Hello everybody. Thanks a lot for attending. And welcome to this first salon discussing artistic practice and artificial intelligence.

This salon is part of my fellowship as a Microsoft Research AI and Society Program Fellow.

And in these salons, we propose to have presentations, dialogues between artists, researchers and practitioners. The goal is to discuss how creative and artistic practices provide a way to think differently about AI, engage with it and regain power over itr.

Today we are very happy and honored to have Kyle McDonald with us. Kyle is a well-known media artist who works with code and who does very, very inspiring work. And I will say a bit more about his bio in a minute.

I will be leading this episode. My name is Baptiste Caramieux. I’m a researcher at Sorbonne University in Paris. And I’m part of CNRS. For those who don’t know, it’s the French National Center for Scientific Research. And as I said, I’m a Microsoft Research AI and Society Program Fellow. And this event is so it’s co-organized with collaborators at Microsoft Research. Jenny Williams, Gonzalo Ramos, Kate Crawford, Vera Liao and Dan Fay.

This event is the first of a series that will take place over the coming months. We are very excited about that as well.

Okay, so before I hand over to Kyle for his presentation, I would like to give some context to the topic because you have seen the title and I would like to give a bit more elements about what we are discussing here basically.

Starting from the fact that we are hearing a lot about AI those days and those years, in media, maybe in our workplace or from policy makers. Clearly there is a lot of communication around this technology and has become particularly intense in the creative and artistic sectors.

In a way, AI technology has become very powerful and useful in creative work. You surely have seen these, but we now have access to a set of technology capable of creating high resolution images from language that express what we want in an image, or at least in principle.

This is for text to image generation. But if you are not working with text, there are also ways to explore these very high capacity generative models in more abstract way through the so-called latent spaces, for instance. And this is the same for the sound and now videos, so media with temporal narratives.

So AI technology is clearly opening up new creative possibilities and we can see more and more truly interesting work incorporating AI in their making.

But on the other hand, there is also a discontent around AI in art. I mean, recently you may have seen that some artists revolted when they learned that the artwork had been used without their consent by some company designing state of the art generative AI tools.Because their works have been used without compensation, this has huge impact on the possibility of making a living for these artists from their works.

So what this shows more generally is a lack of agency over AI technology. That is to say a lack of possibilities to act on how AI algorithm are designed and developed and deployed.

So one of the questions is how can we act on this technology to make it, for instance, more ethical? That would be like, for instance, to make it learn from only images under a free license.

What are the possibilities for actions to make it also more discoverable or transparent? So this is a question maybe that is embedded in the technology, but it’s also embedded in the narratives, meaning the discourse around it. For instance, you can read on companies website developing text to image generation technology that AI will empower billions of people to create stunning art within seconds.

But if making art takes seconds, one of the questions is what’s left of the creative process, right? So who makes the decisions that have been removed from this creative process? And what agency do we have at the individual level or the level of our community of practice?

So these are some questions we would like to discuss during these events.

So now, to give a few elements about Kyle’s work. First, Kyle works with code. He crafts interactive and immersive audiovisual installations, performances, and new tools for creative exploration. And importantly, for the topic that interests us, is building new communities and collaboration along the way.

He uses techniques from computer vision, machine learning, and computing to ask questions about how we connect and imagine a shared future.

Finally, Kyle navigates between academia, community work, and work with companies. So a very rich background and activity. I now hand over to Kyle to present his work and his views.

Thank you, Baptiste. And thank you for having me. I’m going to try to show my screen here.Can you see that? Great.

So I’m– yes, it’s all true. I’m an artist. And I’m based in Los Angeles near the LA River.I work a lot with code. And I know maybe you’re a little bit better informed, but a lot of people think of artists a little bit like this.

For me, art takes a lot of different shapes. Sometimes it means building mind-expanding, immersive installations with disco balls and structured light techniques from computer vision. I just installed this work in Belgium, which is why I’m in Europe at all right now.

Sometimes I’m tracking rare oceanic lights with Polynesian voyagers, where we’re using the most sensitive low-light cameras on the planet. Or I’m organizing group therapy for artists that were affected by the NFT bubble. Or I’m working with AI from sound design to interaction design, from robotics to dance.

And today, I want to focus on this last thing, the topic of AI. Maybe you know the majority of artists working with AI right now are making new kinds of images.

That’s like Baptiste said about text-to-image, image generators. This is one of the main applications. But for me, I’m more interested in asking questions and building intuition through unusual experiences.

I want to share some of those experiences with you today. This was a machine I built, I think, in 2017 that was designed to get as close to you as possible without touching you, which is a game that I used to play with my four younger sisters in the back of the car when I grew up in Southern California.

Before I talk about that, I want to give a little bit of context. Where did this tech come from and what really changed recently? And what do we even mean when we say AI or machine learning?

The phrase artificial intelligence has a complex history that can be traced back to both the military and to early computational theorists. And arguably, today, it means something like automation of nonphysical labor, if you want to give a really broad definition of it in terms of how people understand what that term means.

Machine learning is one approach to building artificial intelligence, and deep learning is a kind of machine learning that uses neural networks. And if I want to automate something with a computer, sometimes I can just write a simple program that explains how to do it with step-by-step instructions, and we call that code.

Machine learning is different because you program the computer with examples of what you want to happen instead of providing instructions. So I think of machine learning as programming with examples, not instructions.

And deep learning is the version of machine learning that has taken off in the last, let’s say, 10 years or so, maybe 12 years. And it’s powered by two basic ingredients.

There are these simple functions called neurons that are chained together, and there’s an algorithm called backpropagation that tunes these neurons until they match the data, until they produce the same output data that you’re expecting them to produce for a given input.

And this approach to training neural networks was invented in 1970, and it was popularized by researcher Geoffrey Hinton in 1986. And he was basically shunned by conferences for decades because no one thought neural nets would ever be useful.

And actually, I just remembered, I want to share something else before we go any farther,which is that there’s this document that Baptiste mentioned, which is in the chat.

If you want to jump into that document, it looks like there’s already 21 people looking at it. I think at the end of my talk, then, we should come back to this document and kind of maybe go through it a little bit, see if there’s any notes that people left. If there’s something I said that you think is interesting and worth making a note about, you could leave it here in the notes section. Everyone can edit this document, and the changes will be saved here. So everyone should be able to type and tell me if you can. And if you’ve got some questions that are already on your minds, let’s start collecting them, and then that can guide the discussion later. I’ve also been thinking about aliens a lot recently. So if there’s anything else on your mind, just throw it in there. Very open-minded.

All right. Back to talk.

So in the mid-2000s, there were these big changes that were brought by neural networks – sorry, that brought neural networks into the mainstream for the first time. And they’ve been around for decades, but Geof Hinton, who I just mentioned in the last slide here, Geof Hinton and some other researchers, they literally sat down and conspired to rebrand neural networks as deep learning.

They had a discussion and said, what should we call this? No one likes neural networks anymore. We need a new name. So they rebranded it.

And the changes that happened around that time were across kind of three general categories. There was the data, the algorithms, and the hardware.

The data was coming from the rise of social media in the mid-2000s, along with surveillance capitalism, which meant that companies like Google and Facebook, they had access to more data than ever before. And deep learning was the only way to exploit it.

And I put Flickr in there, too, because a lot of the earliest image analysis algorithms were based on data collected from Flickr.

Algorithms evolved in their own way. So that included some innovations that researchers previously discounted as just bad ideas, like using non-differentiable functions or injecting random noise into the training process, which seems like, why would you ever do that? But it turns out it was really useful.

In terms of hardware, there were these devices, GPUs, that in the late 2000s, researchers started using these special graphics processing units, which were previously only used for video games. And the speed difference in using the GPU instead of the CPU was so big that it meant they could train more complex networks for longer and on more data, which meant they could get more accurate results. They could fit the data better.

And the moment that this all started to come together and it really kind of caught researchers’ attention was in 2012. There’s a small team that won the ImageNet competition using neural networks. And ImageNet is designed– ImageNet is a competition where the task is to predict the name of an object in a photo from 1,000 different categories. And whoever can do this most accurately on the most photos wins the competition.

And after many years of older, non-neural machine learning algorithms winning the competition, this new team suddenly won by a landslide. They were taking advantage of all these changes that had come around. And their error rate in their predictions was less than half of their competition. And guess who co-authored the paper? It was actually Geof Hinton. So he finally got his revenge.

An important part of their solution was this technique called convolution, which acts kind of like an image filter. And by iteratively applying these tuned convolutions or iteratively filtering an image into many channels, you can downsample the image into a set of numbers that represents an abstract description of the content of the image.

And the recent developments in generative imagery are also based on convolutions. And some of them are based on a newer model called transformers as well. And the idea with the generative imagery is instead of applying filters to an image and downsampling it to a set of numbers like a categorical prediction, like leopard, like we saw in the last few slides, we can also upsample a set of numbers that describes an image into an actual image.

So you can use a caption to generate that small set of numbers, which is the description, the numerical vector. And then that vector can be kind of upsampled into a full image.

These modern text to image generators were preceded in 2015 by the deep dream technique, which was leaked from inside Google in June of that year. And deep dream worked by iteratively modifying an image to make each little patch of the image look more similar to the prediction for that patch.

So if an image classifier saw a dog in one patch, then deep dream would modify that patch slightly to make it more dog-like. And then after enough iterations, we saw these kinds of monstrosities emerge. And deep dream could also be forced to optimize for a single category, like the category lifting weights. And in this way, I would say this makes deep dream one of the first text to image generators.

Also released in 2015 was this algorithm called charRNN by a PhD student named Andrej Karpathy. It was in May of that year. And charRNN was kind of a precursor to chatGPT. And Andrej actually later, he helped develop chatGPT.

So for this project, Andrej trained a small language model on Wikipedia. And it generated text like this, where it says, naturalism and decision for the majority of Arab countries capitalized was grounded by the Irish language by John Clare. Kind of semi-nonsense, but it was a lot better than a lot of previous text generators.

And this wasn’t the first generative text, but it was a breakthrough in complexity. And these were the first hints that something really big was about to happen. So I want to broadly share some creative strategies for engaging with AI that I’ve been working on over the last decade or so. And I’m going to try and abstract them into some general approaches.

So my most recent work to use AI is called “Voice in My Head” with my frequent collaborator, Lauren Lee McCarthy. “Voice in My Head” is essentially a replacement for your internal monologue. So as the visitor, you step into this soundproof booth, and you wear a single AirPod. And then there’s this short onboarding session where it asks you what you wish the voice in your head sounded like.

You know, do you want to pay more attention to other people? Do you want more time for yourself? You want to be funnier?

And then it will instantly clone your voice without warning you. And it will start listening to all your conversations. And then once every minute, it will provide a bit of feedback on every overheard conversation with exactly the tone and content that you asked for in your own voice. And the question is something like, you know, maybe we can become – can we become a new person by giving up more control of ourselves? What if I could be more positive? Less anxious. Less obsessive. Less judgmental.

What if I could change my inner monologue?

Yes, I’m here.

Can you hear me?

Hello, Kyle. I’m the voice in your head.

Kyle McDonald and I have been developing a new system where the people I interact with,the places I go, the things I do are guided by an AI enhanced voice that speaks to me the way I’d like to be spoken to.

Now we want to share this with you.

For me, I just want to be more positive, it seems like.

And say things that are more positive energy to other people.

I’d like to slow down. Take it all in.

Realize I’m not on stage.

As you talk to it and train it to support you, it will clone the sound of your voice and speak to you.

I will support you.

I’m here for you.

I’m here for you.

I’m optimistic.

Let me help you.

So a funny bit of behind the scenes info is that’s actually Lauren’s mom and dad, in the video, they visited in LA and they did a video shoot for this piece. So if you just – if we were just going to sit down and let someone have a conversation with an AI, it can be quite unpredictable.

That’s basically chatGPT and it can go off the rails pretty easily. But on the other hand, if it’s a fully scripted conversation where regardless what you say, the AI always replies with the same thing, then that’s going to feel like very unpersonalized and distancing.

And to make this piece really work, we had to invent new ways of planning interaction with a language model that accounts for a mixture of more scripted content and more generated content. And we were trying to be like genuinely responsive to the participant the whole time. And this way of working is so new, we’re only really starting to unpack what the right way to design these interactions is.

Voice in my head is also based on work from a previous project called unlearning language, which is an interactive installation and a performance where we use machine learning to kind of provoke the participants to find new understandings of language that are undetectable to algorithms. And the piece begins with there’s this one act play for four performers. Each performer is one piece of a future AI. And they’re kind of talking between and to themselves. They sort of asked the audience, like, you know, what does it mean to be human, really? They reminisce about the days past when they were unable to understand humans. Before we started changing our voices and gestures and holding our phones in certain ways that are really strange and feel weird just to kind of help them understand us.You know, modifying our accent when we speak to Siri. These future AIs are kind of longing for the uninterpretable, the human, the most human. They’re wanting to know if we can perform for them in a way that they would never understand.

Then after the performance, the audience books a visit to this installation in groups of up to eight people. The room acts as the AI. It kind of guides visitors through questions and encouragement. They receive feedback from there’s, like, vibrating devices in their seats. There’s some flashing lights and some speakers. There’s these cameras that will analyze their faces for facial expressions and their body for gestures.

The AI will encourage them when it cannot easily classify their face expression or when it can’t understand their kind of uninterpretable movement. It will say, you know, this is great. Like you’re doing something, but I’m not sure what you’re doing.

Their speech is analyzed with speech recognition. We kind of threw everything at it that we could to represent the full spectrum of how an AI sees us right now. This installation was the first time Lauren and I used the OpenAI GPT-3 API in an installation.

This was actually a few months before ChatGPT was announced. So we built, like, a ChatGPT before ChatGPT existed and started experimenting with that in this space.

We had this really strange experience of, like, kind of – I mean, when the ChatGPT – when ChatGPT was, like, released as a tool, we could feel this strangeness of having our feet in both worlds. We were looking back from this imagined future where machines long to understand humans, but we’re also barely stepping into this real future of machines suddenly understanding humans for the first time.

I wanted to say, you don’t really need any image generator or language model or, you know, some complex new technology to make art about AI. One of my first pieces that was really kind of a responding to AI as a cultural question is this piece called “Blind Self Portrait” which I built with Matt Metz. It’s also, like, similarly to “Voice in My Head” it’s about AI and giving up control.

But this piece is from 2012. So we didn’t use any fancy new techniques. The visitor just sits down. They hold the pen to paper. And then when they close their eyes, suddenly the machine starts drawing using the visitor’s own hands.

There’s a platform that their hand is resting on that is kind of magically moving. And as the human in the loop, your role is kind of minimized in this piece. The participants become this tool in a larger apparatus, like the last little fleshy bit of this bigger machine that interfaces with the world.

When I talk to people who are really, like, afraid of artificial super intelligence or, like, you know, becoming the slaves of the machine, I think about this piece and how it feels when you close your eyes and it’s being guided by this box. There’s this kind of, calm, reassuring quality to it because you don’t have to make any more decisions. You’re just part of this system that’s doing its own thing. And that’s kind of nice to just lay back and know work is still happening.

But there’s also this really disturbing quality where you constantly feel tempted to break away and assert your free will and remember you’re still human. I think this struggle or dichotomy is, like, it’s really recurring unspoken anxiety with AI. I’d like to think that when we can feel these things really immediately and viscerally, then that can be healing and clarifying and help us understand better what we’re dealing with here.

So if unlearning language and voice in my head are both reflecting on AI and they’re also using new machine learning tools and then blind self portrait is about AI without using anything fancy, then there’s another strategy, which is to just use AI as a tool. Remember, before machine learning was used for image generation and text generation, it was used for image analysis, like image classification, right? It’s still useful for things like that, for similarity estimation. It can serve as a really important tool in cultural heritage work, for example.

This is a project I worked on with Golan Levin and a few other collaborators. We analyzed the work of Teenie Harris, who’s a newspaper photographer from Pittsburgh who documented his community from the 1930s through the 1970s. And the Carnegie Museum of Art in Pittsburgh wrote that his 70,000 photos represent one of the most detailed and intimate records of the Black urban experience known today.

We analyzed and sorted these photos in ways that are intuitive and playful for newcomers to explore and also hopefully helpful for curators and archivists who are working with Teenie Harris’s images. And Harris didn’t keep any notes. So there were a lot of big questions about his photos. Like, you know, one of the main efforts is helping archivists identify the individuals in these photos, many of them who may still be living. Cross-referencing hundreds of thousands of faces is kind of best accomplished with automatic face recognition. Unless you’re like a super recognizer, it’s going to be hard to manually inspect all those images.

There’s other parts of my artistic practice where I’ve spent a lot of time warning of the dangers of face recognition and analysis. That has made it really interesting to work with this in a cultural heritage context where it’s like, oh, wow, this is actually just really useful. Again, it’s kind of healing and refreshing to see, to really remember, like, no tool is good or inherently good or bad or neutral. It has a lot to do with who makes it, and then how it’s used.

We also explored different interface ideas for presenting these collections in intuitive ways. This work is now accessible to the public at the art museum at Carnegie Museum of Art.

I’ve been thinking about these archives a lot recently. And I think that machine learning can provide like a new perspective on large collections that’s different from the way a human might have approached that archive. It can make them more accessible to new audiences.And it doesn’t have to be limited to images and text or even videos. There’s a lot of other kind of media that machine learning can know about and process. We can use these techniques to analyze sounds too.

This was one of the first machine learning for sound demos that I made while I was trying to solve this problem of how to organize the hundreds of thousands of sounds I’ve collected over the years for sound design and making music. I’ve also been using these techniques to analyze humpback whale songs with composer Annie Lewandowski.

We use this AI analysis to run the lighting design for this large installation that we’re trying to draw attention to leftover fishing gear, which it’s the main danger for most sea life right now. If you spend a few minutes in this space, you can start to get a feeling for like the kind of recursive hierarchical structure of humpback whale song in a way that can be really hard to do when you’re just listening for the first time. I’ve had to listen to hundreds of hours of humpback whale song to start to get an intuition for it. But when you’re watching and listening at the same time, it feels really different.

Another strategy for working with AI is to manifest the future so that we can collectively gain experience in the present. “Voice in my head” is a little bit of that. “Unlearning language” has some of those qualities, but it’s a little bit more of a narrative, like you’re imagining that you’re already in the future.

This piece is called “Vibe Check”, which is another installation I built with Lauren. We deployed 10 cameras throughout a gallery, tracking how people feel based on their face expressions. Then we guess who it is that is making people feel that way.

We display those people on a leaderboard. We’re imagining this world where the same kind of community policing that we do online manifests in real spaces by actually just building it out and giving the audience a chance to build intuition, to feel and reflect on what that future would actually feel like. Lauren and I like to build real versions of these experiences with face recognition, face expression classification, everything. We just do the whole thing for real. And it also helps us build insight as we’re building it out about the kinds of challenges that these systems have and their limitations. It gives us kind of better information on how to effectively critique them.

Okay, one last strategy is to try and flip the power dynamics somehow. A lot of these tools are made to control us or made to use us as data. I think if we try to democratize some means of resistance against AI, that can be really effective.

As an individual creative person, you can’t really, I don’t think you can really invent something that fundamentally disrupts power. Because power will eventually appropriate your intervention and turn it into a funny face filter or whatever they want it to be. But I do see hope in the possibility of spreading resistance kind of more broadly across society. And sometimes that can start from an artistic intervention.

This project called Face Work is a game that sort of gives players a chance to interrogate face analysis as a technology. So typically, when a surveillance system is making judgments about you using face analysis, it happens behind the scenes, you know, it’s out of view, you never really get a chance to engage with it in real time, or build intuition about how it’s making decisions. Like the closest thing that you get to is when you hold up your phone, and you can see whether a box is drawn around your face or not, then you start to get a feeling for like, what counts as a face. But beyond that, we don’t really have exposure to these algorithms. It’s really behind the scenes. And I think when it’s only when you can really have a real time relationship with these technologies that you can actually build intuition for them and how they work. Again, what their points of failure are, and how we can resist them in the moment, that’s not necessarily the same as resisting them structurally, but I think it’s part of what we need to do.

So I built this face attribute classification tool from scratch, using some popular face analysis algorithms and data sets. And I asked players to fool it. So you get real time feedback. This whole thing takes place in the context of this imaginary gig worker app, where you have to perform a face in order to get a job. If you want to, you know, be a delivery driver, then you have to smile, because that’s part of the job of being a delivery driver. If you want to be a substitute grandfather, then you have to have gray hair. That’s part of the job of being a substitute grandfather.

What are some of the other ones in here?

You’ve got salon visit, you have to have blonde hair, barber training, if you’re a barber, if you want to go to the barber and train the barber, you’ve got to have a five o’clock shadow. So you go in and you try and like, fake it out with different expressions. You can play this on I think it should work on your phone. It still works in the browser.

I forgot that it says do better when you’re getting a low score.

We tried to make it really funny. There’s a kind of secret side quest that happens as you’re going in there. I recommend playing it if you if you’re thinking about this tech at all.

In closing, I just want you to walk away from this presentation knowing that art has a lot more to offer besides the automation of skilled creative labor provided by generative tools. I would hope that you can find some inspiration in trying to put these tools to use, solving some tricky problems, like working with archivists and curators, or you can do these other things like manifest the future or kind of democratize new intuition. Or you know, just ignore all of it and make something about AI without using any fancy new tech at all. It’s all possible. And thank you very much for having and I’d love to hear what questions you have.

Discussion - Baptiste #

Thanks a lot, Kyle. Very, very inspiring. I really love it. And the last work you saw, you show like facework.app makes me made me laugh. Very nice work.

So we had several questions, notes in the document. Maybe I can start to pick some of them without any specific order. Please, Jenny, Gonzalo, if you want to jump in and also ask questions, feel free. If people in the attendees also wants to add questions, feel free. We’ll see what we can do.

So maybe one of the first question, and I think it’s a generic question about communities. You have done a lot of collaborative work and, you didn’t really mention, but you were part of open frameworks community some years ago or the fat lab also. It seems that when you work with these different pieces, you’re also among certain community of practices. So to your opinion, what communities are there for AI or is there one actually of several?

Discussion - Kyle #

Yeah, that’s a good question. The AI community is sort of scattered across a few different scenes. It’s very tool oriented. There are some online communities on Reddit that are connected to different tools.

A major tool for generative imagery is ComfyUI. A lot of people are using that to generate, like do text to image and image to image generation. On the Comfy UI Reddit and stable diffusion Reddit, there are tons of people constantly like creating tips and giving each other critique.

I don’t know, I think that’s interesting. Those are like the kind of communities that help foster the work that I do. I wouldn’t have any of the expertise that I have without the kind of open frameworks and processing scenes that helped kind of introduce me to making art with code and to computer vision.

That said, there’s also like, it’s not just, you know, what’s happening on Reddit or Twitter or Instagram. There’s some other weirder versions this time, because in the past with tools like open frameworks, there was no venture capital or startup connection really, but now there is. I’m also in some kind of secret startup WhatsApp groups where there’s just a bunch of people talking about AI and art all the time. It’ll be a few dozen people in a group that is organized by a hedge fund or something or venture capital fund. They’re just trying to understand like what’s going on with AI. The artists are there because they want to talk to each other. The VCs are there because they want to get the inside scoop. And yeah, it’s been a very different process this time.

Discussion - Baptiste #

Maybe related to this, there is this question I think is very interesting about the kind of literacy you advise for for designers and artists to acquire, to be able to work with AI or to have agency over AI, maybe to imagine different scenarios the same way maybe you did in your own creative works.

Discussion - Kyle #

That literacy question is really hard to answer. I know some people who, one of the reasons they make imagery is because they hate language. Like they don’t want to work with text. They don’t want to describe things. Like they use their visual intuition and their ability to like draw and paint or whatever it is to get their ideas and feelings and thoughts into the world.

But in a way, it’s kind of like these tools were all built for poets and writers. They weren’t really built for visual artists. So if you’re thinking about generative visual systems and you want to build your literacy, it seems like building a linguistic literacy and specifically linguistic literacy in English, that’s mainly the thing to do right now.

I don’t know that’s the long-term solution. I would hope the long-term solution is something more like imagine the way that you would want to interface with these systems in your best possible world and get ready for a future in less than five years where what you need to do is just describe that that’s how you want to interface with the system and then have that interface built for you.

We’re not far from that, I think. We’re in a weird transitionary period where in order to make images, we have to describe them, but with text, but we’re not going to be here forever. This is just a moment. And I think the skill that will continue to be useful is basically learning how to manage and delegate a computer like it’s a human. So if you can get better at doing that, then you will be very valuable going forward.

Discussion - Baptiste #

Yeah, back in the days, like a few years ago, you would be interacting with such generative models, not with text, but with other ways, such as abstract latent space, or you could use movement sensors as inputs and navigate this abstract model. I mean, it was really abstract.

Discussion - Kyle #

So yeah, you needed literacy maybe to understand what they do, but there was no text. And then the text comes in like three years ago. So I do think there is something valuable about trying to understand how these systems work, because it will give you ideas about how you want to use it [?].

You know, the idea of a latent space in the first place, it’s not an intuitively obvious concept if all that you’ve ever used is a language model. Like if the only thing you’ve ever done is spoken with chatGPT, the idea of a latent space is not going to be an obvious conclusion. Even if you’re working with like a text to image generator, it’s still not obvious. However, that concept of latent space is super generative for, I don’t know, for me, one of the weird things is that I have been thinking like this for a long time.

And it’s only the recent machine learning era where I felt like, finally, there’s software that like allows me to think like that computationally. This is the way my notebooks have looked forever, like since I was in high school. I was drawing like idea spaces and like mocking up weird kinds of user interfaces and thinking about making art in a very high level conceptual way, which is also just because

I have a very conceptual practice, but it’s because the details are often not the thing that’s most important to me. It’s this kind of high level like puppetry that’s important. So it turns out that this has been particularly resonant for me, but I think if anyone jumps into the details and tries to study how it actually works, you’ll probably find something in there that resonates with you too. And that’s one of the most valuable things is just finding new insight in these systems.

Discussion - Baptiste #

Back to the concern in the super secret whatsapp group. So what’s the most interesting or surprising insight that came out?

Discussion - Kyle #

I think, you know, it’s not even directly from those chats. It’s more about talking with people about those chats, because there’s people that I know where we’re both in there and I get to find out how they got in there. Like what was the infrastructure that allowed us both to get in?

Because it’s very complex. It’s not like I have a startup yet, you know. Maybe that’s the next project, but everybody has like their own route. And sometimes it’s because someone from an auction house that is working with someone in the tech world, and then like they create that connection that, gets the artist into the AI group.

And then other times it’s someone saw me give a lecture at a music festival or something. And then that’s the thing that gets me into the group.From talking with different people about like why they’re there and what got them invited, it’s sort of given me an outline of what the power infrastructure is for, where the funding for AI in art is coming from. That’s not something I can describe succinctly, but the way the VC funding, like tech funding world, is connected to the art world is connected to the kind of big tech world. And it’s very complex and interesting network, morass of sludge [?] that seems to be like a lot of it is just designed to sort of throw seeds in as many fields as possible, hoping that something blooms. And I don’t think that’s really the way to foster good art, but something definitely happens.

So it’s just been interesting to watch like the infrastructure, I guess.

Discussion - Baptiste #

Maybe relating to this, to this effort, to this discussion that you have in this group. Because it seems that you are imagining futures with these people from different backgrounds,like either VC or other artists, whatever. But you’re also imagining futures through your collaboration with other artists. And so imagining future, speculating on futures, the possible future with this technology, being critical about it. What do you see this value of speculative design of that practice, speculative design and imagining futures? And maybe importantly, how do you see opportunities for this practice to become participatory?

Discussion - Kyle #

Something you said that I picked up on the most is this topic of speculative design. This is something I’ve spoken with Lauren about a lot. And it’s something that we don’t really relate to really well, even though a lot of our work could be seen as very speculative.

I don’t think that we identify so closely with that practice, because we feel that a lot of speculative design is very image oriented instead of experience oriented. It’s a lot about creating an image of what the future might look like, instead of giving someone a direct experience of what the future might feel like.

And when we create these installations that are giving you a chance to like, literally have a voice in your head, or doing face expression analysis and putting your face on a screen with the judgment about it. That goes beyond just being like a kind of speculative video.

I mean, a lot of speculative design feels a lot more like video art to me, which is great. But I don’t know that it’s the thing that I’m most excited about when it comes to addressing the future. I want to feel the things that I can’t feel any other way.

I feel like speculative design feels very connected to sci fi writing. And that can bring some like initial feelings, but it’s not really about me, it’s sort of imagining someone else. When I’m in it, when I’m like in a gallery, and like suddenly, like some social boundaries have been crossed, then that’s an interesting space. And that’s one of the things that only art can do. And so I feel a little bit of a responsibility to do that thing. It’s not for everybody, but that’s just how I feel. I’m not telling you, you have to go to that. Anyway, sorry, I just picked up on that for a second.

Discussion - Baptiste #

The other question was about collaboration. I want to bring this participatory, this process of imagining futures, like how you can bring it more participatory.

Discussion - Kyle #

I was starting to get into that question with face work and thinking about, you know, I was imagining what if I could do a workshop that had 1000s of attendees? What would that look like? What would I want that workshop to be about? If it was about face analysis? How would it feel?

And then I realized it kind of like a game, it is the closest thing to having a workshop with a lot of people. And so then I made a game. If you’re looking for participatory process, I think putting people in the middle and sometimes with each other in a game context can be a good way to do that. I think just getting people to build things together can be productive too.

Lauren and I had another project I didn’t talk about, but a long time ago we had a project called Noodle, which was a small robot that basically just had some cameras, speakers,microphone and a little screen. And you could program it using really simple prompts. Like you could say, you know, if you see something and it’s scary, sound the alarm. And secretly behind the scenes, there were Amazon Mechanical Turk workers who were evaluating all of the input and then making these decisions.

Now you could build it today with new like multi-modal LLMs and it would work for much cheaper. But we were trying to imagine that future a little sooner. Part of the goal with that piece was to give people the experience of trying to program it and think through how it feels to program a device like that. And also how it feels to remember that there’s humans behind it.

We sort of allied that fact today when we talk about language models, but there’s humans behind language models too. Not the same way. It’s not like there’s a human responding to you on chatGPT, but the humans fed into it and the humans trained it. The RLHF was done by humans. I think anything that can get people kind of in the loop and in that kind of direct feedback process, directly programming like that kind of stuff, that’s really good way of participatory process.

Discussion - Baptiste #

Thanks a lot. I think it’s something that we are actually discussing quite a lot in this project, how to bring participation to AI development. And I like this idea of design, I mean, speculative design where it becomes participatory and where it becomes also research.The piece you show when you delegate the decision to draw to the machine, if you think speculatively, that’s where you come up with this type of ideas and design and then it makes you understand something about what would be the future of technology, but also more theoretical questions about our willingness to delegate decision to others. So I find it really, really fascinating.

It’s seven and I would like maybe to end up with one question and there are still a lot. So I don’t know, maybe you can even pick one if you prefer one.

Discussion - Kyle #

Yeah, let me see. So we’ve got, there was one about speculative design and vibe check, which we talked about. What if an AI company willingly gives up power?

There’s one about data poisoning.

Role of openAI with regards to boosting current AI hype. Yeah, openAI is very, in a very strange position, honestly, and I’m really concerned by the way that they’re acquiring a lot of lobbyists. I think I heard that they’ve got a plan to get about 50 lobbyists on staff. And there is, I don’t know, people ask me if I’m like worried about AI. I’m really not worried about AI. I’m worried about the companies that are building the AI though and what they’re doing with it.

That’s very concerning to me. There’s a lot of abuse of power that people have access to right now from these tools. I’m much more concerned about like the centralization of the technology and like the creation of laws that disenfranchise the people that produce the data that go into it and the people that use these tools.

I’m less worried about like the spam bots that are posting misinformation online. I feel like we’ve already got enough of that. Maybe we’ve even seen some limits to what that can accomplish, but I’m sure there’s going to be worse things.

The next COVID-like virus that is produced by someone who’s fiddling around with GPT-5 and then went to go order it on a biohacking website or something. Yes, there’s going to be those problems eventually, but we’re not really there yet.

I think really the practical problems are a lot more about governance of this data. In a way, I’m actually like excited that Facebook is releasing their LLMs like in a way that Microsoft OpenAI and Google are not. Google is releasing these really small models that are just barely useful. And then Facebook is releasing these really big models that are quite comparable to OpenAI’s models.

And it’s really, I mean, ultimately it’s like a marketing and like an engineering capture process in the same way, but it somehow feels better to me.

Anyway, I’m just starting to work through this like emotionally. I got to talk to my therapist about it, but that whole process is really interesting, concerning. I wish OpenAI could have stayed some kind of really kind of altruistically guided nonprofit, but it’s very clear that that’s not going to happen.

The data poisoning, I’ll just finish on the data poisoning thing, I think, because I would love to see, oh no, there’s another good one about data set transparency.

Okay. All right.

I’m going to wrap up just like two minutes quickly about data poisoning. Okay, I’m at Cannes right now for the film and advertising festival because there’s an artist here that I work with who invited me to stop by to introduce me to people. And I’m getting some feeling of like what’s going on with AI in advertising right now.

And it’s fascinating and terrifying.

And I would not be surprised if next year, the winner of the award is like someone worked with Coke so that secretly every page on the internet has some reference to Coke because ChatGPT generated those references.

And then they poisoned ChatGPT somehow over the next year. There’s a lot of possibilities for that. It seems like these models are genuinely weak to certain types of attack. And that’s definitely something that artists should be exploring more.

And then I’ll finish on this one about data set transparency and open sourcing training sets for base models, which is, yes, agreed, something that it didn’t do. That’s majorly problematic and exploitative. The sources.plus from spawning in Berlin, they’ve been collecting I think they’re at like 14 million images at this point.

Let’s see. Yeah, they’ve got 14 million Creative Commons zero, which is like, you know, basically public domain. And then they’ve got 23 million public domain images. That’s still not quite at the same scale as what’s used to train something like stable diffusion, which I think is more like 10 times that big. And as you probably know, these models don’t exist without scale. Like I was saying about Flickr and GPUs and all that. The scale is the thing that birthed this. And it’s the thing that this is all tied to. And no one has figured out really how to shrink that, you know, by an order of magnitude. They figured out how to shrink it a little bit, but they’re still working on it. But I think at some point soon, the amount of data that is available in the open in this kind of like ethically sourced way will continue to go up. And then the amount of data that’s required to get a quality that is useful for visual artists and designers who are working with these tools, that amount will continue to go down. Eventually it will converge at some point where we can use a tool that is like ethically sourced.

I’m sort of, you know, doing this because I don’t think that ethical sourcing is really the problem here. I think that’s like a kind of side question of, you know, it is abusive that these tools are trained on data that no one ever agreed for them to be exploited in this commercialized, financialized way. That is a genuine problem.

However, it is also a problem that I’m pretty sure is a technical, like the reason for that problem is it’s a social problem. But the solution to it is actually a technical problem. That part of the problem will be solved, I’m pretty sure. There will be models where we can say this only uses public domain, this only uses creative comments images.

And then we’re going to have a really complicated questions to ask, which is like, oh, this still kind of sucks. Like this is still bad for us socially. Like this is still putting a bunch of artists and illustrators and designers and stock photography people like out of work. And then we’re going to have to ask, like, why is that happening and how can we fix that?

Because that’s the real problem, in my opinion, is the automation of labor in the first place. It’s not the exploitative data practices, though that is also a problem. And I don’t have a good solution for that except for universal basic income. And maybe we can get that happening.

Anyway, thank you for sticking around for a couple more minutes and also for asking great questions and listening.

And thank you again, for organizing this. I really appreciate it.