Today we talk to Celtra Founder and CPO, Matevž Klanjšek, about how AI can be used to accelerate creativity, and what would happen if it eventually replaces humans in the creative space. We discuss the limitations of the tools currently available, why Matevž isn’t interested in teaching AI variance, and how humans and AI need to work together in advertising. Tune in to hear what the future of advertising looks like, and why the human-AI feedback loop is essential. Matevž tells us about the bizarre adverts he’s seen AI produce, and talks us through the evolution of human creativity: from paintings to photographs, and how humans stay relevant when we invent something new. Key Points From This Episode: An introduction to the founder and CEO of Celtra, Matevž Klanjšek. Where the idea of using AI in advertising came from. How Celtra technology helps creatives scale their media, accelerating creativity. Why AI is the right tool for the job. Two dangers of using AI in advertising: impacting the communication strategy, and losing uniqueness. Can you teach AI variance? Why it’s important to leave space for human error. Humanity in advertising: why brands are trying to be more human. What a collaboration between humans and AI looks like. How human genius lies in building communication strategies. The surprising results when AI tries to create adverts. Playing with generative design: how AI can inspire humans. Why AI won’t replace humans in the future.
Today we talk to Celtra Founder and CPO, Matevž Klanjšek, about how AI can be used to accelerate creativity, and what would happen if it eventually replaces humans in the creative space. We discuss the limitations of the tools currently available, why Matevž isn’t interested in teaching AI variance, and how humans and AI need to work together in advertising. Tune in to hear what the future of advertising looks like, and why the human-AI feedback loop is essential. Matevž tells us about the bizarre adverts he’s seen AI produce, and talks us through the evolution of human creativity: from paintings to photographs, and how humans stay relevant when we invent something new.
Key Points From This Episode:
Tweetables:
“It just makes sense to automate [repetitive tasks] as much as possible, and remove that from the equation, let human genius think about big ideas and communication strategies, creativity and so on.” — @hyperhandsome [0:03:47]
“I think on all of the levels, across the creative process, we always try to have humans involved. It’s almost like a basic principle.” — @hyperhandsome [0:14:47]
“So that’s the nice thing, actually, perhaps using pretty advanced AI to really inspire creativity in humans instead of replacing it. It’s kind of beautiful in a way.” — @hyperhandsome [0:17:24]
"I think the whole point of advertising, and humanity in general is precisely to be always different, to invent new things." — @hyperhandsome [0:18:12]
“I think technology always gets to a point where it can perfectly imitate, and do it better than humans can, but then we invent something new.” — @hyperhandsome [0:19:52]
Links Mentioned in Today’s Episode:
EPISODE 19
[INTRODUCTION]
“MK: I think technology always gets to a point where it can perfectly imitate and do it better than humans can. But then we invent something new, and we want to do something new.”
[00:00:15] RS: Welcome to How AI Happens, a podcast where experts explain their work at the cutting edge of artificial intelligence. You'll hear from AI researchers, data scientists, and machine learning engineers as they get technical about the most exciting developments in their field and the challenges they're facing along the way. I'm your host, Rob Stevenson, and we're about to learn how AI happens.
[00:00:45] RS: Today on How AI Happens, we're going to take a look at creativity. What AI can replicate, what it should replicate, and what humans will do when an AI as creative output can replicate their own. I'm joined this week by Matevz Klanjsek, cofounder and Chief Product Officer at Celtra. Celtra utilizes computer vision to help generate online ads at scale. But Matevz spends just as much time thinking about the relationship between technology and artful output, as he does thinking about how to automate image cropping. Matevz explains how he and his team understand the relationship between automation and creativity, how he expects human output to shift as technology improves, and even the results of some of Celtra’s experimentation with generative adversarial networks or, GANS.
[INTERVIEW]
[00:01:34] MK: Perhaps, first few words about Celtra. So we are a creative automation company. We're building software that helps brands, media owners and agencies to scale their creativity, which is kind of a fancy buzzwords to say that we help those companies to create more ads faster, and make those ads better. So it's essentially taking the whole creative process and really kind of accelerating it, scaling it as much as we can.
And we work with a wide range of companies from top digital brands, Spotify, Adidas. One side on the media sites with big publishers like Wise, NBC, etc., etc. So the main goal, the main mission really is accelerating the creativity, scaling the creativity. This is also where kind of my background kind of comes in, right? So before Celtra, my previous life, I was a creative director, and kind of seen – That was decades ago, and really transformation to digital. Already back then, had the sense that the tools that were available, or the design tools, creative tools, can really address those challenges. So challenges of scale, challenges of content used specifically for digital, great advertising, so on.
So I would say perhaps the big motivation or the big mission, at least for me, personally, behind this was – I think, advertising, I think a lot of fantastic advertising. We all think of we're all fans of Superbowl ads. But on the other hand, think of all of these banners back then. A lot of really crap advertising. Most of the advertising, most digital advertising is really crap. And really thinking, could we help bring some of that creativity that we see in Superbowl ads? Fantastic can line sets to much, much, much bigger scale and kind of raise the quality of advertising across the board. That's the main challenge. The main mission really.
[00:03:37] RS: Got it. And so was AI the right approach for you just because of the scale when you think of banners, for example, and all of the places where ads are displayed. In the Superbowl example, you know, it's over a four-hour period on television for 30 seconds at a time, right? But when you think of just the breadth of the advertising industry, it's everywhere. So is that why AI was the approach for you?
[00:04:01] MK: Scale is absolutely the main factor, right? And if we look at the main challenges that we are seeing here, so on one hand, between just an enormous number of repetitive tasks. So when you're producing advertising at scale, you're doing the same thing over and over again. This is a lot of times very routine stuff, cropping images, kind of breaking the text, things of that nature. These are relatively simple tasks, annoying tasks, very time consuming. Something that humans generally don't like to do very much. And it just makes sense to automate that as much as possible and remove that from the equation. Let human genius think about big ideas and communication strategies, creativity and so on. And these tasks are also – Because of their simplicity, very learnable. And really great fit for automation for AI. So that's one consideration. Scale, absolutely. We're all about scale. We're all about scaling the production, scaling the creative excellence. So it just makes perfect sense to kind of look into AI, look into automation, help to achieve that.
[00:05:10] RS: What are some of the more advanced ways you're pairing with, for example, the display of the ad and the influx need to have human creativity in these ads to make them something that can be engaged with?
[00:05:21] MK: When it comes to automation, when it comes to AI, we’re deliberately focusing on relatively simple tasks, precisely because as soon as you try to expand this into more complex tasks, into more kind of complex creative thinking, there is a danger that you will start impacting, potentially negatively, the whole communication strategy and message, right? So we want to focus on scaling sizes, cropping images, treating images.
So typically, we see our approach from two ends. One is, really, this where we're mostly focusing, preparing the ingredients for scaling, right? So preparing texts, or copy versions, preparing visual materials, images, videos, and so on, and really treating those. So it might be cropping images. It might be resizing images. It might be extracting certain properties from the images that would be used later.
So example of that would be you have a product shot, and you want to extract color or texture from it to use it in other elements of the ads. So maybe use the color also. Then the color of the headline, or the background, things of that nature. So these are, I would say, non-dangerous things for AI to do. But when you get to something that starts impacting the message, there is a danger, right? There's a danger of oversimplification. And what we're seeing a lot is that, especially in advertising, to try to fit the content or the outcome of the process into the technology that's available, right?
So we've seen a lot of simplification of advertising in the past 5, 10 years. I think a lot of times, we just end up with a hero image with the back shot, a bit of a copy that nobody reads, a logo. And that's it. And advertising needs to be engaging. Because as you said, needs to evoke certain sorts of emotions. It needs engage consumers. And when you simplify everything to something that's really bare bones, maybe felt as though you're relying too much on the brilliance of one image, right? So that's one danger.
The other even bigger one is really that, eventually, all of the brands start to look the same. And you're using the same tools or using the same automation processes, Ais, and so on. And it starts to look the same. And all of the brands look the same, which defeats the purpose of advertising. Because that is precisely to stand out, precisely to have your unique voice to be the brand. And then you look at the ads within one category. They all look the same. And you end up advertising a category. Let's say all of the direct to consumer cosmetics company, they have the same ads, and you're not advertising your particular brand anymore. And that's the second danger. Kind of related to the simplification, right? With this simplification, also, what you end is kind of unifying all of the output.
So we are careful how far we go and try to avoid automating too much, or relying on the AI too much when it comes to us to do impact communication strategy and message. So help with the process. Help with the simple task. Belief to humans to really think about the big ideas.
[00:08:56] RS: When you give the example of how important variances is, is variance not something you could also teach an AI to do? Is variance – like is it a matter of color? Is it a matter of the particular emotion you're trying to evoke? Is it a matter of imagery? Like could you break that down into building blocks? And is that a long-term goal?
[00:09:15] MK: You absolutely could. And some of that we are covering through our workflow automation, right? So you kind of try to systematize as much as you can see. Bring in the ingredients. You kind of build creative taxonomy around. You kind of try to help with the rules, which ingredients can come together to create precisely that variant through all the different permutations of that. But I think when it comes to uniqueness, you're just reproducing perhaps the same problem of sameness only on higher level.
So we feel strongly that there should be human control. And we should even leave space for human error for, human flaws, if you will, because this is – It's very humanizing, so to speak, right? You're trying to create this emotional connection with consumers. So brands are generally trying to behave like humans, right? I'm a brand. I'm your friend. We have a connection. So you want to have some of that humanity. You want to see some of the things that are not perfect, right? Imperfection, sometimes it's very humanizing. So you don't want super sterile communication. It betrays a lack of humanity, lack of connection. It reduces the opportunity for it.
But back to your question. Yeah, of course, you can get to that kind of variance. You can help inspire that kind of variance. And you can help once you have the idea where you want variants to be. You can absolutely help scale it with AI. And I think this is where you want to come in, right? This is where you want to have most impact.
[00:11:09] RS: When you think of the ideal partnership of this technology enabling scale, and then also the application of human in the loop, what is sort of the handoff there? And what is the ideal relationship?
[00:11:23] MK: Human in the loop for us is almost essential part of that, not just in terms of creativity, but also in those small tasks, right? So you want machine to learn from humans. So even when you're doing cropping, you want machine to learn from humans how to do really great crops, right? It's a trivial task. And you could just have a really good training data and so on. But you want – Things change, really.
Ideally, you build that into feature. So when machine does the crops, we offer random sample to users to really just kind of look at that. Maybe adjust something. Maybe kind of correct the machine, so to speak. So even in those tasks, you want to have humans in the loop, right?
When it comes to, let's say, big creativity related things, then absolutely. And you always have automatic feedback loops through consumers. So creatives in the end go to the market. They perform or not perform, perform better, less, less well. So we have this loop automatically built in just by the nature of advertising. But also, human genius is building communication strategies. But again, machine can kind of help with recommendations. What worked best in the past? What could be the combinations that could work best in the past? Predicting the performance. Fairly standard stuff, I would say.
But here, for us, again, we try to kind of build all of that on top of a very clear, creative taxonomy. And this is, again, something that humans would do, right? So kind of the labeling. I mean, machines can automatically do the labeling. But really thinking of what the taxonomy is on the relevant factors, etc., etc.
And again, I would say, gain nice collaboration between, so to speak, humans and machines, right? Software can also help you find those factors, right? So what could be potential factors that I can't think of, but then human would make a decision? Okay, these are really some good ideas. I will use that moving forward. So I think on all of the levels across the creative process, we always try to have humans involved. So it's almost like a basic principle.
[00:13:36] RS: Yeah, yeah. It makes sense. Companies are using generative assistive networks to assist humans really and learn in more creative ways and have more creative output. Is that something you experiment with at Celtra?
[00:13:48] MK: Yeah, we're playing with generative design, doing experiments with GAN just to see the outcome. The expectation is not that, again, would just go and produce really functional at something you could take to the market. It doesn't, right? So ads are really complex creatures, right? So there are multiple layers of meaning that are very multi-dimensional in terms of medium in terms of – So you have copy, you have the imagery, and this needs to work together, etc., etc. So it's a pretty complex situation. But we are playing with that just to see the outcome, just to see if something interesting, something surprising comes out.
And what we are seeing is kind of expected, but really nice to see. So you get outputs that are somewhat alien, somewhat abstract. But at the same time, very familiar. So they look like ads. You see that image, right? And you immediately say, “This is an ad.” And then you look closer, and it's completely weird. And this is precisely the essence in a way of GAN, right? To kind of extract the essence of what it's trying to generate. So extract the essence of being an ad. That's the beautiful thing about it, because advertising is all about being familiar. But at the same time, always new, and unique and different, right? So you get this combination of weird familiarity, but always something new, something surprising.
So we like that. We like that to lot. And we are playing with the idea to actually bring this into the product just as an inspirational tool. Not really as a production tool where you would produce an end product. But really, when you're facing blank canvas, you need to build an ad. You need to think of something new, right? There's this huge problem of blank canvas anxiety. So what should I do, right? And here, actually, GAN can help, can kind of just based on a few inputs, generates a bunch of ideas that are not that that look completely weird, but familiar and can give you an idea, right? So that's the nice thing, actually, perhaps using pretty advanced AI to really inspire creativity in humans instead of replacing it. It's kind of beautiful in the way.
[00:16:16] RS: Yeah, it sounds like it's just generating like the uncanny valley of ads, right? Where it's like, “Okay. Yeah, I can see that it's recognizable, but there's just something a little off about it.” Is that a result of just where the technology is? Or do you think that's a lasting limitation?
[00:16:32] MK: Well, I think the technology will get better and better exponentially. And eventually, in probably in the fairly near future, we'll be able to produce pretty good high-fidelity ads that will be functional in a way. But I don't think this is the end results we're looking for, right? I think the whole point of advertising, can humanity, in general, is precisely to be always different to invent new things?
And so AI, as we know it today, or it’s function today, is fantastic. It's imitating, right? So we train it, right? It imitates in a way what we do, but it does better, with less errors, faster, more efficient, and so on, and so on. And I don't see a reason why, very soon, even in production of ads, AI couldn't completely imitate all of the work that we do. And we'll just do it better and faster. But this is when we invent something new. I think this is a great example.
So you have centuries of art practice, from Renaissance, to Baroque, and so on, trying to paint a perfect representation of physical worlds, right? You have Renaissance inventing perspective, sfumato, kind of deep focus, things of that nature. You have Rembrandt's painting horses. So you can't really distinguish from the real thing. And then all of a sudden there comes photography, right? And completely disrupts this. And all of a sudden, all of this work is kind of, “Oh, there's technology doing this better than humans can. But then art reinvents itself. And all of a sudden, we get impressionism, we get the whole avant-garde. And now it's trying to show a different type of representation. And we're thinking maybe more about expressionism, emotional level, and so on, and so on.
So I think technology always gets to a point where it can perfectly imitate and do it better than humans can. But then we invent something new, and we want to do something new. So the end goal for us is never, “Now, let's paint the most accurate representation of horse.” It's just inventing something new we haven't been doing before. And technology can always follow and can always do better. But we will always find something new that we will focus on. And it will be the same in advertising. Maybe AI will do perfect ads. But then advertising will change. It will reinvent. And we'll be doing something else. I think that's kind of nice about this dialogue.
[OUTRO]
[00:19:19] RS: How AI Happens is brought to you by Sama. Sama provides accurate data for ambitious AI. Specializing in image, video, and sensor data annotation, and validation for machine learning algorithms in industries such as transportation, retail, ecommerce, media, medtech, robotics and agriculture. For more information, head to sama.com.
[END]