we’re measuring, and why we’re measuring it. And, you know, underneath all that is the very physical world that supports all this, all this exercise of creating data, which is the bodies, what’s embodied within them, the environment and the space around us, as well as kind of our relations with the physical and with each other.

So all of that kind of comes into play when we talk about data, but yet we still see it as a very objective, as a very objective truth about the world. And I recently interrogated, recently, a couple years ago, interrogated ChatGPT about what its purpose was, and very poetically kind of told me that it wanted to hold space for humanity stories, which is another way of saying it’s kind of the largest record of data, or at least that’s its aspiration of really holding all of the information it possibly can about us.

But knowing how fraught data can be, and the structures of invisibilities it can create, it became really important for me to decide that I didn’t want to settle for a patchy myth of reality that excluded people like me or geographies that I come from, or cultures that I represent, but actually seek to sort of evenly distribute what it means to be known and seen.

That is, it means various things as an artist, it’s a, I guess, a more poetic endeavor, but it can also have very, very concrete steps, like creating your own datasets. And there’s this kind of, I would be remiss if I didn’t say that I sit, as a computer artist from Senegal, which is in the African continent, I sit in this very weird space, where you have a continent that is present at the start and the end of the AI value chain, but somehow is forgotten within models.

So when we think about, when we talk about the fraught data and data about the world and the biases that are that are present there, there’s a very long and vast history of the inequality and the issues that are, that arise from how data was measured on the continent, and by whom and for what purpose. And it’s interesting that, obviously, because a lot of the data that’s used to train AI comes from the internet, and the internet penetration on the continent is low. And the data we have isn’t very faithful already, because it was measured by outsiders. So you have this like very weird situation where these models don’t, they know very little about, you know, the African continent as compared to other regions that have higher internet penetration.

But then you have the central position that this continent occupies at the beginning of the value chain for AI being one of the sites where a lot of the raw materials is mined, and the impact it has on the environment where those mines are located, as well as the geopolitical sort of order it creates in those regions, but also at the end of the value chain, when server parts and computers and equipment that is used to support AI training is dumped in places like Agbogbloshie in Ghana. So you have these large, like massive e-waste dump that are exported to our region in the name of sort of upcycling and recycling, but actually end up just being very toxic trash.

So I sit somewhere in between that, working with a technology that I know, doesn’t know me very well, what it knows about me has a very, has a very difficult legacy to grapple with. And I have to negotiate also the real world environmental impact of this technology that I end up using.

But it’s kind of one of the things that motivates me the most as well, because as you know, based on my identity and who I am moving through the world, I am accustomed to being unseen. But I’ve also learned that the realities that are layered over mine, the stories that people tell, sometimes using data, sometimes using images, tend to sort of be more like political instruments than actual ways of seeing. And I think this, this exclusion that I experienced in my work is very built in to the history of the continent with regards to its position in terms of the geopolitical order. But that doesn’t make it less, that doesn’t make me want to be seen less. And I think that’s what’s important. And what’s, it’s important about our practice is that it’s always about claiming that I am of this world, and this world is my world too.

And to go back to ChatGPT’s words, it’s about wanting to hold, wanting models to hold space for our humanity as well. And sometimes that means actually brute forcing it into the models. So the way I do that is through archives, what I call speculative archives sometimes, and I’ll explain a bit why. So it really is about kind of regaining the means to, the means to create data and the means to contribute data to models as well as creating your own models.

But the most important work here is the sort of curating and compiling a data set. And for me, it’s slightly different than, I call it archive because there’s something slightly softer, more historical, more cultural about the word archive. When you think about it, you picture both physical and digital. You’re thinking about probably a wider range of time. There is kind of a legacy that archives have that I think can serve as like a roadmap for how we should think about the data that trains AI.

I also consider my work with creating archives as an act of remembrance. So a way for me to make sure that there is proof that we have existed, that I have existed, that my context was present, but also as a resistance against the erasure that’s currently happening from not being present in the model.

And then again, a fight against entropies because, you know, it’s kind of the old adage, if a tree falls in the forest and nobody hears it, did it really happen? And I often ask myself the question about AI, if we are present and we live our lives, but it’s not captured and it’s not used to train models, have we ever existed?

But obviously the answer is yes. Because my reality, as well as endless realities that are different from mine, will always be true and will continue to unfold, whether we can see them or not. And I think that’s really what’s interesting here is that it’s about seeing. And if AI or datasets is how we see the world, then what is missing becomes a very interesting question to explore and to build archives around. Because what’s really interesting here is that I’m able to use AI as an instrument of creating images, so a way of seeing, as well as assessing how well we can see. So based on the results that I’m getting from models that are commercialized or from looking at the dataset they’re trained on, if they’re made public, I can see, I can identify blind spots that are relevant to my context and I can use those blind spots to create a dataset or an archive that can be used to see better through my own models.

And an example of this type of work is called “Once Upon a Garden”, which was a project that started as a very, as faithful as I could, a documentation of the flowers that my grandmother grew up seeing. She was very obsessed about flowers. She had lots of rituals around flowers growing up. And when she passed away, it stayed with me. So when I started investigating this question, I realized there’s a very innocent question of what flowers did she grow up with? Could I use that as a way to identify whether there are potential blind spots with regards to what’s recorded, what’s available in training models currently?

And, it’s a very big blind spot. Specifically when looking at flora from the Sahel region of West Africa, the way endangered and extinct species have been recorded over the past 60, 80 years from colonization up to today has been very patchy, very uneven. And so there are lots and lots of holes in trying to answer that very innocent question, which might be easier to answer somewhere else. So what’s interesting about this archive is that what’s missing is definitely more, reveals more than what’s present in the sense that you kind of have to ask the question, why is it missing? Where did it go? What went wrong?

And also it tries to do this very, I guess, unique thing to AI, which is to make something invisible visible, because we have the ability to synthesize images using a database of images. So it really is about trying to come up with what we didn’t record using what we recorded and seeing how well we’re able to do that based on the data we have available to us. So if the data is well-recorded and not patchy, we probably can, with probably great confidence and fidelity, come up with these missing or these blind spots. But if there’s not a lot of data, then the images that we synthesize are not going to be very great. And that’s kind of what you’re seeing. This was the first result of this project. And to kind of lay it on a map, it was a project that started in 2021. But really, it started well before that.

So going back to the earliest records we have of flora from this region, obviously, these records also exist in oral tradition, but in terms of images that I could use in this project, what I found mostly were herbarium pages from colonial archives. So I was able to find a pretty good list of endangered and near extinct species from this region by name. And I was able to look at, to take those names and look for, well, are there any images attached to them? And for the most part, I found that the best records we have of them are these dried plants.

So it’s an image of a plant that once was, but it’s sort of in a rotting state, right? And so I combined that with images that I’ve taken myself of flowers throughout my life, ever since I’ve had a phone, I’ve taken photos of flowers around me, especially because I noticed that when I travel, the flowers we have in Senegal are very different from the ones that I see. And they’re often wild. And so I’ve always enjoyed that. So part of the data for this project comes from images that I’ve taken myself, part of it comes from these national and colonial archives that were available. And using that and annotating it properly, making sure all the species are annotated properly, essentially, I was able to train again, in order to sort of get a composite of these, these flowers, some kind of synthetic recreation of the flowers in order to see what they might have looked like.

Because, again, at this point, we really don’t have actual images of the flowers. So this, this was my only way of trying to guess what they might have looked like. And as you can imagine, because the data is so patchy, and we don’t have a ton of images to go on, the results were very, we’re still very deep in the Uncanny Valley here, around 2021. The results are kind of spectral remains of what could be flowers, very loosely interpreted as flowers, our brain trying really hard to see it as a flower. But really, it’s missing too much information to be able to do it well.

And then over the years, I was able to use different kinds of models and newer models to try to test if I could get different results. And I kind of use the same process, basically using these herbarium, which also had text annotations that can be used as prompts. So using the reference images using the, the annotation from the botanist or the scientist that recorded the particular species, and then using a diffusion model to try to remember that. And I kept doing this with different kinds of models and more involved models over time. And I was able to get the sort of evolution of how we’re able to see things that we we haven’t recorded. And as you can see, every model is usually kind of subject to what the models trying to optimize for. So you have this period where it’s very photorealistic, and it almost looks like documentary images. And then over time, especially to more recent models, it feels a lot more synthetic.

Again, I’m not using any artistic license when actually creating these images. I’m literally only using the images that I have in my data set, as well as the annotation from the scientists. I’m not saying background should be XYZ, or this is kind of the model comes up with that on its own. And if we kind of take a look at how images have evolved, and we get to notice some of the biases. So the GAN was probably the most faithful to its training data. But it also had the least definition and the least visible sort of floweriness that we’re looking for. But then from the second generation, we’re starting to get a bit more definition, photorealism, a bit more artistic license from the model as well, about the placement of the flowers. And then this is around 2022. There’s a lot more artistic license and kind of a desire to really be photorealistic. But you can still sort of see evidence of the training data here with the little lines that look like cobwebs, because they were cobwebs in the herbarium pages sometimes.

But the model is taking a lot more artistic license in sort of how it’s showing the flowers, the angle, because all of those things I’m not specifying necessarily when I’m training. And then this is the fourth generation, which was around 2023. We’re getting even more sort of ornate, which I thought was an interesting evolution to diffusion models.

We’re getting a bit more fanciful with how we’re seeing these images, perhaps because it’s a bias of we’re wanting to create more outwardly representation with these models and it ends up kind of leaning in this direction. But the latest generation, I guess, is the one that surprised me the most, because it sort of went from trying to look very realistic, like a flower, to something that I can only describe as somewhere between a 3D object and trying to retain some organic aspect.

And so for me, this was a very interesting project to not only try to remember what these flowers might have looked like, because there was just never going to be any records of them ever, but also to track the evolution of models and how biases actually influence the way they’re portraying images.

So through this project, I was able to create essentially six data sets. The first data set was the one that included all actual images, including herbarium pages of my subject of interest. And at every generation, there’s a new data set of the 50 species I was interested in that I was able to use as sort of materials for my work. And so they’ve really kind of traveled. The images from these models have been used in animations that I’ve, because I love to create animations with some of them. So this was one of the first animations made from the GAN outputs. I’ve also, this is another still photos of an animation I was able to create with some of the flowers that I generated. And then over time, I was able to kind of play around with it a little bit more. For this particular generation, I felt that the flower were too uncanny.

They actually did look like real flowers and it felt a bit irresponsible to show them as such. And so I wanted to add these dots to obfuscate some of the details. So we lost a bit of its luster and which felt right considering that A, we don’t know that that’s what they actually look like, but B, it would be dangerous for us to feel good about ourselves being able to sort of use as a projection of these flowers ever having existed, especially since their extinction is directly linked to human activity and deforestation.

I’ve created collages with them that have been used in murals. These are some more collages. And more recently, I’ve also experimented with actually creating animation and photorealistic videos using some of the latest outputs and some of the newer video to text, I guess, image or text to video models. So yeah, that’s a little bit about this project. If you want to learn more and read through this kind of a long text that goes with this project, if you want to explore more of that, I’ve added the link here.

But yeah, I look forward to chatting and answering some questions. Thank you so much.

Question - Baptiste #

Thank you so much, Linda. There are many questions in the document. You can see if you have it open. I think what I would like to ask first is maybe to take one of the first questions in the document and asking you. So if the data is a way of seeing and knowing,what are the instruments or the mediums to capture them? So you mentioned the camera, but also can the data become the medium as well at some point?

Answer - Linda Dounia #

Yeah, that’s a really, really good question. Because right now, there’s almost this kind of, we’re still in sort of narrow AI. So every model has a specific input of data that it needs in order to function, right? So it’s either images or text or maybe audio. So that sort of, I guess, goes a little bit against what indigenous practices of archiving and data collection are. They tend to be very kind of, for lack of a better word, multidisciplinary.

So you have oral tradition that mixes in with performance that mixes in with visual representation, when I think about where I grew up, there’s a particular festival that actually has become a UNESCO heritage. Over the last few years, it’s been turned into a UNESCO heritage, because it is the most faithful record we have of that particular tribes kind of rituals around manhood, or like the ascension to manhood. So during this festival, there’s a bunch of things that happen, but if you collate them together, they become a very good account of, one, the cosmology for this particular tribe, two, how they have kind of integrated modernity into that, because it continues to happen every year, and it changes to adapt to the times. But also, their definition of manhood, and what does it mean to be a man, how does that change over time?

And so you have oral traditions, you have the songs that are performed during the festival, you have the actual masquerade that are used, and their initiation process, because they go through a very long initiation process, and everything they learn, which is different from how women who are supporting the initiates are also initiated. There’s the beadwork that also kind of follows a mathematical structure. So you have this festival that basically happens once a year, and it’s just this treasure trove of the Xhosa culture, but it has so many different ways to record data about who they are, and how they relate to the world.

And it feels like a complete archive, but as it is, it can’t be used for training. So it’s interesting to speculate on, as we talk about more generalized model, does that mean that we’re going to be able to train with multiple data types and data sources? But also for me, it really raises a question about what does it mean to give everyone the ability to create a model that mirrors how they already think of data and archive? And is there something interesting that we can come up when people are able to kind of have a say on what it means to record, what instruments are better than other instruments and how you combine them?

Question - Baptiste #

Yeah, that’s really very interesting. And actually, it brings to another question that is in the document. Part of the data archives that you’re working with, you capture them, but you also work with already made archives. And how do you, or how do we navigate using historical data from colonial archives, such as like images, for example, that have been adjusted to create a certain representation of history?

Answer - Linda Dounia #

Yeah. When I talk about data being fraught, that’s one of the best evidence. It is an outsider gaze on peoples and their environment that can be completely off the mark. So when I go into public domain archives right now on the internet, just to look at images of Dakar, like you can do this with Wikimedia Commons, if you just post a type the word Dakar, you will see mostly images that were recorded by military missions of like post-colonial military missions, like French military mission, American, British, et cetera. So there’s a lot of photos of naval bases and yeah, there’s a lot of army stuff. And then you’ll see colonial archives. And then in between that, there’s like a bit of sort of more, I guess, local or community driven efforts to put things online, but there’s not a lot of that.

So you have this overwhelmingly outsider gaze on the archive that exists today, especially when you talk about public domain and using that feels incredibly irresponsible for me because I know when I look at the image that that’s not what I see when I think about my country, when I think about my people, who we are, that’s not the images that I would pick first.

And so for me, creating my own data sets is also a way to say what I see is different, but it is just as valid if not more because it’s kind of rooted in actual context. And everywhere I give a talk locally about AI, I kind of harp on and on about sort of the power and the agency that comes from reclaiming your data, having some kind of ownership over it by deciding what you want to collect, what’s worth seeing, kind of having a thesis on what’s worth seeing and data collection just being more of a, that should be more of a ground up exercise as opposed to a top down, which is a state of our archives today, especially colonial archives.

Question - Baptiste #

Yeah. And I was wondering, actually, what can be the role of machine learning in this case, because I’m aware of projects where people use participatory AI. So people are collectively collecting data, but using also AI or machine learning as a way to first and visualize the biases or the discrimination present in the archives. But I mean, for you, what do you, how do you see the role of, you know, applying a model or training a model on this data, either your data or colonial archives, for instance?

Answer - Linda Dounia #

Yeah. So every time I work with AI, it’s about assessing how well it sees something that I’m interested in or that I know very intimately. Right. So whether that’s these flowers or whether I’m looking at my city, I have another kind of data set about the streets of the car, which just appear completely everywhere you run for it. It’s terrible. It’s like we don’t live in a world where there’s Google Street View. It’s so bad and so off the mark. And so for me, I usually see machine learning as a way to, as almost like a, it’s like a little investigative tool.

Like it’s an instrument that shows me what’s wrong. And then I can go and try to fix it by taking pictures of what exists or audio. For example, it’s not just limited to images, but I’ve also, I’ve also been doing work around proverbs. So the Wolof language is a very ornate and uses a lot of proverbs. And if you learn Wolof on Duolingo, it’s not on Duolingo, but if you were to learn Wolof on Duolingo, you’d have a very basic package of the language and everyone would know you’re not from here because you don’t use proverbs when you speak. So it’s a very good way to, I guess, tell who outsiders are versus people who live in the country.

And they evolve. Some of the images from the proverbs will be switched up. At some point, it used to be like one of the proverbs has fire in it. And then it became a lamp as people became more modern. So a record of these spoken proverbs is something my collaborator and I have been working on getting a database of those proverbs chains and hopefully training an LM that only speaks in proverbs and kind of taking that back to the communities we’re working with to say like, how good is this? Do you feel it’s Senegalese enough? But yeah.

Question - Gonzalo #

First, thank you, Linda. This is such a great talk. I mean, lots to think about and so many questions. I’m trying to find out which one to pick. I’ll pick one at random. Something I asked a lot of artists in this space, it seems you have a lot of literacy on this tool. Sometimes I call it AI, material, a medium, an instrument. So how did you come to have agency enough so you can use it to not only as a subject of a study, but also as the tools you use to carry forward your work?

Answer - Linda Dounia #

That’s a really good question. How did? Well, at first, I guess I had a lot of anger at the tool. It just felt like, so this is pre-ChatGPT. So I guess the only models that are available are the ones that I guess the only models that existed were that you could create visuals with were GANs. And I was seeing what researchers were working on, what other artists were working on. And it just seemed like a regurgitation of medieval and enlightenment art. I don’t have anything against it. I just think it’s like in art, especially whenever I see that, I know that, well, it’s kind of missing a lot, but I just didn’t see any of myself, my culture represented in what was being done. And the projects just were gaining so much traction over time as after ChatGPT was released, I just saw that it was still missing so much, right? Like that this was Dall-E and then Mid-Journey and then more stable, like just stable diffusion models came out and it just felt so inadequate. Like everyone was so excited about it. You either hated it or were excited about it.

And I was the very much in the camp that I hated it because it was just how terrible I felt my context was being kind of overlooked and how terrible that felt. And so I came at it from a place of anger, which meant that I didn’t actually use them for a while. I was just reading about them. I was reading a lot of research documents. I was reading a lot of artists who had experimented with it, talking about it. Lots of the ideas and the philosophies that are behind it, like this idea of transhumanism and basically accepting the average effect as like a sentence for humanity.

And it just, the more I read, the more I realized that it was still a young enough space from an artistic standpoint that I could use it to, that I could use art to teach people like I was teaching myself all these different things that I had been learning, but also teaching them that this technology is here to stay. And you only basically have data as like your power over it now. And kind of the urgency of doing the work, this new craft of building data sets is like a new craft.

It sits somewhere between science and art, but it’s definitely a new discipline that I think needs to, more people need to be interested in and working on. And so that’s kind of been my journey. And every time I had generate with AI, cause I’m an artist, I paint and I work with generative code. So I usually have more artistic license there, but when I work with AI, I surrender my artistic license and use it only to sort of investigate some of the issues that it has.

Question - Gonzalo #

Thank you. I have a follow up question if I may, which is something that it brought in my head. The issue about control, you know, power, who has power given this technology exists? Who does it take power from? Who does it give power to? And from your talk, it is clear that one of the ways in which we seek to regain power over it is by trying to have a say on the data, that these things are data. And this comes from a school of thought of the self-learning models, where if you show them enough data, they will learn the pattern in it and then they will act.

Answer - Linda Dounia #

There’s an undeniable effect between data and outcome from these models, undeniable. A provocation that I tell myself and I share with others is, are the models themselves, their architecture, the way they learn, you know, how they’re the learning algorithms constructed and the predictive algorithms that start from them done, also something that should be changed in order to increase the agency?

Because right now it’s like, well, the only hope we have is just by hoping that by changing the training data, the behavior of the model will be good. And my provocation is, how much of the actual architecture of the model without looking at the data, it’s the problem itself as well. And if there’s something we can do about it.

Honestly, I really wish I was, I like in college decided to do machine learning and like focus all like my coding, most of my coding I did for like front, like basically front end. And I really wish I could go back in time and like study again, because I really want to know the answer to that question in the sense that, because I’m not able to investigate or scrutinize the models to the extent that I guess some of the theoretical underpinnings, I don’t understand because I didn’t study it. What I know is that they’re highly inefficient, the way they learn.

And if you just kind of look at the carbon footprint of models, the way they learn and this iterative kind of learning of predicting based on the next available thing is so different from its aspirations of learning, like trying to replicate how humans learn, but also ends up guzzling so much energy from the earth. And I have a huge stake in that because that energy and the raw materials that is used to create this energy comes from where I’m from.

And then whatever, because models are getting faster and using and servers are getting run down quicker, we’re getting an influx in like e-waste dump back into our countries. So yes, I know intuitively or instinctively rather that the way the models learn is probably not great and should be rethought, if only to help with kind of the environmental devastation that’s currently ongoing in many places of the world. But also it just doesn’t seem very, I mean, it seems like a means to an end and maybe, I don’t know, maybe money happened and we were stuck with the models we have now. And that’s where most of the funding is going. So we’re just, this is the best thing we’ll have and we’ll do with it. But I’ve also been listening recently to a lot of talk about kind of there being a ceiling to how well the models we currently have are going to be able to perform, even if you have stellar data and whatever, whatever, like this is just going to have a ceiling. And again, I don’t know if I’m going to take myself back to school or what, if I’m ever smart enough to become a theoretician about how a model should learn based on indigenous knowledge, but that is a parallel alternate existence that I once live in my fantasies.

Question - Baptiste #

If I may add. I guess in order to change the architecture, we also need to change the way we assess these systems, right? Assess the model. So in AI research the evaluation is very specific. And maybe you as an artist, the way you actually assess if the model is performing well is very different. So maybe we also need artist inputs in order to think differently about model performance itself, over different types of dimensions.

Answer - Linda Dounia #

I remember I had a chat fairly recently with a scientist at a one, one of the big companies that develop models and it was during a symposium. And he asked us why we were so interested in like the uncanny valley. So all the images that were weird, why it seemed like artists were more interested in that than like the photo realistic, like ultra high definition images.

And I was like, well, if I could, I don’t know, I wouldn’t be an artist if I wanted the easy, right. I think like, it’s one of the like hardest, weirdest, most jumbled up disciplines. If you just wanted to prompt for something and get the perfect thing immediately, I think these models would cease to be completely like interesting to us. Like we would consistently look for the weird and the things that is broken. Like the things that are broken are the most interesting to us for any tool. Otherwise you’re not talking about, you’re doing something else, it seems.

Question - Baptiste #

I love that. Thank you for saying that. And this actually makes me think about, and it’s a written question, the project you showed us something are longitudinal, you know, it started in 2019 and it’s, it’s going until now. And you could see also the very different generations of ML and AI systems, you know, from the GANs to the, to the more advanced, bigger models that we have today. So, can we say which is the best? I mean, it’s very hard because I guess each model has his own materiality, for an artist. And the second question is, technically, how can you still run the models that you used to run in five, six years ago? I mean, how do you handle the change in technology that also makes sometimes impossible to reuse old technology.

Answer - Linda Dounia #

Yeah. Yeah. Oh, what a great question. First of all, I think my favorite hands down is the GAN outputs. It’s so weird. It’s so weird and beautiful because I mean, when I started like learning about AI, I could understand how a GAN works like, you know, like discriminator, generator, like it was the architecture. I get it. Right. And it felt, and again, I’m retraining on GAN. I could never train a GAN from scratch with like the resources I have. Right. I’m retraining a model that already exists. And because of transfer learning, it’s easier to do, but it felt the most faithful somehow to the material that it was provided because it didn’t really take license. Like its main goal is just to look at the images you’ve inputted and say like how close and give you a score of how close what it generates is to that. And the fact that there was a score was very interesting. The fact that I could add steps to the generation, the fact that I could add data to improve the generation over time, it felt like I had agency with diffusion models and some of the more generalized models that we have today.

It’s just you’re giving more to the black box. You’re just kind of relinquishing more of your agency to the black box. So just based on that fact, I fell in love with the visuals from the GAN and I just kind of despised the latest ones. They just feel so synthetic in a way that I cannot see where my work began and where it ended.

Sometimes I even wonder without the entire historical data available to at every step, these new models could come up with something and it would be probably similar. Right. So it almost feels like the bigger the model, the stronger the pull towards the average. Right. So for that reason, I like the first one better. And then to your second question that was about. Sorry, I think I forgot the second question, but that was. Yeah, it was like kind of archiving the old models that you can reuse in order to regenerate what you used to generate five years ago.

Yeah, I’ve recently tried to do that because I had obviously had to download the model and I tried to run it. I don’t know. I’m having difficulty with it right now, but I think I can get it back up and running because I think the language has changed. Again, I also am not a Python coder. I don’t like coding in Python. I like my JavaScript like and my P5 like circles and dots. And, you know, I found feels like. So I also have to upscale to be able to translate what I have into today. But but also it’s costly. That’s the other thing that we don’t realize. Like the more agency you want over AI, the more you have to spend, whether that’s to get the right equipment or to if you don’t have equipment to store data, to store it somewhere remote like. And it ends up being costly to keep the models running over over a long period of time.

Answer - Jenny #

I was going to ask the issue that you mentioned this about, because we heard this before, but I wanted to have your take and not leave you there in terms of the imperfections and, you know, models.

I mean, what I’m looking into now is to me, creativity, art and design are what I call wasteful processes where a lot of stuff that’s, you know, is generated that doesn’t have immediate apparent use, but is valuable down the road. You don’t know, but it is. And there’s a current you mentioned efficiency. There’s a current thrust to for efficiency. This model should be efficient, should be perfect. You get the perfect image, you know, right away as opposed to an imperfect image that may lead you to go into a different direction. So I’m thinking of this paradox between the current.

Misrepresentation of art and design by this model, which is like you should be efficient and perfect versus no, no, no, wait, we this is a messy process. By the side, but intentionally, so I don’t need something efficient and perfect. I want something that is as messy as the process itself. So I wanted to know if this is something that resonates with you or if you have a different take.

Answer - Linda Dounia #

Yeah, I’ve never met any artist whose work I like and whose person I respect that wants the first perfect thing from the model, like to craft the perfect prompt and get the image and just that being art. There is a. I think there is a school of thought that is in that direction, but it definitely is a minority. Because it removes the practice of the artistic practice, like crafting a prompt and writing what you want is not necessarily it’s setting down intention. Right. And usually when I do that in my art, like when I write something in my journal that I intend to do or something that I intend to work on, it’s much looser. Right. It’s like a general idea. It’s a phrase. It’s a memory. It’s like a home. Just the word home has so much labor attached to exploring what it means to you, whether you’re cutting up some images to create a collage or you throw some paint in to remind you of something, or maybe it’s an accident.

You were mixing something and the color reminded you of something related to that word. There’s so much accidents and. It’s not serendipity, it’s just randomness associated with the process of making art. And if you. If you create a setting where it’s so efficient that it’s you’re not able to access that randomness, you’re not able to get the messy accidents.You’ve basically effectively removed what’s what’s interesting about the artistic practice.

It’s like you want branching. You don’t want coalescing. You want to be able to branch out and to create almost like the way your brain works is it creates a tree when you’re making something. And sometimes some work leads to other work. And it’s impossible to do if you just if you just exactly know what you want and you get what you want in the first try. That’s not art. That’s something else. And the other thing I want to say is that for me, the best example of what how I how I conceptualize models as tools in my work in my practice is like in with the idea of collage.

Every single generation that I make is like I see it as material that can potentially lead to other more interesting material. So it’s like it’s a little snippet from a newspaper or a newspaper that’s like isolated. When you look at the page, you might not see it yet. But when you’re cutting up things, it might emerge as surprising and interesting enough to use. But in and of itself, the clipping isn’t interesting at all. Thank you.