AI is rapidly changing user acquisition (UA), not just by speeding up creative production but by expanding how campaigns themselves can be managed. Host Devin Becker is joined by Amit Carmi, co-founder and CEO of Sett, to explore how AI is being used to combat creative fatigue, generate video and playable ads, and increasingly automate parts of the UA workflow. Amit breaks down where AI creative works well today and where it still falls short. The conversation also tackles the risks of increasingly homogenous or misleading ads, how platform-owned AI tools from Meta, Google, and AppLovin affect third-party providers, and whether creative teams are being replaced or simply evolving. Finally, Amit explains Sett’s move from data-driven creative generation into UA management agents and shares his view on what AI-driven UA could look like over the next few years.
Big thanks to Heroic Labs for making this episode possible. When MADFINGER expanded from mobile to PC they used Heroic Labs to build a preview mode with full progress carryover. Read how the work led to increases in every core KPI from CCU to DAU/MAU, retention and session length, at launch.
We’d also like to thank Medal.tv for making this episode possible. If you're a PC gamer and want to clip your moments or a studio, publisher, or marketer looking to reach a high-quality gaming audience and get your game in front of the right players, check out all Medal has to offer at https://grow.medal.tv
This transcript is machine-generated, and we apologize for any errors.
Devin: Hello, everyone. I'm your host, Devin Becker, and today I'm delighted to be joined by Amit Carmi, co-founder and CEO of Sett. Sett optimizes user acquisition with AI agents, including creating and managing UA creatives for games and apps. And today we're gonna explore the topic of using AI for UA creatives.
So just to start things off, could you give a quick overview of the origins of Sett?
Amit: Yes. First of all, thank you for hosting me. It's great to be here. Yeah, so the origin of Sett started with myself and Yoni, my co-founder. We have a big passion for tech and huge passion for games. Eh, the real dream was to open our own studio.
Eh, we didn't, eh, really aim for mobile, but if you would ask me what is our biggest dream is to have like a PC or a, like an open world MMORPG kind of game. Eh, but we understood that we don't have a, a real advantage there as founders, so we took the B2B approach and, we came to one of the, I don't know, biggest, spaces of issues which is user acquisition for mobile games.
And basically, we have a goal to optimize that via owning the entire workflow of anything besides the media buying, so including data analysis of the creative's campaigns, a creative idea generation, the gener--, the generation itself of the creatives, including like high diversity of new concept and maintaining the high quality of that, out-- deployment to the campaigns and managing the campaign.
So, the goal is to really take control of the entire loop with just to optimize the performance. Like in order to improve ROAS, we believe that agents can help at each step of the process to reach the final goal. So, we're 70 people now, most of them based in Tel Aviv, raised approximately $60 million from great investors and great partners.
Yeah, we have a lot of goals. We're only halfway through the current, uh, vision in terms of product.
Devin: Nice. Well, I'm sure we'll get to talking about some of the other half there. But what does Sett actually do for a UA team on like a day-to-day basis? What does it look like?
Amit: So, imagine that you're in a UA team, so either-- usually we work with the UA managers or the creative strategist or creative managers, depends on the, the company and the structure.
Usually, we don't work with the marketing artist or motion designers. What is happening, basically, we have a few agentic system one that is kind of the brain of the generation that analyzes data from your competitors, your campaigns, the in-playables analytics, and a general ana-- , like continued analysis of your target audience.
So, you as a UA manager/as a creative strategist, you have a way to impact the brain of what our system is going to create, whether with specific briefs or most likely defining the strategy and kind of directions to guide the agents to go forward to. And then we have the part of the generation itself, which we have our own internal tools that I would say in terms of agentic harness and, capabilities, I believe that probably the best tools in the world for generational video and playables.
We have our own internal team that are working with the tools in order to complete what AI is not doing well, well enough, and then this is how we train the model of-- to optimize the generation. And at the end, as a UA manager or a creative strategist, you get from us pretty high scale of video ads and playable ads at the moment.
High diversity, meaning we're only focusing on new concepts and to diversify as much as possible because we believe that what is impacting the campaigns are winners that are new concept and not necessarily iteration. We also do iteration, but it's the not main thing. But you're getting a lot of fresh new concept of video playables in a considerably very high pace.
Of course, lower prices than the market. And in the future, we will include more UA, UA automation. We just started with first design partners a month ago, our own platform that we also give access to, for you to generate, not only us. But yeah, a lot of things to come, but at the moment is think about it kind of a data-driven creative's flow for your team.
Devin: Nice. Well, appreciate the extra detail on that there since there's definitely a lot to that. One of your big things about your, your pitch is beating creative fatigue. So how would you define that, and how fast do you think that's kind of currently hitting? You talked about like looking at new generation versus iterative and things like that. Does that factor into that creative fatigue?
Amit: Yeah. So, first of all, you know, it's one of the biggest question and, you know, the real answer about how to find the winner and how the deep learning algorithm of the networks, you know, everything from Meta, AppLovin, Google and et cetera, is working, no one have this issue.
It's a bigger let's call it area of research that is currently happening in deep learning that no one knows how the systems that are so deep learned are making decisions on, you know, milliseconds based of billions of billions of every second signals, like data signals. So, there is no real answer for that.
But I would say that you usually identify the fatigue itself. You're seeing like either top line metrics like upper funnel or the ROAS itself getting hit in your campaigns. The usual, let's call it a medicine for that, is starting to iterate on the concept that is already working, but it's kind of a race to the bottom because you might, in a short term, impact the performance of the campaign in a good way.
Like maybe you lift it up back to your original KPIs, but in the mid/long term, which a, a tough creative fatigue from the data we gather and the partners we work with can... Like a tough one that is kind of a loophole can be, I would say month and a half or maybe two months without a real, real, real new concept winner.
And what usually revive the campaign is like a strong new concept winner. It can be both video and playable. You know, if you have both of them is great, but even one of them, uh, should be enough. And today, because of the targeting capabilities of the networks is down, so the creative fatigue is actually faster than like significantly faster than the past.
I would say that from few months, I don't know, I heard partners that five years ago had the same winner for like six months. Today, at high scale, depends on the network, it can be like few weeks. So, you consistently need to fresh your creatives and to find a new winner that engage with your, uh, with your, target audience and the user themself.
So, I would say the way to beat it is consistently feed, the algorithm with new concept of video playables that of course match somehow your target audience, like, uh, let's call it, I don't know, passion or needs. But you never know what is going to stick. So, we just need to try to try to try and this is kind of a consistent fight.
I would say that Every month, month and a half on average, you will h- face a real fatigue, and this is like how you need to tackle it. Like every month, month and a half to find a new winner, and for that you need hundreds of, hundreds of new concepts if you're like in a very high scale.
Devin: Yeah, it definitely sounds like that's a big part of like using this technology, right?
Is being able to kinda keep up with that scale. And if the fatigue gets shorter, as, as you're saying it's trending shorter, that becomes even more difficult for humans to keep up with. In general, though, where do you think AI creative is still kind of falling short? Or what, what can't it make or do? Like where are you seeing kind of the edge at the moment?
Amit: So, I would say that in general, LLMs has like pretty big, big issues that are not actually getting better. You have a lot, a lot of stuff that LLM is getting better because inherently how it's being trained. For example, code generation, like we see it in every new model, code generation is like insane.
And it's pretty reasonable to understand why, because you just see-- you can just imagine the, the number of the data you have on code generation is like insane. It's a community that has been open sourced, you know, uh, inherently in the user base so much, so you have huge, huge, huge amount of data, and this is why LLM is getting better.
I would say that where AI is currently and LLM are getting, let's call it more stuck and not improving as fast as, as the other stuff, is everything around spatial intelligence and understanding 2D and 3D environments. Of course, 3D is harder. Everything there is a big issue, like from the asset generation itself, but mostly the compositions and to fine-tune and make it perfect, like pixel perfect, is very, very, very far away.
So, I would say 3D is still a, a huge issue that a lot of the big tech companies are trying to solve, and a lot of research labs are trying to, you know, train the best model based on actually games, a lot of open world like 3D games. But it's not really getting significantly better, and this is like where we do most of the hardness of the agents and most of our manual team that is working with our system, this is where it spends the most time on.
So, I would say assets, composition of 3D scenes, gameplay in general is not very high quality. It's getting a lot of issues, so it still needs a lot of work around everything, uh, in th- these areas. Everything around innovation and novelty of even ideas, and I, I'm sure that you know it as well, it's not, it's not a game issue.
LLMs are-- it's a statistical model. Again, inherently, it's not, something that going to be innovative. It's statistic. So, you see it, for example, with idea generation, if we're taking it to the creative world, you're just getting more of the same, kind of. And to, to extract real novelty from LLM today, it's, it's very, very hard.
We actually have-- We have a pretty big research team, like over 10 people, like PhDs, that are working on two of these issues. One is the generation capabilities, but the other one is, the ideation actually itself, which is I didn't know how big of a problem it's going to be, like to extract innovation and novelty from LLMs And, yeah, it's like an open research topic and a lot of stuff that we are working on in our internal lab.
But yeah, I would say innovation, spatial intelligence, like under that, everything around the 3D world is still not good enough. And, you know, at the end, you want the users that watch the ads and the creatives, they need such a high quality. Everything should be perfect. Like, unless it's a video, you know, you have this very cool intro cinematic stuff, but at the end, you, you always go to the gameplays, like usually, and the gameplay generation, you know, it's so high quality.
It's so high quality, it's so on details, everything is pixel perfect, and this is-- sorry for my language, but this is where AI usually fuck the deal. It just, you know... It's not 100%. I would say it's like between 60 to 80%, depends on the game, and the real hard part is the 20 or the last 30%, and this part is not really getting better and better and better.
Devin: Well, that makes sense. It-- I, I kind of wonder then, so if you're talking about wanting to be pixel perfect and, needing to kind of work on these, these rough areas, do you have a process then where you're kind of like throwing away videos or, or checking them, like with some automation, or is it just more manual, like, "This doesn't work, throw it out, do it again," that sort of process?
Like, how are you sort of doing the quality checking of these videos for that perfection?
Amit: Man, we, we have a manual teams that are doing that. Areas that AI is not good enough, we have people that are doing that because think about it, at the end, our partners, they need results, right? They don't care if we have a huge lab of AI, a huge number of people.
We, in this area, we have both of the com- like both of them. Like we have big number of people that are validating the process all of the time and teaching the model, and we have, of course, a lot of AI initiatives and products internally. But we have teams that are validating everything, and we are kind of trying to, I don't know if this is a real world, but… uh, world, but agentifying them over the time. For example, we have a creative producers team that are reviewing every playable in the playable specific process we have. Every playable that goes out of the, of the system, you have a creative producer that watches, plays it, give its feedback to the system So we're trying with all of the data we accumulate from the creative producers, we are building our own creative producer agent.
And then you have the evals of how to measure quality, which is again, it's a huge, huge, huge, huge research topic. Like I don't know how to measure quality. You know, it's such a subjective area, it's basically impossible. So, you're trying to do evals for that, for agents that are trying to, to kind of hit high quality.
But it's, we're still in the people phase, and I would say really that I believe that we're probably, I don't know, the top three advanced companies in creative marketing AI space. And man, we still have a lot of people and AI is very far from this, I don't know, this trend or hype you see, like AI is, going to replace us all and no, no, no, it's, it's still not there. Like you still have huge amount of work.
Devin: Well, on that topic then, you mentioned playables a few times. I got to imagine those are a bit more difficult to create, like consistently than the video. What stuff are you seeing that's working well with creating playables and what's not, especially as playables are kind of becoming, I think, a little bit more in demand than just straight videos?
Amit: Yes. So, it is a bit harder. It's still, you know, m- significantly if you are a studio and you will work with agency, like the cost per playable and the cost per video is like, you have a huge difference. So playable is still very costly and the demand for that in terms of the network side is like- It's insane at this moment.
The number is, they require you, it's, it's really insane. And if you work, let's call it with traditional agencies, and you pay like three to four to 7K USDs for one new concept, and you need 20 new concept a month, and,, so you're reaching numbers that you can't really support. I would say that the first shot generation of playable ads is, like, getting better because it's mostly around the code generation.
It's-- Of course, it's depends on the complexity of the game and the environment and the gameplay and the assets and what stuff you have and what stuff you need the AI to complement. But considering you have huge amount of assets and you know how to structure the folders correctly, it's a lot of smalls fine tunes in tons of different areas.
But considering you do all that, I would say the first shot generation of playable is nice to reach to the 60, 70%. The other parts, you know, it's a nightmare. , You still even-- Maybe you experience that with your own, Claude code or I don't know what specifically you use, but sometimes you can tell an agent something so easy to say, like, let's take, "Make the text larger," and ma- You even, you know, let's say you have a, a playable ad, you have a text, "Hello," and you want to make it a bit larger, can even sometimes say, "Make it 20% larger," like you give it the exact number, and it just, the output is like, I don't, I don't, I don't know what the fuck it's doing.
So, you control the input process, but not the output process. So, the 70%, the 60 to 70%, it's pretty good, mostly for 2D games. 3D, still far away. The last 30% is very, very, very tough, and you can get to a lot of loopholes of, , like input/output misunderstanding and agent hallucination, that it's getting better again, but...
And again, to be honest, it's like I'm still shocked in a bad way what mistakes agents are still making. Even for me personally, it's like inventing stuff. So, if you try to build a system that is generating playables that are hitting the campaigns, like this is the real goal, not to generate playable, to find new concepts that are high quality, that move the needle in the campaign.
It's far from being fully automated, and the easy part is the beginning, and the fine-tune is the hard part. It's like the real, real, real hard part and everything around, you know, asset management, even the organization of the code. So, because you need to control a lot of about the size of the file, because you need to be under five megabyte.
And again, if you have huge envir- 3D environments, you need to fine-tune every assets. So yeah. So, I would say the beginning is getting easier and easier, but it's reaching to a cap that it haven't been able to break yet.
Devin: Well, hopefully the process you're going through helps improve that just because I imagine you're, you're working that out, figuring that stuff out by actually having to do it on a day-to-day basis.
I, I'm kind of curious on something what you mentioned around the, the having all the assets. Are the assets provided by the game developers? Are you generating your own assets? Like, how does the whole asset side of it-- 'Cause playables obviously need assets in a different way than video does, as you, as you touched on a little bit.
Amit: So, both of them, of course, need, we need assets. The question is, like you said, like what is the quality of the assets, and then where do we get it from? You have some IP kind of games that we work with, for example, so we can generate stuff with AI, so putting that aside, so they provide everything, and if we need something specifically, we can buy it, us as a company.
We will co- , confirm it with the customer, and this is the process with IP. Usually, we get a big, let's call it a chunk of assets from the customers, but then it's mostly, uh, like our own models that are generating assets. In 2D, it's like, again, way easier. 3D, it's still complex, and if you will do like Let's call it idle arcade, all of the Playrix ads and, you know, the complex 3D ones with a lot of progression mechanism and a lot of animations, a lot of different assets, a lot of things that are happening at the same time.
Everything 3D and beautiful and on point. Th- this is the area where AI is still not there, so we have two options. Asset, one, we get the assets from you. The second one, we, we have people that are actually d- creating them. And again, I'm speaking about the perf-perfect stuff. We, we utilize AI in the most customized way possible for each of our customers and games.
But sometimes again, you reach to a limit that you must have this final, or a lot, or not only final, but like a lot of manual work. So, we support it with manual team and, like getting the assets from the customers. Like animation and stuff like that is still, again, far, far, far, far away from being solved, mostly like in 3D characters, to being solved with AI or Mixamo, like general marketplaces for animation.
It's, it's good, but imagine that it's so easy today, like everyone is saying, it's so easy to create ads, right? And the market is kind of flooded with tons of numbers of games, and so the competition of each user is getting harder. And if you will do an ad that is just okay in terms of quality because it's easy to create, so it's not going to last because your competitors are understanding that you must, getting the bar of quality and power-- , bar of concept and the ideations and the level design way, way, way, way, way higher.
And the only way to do it today is with real processes and real validation of people that are still in, you know, subjective judgment of what is good quality and what is not. Because agents and general LLMs are still not there.
Devin: Well then I guess that brings up the next question then, i-in the sense of like why it makes sense for all the processes you have being like, I don't wanna say boutique, 'cause you're not boutique necessarily, but like being a little more hands-on, a little less of a giant enterprise sort of thing right now, compared to Meta, Google, and AppLovin that all ship their own AI creative tools now.
So, I, I'm trying to figure out where, where do third parties fit in? And it sounds like some of it's having that, that sort of human supervision and sort of client by client handling. Like, what do you see in general in terms of the, the overall market growing with, you've got the, the super easy ones shoved into every ad product, and then you've got the ones that are like, "Hey, we understand games. We understand how to do all this."
Amit: It's a good question. I would say, and again, this is a, you know, I don't have the truth, right? So, I can-- I believe that we will win this competition based on few assumptions that I have. But of course, I, I mostly hope that, and we're working towards differentiate ourself.
But I would say that my assumption is that the networks themself, the Meta, the Google of the world, the AppLovin in the world, it's very, very, very smart from them to generate their own, generation tool, right? I would say that I know how big of a challenge it is to reach to the really high quality of the ads, and I think for them, and, we're in deep touch with them, so I also know this is the reality, they are focusing on the 90% of the studios or the apps that are kind of in charge of the 10% of the revenue.
They want to lower the bar for interest for their network. So, they want to come to a studio, very small one, an indie one maybe, that has a new game or an app, you know, they're just outside and say, "Hey, open a campaign with us. Let's do a quick AI interactive playable." It's not, not complex, not 3D. W- you don't need this stuff for the beginning.
So, I would say they're lowering the threshold for hand-- interest, and this is where AI is, is kind of completed, let's say it like that. It's easier to ship tools that are AI generation-- , generated videos or playable ads that are just enabling you to be part of the game. But to win the game or to, to, to service the 10% of the studios that are in charge of the 90% of the revenue, that the competition between them, you know, is fierce.
There you need the real tools, the real supervision, the real validation, the high, high, high quality. They don't, you know, they, they don't compromise anything. Everything must be perfect. And this is where I believe the networks won't focus on, because it's too big of a focus. It's-- I know, again, it's a huge technical issue to solve.
It's not something that you put an engineer on, or 20 engineers and you solve it. You need to put ML engineers, you need to put AI researchers, you know, it's, it's just a lot of stuff, and it's not, in my opinion, the main bus-- the biggest business value they can get because they can get more value from more, studios/companies that are now advertising on, AppLovin, Meta more easily because the interest is, bar is, kind of low. So, this is my kind of prediction/assumption there.
Devin: Oh, nice. Well, well, relating to that then, like if everyone's just generating thousands of different concepts, like, you know, all these different companies, more c- ,other companies such as yourself that are doing it specifically for games. If everyone's generating these, these thousands and thousands of ads and concepts all the time, does that end up just kind of homogenizing or making everything look the same?
Especially since AI is good at kind of average, but not great at like pushing past that. 'Cause you mentioned a little bit of that earlier, being kind of problematic with being innovative, and everything. I imagine especially if you're trying to train it off of what's trending or whatever, that everyone's doing that at the same time, and everything kind of starts to look the same.
Like, has that been a problem for you guys? Are you-- How are you kind of working around that?
Amit: You know, it's a, it's a funny question because I think when we started Sett, when we did the discovery process, I think, I assume maybe four years ago. So, we spoke with guy before AI, you know, before diffusion models, before LLMs, and I don't remember if it's Rovio or other company just, you know, came and told us, ", look, so we will see the same ads with everyone, like with Sett.
Let's take it four or five years from now." So, it was funny. And what we answered then, which is still relevant, is think about it as the world of options, of concepts. Think about it statistically. It's like a huge galaxy with tons of different parameters and areas to explore that You-- I just don't believe that you will see the same stuff.
Like when you have a concept that is working, everyone is trying to utilize that, but you see that not everyone is making it successful. So, you will take Match 3 and Match 3, you will do the Save the King, and for you it's working, for them it's not working, and you need to tweak a lot of stuff. In under Save the King, you have billions of different options.
Eh, so I would say that the market will probably be more flooded with a lot of new ideas. I don't see a situation where the, the bank of ideas ends because the market is evolving, everything is moving. It's not, eh, it's not static. It's always you have a new thing because you think out of the box. And I believe that if LLM today and content generation for games or high-quality ads for apps was commoditized completely and LLM were still not innovative as is...
So, I, I, I would have this, eh, I would believe that we will see this kind of issue, in the market because everyone will do thousands of ads and they will be the same because the LLM comes up with the idea and et cetera and et cetera and et cetera, and you will get into this never-ending loop. Eh, but, fortunately, eh, we're still not there.
Devin: Yeah, I mean, I, I just imagine that maybe it's just very, very fast cycles now, right? Where it's like because you can create the playables much faster, you have the ability to fast follow very successful ones, but then that probably doesn't have the same impact when you're fast following it, and then you maybe drop it because it doesn't, doesn't have the same impact you were hoping for.
But it's funny thinking back to like, you know, if you watch, like YouTubes of old stuff from, you know, previous decades for nostalgia or whatever, and it has all the commercials in it. You see back then commercials were fairly homogenized in trends very-- because it takes a long time to make them.
People are kind of looking at other people's things. There's very limited, broadcast channels, and so you kind of end up with that. Whereas now we're in an environment where maybe you can try, you know, thousands of things very quickly and see what works. But that seems like you may have trouble then with the data part of that, right?
Like if you're fast following, is it, is it because the crea- like if the creative doesn't work, is it because the creative didn't work or is it because that trend's over? Like how are you kind of looking at like measurements for success in such a hyper rapid environment with lots of contextual clues that may differ, especially like, oh, I'm having to be targeting the same thing but a different country that may have a different demographic.
Like how are you kind of managing it and understanding the metrics that you're getting back?
Amit: Who-- So attribution for creatives, it's an unsolved issue, like to know exactly, you know, what is happening there, and you're right on point. It's, I would lie if I say that it's so obvious to see whether something is good or bad or the causality of that.
I would say we, we try with our partners to look at the more high level stuff like share of spend and general loss and see the campaigns. And again, even on that, there are too many uncontrolled parameters that can affect that. So, I-- to be honest, it's still a big issue of the attribution of creatives and even more to understand the why, behind each thing.
Like even if you have a new winner, you can't really know why. You can assume a lot of different thesises, and yet then you can exploit each thesis until it's proven right or wrong. And the only way to prove it right or wrong is if you-- if it's generated a new winner, and then even based on that, you can have different type of assumptions.
But yeah, it's like a very, very, very big issues that it might happen, it's just going to be bigger and bigger because it will be easier to create. You will have more, again, competitors in the market in terms of apps and games The market will be flooded with more ads, meaning the fatigue will be higher because the number of users that are watching the ads are, is not changing as fast as AI.
So, it will be harder for you to, to find winner. It will be faster that you will have fatigue, so you will need new winners, and it will be even harder to you to exploit thesises because you're... The only way to have a successful mechanism here to find winner is by system, like technical agent like agentic is my addition, but tech-oriented, machine, a tech-driven, sorry, machine with very, very strong algorithmic data feedback loop and optimization models, very strong generation capabilities and understanding capabilities.
And, yeah, it's, you can't do it by humans. Just can't.
Devin: Yeah. Just gotta find that, that blend there, right? Where humans are steering it, but then the AI is doing the, the hard detailed work of combing through everything just 'cause, yeah, as you said, it starts to become too much data for humans to process.
You just have to kind of steer the boat at that point. One of the, one of the things that, that does come out, I think as a result, besides just all the flood of data, is the sort of ease of making ads for gameplay that doesn't actually match the game. Partly because like, you know, previously like making actual fake gameplay would take a lot of effort.
Or, you know, like you, you'd usually use actual gameplay a lot of the time 'cause it's easier to record and represent that. But now it's trivial to make really fake easy, you know, copies of other gameplay, things like that, whether that be the playables or video ads. Do you think that's exacerbating some of the problem of creating a lot of fake-- I mean, do you, do you guys even engage in that yourself?
I mean, you don't have to say if it's, if it's something you don't engage in, but- No, of course, of course ... but is that something that you're, you're finding people are wanting to do with AI a lot? Is it helping or is it kind of misdirecting users? Like how do you see that whole thing? Because it's kind of a w- a weird series of things that we have in this, this advertising industry that I don't think typically you see outside of mobile games.
Amit: Yeah, so it's, the-- First of all, it's, it's a must. Fortunately, unfortunately, it's a must to have different types of gameplays that are unfortunately not necessarily represent your own real gameplay in order to attract the users that you really need. So, this is the funny part here. You're displaying something that you don't really have, or you can have as a mini-game, as, you know, your main gameplay.
And then- You try to find the users that say, "Hey, this is what I want to see. Ah, this is what I got, but it still matches my needs," that the relaxation needs, the competition, like the user motivation core psychological attraction for, for the ad itself. It's a must today. We see it with-- Again, I don't have any, any studio that we currently work in besides the IP that we're not doing misleading stuff, and everyone is trying to, to take it to the edge.
With AI it's, it's easier today than in the past to create misleading ads, but AI is not really helping there a lot if you're going to the complex misleading ads because you need new assets, you need new mechanics, you need a lot of usually 3D stuff. And again, this is what we said, like 10 or 15 minutes ago, AI is still not good with it.
And the-- Before that, you need the innovative idea, and AI is shit on that as well. So, when you have the idea and you have the assets to generate the stuff, to code them, to fasten the process, it is easier, but it's still the-- It's still a bit hard with AI compared to like usual gameplay when you have everything you need to do it, uh, in a good way.
But yeah, it's, it's a must. We are of course very, like I, I'm not proud to say that, you know, we are doing that, but we are doing that because at the end, you know, uh, the system, the agents understand the target audience, and based on that, you generate different type of mechanics. So, it's a must today. Again, I just-- the trend that I see with all of our partners, like 100% of them besides already the IP, and there are not a lot of that, is doubling down on misling- misleading as much as possible.
You even see it now because a lot of the Chinese studios or the more Eastern, let's call it ones, I don't have, let's call it a shame in doing misleading stuff. So-some of the time they're taking it to a more ethically problematic level that this stuff we don't do as a company. But the misleading stuff is something that everyone is taking to the edge.
And now we see the Western studios are trying to do more. And then you see the IP games are thinking about more like, how can we approve to do something misleading with the IP? So, everything is kind of gathered around the misleading part of the ads, and I just see triple, quadruple down on that. It's like getting even insane and insaner.
Devin: Maybe it'll eventually have to hit a point of just oversaturation, then it goes back the other direction, as a lot of- Yeah. Then- ...trends do, right? So, I mean, it's funny because like y-you know, as we trend into playables, you're actually having to make more of that an actual game, right? You-- With the misleading stuff you, you maybe were before, "Oh, let's kind of fake the, the concept," or, "Maybe we have this somewhere in the game, maybe we don't."
But now you actually, like, if you're making playables of it, you're like closer and closer to saying, "Well, maybe we should just actually make that, if it works," right? 'Cause we already have like something somewhat playable as, at least a prototype. So, it's interesting to see it, like, go that direction. As you said, doubling, tripling down is, I mean, I guess a lot more feasible.
And if, if it's trending westward, I can only imagine that that is going to get worse for quite a while. I mean, assuming you, you see that as a worse thing. I would generally see it as, as misleading and probably not great But, you know, it's a market reality right now that we kind of have to deal with.
On the other side of the market reality when it comes to the, the game companies, um, I'm kind of curious what you think how this is impacting the creative teams, like inside companies or even outside companies that are like consulted or worked with. Like what happens to those creative teams? Like are UA artists or other creatives in the, the UA teams replaced, repositioned?
Are they kind of managing AI projects? Like where-- how is the impact that you've seen so far on this process?
Amit: So, I do see the smaller studios or the new ones are really enhancing AI in a beautiful way. So, I would say the artists there are kind of repositioned A motion designer in a 20 people studio is not the same as a motion designer as a 200 people studio.
So, I would start maybe with that, and Playableets developers is kind of the same. On the more, you know, complex stuff, animations and maybe 3D artists, I see them also-- Mostly 3D artists are not, from what I saw, not getting replaced but them are w- like way, way, way, way, way faster and better using-- they're using a lot of AI.
Ani-animation is kind of the same, and even animation is, like, less good, like way less good from what I saw in te- this, like, compared to 3D generation, which is considerably also hard. I would say that the general thing is that you see a lot of-- I speak a lot with the high level, you know, the C-levels management.
A lot of them want to decrease, to reduce the number of people, you know, to be more efficient. No one has something against the creative team. But in reality, you don't see that. You see some, some that are doing, but with some of them that, you know, you have operations of hundreds of people in creative team, hundreds of hundreds.
I didn't see a huge deduction there, like in the last three years. I now hear studio, because you need to generate more now, because you have more competition of games. You just see people hi-hiring more and more. It's kind of what I think about code generation today and what Yoni, my co-founder, and I discussing a lot.
You're saying, "Okay, now you're way more efficient. Let's be more efficient as a company. Why do we have thir-" We have a lot. We have 30 AI engineers. Because we want to do more instead of making everything less efficient because the competition is harder. So, y- if you will be more efficient, you might lose the game you're playing.
So, I think it's on steroids in games. So, I would say the creative team probably are doing pretty tough repositioning. You, you s- you have a mental state that is a bit more frightened, let's say, with some of the studios. Some, some really not. - But unfortunately, I don't see anything, you know, either here or either here.
It's still, it's still kind of moving. I, I, I feel like the market is still trying to understand really the impact, the ROI of AI, and how you really use that, and where are the limitation. And you have, you know, the people that are, I don't know, super tough and say, "Yeah, with one-- I will do it with one people, and, I don't need anyone else, and AI is the best."
But then you see this one people is doing one concept a, a month or, two concept a month. Then you say, "What the fuck are we doing here?". So, it's still not there. You have a lot of rumors. You have a lot of, let's call it a thought of it, but nothing practical that you say creative team is now 60, 70, 80% smaller, and you're generating at least the same number of ads that you used to.
Devin: I mean, it makes sense, your, your point about if you're suddenly able to do a whole lot more, so is your competition, so you kind of have to do a whole lot more. It becomes kind of an arms race situation where it's like, yeah, that's, that's great that you could be more efficient, but you kind of have to now since everyone else is.
And also UA is also at a tougher spot because of just all the situation that happened with privacy and, and other factors in UA. It is even more important that you do a good job rather than just a job, you know, scaled higher, higher quantity and things like that. So, it does make sense that, like, the quality would, would be the focus then, to make sure that that quantity's paying off and that, like, you're spending money to do these, right?
Either way, and so you're either making money off of it or you're just wasting a lot of time. So, it does make sense- Mm-hmm ... to like at least have a good, at least have good supervision there in terms of making sure people are, are making sure this stuff's working however they can with whatever the metrics you get.
And un- until like, you know, recently you, you've had more like data-driven creative, like video and playables and stuff, generation agents. But from my understanding, you're, you're sort of shifting also into like management agents' side of things as well. How is that gonna work in terms of like not just generating content, but also doing more of the management of the, the campaigns? Or like what does it look like on the agentic side?
Amit: Yeah. So, I would say you have like three steps that are still missing in our value proposition. I would say, like in our, let's call it product capabilities. It's not value proposition, because the real value proposition is let's find winners faster and cheaper.
So, the time per winner, the cost per winner is probably, probably, probably in the long term the best way to, to, to measure Sett. - To complete the winner search loop, how we call it internally, we understand that the pace of the data that we get and the signals must be faster. So, we are doing this progress towards the deployment of the creative, which you don't need us basically to do that because you can build it with the, the own network MCPs, but you need to do it in an intelligent way.
It's not just automation. You need agentic workflow that really understanding the campaigns, the best performance, performed creative at the moment, where to do the attachment of the video and the playable, what happens in the different campaigns, where do you see fatigue, and it's not the time to replace and stuff like that.
But we're going to handle the deployment part of the creative and the management of the campaigns that will include, I believe, everything besides maybe at the beginning, you know, budgeting and stuff like that, that we will make it also supervised and also in way more controlled by our partners and with their confirmations to probably every step of the way.
But we must, must, must, must, must have it as product capabilities as a company, because first of all, it's, it's enhanced the value proposition. Also, we're getting in part of the workflow, so we're not measured only by creative generation or creative ideas. And it will enable us at the end to build the algorithm that we already started to build, that from the data that is now way faster, we will be able to build this algorithmic layer that really generate the best ideas and best creative that at the end you, you reach this KPI of cost per winner and time per winner way, way, way faster.
So, it's a must for us. We, we are doubling down on that, of course, internally, and, uh, it's taking a not short amount of time because it-- you need to do it in an intelligent way. Again, not only in automation, but we will get there. We also, we, we l-- we have the platform. Until now, we are working with our customers only via Slack.
Now we have... And with kind of agentic conversation and with very, very s-strong and high touch customer success team. But now we're doing the transition towards platform that you really speak with agents, you feedback the playables, the agents speak with you and then fix them. You're able to generate your own stuff.
And the next step, which is going to be by the end of the year, in a holistic way, is to deploy the creatives, but also manage the campaigns from our platform.
Devin: Well, well, do you think it'll eventually get to the point where essentially you're able to kind of give it a budget and say, just run a whole, like, campaign that you do and, you know, try-- optimize it on the fly as you see fit, but based off the metrics and things like that, like fully hands-off in the future?
Or is it more like there'll always be a certain amount of needing humans in the loop to manage those things, to make sure that it's not just going off the rails? I imagine like obviously you can constrain budgets and things like that, but do you think it'll, it'll get to that point even within the next year or so?
Amit: Short answer is yes, but you will always have some kinds of controllability and supervision. But I believe so. I kind of know so, but I don't really know. So, I, I strongly, let's call it, believe so. And, uh, it will have to be with supervision because again, I still don't know how the big... Like, you can do it with rules and more deterministic flows, but again, LLMs at the end, you put an out-- input, the output, unless you're doing a lot of stuff, it can be different.
And when I'm thinking about, let's say Sett is the best of the best, we're managing millions of campaigns at the same time, an agent can make a mistake, and you want to remove that option of mistake from the room, and the only way to do it is via supervision of the customer that at the end, it's his, his bank, his or her bank account. So yeah.
Devin: That makes sense. And, I mean, well, given-- as-assuming you're, you're correct then that will, that will likely transpire in the, in the near future, what do you think just UI or UA, sorry, in general looks like in the next three years? Or what is your vision of it even that you're driving towards and hoping that like AI can help with or people help with, however you think UA looks, especially sort of in this difficult environment right now as that escalates?
Amit: I would beli- I believe the role of user acquisition will change drama- like dramatically. I do believe that in the three years the creative team will be the number of people will be reduced. I believe the UA even managers number will be reduced and the position itself will be repositioned as the UA strategy and how to work with the agentic system in order to find the best results in the campaigns.
So, I believe that everyone will become the strategist with a lot of A-AI, let's call it powder in their, like fire powder in their hands. Eh, I s- I expect it to be way more competitive than now. I expect the volume of creatives will be like, I don't know, maybe three, four, five times more per studio in terms of new concept and high quality, not like AI slop.
It will be, I don't know, it will be-- It feels to me that it will be tougher in this space and people will, will kind of reposition their mindset about how I can kind of control AI system and utilize them in the best way to impact my strategy. I'm thinking a lot about brand marketing or product marketing, people with these kind of capabilities of understanding the target audience very deeply compared to UA managers and motion designer that not necessarily understand them like product marketing managers.
I believe those type of roles will be more significant and will take A kind of authority and responsibility under campaigns management as well. Because campaign management will be more like of an execution mode, in my opinion, less than a thinking mode, besides a high-level strategy. So, the high-level strategy and how you direct and guide the system will be the main, like the main le- the, the biggest lever, and the talent that will lead that should be very s- like on top of the target audience, on top of the product capabilities and features, on top of how the trends in the market looks like.
So, I believe product marketing kind of persona will be more impactful in actually the campaign management or brand people.
Devin: Do you think overall that like human creative directors will still have a place in, in three to five years?
Amit: Bigger, bigger even. In my opinion, bigger because they will have more responsibility.
It's not necessarily that the team will be, 50% bigger, but they, in my opinion, will have more responsibility and authority because they will be in charge of the strategy of the execution layer. The execution layer, in my opinion, will change, but the strategy and how you guide your systematic workflows, that will be a real tech system, will have to have more input of thought.
So, you need to do more research, you need to have more context in your mind, you need to work with more teams internally in the studio to understand more, you need to meet and speak with your, the users of the games more, more. So, the direction roles, including art, creative, will be, in my opinion, bigger and bigger in terms of responsibility and authority.
Devin: Well then, I mean, I think all, all that's been... A lot of the things you've said have been good news for, for humans and that, that are, that are worried. As you said, some, some worry more than others about what AI means for their jobs, and it's more like, well, probably just means more skills to learn. It, it reminds me when everyone had to learn how to type.
You know, it was one of those things where it's like typists aren't replacing you, but you better learn how to use the tools available- Yeah ... if you wanna keep up, right? Same kind of thing in that sense. Well, so we just wanna finish off with three fun little questions here. So, yeah, just, you know, off the cuff what you, what you wanna say on these ones.
But, uh, the first one I think a little more opinionated, which is a, a game whose, whose ads you think are better than the actual game.
Amit: Wow. Look, I am a sucker of, like Whiteout Survival, King Shot, Lord Rush, Doomsday, Lords Mobile, those type of game ads. For me, it's like the best of the best. I'm truly addicted.
So, like 4X genre is something that I used to, like Clash of Clans usually in the past I do significantly like the ads, the playables more than the game. The game is great, but the ads are, like, freaking amazing.
Devin: The, the game kind of drags down in comparison.
Amit: Yeah, yeah.
Devin: So just get to the good part. That's fun. What LLM models have worked the best for you or are trending towards working the best for you? Like, just give me an idea, like, how you see the LLM models in your work specifically.
Amit: So, this is actually a super interesting part that is becoming... I will do it, unfortunately, maybe to the listeners, a bigger, longer answer because it's a really interesting part.
LLM is actually coming more and more as a commodity, less that I've expected in the past. Internally, like, Yoni, my co-founder, is in charge of the technology, he's the CTO The costs are so high, so high, and the d-differences between the models is not that big and in different areas is different, so it's not, let's say, Fable is not the best at everything.
Gemini is not the best at everything. So, they might be very good at something, but not very good at other things, and it's kind of become a real commodity that, look, we're heavy users of Claude. Like all of our people internally are using Claude code. But today, the m- the other models, including open source, including Chinese models, are pretty, pretty, advancing the race, let's call it like that, and we're using them as well.
Like you see a lot of codecs lately, like OpenAI. We utilize a lot of Gemini, so we're kind of, we don't care. We're kind of LLM agnostic. We don't care what we use, as long it keeps being, delivering us the best results. And I believe now you see the shift of the big, research labs that are focu-focusing on cheaper and cheaper and cheaper LLMs instead of better and better and better.
Devin: Do, do you find, like, the, the CTO or other people in the company are having to constantly reevaluate and retest as new models come out or as pricing changes? So, you even see, like, pricing will change for, or token usage allowance will change it for like brief periods of time as the competitive landscape keeps shifting.
Amit: Keep. It's, it will keep, it will keep. I, I see it, like, it's being in everyone focusing in un- not unrealistic way because we're spending a lot of money on this, but you're kind of thinking to yourself, "Wait, why am I real with, Sonnet or Opus now? If, if it's so expensive?" Right. So, you all of the time on Twitter understanding the new models that are coming out, you, you, you...
The real way to understand if it's good or not, you must see it on Twitter and people that are tweeting about it, and the evolves are not necessarily representing the reality of today. Right. So, so yeah, it's will keep changing.
Devin: I, I imagine at some, we'll get to at some point where instead of depending on benchmarks or social media, you'll just have an agent itself that tests agents- Yeah, and then, like, against, like, sort of expected feedback and sees how well they execute. Yeah. So, it can just in real time tell you, like, "This, this new agent's better, cheaper," all that sort of thing.
Amit: Yeah. The, the agent, your agent that is based on Anthropic will say, "Hey, probably Anthropic is the best."
Devin: Right. Exactly.
Amit: The best.
Devin: B- better use an open source one just- Yes ... just for less bias, right? Of all the stuff you've worked on, as well as other stuff you've seen outside of what you worked on, what's been your favorite UA trend? Whether that be like a specific like s-style of UA or just like some, some ways people were thinking about from a strategic perspective?
Amit: Again, I think it's like an obvious trend, but I personally, I really love ads that has a cinematic story as an intro. Like this is a person... Like it's been trendy, let's say, since AI improved. Like you saw UGC is still like gaining a lot of spend in the campaign. But if I compare it to UGC misleading Save the King and those stuff, I personally really love cinematics intros because it's bringing me back the world of advertising and target audience thinking and stuff like that, and less gameplay, gameplay, gameplay, misleading, misleading or UGC that is so obvious because you speak about the game, the, the advantages of the game. But the cinematic is kind of getting you into a story, a bit of a fantasy, and the reason that I really love games from the first place, creativity and stuff like that.
Devin: I think that's a great mindset to, to have when you think about this stuff is, is why that, that makes sense to you and works for you.
But yeah, good point about the, the psychology stuff rather than just about the execution. It's about where is it coming from and where is it trying to go, with the audience. So cool. Well, I wanna thank you for, for stopping by today. A lot of great stuff happening. Sounds like a lot of work in progress that is trending in the right direction, but will take time to get there, not just from you, but also from all the AI companies trying to build all that stuff.
Maybe World Model's helping as well. So, looking forward to, to seeing where a lot of this goes. So definitely make sure to, for those of you out there wanna check it out, it's Sett with two Ts. So don't just type S-E-T. There's a second T on there. Just, just to make sure those who are listening know how to find you.
But definitely recommend checking it out. A lot of cool stuff happening out there. So, thanks again for, for coming by.
Amit: Thank you very much. Thank you for having me.
Devin: Cool. And thanks to our listeners out there, and we'll catch you guys on the next episode.
If you enjoyed today's episode, whether on YouTube or your favorite podcast app, make sure to like, subscribe, comment, or give a five-star review. And if you wanna reach out or provide feedback, shoot us a note at [email protected] or find us on Twitter and LinkedIn. Plus, if you wanna learn more about what Naavik has to offer, make sure to check out our website www.naavik.co there. You can sign up for the number one games industry newsletter, Naavik Digest, or contact us to learn about our wide-ranging consulting and advisory services.
Again, that is www.naavik.co. Thanks for listening and we'll catch you in the next episode.
