About us
Stay in touch
Peter Bauman: Hello everyone and welcome to this Le Random artist discussion. I'm your host Peter Bauman, or Monk Antony, the Editor-in-Chief at Le Random, and today our talk is about the artistic, philosophical and cultural implications of GANs and deep generative models. This topic stems from some of the research that I did earlier this year about the history of GANs and deep generative models, and I'm ridiculously fortunate to have on two of the most prominent people to feature in that research. That is Gene Kogan and Tom White, or Dribnet.
Neither of them needs an introduction, but Gene is a pioneering figure in generative AI known for his open source initiative ML4A, which was highly impactful in teaching the early GAN community and the early deep generative model community about these tools. And then Tom White, also known as Dribnet, he began his career at MIT working under John Maeda with Casey Reas and Golan Levin at ACG, where he developed ACU, which was this really important predecessor to Processing. Since then he's continued to build tools for computation and creativity. His work reveals the strange logic of how algorithms interpret the world. I can't believe I can talk to both of them about this particular subject. I can't wait. I'm so excited. So let's begin.
Thank you again both so much for joining. Both of you have such long backgrounds in deep generative models, or just machine learning. Tom, you've been in this for decades. I mean going back to ACG at MIT. Gene, was that something you just learned about right away and were excited about, back in 2014? Or how did you first encounter GANs and what they were?
Tom White: I'll say one thing, and Gene might correct me, but I think one thing that we also have in common is that we don't only have a machine learning background. We have a background in creative coding and just creativity in general, creative computation and what's now called generative art. I think that it was pretty obvious to people that were immersed in that at the time that machine learning was going to become a big part of that.
As far as GAN parts of it, that was one of probably the three big things that was happening in 2015. So 2015 was kind of a breakout year, I think, for this, where we had Deep Dream, we had style transfer and then we had generative networks. The generative networks were very primitive. I was interested mostly at the time in variational autoencoders, and that's a generative network that still exists now, kind of inside of models like Stable Diffusion. The other one is the generative adversarial networks, the GANs. Both of those sort of have very similar properties. They both can generate graphics and they both have what's called a latent space, a space of possibilities.
I think it was pretty clear when they were coming out, even though they were very, very limited at the time. They could only make essentially images much smaller than an icon, like 16 by 16 or 32 by 32. It was pretty clear that they were going to be part of the vocabulary of the creative coders going forward as machine learning became more important.
Gene Kogan: Yeah, I think that's exactly right. It's correct that probably one of the reasons why we're associated is that we had this coincidental hybrid where we just happened to be interested in generative art and creative coding, but also machine learning, which even most of the people in our creative coding cohorts weren't so interested in. For me, I actually started out in machine learning and I did digital art in a very, very not particularly active way until actually after that. A big part of the reason why I got into creative coding was a little bit of boredom with sort of pre deep learning machine learning. GANs and style transfer and Deep Dream, as Tom mentioned, were really watershed. I think they were a big part of the reason why I really doubled down on this intersection.
I think the original question, Peter, correct me if I'm wrong, is where did we first encounter it? I first encountered it with the DCGAN implementation. So Ian Goodfellow, one year before that, had published them. I think they were fairly obscure for a good year until people started messing with DCGAN. Before that, even after Deep Dream and style transfer, they felt like they were kind of limited, and they were limited because of technical reasons of how they actually work. I think GANs and VAEs were really the moment where I thought, oh wow, we could maybe actually be doing what we're doing today. I thought maybe that would take like 20 or 30 years rather than just 10, or six even. But that did feel like the moment where I felt like we started making progress on image generation.
Peter Bauman: Well, yeah, that's really interesting. It was DCGANs and I guess I would presume Alec Radford's tweets, and maybe even some of Tom's too, because Tom, I think you were maybe the first artist to respond to Alec. I think maybe also Kyle McDonald's. You saw some of his first tweets, I guess, about maybe VAEs and GANs in the middle of 2015, and then you also started playing around with some of it. So was your first encounter Ian's paper in 2015? Was it 2014 or was it Alec's tweets? I guess those were probably mid 2015.
Tom White: Yeah, about a year later. I guess it depends. If encounter is reading about it and thinking about it, then it would have been a year before, when the paper came out. But if it's actually getting your hands dirty and training your own models and doing your own work with it, it would have been about six months or so before the DCGAN paper came out. But it wasn't with DCGANs. It was with variational autoencoders.
The person in that space is Durk Kingma, who ended up going to OpenAI. He's a super genius. He also did the Adam optimizer and a bunch of other foundational stuff at the time. But he did the variational autoencoder. Like I said, it has many of the same properties. So I was working with that model, training my own weird data sets of stick figures or letters and numbers. But what we saw, and one of the reasons I was excited, is I saw the results that Alec was posting were much stronger than what I was getting.
Variational autoencoders treat the image a little bit differently. Their weakness is they tend to blur the image. They tend to represent the data set a little bit better, but at the cost of fidelity or believability. They don't give you crisp images. What Alec was able to do, he was able to come up with a way, with his team, of scaling that up and making big images. As soon as he was posting these online, I think everyone could see the promise, that these things could scale bigger and could be bigger than 16 by 16 or 32 by 32 or whatever one was using at the time.
One other thing I want to stress about this period, and Gene can probably speak to this, is it was really nasty messing with these systems. People complain today about installing CUDA drivers and learning PyTorch. But at the time, many of these systems were running on the original version of Torch, which was in its own language, in Lua. It was the dark ages of getting this stuff. So I think a part of it too, and I know Gene has this too, is just the willingness to be frustrated and get your hands dirty and spend days configuring your machine to get any of this to work. It's easy to forget about that. There weren't tutorials online. There weren't trained models even. I remember if you wanted to do style transfer, just getting your hand on the trained models was a puzzle, because that's one of the things you had to do. So it was before you had the kind of support structures and the language support in the libraries. I think that's the challenge that we have now.
“But at the time, many of these systems were running on the original version of Torch, which was in its own language, in Lua. It was the dark ages of getting this stuff.” — Tom White 8:02
Peter Bauman: I'm also curious just how many people were doing that, were getting their hands dirty with it at the time. I mean, how many people that you knew of?
Gene Kogan: I think we'll never know because, if anything, there might be more people who were doing it, but not necessarily active on Twitter as we were, or really even thinking that it's significant enough to talk about. Of the people that I knew, I don't know. The people that I knew was kind of Tom and a handful of artists, the ones that are coming up that were in my world that I was checking up on. And of course people like Alec Radford and Durk also. I was also looking at VAEs at the same time and just being active.
There was kind of, I think, like deep learning had crossed over into Twitter not long before then. I think there was this emergence of AI Twitter in that year or maybe the year before. It was mostly the young researchers who were adopting social media that were active. It was really a big change in the culture of machine learning, because I had studied machine learning in college. It was seven, eight years before that. It felt like, especially from a kind of sociological perspective, it felt like a completely different world. It was sort of pre deep learning. It wasn't even really very programming oriented. It was very theoretical and very mathematical, decision boundaries and proofs and things like that, and not so much getting things to work. It seemed like most people were using things like MATLAB and most people were publishing to academic journals and conferences. There wasn't like arXiv yet, or arXiv was new at that time. The open, very public, deep learning oriented world really emerged around that time, and it was through this that people first started discovering these things.
So, yeah, I don't think a very large number of people were trying to do more than install the software and maybe run the examples. But definitely, I would think that there were more people than we realize actually plugged in and trying to do some of the same things.
Peter Bauman: It's interesting that when that open, deep learning period emerged that you talked about around that time, yeah, it makes sense that it would have coalesced around Twitter, because Facebook wasn't really the popular, I guess, social network anymore for people of that age, for younger people. That was kind of when our parents started using Facebook. Even if you just do very, very specific Google searches, which maybe you know that I've done some of those, it's obvious that it does start right around that time, kind of early 2015, arguably with Alec and some of his early VAE tweets that he put out.
I mean, I think, Gene, you also called out some of the first latent space art, and I do want to talk about that. But I'm interested, Gene, that you talked about DCGANs as the catalyst for at least your interest in GANs. Was that the catalyst for ML4A? Was that what inspired you to begin those workshops even?
Gene Kogan: It was one of them. I would also put the style transfer, Deep Dream category in that same group. I think Tom identified those three specifically and I would agree. I would draw a circle around those. There was sort of variants of all three of them. There was all these kind of no content texture synthesis, which I would also put into that umbrella. There were a few things going on at that time. But I think those three definitely felt like a critical mass of things that you could put together. I would also maybe add things like t-SNE, which was a fun technique that I had a lot of fun with and I made some tools for. t-SNE is this technique for basically taking a data set of images, or anything really, but mostly we were doing it with images, and then organizing them in the 2D space such that the similar data points are arranged near each other. You can actually do this with the VAE also.
Basically, those were the sort of things in 2015, early 2016, that there was just so much relevance to electronic computational digital art, that ML4A made sense as a category of things that you could teach the students that were interested in the electronic art.
Peter Bauman: I'd like to hear both of your takes on this. Today with public discourse and AI, I feel like it's increasingly becoming fearful, or maybe the fearful voices are becoming louder, and maybe rightfully so. Yeah, I'd love to know what you guys think, but it seems to be more dominated by fear, downsides or negative sides to AI and machine learning. Both of you are involved with educating people about these tools, and I'm wondering, how do you see your mission to make these tools more accessible? I mean, especially Gene, that was your mission with ML4A. How do you see that mission fitting into or challenging this darker narrative that seems to be emerging around AI?
Gene Kogan: I think a lot of the fear is born out of a feeling of alienness, that this doesn't have to do with me. It's kind of an alternative to me. It doesn't need me. So I completely understand. There is a sense in which it's profoundly changing a lot of the things that we do. The funny thing, the ironic thing I would say, is the first thing that you can automate, I think fully automate, is a creator. It's like an artist, a digital generative artist using AI to make art. I've been automating myself for basically my whole career. So who will need me? At that point, then there's a whole layer of fear, of course, of the broader implications.
Something you mentioned earlier in the talk is the crazy... It's going to be the currency of how we make technological change happen on Earth. And if you know more about it, you're going to be better off. We don't know in what sense you'll be better off yet, but you'll definitely be better off. It's just better than not knowing.
It's hard to learn about these things, obviously. People have a lot to do in their regular lives. So trying to reduce it to the elements that are most important and sort of digestible to them, most relevant to them, is, I think, the way to try to squeeze in a little bit of awareness and some dexterity with the tools. Even though the word is overused, I'm a big fan of democratization and just open source permeation of these tools. I think it reduces a lot of the risk in ways that, well, that's a whole other conversation, so I probably will just leave it at that. But basically dissemination of the tools and the techniques and the knowledge is a way to, I think, reduce inequities that result from the technology.
“The funny thing, the ironic thing I would say, is the first thing that you can automate, I think fully automate, is a creator. It's like an artist, a digital generative artist using AI to make art.” — Gene Kogan 14:10
Tom White: Yeah, I'll make three small points on that. I think Gene summed it up really well. One thing that I think is true is that one of the reasons this era is kind of fun to think about is that you didn't have a lot of that background at the time. So in 2016, 2017, 2018, this was a weird, non-threatening, quirky technology that people were experimenting with. I think people were approaching it differently.
Gene talked earlier too about how machine learning, the personality types had changed, and how people that were doing this 10 or 20 years ago, they're very analytical. They're very top down. That's very true. But that's also true, I think, of creative coders. I think that a lot of the programming libraries, Processing or p5 or openFrameworks, they all kind of, you decompose a problem analytically and then you reconstruct it. Machine learning is coming out of this from a different set of skills. I think there are a lot of people that are interested in this, but it is a different personality type that's interested in this, where you're teaching a computer, for example. You're learning the importance of data sets. You're learning how to clean data. So I think it's a different skill set. But like Gene said, it's one that once you understand how the systems work, then you kind of see the creativity. It's a matter of how you do it. It's not a black box. It's not something that's always going to spit out the same answer. It depends on how it's configured and importantly on the data sets that it's trained on.
Peter Bauman: Yeah, I'm really interested in that, how maybe the philosophy of computation is changing. I think what you both were talking about is hinting at that, how working with these systems now does require these new skills. Do you think that machine learning is going to eat the other, like the hard programmed skill base? Is there still going to be a place for hard programming in this machine learning world?
Gene Kogan: The thing is that what programming means has always been changing, not just now. Programming like very framework driven kind of high level languages, interpreted languages that didn't exist 30 years ago, let's say, would have looked to people who were programmers back then as a completely different layer of abstraction.
To me, I don't think it actually changes the really fundamental things that much. The next version of what we're doing, the next generation version, I think still has a lot of the same aptitudes for breaking technical problems and systems down into parts and thinking very rigorously about how they all connect together. Maybe the actual sort of syntax changes around that.
Right now, although I love vibe coding and everything, I think the ability for an average person who isn't thinking deeply about computer systems, it's still kind of difficult to do a lot of the things that it's hyped up for. Really good programmers are the ones that are benefiting the most right now, because it gives them almost like an army of interns that can work for them. That'll change more and more, but with it the goalposts will keep moving, and you'll still have to be analytical and fundamentally breaking problems down in a very engineering oriented mindset. I don't think that's going to change for a really long time.
Tom White: Yeah. I'll just briefly say what Gene said is exactly right. I think that what will happen is that the definition of programming will continue to change. My favorite example of that in the past is that it was really artificial intelligence in the seventies that was even thinking about how to represent information or knowledge. Marvin Minsky in 73, 74 came up with a system of doing that called frames, and frames was you have a template and you fill in the template with, like a chair has legs and this chair has this kind of legs. He came up with this whole formulation of doing that in a real paper on it, but it looks very much like object oriented programming now. So we didn't have object oriented programming in the seventies. People weren't sort of thinking about how to represent things or model the real world like they are today. So if you were to revisit that paper, it would be very hard to understand it in the lens of object oriented programming as it exists today. I think the same thing is happening, where we're going to just keep incorporating these improvements into how we make things, and that'll just be part of the vocabulary.
Peter Bauman: Yeah, because Tom, your practice, it was a gradual transition. It's not like a black and white thing. But you went from focusing a lot at MIT on the hard programming side to now, at least creatively, more on the machine vision side. So yeah, I'm really interested in whether that will maybe mirror, whether you were kind of first to also go through that transition.
OK, so we talked about pertinent societal critiques of AI, but there's also critiques of AI that are more art world based. I'm sure you're both familiar with Joanna Zylinska. She's an author and a critic, and she wrote a book about AI art. So you would think that she would be fairly sympathetic to machine learning. It turns out she's not. She refers to deep learning based art as glorified Candy Crush, "a psychedelic sea of squiggles, giggles and not much in between." And she calls it art as spectacle.
So I'm wondering, as artists who have devoted your careers and decades to these systems, and you're deeply embedded in their inner workings, and you've promoted them to other people, how do you respond to that framing, first of all? And do you see spectacle as, I mean, does she have any kind of point, really?
Gene Kogan: I don't know the author, so I can't really comment too much with respect to what she said. My sort of best guess is that she's maybe referring to AI art as like almost like a sociological term. It's like a group of people who have a particular shared aesthetic. It's like a scene. Punk music was a scene. There was no scene when Tom and I were doing this, and we don't think of it that way. We think of it as a technology. So for it to say deep learning based art, it's kind of to me sounds like painting based art or sculpting based art. It's so broad that I sort of don't know what you're referring to. So the critique, I don't make much of it.
I'll say probably what she is kind of referring to is the fact that it's very possible to mass produce very homogenous, very sort of average looking stuff because of this, because of the technology. So in that sense, it's for sure. I mean, it's not all like that either, right? So you can definitely find interesting stuff.
Tom White: Yeah, I think what Gene has said is spot on, that it's a really broad set of interests of artists and people doing things. I think what's happening is that in society, though, there's a dynamic. Of course, most people aren't familiar with the techniques and aren't familiar with the philosophy of machine learning. So in the early days, I think probably the people that do sensational work are going to be the ones that bubble to the top. For better or worse, I don't think most people have the critical skills. Maybe that's not the right way to say it, but I think that there's a certain tendency, and this probably happened with photography and probably happened with other mediums, to kind of embrace the kind of flashy or some of the surface level qualities of the medium. I think that's probably playing out now. Or if you're not familiar with it, you don't know the deep bench of artists that are working on this. Even Gene and I have very different interests when we approach this. And if you look, you can probably find artists that are kind of quietly going into deeper issues of how these things work, which might not make things that get projected three stories high. They might be looking at specific issues.
Part of it is it takes a certain literacy from the public to understand how these systems work and what kind of piques their interest in certain ways. I think we're still kind of working through that. I would encourage people, if they're interested in it, to look deeper at artists that both are getting publicity and aren't getting publicity, because there's lots of people doing this and they're not all taking the same approaches.
Peter Bauman: Yes, certainly. And I think also people are not familiar with the history and think it's all just Silicon Valley and it's all just text to image. And Joanna Zylinska, I mean, she's an artist. She wrote a book called AI Art. And before that, she was a really big art and technology advocate. So it's disappointing, I guess. That's kind of why I find her critique so interesting, because it is from an informed opinion, but also it's a bit of a hot take opinion, like just even comparing it to Candy Crush.
Tom White: Well, there are lots of artists doing superficial work. I think we could probably make the same argument if we cherry picked certain things. So I think that it's probably a valuable addition to the art world. But I think that smearing a whole medium is probably not the way to go. I think what I would critically look at is what her selected examples are. There are certainly very popular ones. There are certainly people that aren't engaging deeply and they're using it. It's very easy to make bad stuff with machine learning. And one difference between that and kind of traditional creative coding is that there's a certain kind of skill level you have to get to before you can do anything with creative coding. But with these tools now, anyone can. It's almost like when photography was introduced, anyone could push the button. I think she probably has a point for some work, but I'd hesitate to use that to color everyone working in the medium.
Peter Bauman: Yeah, definitely. Tom, I have another question for you. I think since at least 2017, you've talked about how your work has investigated the algorithmic gaze and how machines see, know and articulate the world. You invite viewers into the perceptual space of machines. That's really interesting because, you guys know Phillip Isola. He has this Platonic Representation Hypothesis, which argues that neural networks, even if they're trained on different tasks like classification or segmentation, even if they're trained on different modalities like vision or language, and architectures like ResNet or transformers, converge on the same representation of the world. So I'd be interested, from that brief explanation, what you guys think of it and if that makes sense.
Tom White: Yeah, so a lot to unpack there. I'll try to do it. If I start ranting, Gene, just jump in and shut me up. To go back, you were talking about the algorithmic gaze originally. That is my main interest, how are these machines working? What's inside of them? Even going back before 2018, the way that I was looking at GANs and variational autoencoders, I was mostly interested in the latent space and kind of that space of possibility. So if you look at work that I present, it's almost always displayed in a grid. And Gene was talking about t-SNE and ways of organizing space. I was mainly interested in not any one output, but the space of outputs. And that's what I was trying to communicate artistically, is look at these faces that are all in the local manifold of possibilities. That's kind of what my core interest remains even now, and that's what ties my work together.
As far as Phillip's particular hypothesis, I think you summed it up well, the paper he wrote. I have a lot of opinions on it, but what I'll say is first, I think that what he's saying is true for this point. I think that his grand theory on this is a little overblown. I think that what's happening is that we have language models and we have vision models and they're both trained on the internet. And so, yes, if we learn that a basket can hold eggs in a language model and we see a picture of a basket holding eggs in a picture, these representations are currently converging. We can answer the same kind of downstream questions in both cases. And so the systems independently are learning kind of the same knowledge. And when you put them together, you get something stronger. I think that's kind of true right now in this moment. I think as we learn how to scale out into different domains where we're not training on a known corpus, or in different modalities, we might see that that hypothesis gets stretched a little bit.
But what I want to say is that I really like, and I've talked to Phillip about this, he's a great guy, I think I like the hypothesis because he's asking exactly the right question. And I think that in AI there's so many different questions you can ask. And I think he gets people thinking about the right things and talking about the right things in that. Like, well, what's inside these systems? And is what's inside a language model the same thing as what's inside of a vision model? Is it changing over time? I think it's a great overall discussion to have. And I think that the hypothesis he puts forward is really a powerful way of structuring that conversation.
Gene Kogan: Yeah, I also know Phil. I haven't read the paper or watched the talk, so I'm not familiar specifically with this. But it looks very interesting and I'd like to give it a read after this. But maybe just to add, I think we're all fascinated in a sense by maybe the same thing, even if we express it differently. And I think there is something about these models that connects to the bigger picture of human consciousness and communication. That's always fascinated me.
So I think early on my read on these generative models was that they felt like, in a very literal sense, they're a portrait. They're sort of converging on some kind of like a median, like almost like the average of all sort of human expression, expressed through images and text and audio, whatever else. And I think that you can draw analogs, for example, to deeply psychological concepts. I've always seen analogs between that and Carl Jung's kind of theory of the mind as having all of these highly evolved shared archetypes. We have these kind of categories very deep in our psychology that are shared, that we can communicate about, and that we find ways of talking to each other about even very abstract things like the self and the other, and God.
That was kind of the big reason why I did this Abraham thing, which is kind of trying to manifest an artist that is expressing these things. And I think that we're all a little bit, I do feel like I'm kind of grasping at it a little bit from another direction from Tom and from Phil. But we all sort of see this. It really shows us something about what it is to be human. That's probably the thing that I find the most drawing into it on a personal level.
Peter Bauman: Yeah, that data is, I mean, like you said, it is a representation of humanity, all of the data points that it's trained on. And then, yeah, Tom, so again, back to your practice and how it relates to this, to Phillip's hypothesis. Tom, you talk about in your practice art by AI for AI. And I'm wondering, how does that engage with this idea of convergence? And then do you see your abstract prints as participating?
Tom White: Yeah, I definitely think it fits in with the larger picture of what the Platonic Representation Hypothesis is trying to get at. Gene just talked about averaging things together. That might be a good starting point. So when you're in the latent space of possibilities, it's not clear how to represent a population. And one of the things that these generative models trade off a lot of times is they trade off fidelity for, the technical term is, capturing all the modes of the data set.
But let's give a concrete example. So if you're training on a data set of faces, probably very few people in that data set are wearing baseball caps, maybe three or five in a data set of a hundred thousand or something like that. And so a good generative model will be able to draw, you have a trade-off. You can just draw a few of your archetypes or your prototypes very well. You can draw kind of a stereotypical guy or stereotypical woman, or whatever the mean there is. Or you can try to represent the entire data set. And these are trade-offs you can kind of make in how you represent the latent space, and different models will make different trade-offs.
The thing that I'm interested in is internal to these machines, these trained models. They kind of have these trade-offs when they're recognizing things. So I was approaching this as a study in abstraction, and what abstract forms could the machine create that they would then later be able to recognize.
And it turns out that many of the mechanics of how machine learning works, so for example, when we train machine learning models, they're blind to left-right symmetries. And so a lot of times when they draw things, they draw them not kind of discerning the left from the right. Or if there's pollution in the data set, these images are sourced from online images, a lot of those are coming from eBay and other places, and so their idea of what a particular item is, is colored by that. And so the representations in these systems often diverge from human expectations.
And that's something that I've looked at kind of passively with these objects, but also actively. So one of the series that I did was a series called Synthetic Abstractions, where I basically made abstract pornography, but it doesn't really look like anything. It's these abstract shapes. But they were made by these filtering machines looking for objectionable content online. And if you take a picture of them or you upload them to an online platform, a lot of times they get flagged, even though they don't have anything objectionable in the image.
The reason that that's intellectually interesting to me is that points to the fact that it kind of goes against the Platonic Representation Hypothesis in a way, if you take it in the large. These are images which are not objectionable to people, but whatever it is that we've taught these machines to recognize must be divergent in some way, because they're able to draw things which are triggering these filters. So they have a different, broader sense of what these filters are supposed to do. So yeah, that is my interest. I do think it ties into this hypothesis and it kind of points to the representations that we're all carrying around kind of subconsciously and not thinking about.
Peter Bauman: Gene, you gave this talk about decentralized AI, but that was about eight years ago. I think it was 2017. And both of those things were in very different places, that the kind of these concepts of decentralization and also AI, they were both on the kind of the cusp of broader cultural adoption. So I'm wondering, in 2025, how has your thinking shifted now that, I mean, these things are a reality in a lot of ways? And what part of your original thinking has shifted? What part of your original vision has held up? And then what has been a surprise, maybe in a good or bad way?
Gene Kogan: Yeah, the original kind of interest in decentralization was that a lot of the technology that was being built, these decentralized computer networks and blockchains and things like that, they're creating a system for communication, value transfer especially, that is basically permissionless and borderless and was so much more amenable to automation than any of the pre sort of legacy systems that do that. And so it felt like a natural convergence for AI. AIs really want these kinds of things to be able to transact and do sort of value transfer.
And so in the sense, one part of it is just, I'm always curious about new technologies, new computer technologies, but especially in this case, it felt like these two, AI and decentralization, which feel kind of very, and still are to a large degree, they're very divergent fields, and people don't generally work in both, because in the sense AI is intrinsically centralizing more than anything. And so there's kind of that friction there. But AI really wants digital currency. It's the way that AIs can become, you know, for something to happen. It has to be able to participate in the economy.
And so now you're seeing a lot of, the way that's manifested in 2025 is the sort of rise of this idea of autonomous agents and so on, which felt very much like on the horizon from back then. I certainly couldn't have told you about all the things that happened in between then and now. But yeah, of course it's a can of worms, and so we would really need more time to unpack it. But that's what I could say right now.
Peter Bauman: Yeah. Yeah, I just had, not an argument, with a curator about that. I was kind of disappointed that they didn't really see that relationship between those two things in a bit more sensitive way.
OK, because I would really like to ask this last question to you, Tom, about your time at MIT. You were there at this incredible juncture, I think, in art history, at MIT at ACG under John Maeda. And so you worked on ACU, which, for people that don't know, was a toolkit that was really important for the development of Processing and p5. And so I'm wondering how those early experiments in interactive graphics, how did they shape your later approach to neural networks and machine perception?
Tom White: Yeah, I'll say two things on that. So one is that even when I was there, my interest was AI in a sense. I think the artists of the time, like Harold Cohen, and one that I think is still doing work and doesn't get enough people covering, is Karl Sims. Karl Sims was doing amazing work and still does work, but the language was different at the time. So the language was genetic programming and things like that. And so I was dabbling in those things.
But to your point about how the work specifically at John Maeda's group primed me for what's happening now, I think the main thing is that it was really freaking hard to draw anything at the time, and people don't know that. So like I talked before about how terrible the toolkits were. At the time, we're in the late 90s, the toolkits for drawing were just as bad as the early deep learning toolkits. And strangely, a lot of the core competencies are the same. Like if you were good at linear algebra, you could probably do some pretty good graphics programming. But it was really hard.
And I think what Maeda showed is that if you can build a system around that, you kind of trudge through the forest. Somebody has to be the person with the machete going through the forest, kind of figuring out the paths. And I think maybe the philosophy of the approach in general was the main thing that primed me, in that deep learning was kind of the same way. When I saw it coming out, I saw like style transfer and GANs and generative networks. Those were the kind of, I saw them as the primitives, similar to how 20 years ago arcs and lines and pixels and anti-aliasing and these kinds of things existed. But they were very inaccessible, I guess.
So I guess that the main thing that I carried forward from that is just kind of the perspective where I could kind of see that these technologies that seemed very hard to approach, and that it took a lot of technical skill, were going to become more mainstream and were going to become more popular once people kind of figured out the conventions around them to make it approachable.
Peter Bauman: Yeah, I like that analogy of you in the jungle, in the forest with the machete. And I think both of you have done that, and I think to everybody else's great service. So yeah, thank you both so much for that.
“Somebody has to be the person with the machete going through the forest, kind of figuring out the paths.” — Tom White 42:11
Tom White: You point it in the wrong direction sometimes. So we're not always going in the right direction and there's many dead ends. But I think it takes a certain personality type to even pull the machete out. And I think that John Maeda's group did that in one domain. And I think that that's maybe the thing that it's useful to do, is to have that mindset.
Peter Bauman: Yeah, I think it's a really good point. Well, yeah, thank you both so much. Again, I wish we could do it for several more hours, but yeah, we barely touched the questions. This was really great. I really appreciate it.
Tom White: Thanks, Peter.
Gene Kogan: Yeah, thanks.
