Three Steps Every Creative Professional Should Take to Prepare for AI | AI in Action 2023

Hey, everybody, uh, excited to share a little bit more about Creative ai. I know that this is, uh, probably gonna be a fun topic and maybe a little bit different than, uh, anything else you've seen today. Um, I'll give you a little bit of, of context about myself, but then we'll dive in to the technology and what's happening in the space.

Uh, I'll give you my perspective from a, a strategic, uh, viewpoint of how I, I see companies starting to adopt generative AI in content production and creative, uh, pipelines. And I think there's been, uh, over the last year, just kind of some of the most rapid advancements in technology that I have ever seen. Uh, I'm sure that there are, uh, a lot of people who would agree that, you know, this technology is, is changing really fast, and what we think we are capable of is kind of being, uh, disrupted almost every day.

Uh, so excited to share a little bit about some of those cool things that are happening. Uh, but first I will answer the question of whom I, uh, my name is Kent czi, and I am the CEO and founder of a company called Invoke Ai. Previously, uh, prior to, uh, invoke, I worked at a variety of, of SaaS companies working both with consumers and businesses enterprise.

I was also a consultant at Ernst and Young. And so I've got a, a pretty wide, uh, degree of technology experience in building, uh, both enterprise grade, uh, technology as well as consumer grade experiences. And, you know, I grew up a very creative spirit.

Uh, I don't know if you can see the guitars in, in the back there, but, you know, I've, I've got, uh, a little bit of a creative streak. I've got the hair to, to kind of attest to that as well. Um, and, you know, I, I fell into generative ai.

I started working on an open source project. Uh, there was a gentleman, uh, who actually is the head of adaptive oncology in, uh, the Ontario Institute of Cancer Research. Uh, and he threw together a quick stable diffusion CLI app.

And it was one of the first, uh, tools that was available in open source to be able to use and run stable diffusion locally, which is an AI image generation model that was released, uh, with an open license. And so I started contributing to that. Um, I started using it locally, and I fell down the rabbit hole just became my, my full-time hobby.

Um, and, uh, my, my wife will attest that I stay up way too late playing with this stuff, uh, even today. And, you know, in the months that followed that, we built a very large community and built a tool around this growing ecosystem of image creation. And we watched a lot of the technology unfold, and we were building a tool that allowed artists and creatives to do a lot more with it than, you know, people were kind of widely talking about.

I think, think a lot of people think of this technology as you type in a text prompt and you get a picture out. And we were really challenging a lot of those assumptions about where this is gonna fit instead of a professional workflow. And that eventually turned into Invoke, uh, became a, uh, startup earlier this year.

And we are, uh, actively growing and, uh, working with primarily kind of the, the creative, uh, fields, entertainment, uh, things like that to help them run and deploy customized models that are trained on their IP on top of this openly licensed model to their teams. And I'll share a little bit more about the technology and the reasons why those companies are starting to look at this as kind of a fundamental aspect of their business and what they're doing. So let me talk to you a little bit about the state of creative ai.

Where are we at and how did we get here? If you go back to the seventies, you know, people have been trying to create stuff with computers for a long time in this kind of generative way. And you have, uh, Aaron from the seventies, which is to me is remarkable given, you know, that the amount of technology, uh, advancement that's happened since the seventies, it's still pretty cool to see something like that come out of a computer back then.

But I think it was really when we started to see the generative, uh, adversarial networks, uh, with the gans in 2014, where we started to realize that this technology and machine learning generally could be producing images that to us look authentic, look real, look, look like, they, they're useful and valuable, they're aesthetically pleasing. Dolly, which is open AI's image generator, uh, first hit the scene in 2021 and offer this ability to prompt, uh, the model to, to say, I want a fox in a field at sunset in the style of Monet. And, you know, the quality was relatively poor, but the controllability through a prompt was very exciting.

And a lot of people were very, very impressed by that. Dolly, too, in 2022 really kind of advanced the, the state of the art in that sense. But this, uh, this tool, midjourney was kind of evolving alongside that.

And Midjourney has, I, I think probably one of the best image generation models to date. Uh, and they still kind of maintain that lead today. Um, but all of this technology was kind of moving in a closed source and kind of proprietary model ecosystem.

You had these kind of large providers who were offering either an API or an interface that would allow you to use these models, but you weren't really getting exposed to the underlying mechanics, the code and how it all worked. And that all changed in August of 2022. Uh, stable diffusion is a model that was released by a couple of, of researchers, runway ml and the company Stability.

They released this model openly licensed. You can look at how it runs, how it runs, inference, you can take the model weights. There are like a couple of gigabytes of a file that you just sit on your hard drive and you run inference on it and it produces pictures.

And that was kinda game changing. It, it made this accessible in a way that was kind of before then, uh, not seen at all. But it also added the ability for researchers to find new ways to use the technology, ways to advance the capabilities of that model.

Because the model offered this kind of evolving capability set when you went in and tinkered with the software that runs the model. 'cause the model is effectively just kind of an understanding of words and pictures and all of this kind of stuff. It's this very condensed, uh, dictionary is the, the analogy I like to use of what do all of these words mean in a visual, uh, space.

And within weeks, I think the, the example I have here is seven days later, people were coming up with new techniques to inject content and, uh, new concepts into this model or, or pull out those concepts in novel ways. Um, you know, so we, we saw the ability to create a text embedding, and that text embedding is essentially like I'm creating my own, uh, dictionary entry, a word that means something to me, and then I can use that in prompts, right? So I'm, I'm kind of creating this symbolic representation of what I'm trying to invoke from the model.

And I'm able to get that there was also, you know, new capabilities of, you know, how the text prompt could be manipulated and how the attention of the model could be manipulated to get different effects in the image. And so in the example below, we've got a forest scene. What if somebody wanted to see a winter forest scene instead, being able to kind of go in and augment that generation, uh, kind of point in time was, was relatively new.

And you'll look at the dates and you'll start to see what I'm talking about when I talk about the speed of technology advancement. Uh, August 29th was when the first textual inversion paper came out. Uh, SSD cross attention was on September 9th.

So we're talking about like major, major pieces of research coming out within a matter of weeks. And that type of research has continued to come out, and it, it's kind of continuing to advance. And, and I, I won't go into every research paper that has come out since then, but I'll just say there's a lot of lot of them.

And they do a lot of really cool things with this technology. Now, where we are today, uh, Dolly three has been released. Uh, you'll notice here that this is, I, I think it's hilarious.

It's a comic of, I just feel so empty inside with like this avocado's been pitted. That image was generated by ai. The text was generated by ai.

Somebody said, I want this scene, and I want this to be the text. And the combination of large language models and image generation tools that OpenAI has access to, they're able to compose something really, really good. And that's great.

That type of controllability is awesome. Mid Journey has one of the most aesthetically pleasing models out there. It just kind of produces beautiful pictures.

Even if you give it very simple prompts, they look real and stable. Diffusion has evolved as well with a new model called SDXL. Now, again, I think the most important thing to recognize here is that SDXL can be fine tuned.

You can train new concepts into that model. And when you think about what that means is content that it wasn't exposed to previously in the wide world of the internet, that all of these models are being created from that content can now be accessed by the model when you train the model on top of it. So what that means is, for an artist or for an enterprise, if your, you know, primary function as a creative is producing new works and you have a specific aesthetic, or you're working on a creative brief, and it needs to match that creative brief, and it needs to be consistent and coherent.

The problem that you'll find with other solutions is they're kind of high variability. They're a little bit random. They don't have access to all the concepts that you're thinking of in your head.

And so this fine tuning process really changes that, and it gives you an asset that you can own. That model is something that you are able to train and make your proprietary asset, right, something that you can create with. And so we're seeing across creative disciplines, people fine tuning these models and creating new, uh, con conceptual understanding inside of the model for the use of creating with ai.

But that's not all image quality is, is kind of becoming commoditized. People are starting to expect that images just look good when they come out of an AI system. Like that's just kind of something they expect.

But you can't, you can't commoditize the creativity of producing an artistic image, right? Something that was designed well, something that evokes emotion or is aligned with kind of the intent of a project. You can create all kinds of cool pictures, but if you talk to many professionals who are using tools like midjourney, they'll tell you it's like, this looks great, but I don't really feel involved in the process.

I don't feel like I've got a lot of control. It's, it's a little bit random, and a lot of times it doesn't really match what I'm trying to go for. And that really means this kind of direction of creative control over AI is where the real opportunity is for professionals.

And that is what we have been focused on. And really looking at kind of incorporating that type of research in a usable way for artists and creative teams. So I'll talk a little bit about what that means and how we can control creative AI today.

And then we'll talk a little bit about, um, how, you know, teams are looking at managing these assets from an IT perspective and looking at this as kind of like information and intellectual property that needs to be securely managed. And then I'll just share a couple of use cases that we're seeing kind of in, in, in the wild, uh, to, to maybe stir up some, um, creative juices in your own head if you're interested in it. And, uh, then we'll conclude.

And if I, uh, am able to make the live session, I will answer your questions. Um, so the first thing I want to talk about is the core concept of diffusion. Um, all of these models that are kind of the modern state of the art for image generation follow this diffusion process.

And diffusion is really a way that describes how these models are trained and how they generate new imagery. So the way that the model is trained is you have this picture, this large data set of pictures, so not just one picture, but you know, hundreds of thousands of millions, billions of images. And each image has a paired text description of what is inside of the image.

It's basically a image and text pair the model or the machine learning, um, process looks at the image, slowly get transformed into noise. It's, it's essentially creating a relationship here of this image looks like this noise once we get to completely like disintegrated imagery. And then it's challenged with the, the task of recreating the image that it saw with only a text description and the noise.

So basically it says, here's a picture of noise. Here's the test text description that describes that image we just saw, get disintegrated, go make me the original image. And that happens over and over again.

This is kind of iteratively training on how do I get really good at turning noise and a description of what that noise should look like into an image. And it gets so good. It's kind of like, you know, the magic of machine learning.

It gets so good at doing this that you can pass it a, you know, random set of noise and a text prompt, and it's gonna squint really hard and figure out how to turn that into an image. Um, and that's, that's kind of like the, the, the crazy thing about all of this is it's just, it's gotten really good in this, this property of being able to create something from nothing. It's kind of emergent from this diffusion, uh, process and, and training.

And so when we ask ourselves like, what is a text prompt? What exactly am I doing? Well, I'm passing it a set of words that the model has developed an understanding of.

I, I used the analogy earlier, uh, earlier of a dictionary. It's looking at each of the tokens here inside of this prompt and trying to remember, okay, like, I've seen a hundred sunflowers, I've seen a thousand cars. Like I've got a good idea of what those types of things look like.

And so if I'm thinking about what a photo might look like and what the sunflower on a car might look like, I'm gonna squint at this noise and try to like, you know, make an image out of it. And I'm anthropomorphizing obviously a little bit. Uh, but it, it is sometimes hard to escape doing that when you see stuff that, that works that way and, you know, over a series of steps.

It's, it's looking at that noise and creating the image. And I'll, I'll actually even comment 'cause this happened within the last week. Uh, that series of steps has been reduced successfully to one.

So it'll literally take noise and go straight to the image. Um, it's kinda like one of the speed and optimization, um, technologies that's, that's being, uh, developed right now. But all of that's happening because it's been trained on the ability to, uh, take the noise and create the image.

And that's really what's happening here with the text prompt. So when we ask ourselves like, well, how would I, how would I get words that it doesn't know into that process? Like, how would I be, um, able to, to train it to understand my project, the thing that I'm working on, if I've got, you know, a specific game or a movie and there's a place and I want it to learn about that place and show me more of the things that might come from it, or in that style, I kind of need to be able to teach it what those words mean.

And if I'm creating something very specific, if I've got something in my mind's eye, I probably also want to be able to control the generation in some way. I wanna be able to say, this is where things ought to be. And, you know, a lot of, of, a lot of the early technology here was trying to tease out, can we get better, you know, uh, text prompts that do that type of compositional control, you know, car way, way in the back and, you know, person way, way up front and, you know, close up portrait or whatever.

And those words work like they help you compose, uh, a little bit in the image, but most artists really want kind of that tactile, very specific, uh, degree of control. And we have that capability now. So there was a technology, uh, developed in, I think it was March or April of this year, uh, by a student, um, or a group of students at Stanford.

And the title of that paper was called Control net. And what they discovered was that you have the ability to train sidecar machine learning models and inject that as conditioning in the generation process. What that means is you can take a sketch or a rendering if you're an architect or you know, you're working on a 3D game, you've got like a 3D model.

You can take these, these, um, what what you might call prompts in some sense, the visual prompts of like, I want something that looks like this or is shaped like this. And you can pass that into the generation process with a text prompt and have, you know, completely finished products come out the other side. So when you think about what this might look like in an artist workflow, if they train a model on their style, uh, which I'll get to in a second, when we talk to, uh, the second, uh, picture here, if they, they take that capability to generate in their style, and they pair it with the ability to, you know, finish out a sketch or, you know, you know, finish my idea.

Here's what I'm thinking. You, you talk about shaving weeks off of production timelines for creatives, and you increase the amount that creatives can produce, which increases the quality of creative projects. If you've ever worked on a creative project, one of the biggest challenges is finding the thing that really feels like it's gonna work.

Um, you know, a lot of times you're working through thousands of ideas, but not all of those thousands of ideas ever make it to the presentation, because you actually have to spend a lot of time to make any one of those look good enough that the creative ideas actually conveyed. Now, you can actually do that really, really quickly, and so you can increase the, uh, amount of diversity in options, uh, but you also reduce the timeline to get the thing actually made. You've got a lot of people who are using this to, to create the asset that's gonna actually go into the game or the movie.

And so there's this, this rapid, um, kind of accessibility of creativity. And the, the productivity of creative production is, is being, I would say, is in some ways revolutionizing what it means to be a creator. What it means to, um, be able to produce something that people want to consume.

Um, if you think, if you think about like a lot of indie creators, especially, uh, you know, it, it takes a lot of time to build something that has music, has moving animations, has good art, has, you know, a, a fun game. Like we're talking like an indie game developer. And now they actually have these tools that turn that from like a five year gamble into some something they can really realistically do in a short amount of time.

And I think overall, what this is going to do is increase the, uh, increase the supply of high quality creative content that is specifically interesting to us as individuals, rather than kind of like mass media, lowest common denominator. Uh, formulaic production will get these really, really interesting ideas that are able to be, um, made because they're accelerated with the power of ai. And again, it's not replacing those creators, it's giving them tools to accelerate their creative vision.

Now, on the right hand side of this, I've, I've got this note of semantic and conceptual control. And this goes back to that idea of being able to inject new content into these models. Now you, you've got people doing this for any type of image.

So you can do this with photographs, you can do it with, you know, fashion. So if you're a fashion, um, company and you're producing like, you know, new, new assets, whether it's, um, designs for fashion or you know, a product photography that's gonna go on the e-commerce website, you can use this generative technology to produce those assets. You just need to train it on the thing that you are trying to create.

So, um, you know, I'm just gonna go down the rabbit hole in this example. You might have a new product, like a new backpack that's coming out in the fall, and it's the best backpack that's ever been, uh, created. And your options to, you know, create assets that are gonna go into the marketing for that are go out and take a bunch of pictures of this backpack on a mountain and in an, you know, beautiful lake and all of these places around the world and on, you know, 45 different people to show all of the different options that you might want in your marketing materials.

Or you could train an ai, uh, model to understand what that concept is. This is our backpack concept 55, you know, if you want to like, give it a, an a product ID or something like that, and then you can prompt for that on a mountain in a lake, um, you know, being worn, uh, to school or whatever it is. That type of creative flexibility is what AI allows for.

And when we think about the type of control that that can offer to any type of creative production and pipeline, that's what's really interesting about this technology, uh, to me. And, and I think a lot of where we are focusing our time is, is providing the tools to allow enterprises to deploy that to their teams with flexible pipelines, flexible models. You know, I think a lot of the AI startups out there right now are trying to build some, you know, big mega proprietary model, use their customer's data to make that model better and kind of like just sell access to it.

And our view is that most enterprises are gonna wanna own this asset if it's, you know, their, uh, generative capability, especially if it's like sensitive intellectual property in the entertainment industry. You want to own that asset. You wanna own that model.

You wanna have that something that you can reliably use in perpetuity, especially if you're gonna invest thousands or tens of thousands of dollars training these things. And that all means that you need to be thinking about how you are managing base models and other adapter models that you're going to use in your creative pipeline, because these are what control the language and concepts that are available to the system during generation. And they're also the kind of like the asset or the, uh, you know, the, the workhorse engine behind the generation of these images.

Uh, and like I said earlier, this is kind of like a dictionary. So if you've got a dictionary that's trained on all of your content and your ip, is that something you wanna be licensing from someone else? Uh, my my thought is probably not, that's probably not true in all cases.

I think it's kind of an, it depends, is this a core capability for the business? Is this something you really need to own and think about as an enabler for the business to be successful? Or is this kind of like a nice to have, it's like the marketing team just needs, you know, stock image replacements, uh, stuff like that where it's not super critical to the core.

Um, the core of the, the business, and, and this is kind of where I start to, to piece out this, this model of closed models and open models in the closed model ecosystem. If you look at something like open ai, the base model and all of the pipelines and workflows that are generating images with that, it probably the control adapters as well. Things that like allow you the ability to say, um, you know, more detail, less detail or, uh, you know, here's a picture, kind of take stuff from it.

All of those things are gonna be platform proprietary. They're gonna be locked up and you're gonna be provided access to maybe augment those with embeddings and concepts. So there is the capability of, you know, maybe plugging in your concept into that base model.

And so this notion of like, my product or my style is kind of like a tinier model that plugs into that base model. But the challenge from a strategic perspective, when you're looking at the technology that way, is that, that concept model is extremely compatibility dependent. It has been trained on top of that base model.

So if you no longer have that base model, that concept model is worthless. It doesn't plug into anything if you don't have it, right? So this is where when you're thinking about, well, how do we maintain this capability and perpetuity?

You have to have the base model, you have to have the model weights, you have to have that file that this thing's gonna plug into. And you might want to train that too, if you want to spend the time and like really building out a very strong model that's capable for your business. It is a differentiator.

Um, but you need, you need access to those model weights in order to be able to have reliable, uh, functionality of these concept models. But when you own that, that model weight, you also have the ability to augment the workflows and the pipelines and the controls that you have, uh, available to you. So let's say for example, you know, I mentioned an artist taking a sketch and passing that in and having it finish it out.

Well, the control adapter has been trained on what sketches should look like when they're finished, right? Roughly approximating what it does. Well, what if your business produces something early on in a conceptual pipeline that has some standardization to it?

Let's say you're an architect and you've got blueprints, right? And you want to train a control adapter that turns blueprints into the final rendering, those types of control adapters could be created, right? The, the control adapter is really just a sidecar machine learning model that says before and after, this is what it should look like.

So if you wanted to, you can invest in that, and that can be a differentiator for your business. And those are the types of things that, again, are only available to you when you have an open model and, and kind of a tool that allows you to deploy those. So when you think about the process of creating those models in the open ecosystem, you're curating a data set.

This is what we want to train into this model. This is the capability. We want to, um, empower this model to be able to generate with you, develop the adapters.

You might develop a concept adapter, you might develop a control adapter, something that's gonna help you with that generative process. And then as you expand your data set, because you're able to generate now with ai, your data set can get bigger. You actually have the ability to using human feedback, curate that data set and create an entire model off of that.

And then it becomes really, really good at doing that thing. It becomes very specialized. And this is like how most of these creative models are being developed.

And, and there's a lot of, um, there's a lot of research going into does synthetic data, data that's created by ai, uh, is that something that can be used to further train it? And, you know, some research papers come out and say, no, if you feed itself on its own generations, it collapses. The model just gets to garbage.

And that's true. If there's no curation, if you're not fixing the imperfections, if you're not, um, you know, correcting for quality, if you're like picking the top 10% and you're making sure that it's a really good data set, and you're also kind of like going in and augmenting that in some capacity, you're able to increase your dataset pretty, uh, substantially and therefore increase the capability of really honing in and tuning this model to a specialized use case. So we think now of this model as the next generation of intellectual property, and there are a lot of open questions around whether training data can be used the way it has been.

A lot of these models are trained on like the public internet copyrighted materials, so and so forth, GT four for example, trained on everything under the sun. But regardless of how prior models have been trained, future models will be trained and will offer these capabilities. And, you know, again, you're gonna have to think about this as, as information that has been consolidated and compressed into an AI model or a machine learning model, and it is now able to generate with that knowledge.

But how do you protect the assets that are trained on your private information? That model is now a very valuable piece of intellectual property. Are you gonna want to just deploy that, uh, everywhere?

Are you gonna wanna send that file to every employee on a USB thumb drive? Probably not. Um, how do you ensure that you can use that model forever?

What if one of these services goes down? We had an open ai, uh, you know, potential implosion, uh, a week or two ago and, you know, people were asking the question, well, I built my business on open AI's APIs, what's gonna happen to them? So you, you have to look at this as like a core capability for the business and ensure that the long-term operability of these investments is, uh, protected.

And you also wanna make sure that the information generated by the models is protected. You know, a lot of people have started to do, uh, red team attacks on GPT-4 and have showed that you can actually get a lot of the training data out by using very silly tactics like having it just speak the same word over and over again until it starts breaking and just starts spewing training data. And the same goes for, for images when you're, uh, training this model on, on, you know, sensitive ip.

Let's say you're an artist and I'll, I'll just like go down to the individual level. You're an artist and you have your style. Well, if you allow somebody to generate with enough like, you know, intelligence and, and kind of, uh, how the prompts should be structured and how they're able to kind of like hone in on certain trigger words they could potentially generate with your model, a lot of images that they could then use to train their own model on that thing, right?

And so like, when we're talking about capabilities and differentiators as a business, if this is a, uh, an investment that you have made that you believe is like a significant reason why your company operates better than someone else, you have to protect it on all of these diff different dimensions. And you have to treat it like the confidential piece of, of technology that it is. This becomes something that you want to have in Fort Knox in some sense.

And so, you know, the National Cybersecurity Center in the UK and the cybersecurity and infrastructure security agency in the US as well as I think 16 other countries all collaborated on joint guidelines for secure AI system development. And I have included the QR code, so you can read those guidelines. Um, but effectively the four big bullet points are secure design and development.

How are we ensuring, uh, early on in the process that these AI models are, uh, gonna be okay and secure for us as an organization? But the real piece that I think was maybe not on everyone's radar was this deployment operation and maintenance of this that kind of like life cycle of what happens once we've made that model. How do we make sure that we're deploying that securely?

And, you know, they talk, they talk about things like ensuring that you've got the ability to do audits on what something was trained on, uh, audits on how people are using the model, how information is leaving that model, and like what, uh, generative tactics are people using because you might be able to find something that's maybe malicious in pulling information out of that. Um, and so this becomes a new data exfiltration vector for attacks. And operationalizing that in the business is not just throwing that model out and hoping that it works forever.

'cause you're gonna have to maintain it, it's gonna have to get updated, it's gonna have to get improved as, as more data gets created and human feedback comes in. And so you really have to look at that, uh, as a holistic strategy that you're taking and make sure that you're working with a partner that can help you with these things. Um, and that, that's kind of one of the big areas that it, at least in the most, um, recent few months has been getting a lot of attention, especially from investors, is companies that do this with both large language models and images.

And, you know, this is kind of what invokes, uh, doing on the image generation side, but there are companies that do this with large language models and other modalities are coming soon. So really any type of secure deployment and operation, I anticipate that there's going to be a partner that you can work with that will help you with these things both in, um, ensuring the quality of, of the models that you're training, but also that you own the asset, that it's your model and that you're deploying it in a secure environment, uh, in a usable environment for that team. So, uh, I'll conclude by just sharing a couple of fun examples of how creative AI's being used, different use cases.

Um, you know, we've got an open source community of people who are using this, uh, to create everything under the sun, and we also have a lot of professionals who are using this in their work. Um, you know, we, we, I think our core, uh, user right now from a professional perspective, at least the, these are the earliest adopters of ai and they're really pushing this forward in a lot of ways, uh, are game concept artists. And so, you know, you have, uh, accelerated concept and capabilities where an artist can pass in a sketch and you can turn that sketch into something that looks like, oh, I get it.

I get what you're going for. This is super cool, right? Like, there, there's a whole lot of of capabilities there.

You can also do, you know, 3D texturing using these tools, throw in a model and say, this depth map, this rendering, help me texture it, right? Help me get the, uh, that, that final texture on this so that we can really see what this would look like as a final asset. You know, these, these tools and technologies are again, fully customizable.

So if you want to go into a, uh, model pipeline and build a model for your business that does 3D uh, texturing really, really well in the style of your games or your movies, whatever it is, uh, you're an architect, right? You can do that and, and really control the quality of that output. Uh, an artist in, um, Hamburg, uh, recently did a couple of exhibitions using Invoke to generate, um, his art and photography.

So the two images that you're looking at here were generated with ai. Uh, the first was trained on all of his artwork in this kind of series of, um, kind of like forestry that he does. And he produced an entire exhibition that was produced completely from his AI model and kind of curating that model, teaching it what his art style was, and then using that to invoke new generations.

He did the same thing with his photography and did this entire like, photo, um, exhibition, but none of it was real. Like none of those photos were real. They were all AI generated.

But, you know, the people who kind of interacted with him were, I, I think probably pretty astounded by the amount of emotive capability that he was able to instill. And it's because he knew how to use these tools to control the generation and really get what he was going for. Uh, and, and I think that's kind of the, the big takeaway from a lot of this is you do have control in the creative process of using ai.

It's really just understanding how the tools work together. Uh, there's a content creator who's producing an entire series, um, you know, using AI training characters in and using that to produce, um, kind of like comics and movies. Uh, you know, we had a, a comic artist who said he's using this in his work, uh, especially in creating kind of like backgrounds and supplementary assets and kind of composing those together into the final comic.

Um, and I think that this is really just kind of a, a scratching the surface of the use cases. There's, uh, like I said, fashion, there's architecture, any, any type of pipeline where you have to render or produce an image of something, uh, you know, that that's capable, uh, of being trained into these models. And we're starting to see really kind of the first, uh, hint that we're at that same point of stable diffusion from August of 2022 now with music, with video, and those modalities are coming in, uh, and I think we'll see in 2024, uh, probably the renaissance, uh, that, that I've been dreaming of, of everyone being able to create the thing they come up with.

So, uh, appreciate the time, uh, if I am around, which I will, uh, endeavor to be. I will try to answer some questions, but if I'm not, uh, I appreciate you listening in and you are welcome to reach out to me. Uh, as you can tell, I probably am the most nerdiest, uh, most nerdiest of people when it comes to talking about this stuff.

So I enjoy it. Take care.