Season 7, Episode 11: Understanding Unity's Vector (with Felix The)

Branch is an AI-powered MMP that helps growth teams see exactly where to put their dollars to bring users in the door and keep them coming back. If you haven't had a chance to check them out, now is the perfect time to do it. They've recently updated their look and feel to reflect the innovation that's happening in their product and they want you to celebrate with them. Head to branch.io/about and sign their virtual guest book. You'll be entered to win one of a number of prizes, including a $1,000 Southwest travel voucher. A personalized growth strategy session with their consulting services team or some sweet premium branch swag. Go check it out. That's branch.io/about. The problem is that the distinction needs to be drawn between the confidence of the economists and the correctness of their analysis. Welcome to the mobile dev memo podcast. I'm Eric Suford and I'm joined today by Felix Tay. Felix, welcome to the podcast. Thank you, Eric. Thank you for having me. I'm very glad to have you here to discuss all things vector, all things Unity ads. I'm really excited for this conversation. I've been looking forward to it. Yeah, meet me. Before we dig into the meat of the conversation, can you please introduce yourself to the audience? Sure. Felix, say I've been at Unity for nine years. So I've seen the different stages of Unity, the ups and downs. I'm excited to be here. I have a background in both, I guess, optimization sciences, if you want to call it that, and game development from my undergrad. So Unity in a way is kind of the perfect combination of the two. I am passionate about video games. I am passionate about our creators. I am passionate about finding ways for them to be able to monetize their creation and help them getting discovered. So I think Unity plays an important role in the ecosystem and I'm happy to be a part of it. Maybe just for some context. We've known each other for quite a while. My wife works at Unity. She's been at Unity for over a decade. But we met. The first time we met in person was almost, I think, I think it's probably like over seven years ago now because we were doing the class that people do in San Francisco when you're about to have a baby. It was like the CPR class, right? It was taught by the fireman. And that's the first time we ever met. I think we both lived in the same kind of area. But that's the first time we met in person. So we were both of our kids one week apart, I think. Yeah, right. Yeah. And also full disclosure. So, I mean, this is public, but I am an advisor to Unity. I'm advising on a lot of the AI initiatives that Unity is undertaking and those are sort of diverse and wide ranging. Today, I think we're only going to really be talking about the ad network initiatives, but we also kind of interact in that capacity. Yes. All right. So I always like to start these conversations off with like a big kind of broad conceptual question. Unity today, talking about Unity ads, I'm going to start off with Vector. So what is Vector? How does it work? And what is its purpose? Vector is Unity's AI-powered growth and user acquisition platform. It is the fastest we're in part of our business. It represents a full rebuild of our machine learning and data infrastructure that powers our advertising platform. At a high level, Vector connects gameplay, behavior, monetization signals, campaign performance into one continuous learning model. What really sets it apart is the architecture factors specifically designed to ingest and interpret very large volumes of gameplay related data. The model we use is also very centered around games and application discovery. It makes a natural fit with Unity and our engine business because we understand how games are built and how players actually behave inside them. By pairing the gameplay inside with our ad network at scale, Vector can learn faster, make better predictions, translates into stronger results for developers and advertisers. Since launch, Vector has grown rapidly. Revenue was in January as you heard in earnings 72% up year over year. It's been growing quarterly, sequentially, quarter over quarter. We believe we're still in the early adoption curve of growth. Now our focus is to work on the core engine driving performance across the grow business by making sure that all the constituencies, data elements that we are partnering with the developers to use can power Vector in ads and also beyond. That's Vector. We invest in Vector because we understand games. We understand games not from the perspective of just the machine learning elements and the AI elements, but also the way games are built and the way games are played. And this has been an ongoing initiative for some time. Do you have a start date or something like a birth date? Why does this initiative take shape? The most big-backed moment would be somewhere between Q2 and Q3 2025, I guess, because that's when the industry actually saw the impact from last year. But in R&D, typically you don't really have a single release, right? Like in AI and machine learning for ads, there's elements of the model that you need to update sequentially. There's the data architecture that you need to update sequentially. So that work was ongoing, but we just saw the majority of the benefit happen in this big moment around Q2, Q3 of last year. So that's when we went public with a message that this is the impact factor. That's a really good point about the R&D process. Maybe we could just kind of hover there for a minute. So like, talk to me about that because I think particularly with respect to ads, right? I mean, like, so I was on this crusade for a long time against the use of the phrase AI in this context because my argument was like, this is kind of this is ML. It's not AI. But I gave in because it's more exciting. It's more exciting to just say AI. But nonetheless, right? So like, that's become like the big topic of discussion across the industry right now, across the entire landscape is like these applications of AI to advertising and with Facebook, I mean, or meta, they talk about gem and drama. And you know, kind of like every time they do, they'll say something in earnings, right? And then you'll get kind of like these LinkedIn influencer posts about like, oh, the new meta, cool or the new meta AI algorithm that is changing advertising wholesale. And then they'll kind of just regurgitate what someone said about Andromeda, right? But the reality is Andromeda has been in place for years and they make sequential updates to it, right? This is a research project. It doesn't happen, as you said, in one big release or something like here is Andromeda thrust onto the world. It's introduced. It's improved. It's tweaked. It's tuned. The constant improvement and iteration process in these kind of systems like vector. There's two. One is you can basically tune based on the scientific advancement. Like there's of course a better model. There's a better way to basically develop certain protocols, parameters or adjacency to the model. Like for example, do you want to basically tune the model based on a certain outcome? Do you want to tune the model based on actually understanding multiple outcomes in a multi-head? There's pros and cons. Those are like just modeling techniques, modeling techniques basically improve over time. That's one path. The second path, typically when it comes to the iteration is like, you don't have to limit yourself on model development based on just model development alone based on the availability of data that you have and the way that data is represented and how complete it is and the nature of the sources of the data. It opens up avenues of outcomes for modeling. So in a way, think about model as a two track, right? The modeling team can basically think about improving the model independently. And the second thing is like you want your modeling team to actually understand the nature of what's possible based on data and based on what's possible based on data, the paper you use as the base, the technique you employ and even the approach itself will become fundamentally different. So those are the two things that tend to be iterative in nature. One is the modeling itself, but secondly data changes and with data changing, modeling also change. Right. And then within that sort of that data, you know, sphere, you've got a number of different opportunities available to you with like data augmentation, thinking about like noisy the date, like purposefully noisy the data just to add more diversity to it or to train the model to pick up on what's noise and what's sexual signal. And then you know, we on the modeling side, I mean, you could go down to like, you know, just changing the optimizer with some different hyper parameter, but like this is just to make a point that like you're tweaking things, you're making adjustments. Some of these adjustments could be big broad adjustments, like you said, a multi head architecture. It might be just like we're changing the, you know, the momentum in an atom implementation and that might actually have, you know, meaningful effects, right? But again, this is like a regular ongoing process. You've got researchers focused on this doing this iteratively and then they get shipped and it's not just like, okay, V2 is out and that was the first time we've sort of deployed this, you know, this very dramatic change. That's right. And the way like the factor that we released from 2025 is not the factor that people experience today. We didn't really say like it's factor two, but in reality, because we have researchers and scientists that actually actively working behind the scenes, every time there's a better way to do something, it is through modeling or data with modeling, we release it and we just let the customers experience the benefit. That LTV model you're using to optimize your campaigns, even the most accurate one in the world won't guarantee better performance. Google and meta run on signals, developing and deploying them at the right time in the right format for every platform, that's its own discipline. One that has to learn alongside the auction, not behind it. Open door, SCM rush, Miro, they all work with Voyantis for exactly this reason because the model was never the hard part. It's everything that comes after signal engineering, testing, maintenance. If you're ready to close that gap, visit voyantis.ai. That's V-O-Y-A-N-T-I-S. It's talked to me about using the engine data for bad start-getting. How does that data contribute to these better outcomes and maybe kind of talking to me through what that data might look like? Yeah. I'll talk about how that data will look like as a second topic. The first topic is Unity sits in a unique intersection of creation and growth, unlike most companies that squarely represent their service in the growth area of the business. We have a lot of developers, especially in mobile, who make their games with Unity. More than 70% of the top 1000, our engine has the coverage of that developer segment. And there's a portion of that developer. They're not really using services around user acquisition, ad monetization, or services around in a similar category. What the engine data will allow us to do with the right developer framework and consent from developer is these are gameplay data that we could actually use to benefit them when they're ready to actually engage in that type of services. This is also the data that they will get in return of getting products like diagnostics and all these things. What this means for Unity now, though, is that is the footprint of data that I think Unity potentially would have access to that typically you don't get in many other sources out there. So that's exciting. The second part is our runtime data from a data quality perspective, we are the it, right? Like the game boots, that's the Unity runtime. The first thing they got insensiated. So our runtime data technically doesn't have to worry about quality issues that SDK get initialized late. Sometimes they get skipped that the coverage might be missing certain things. We don't have to also worry about joining data from like disparate sources of third party in which like it only takes one of the third party sources to actually not do their job in a high quality for the whole thing to not create that much of a meaningful value. That's the nature of data, right? Like it has to all be excellent across the board. So instead of pulling from a patchwork of sources or manually export data sets that have been stitched together, we get to get access to the cleanest quality data straight from the runtime and the boot time when applications got loaded. That in my mind is the power of the Unity runtime. So one, coverage to quality of data signal. In terms of the data itself, like, yeah, we probably have quite a bit of elements of data that you can't find outside. But I think sometimes focusing only on that detracts from the primary benefit of runtime data, which is the first two that I mentioned. And a point that you had made to me that I hadn't really considered, but I think is valid in retrospect, is like, a lot of the stuff is that having a time series is really important and knowing the fidelity of the time series is also really important. And when you're dealing with like SDK data, I mean, you don't actually know if like event A happened before event will A minus one. And you might just be getting it in the wrong order. But when you've got the runtime, you know, with sort of certainty that, you know, these events happened in this sequence. And then when you're doing sequential, I mean, that's, it's, it's narrative. If you're doing any sort of sequential modeling that you believe that the sequence is, is real, right? Yeah, causality is important. If causality is broken, then like you said, and events modeling techniques like understanding sequence of events become moved. Also, the lack of causality, sometimes you, you didn't think correlation is causation. That's also dangerous. So the idea of having a clean data with the right sequence of events being represented correctly in the same order without joins is really powerful. Right. The only way to to truly know that the sequence is correct is to have access to the runtime. Right. Because then S, I mean, with the SDK, you're always dealing with just transmission. Right. So there's any number of things that could cause that to be out of order. Right. And SDK needs to be initialized. That's the, that's like, so. And I mean, you know, the other issue is developers have to know what to send to you for you to help them. Right. I mean, that's not always clear to a developer. I mean, what should I instrument? Right. That's right. Everything, you know, valuable. And no one may know. It may take that experimentation process that we just talked about to unearth something that's actually really valuable. This was the whole, this was the whole process that took shape with a lot of developers. And I worked on like, I don't know, a number of these projects. When SK ad network was first launched, it's like, okay, well, now we have to really be thoughtful about the conversion values. Do we know what's valuable? And a lot of developers had no idea. I mean, once you went through that process of actually trying to like, adjudicate the value of these conversion values of these events that you then encode into conversion values, it became clear that like a lot of developers prior to that had never really thought about it. And they were sending things that they just assumed were valuable. But in fact, they weren't. And actually, going back to the sequence idea, encoding a single event wasn't really that meaningful. What you really needed to do was encode an event onto a sequence. When the sequence happened, that was the event and that's what unlocked a lot of usefulness in the conversion value. Right. And to build to what you said, not just the manual work by developers. I mean, like people can interpret things differently. Some people do it correctly. Some people do it wrong. But the lack of standardization across events makes it really hard for AI products to use. So I think standardization and consistency with a manual intervention that can cause error is what I would say one of the benefit of runtime data because it's high quality. Well, I am just a volume, right? Because if you think about like if I'm a game developer and I've got a game, I've got one bank of data, right? For, you know, and I don't see that much. I see my game. And, you know, we know that conversion events are rare. If people buying stuff in a mobile game is rare, right? When I wrote Freeman Economics, the first chapter, I talked about this 5% rule. And it's like, okay, as a rough heuristic, a rough rule of thumb, you can expect 5% of your users and a Freeman product to monetize. And that was me bucketing all types of, you know, freemium apps or products into that, into that freemium category. But the reality is with mobile gaming. I mean, I would say now that number given, you know, some categories could be sub 1%. Right? So you're talking about like a very tiny minority of users actually doing the thing that you care about or, you know, doing one of the things that you care about, which is making some sort of purchase or contributing revenue in some way. That's obviously that's changed a lot because the ads are a more material portion of mobile gaming revenue than they were when I wrote that. So those users contribute revenue. But nonetheless, you've got this broad stratification of revenue profiles, right? Like if you think about LTV, it's much more stratified, I think now than it was 2014. And that matters because that long tail is where all the value is. Yeah. And you write about that. Sure, let's say like ads become more relevant. There's a bit more diversification. But modeling was also about like that long tail. It's about the signal versus noise. It's the power differentiation. And sometimes the power differentiation is less about like understanding like, look, CPM varies, but you know, CPM is like high low CPM. They're a good number. It's a very different thing than a whale purchasing $10,000 in a game. So that signal differentiation coming from transactions still matters a lot in our ecosystem. Right. And anybody that's ever worked on classification with extreme class imbalance knows. Like if you've got 1% of users purchasing, making IAPs, right? If your classifier just says everyone's not going to make an IAP, it's correct. 99% of the time, right? I mean, so like it looks really good. If you said, hey, my classifier's got a 99% correctness rate. So wow, that's pretty impressive. Except it's wrong for the group of users we care about. All right, correct, correct. All right. So why now though? Why I mean, so unities, you know, when was unities? Unities, what, 15 years old? It's something like that. Unity ads was amplifier acquisition. That was, I remember when that happened. UC Luckinan was brought in. Was it 2012 or something? Yeah, someone's about right. So unities had an ads product for quite some time. Why now? What was the impetus for vector? Prior to vector, the team did work to improve our system. Typically, when you are in that crossroad, when you already had say a working deep learning model in the past, it's always a big decision when you have a viable business that is running. But then you're wondering, is this the system you want to replace? Essentially, you're flying a plane when you're building and you want it at the same time. And you need to find a way and chart a course for a smooth cutover. It's not simple. And any R&D rebuild comes with a real risk. So that risk, when we evaluate it, we got really comfortable in taking. When we saw a few things line up, new modeling techniques, new information that is available for us, new machine learning frameworks that we just so happen to use in vector that we didn't use in the past. Next, Gen inference tech that simply didn't exist when the original system was built. Combined that together with our goal that we believe this is the pivotal moment where we can actually build the best user acquisition solution for games, it became clear that rebuilding was the right course of action. So that's what we did. And needless to say, we're pretty pleased with the decision we took and we're pretty pleased with the outcome. You talked to me about some of those new techniques. I had met SVP of AI ad products on and I was asking my pretty pointed questions. I think they weren't happy about that. But I'm not going to ask you for any secret sauce. Other to be fair to him, Matt Steiner, he was pretty open. But just talk to me about some of these new techniques. I mean, what's been like an unlock? Generally speaking, I believe in big models, not small. So I think what I mean by big models, not small is like you could create models where they can be like an army of smaller models trying to do different things. Or you could you could combine them that they can actually understand what each other needs to understand. You know, out of correlation is the thing, covariance is the thing, same training pipeline is the thing, understanding causality of decisions that are made by another model that influence another model is the thing. So I feel like the future is about big models. That's one. The second part is even the data a platform world and inference world has changed quite a lot. So something that we could not imagine in the past could be something that we imagine now. On the data side, I think as we talked about, having a full sequence of data about a player will unlock new modeling techniques you mentioned about sequential modeling. That's certainly one area because I believe like the future of understanding a player is not about just understanding like at a certain point what they want to do. But to actually be able to predict what's next. That is so powerful because then you can actually think about the application of what you could do into multiple uses. And then let's see, third, big models should not also scare us. Like I guess in the past like big model by itself. Well, one, without the utility of data, what is the purpose of a big model? But we surpass that. But inference techniques, there's a lot new papers. There's a lot new techniques being done so that even in a big model for a company like us, we can actually do it in practice because we can manage the cost of a big model. Like you could optimize inference. So I believe like the future of the space is about big models, how to optimize your costs on the big model and not be scared that there's going to be new data sources that you can use to actually make your model even bigger and bigger and bigger. So we talked about the kind of tuning process. Like what's the kind of I don't know, like the midterm roadmap look like what's being developed, what's being deployed. How do you expect Vector to evolve in the kind of midterm? Generally speaking, like outside of vector, we've talked about this in our earnings too. We're very keen on starting to utilize the runtime data. So a lot of the vector performance that we've celebrated to date is without that. We're pretty confident based on what we are seeing so far, the early signals that this is something that we can utilize. And that will also unlock new modeling techniques for us to think about. So that's one on runtime data. The second part is even outside of advertising, you could see like how AI has shaped the world of development. AI has shaped the world on how people interact with products. We just don't think like the current classic way for people to interact with the product will matter that much in the long run. What matters is interacting in a product where you can express intent of what you want to do. Like whether it's a marketing campaign, whether it's like the money you want to make from ads or IEP, a lot of this hinges upon our ability to actually embrace AI beyond just the predictive ML that we offer at the core. So we believe like the new workflow for Unity As will be centered around like embracing the notion of AI that can actually influence your campaign performance. Those are top of mind for us. And then the third is what I mentioned earlier about big models. What's kind of nice about the mobile gaming context is like, you know, you have probably like a clearer and for, you know, correct me if I'm wrong too, but you probably have like a clearer picture of like what you want. Right. So if I'm thinking about like a D to C advertiser, that could be setting some number of objectives, right. And they could sit across different points in the sort of funnel. But for mobile games, I mean, you want the user to install and then you want them to enjoy the game and to be engaged, right. I mean, that's ultimately what you want. I mean, it's pretty clear. You're probably less like to the point where you don't even really know how to price it from the advertiser's perspective. You're probably like much less interested in just somebody going to the app store page, which you have no visibility really into anyway. Because once they click out, you don't know what happens. You don't know what they're doing on the app store page. I mean, that's just the way the app store works and Google Playworks, right. You don't have any visibility there. You know those channels your colleagues keep bragging about the ones getting all the credit? Yeah, they might be doing squat. Attribution makes every channel look like a hero, even when it's a zero. Incremental tells you who's actually doing the work. It's like a lie detector for your marketing budget. Start using incremental today. Get your demo at incremental.com. That's I.N.C.R.M.N.T.A.L. com. Mention that you came through the mobile dev memo podcast for a special 15% discount for the first six months. Talk to me about how AI is impacting creative strategy for gaming advertisers. I'm seeing a lot of really, really fascinating stuff happening. I imagine you're seeing more of that fascinating stuff. I just just talked to me about like the kind of new creative production workflows that you're seeing and the new types of creative that they're yielding. It's very liberating for advertisers to be able to generate creatives very quickly. A lot of them look really good. I think AI reshapes the creative strategy for gaming advertisers, but it can also make things messy. In a lot of ways, we basically just embrace Gen AI for creative generation by itself. It basically shifts the cost. I think you talked about this in one of your articles. It really shifts the cost of creative production to exploration really. When you're trying to do performance advertising and when you're seeing something new for the first time, you don't know what that thing does. You're going to have the algorithm basically check it out and see is it good? Is it like performing? Yeah, yeah, yeah, yeah. Technical speaking, in this new era where AI is impacting creative strategy, we don't want to be in a scenario where we spend less time making assets, but we spend more costs in marketing them. You want to have both. You want to actually get productivity without the added cost of marketing. We think the right kind of AI driven creative is something that is not just built with a visual understanding about what is acceptable and what's nice, but also built with performance advertising in mind, especially in the realm of ad network. It's not enough for an app to just look nice. It needs to be tuned and understands about like, what is the potential conversion of this thing? Because it's an ad. It's not like an interactive experience for you to just take a look at. You want people to take action against it. You want people to install you want people to click. So the core of our strategy when it comes to understanding AI and gaming advertiser is we don't want to get ourselves away from the root like we understand performance advertising. So when it comes to the GNI, it has to be built with performance advertising in mind. I guess in that way to answer the question directly is GNI that we hope will work in the factory network will be the one that is not just visually stunning, but context aware about what people like. What's the most innovative stuff that you're seeing coming out of advertisers in terms of their creative and anything, I mean, anything that is sort of like unexpected or new conceptually and anything that is driving performance in particular that wasn't being done before that's been sort of made possible by GNI hours. It's just higher volume. I think for me like the way I think about most innovative is really the same with my answer to the previous question, which is like, what's the one that is most impactful? So rather than the outlook, the visual outlook about what the creative is, it's more about combining the two understanding of the systems on how creative is made and how it creates conversion. So I feel like the best way that this industry can basically embrace innovation in the creative space is to make a creative that really understand a smaller and smaller cohort of users. Imagine this today, a creative today, like let's say you put it in this campaign, it's targeting everybody globally. That's pretty broad. Let's say you do localization, then you basically target people by region or by local language. That narrow is it down, but I think the holy grail of creative is like creating a combination of creative and the back-end understanding of that creative into a smaller and smaller cohort where one day you could narrow it down to a very small cohort and it's truly personalized because then people will convert. It's hard for me to describe whether there's something out there that actually has established this because I don't really know what's going on in the back end, but I can tell you from our side, I believe that's what the future of creative will look like. Yeah, I agree. I'm going to have to talk about this a lot about how these tools, they're tools, and their production layer tools. So you could use them to materially grow your output cheaply, and that's great, but you could also do that in a very undirected way that just shifts the cost from what was previously in production being a barrier to creating that many assets. So now being just in testing because this was undirected. It's just outper for the sake of output in a lot of cases, and I think that's a mistake. That actually doesn't, you know, it probably just nets out to the same outcome. If you're 10x in output, but your testing budget becomes unsustainably large, where you can't test that output because you're constrained, then you haven't really gained anything. And I do think that a lot of advertisers are finding themselves in that trap. It's like, well, we've just got this machine gun now, and we're, I don't really want to extend like the gun metaphor, but like we're not being very strategic with how we deploy it. And I understand your point, no, meaning like just spread, and then hopefully one of them will land, right? Right. The thing that we are doing deliberately, I don't want to comment what other people do, but the thing that we're trying to do deliberately is we don't want to bring our advertiser customer in a position where we have a Gen AI tools that they can create a bunch of creative for us to just benefit because there's a lot more money trying to do marketing with us. We are very conscious because our relationship with the developers are not just growing their game, but like a lot of these their games are made with unity. We want them to genuinely be successful. So what we want to do if we are to launch a product when Gen AI is not something that transfer their cost of creative to the cost of marketing, we want them to actually be able to spend and achieve success. So for us, like the KPI for a good creative Gen AI is better OS. That's simple. Yeah, I mean, I just see a lot of people diving headfirst into the Gen AI tools and they're not like it just, they maybe have convinced themselves that this is very productive because just look at the number of assets we're creating. How could that not be an increase in productivity, but actually just shift in the cost elsewhere and you're probably not better off. Yeah, I agree with that. What emerging trends are you seeing across game design and early soft launch? testing across the mobile gaming space. Because it's off launch in particular, I think, it became a lot harder post-ATT, which is because of the lack of the transparency. How are you seeing people approach soft launches now? Have you seen anything that, from a design perspective, that kind of unlocks, the potential for spending on UA? - In mobile gaming development and gaming development, general, the thing that will make a lot of impact positively for the industry is minimizing your pivot and your change at the 11th hour. The issue sometimes why that cannot be avoided is like, you know what they say about, how can you model fun? How can you design fun? Game industry is a hit driven business, right? You never know until you try. So that's true to a certain extent, but I think there's a creative way to minimize the risk. So one way that some developers have started to minimize that risk is to not actually follow the traditional soft launch to a pivot, spend, you know, tweak, global launch, spend marketing, see if it works. They try to minimize the cost of development by hoarding some existing title they have, re-skin it, change the game mechanics by a bit, but like literally you have a vertical slice of the game, but not the full game, and then try to launch that as a playable unit in which they don't actually measure for install. I mean, there's nothing to install, really, at this point, but measure engagement. When you measure engagement, that basically tells you a lot about marketability, viability of the game concept. And when you see something's wrong, that actually is the signal that they used to say, you know what, this is a great idea. Let's double down. Let's make the full game out of this. Or when you see like the engagement score is pretty weak, even with a game loop being represented correctly, it is the right time to pivot. And when it is that time, when you pivot, the cost of pivoting is very low. So essentially, they completely flip the traditional model of expensive 11 hour change after a failed soft launch. They tried to basically get that signal early as they're still iterating and developing the game. I think that's a really healthy shift for the industry. I think like, you know, if more and more people adopt this, I could see a world where there's less R&D cost that is spent too much down a rabbit hole that we don't see any viability of success and more people failing fast and making successful games. - Yeah, I think that's been a fascinating development to watch. So this idea that you're describing, it's just like, "Okay, we're gonna take this and we're gonna package it into a playable." That concept just accelerated significantly. And it's like-- - Accelerate. And instead of stipulating about like, this is what the game will do out in a while, you get that data of what the game does out in a while. - Right. Right. And then just determining, okay, is this worth developing or not, right? And then you're getting that early feedback and just making that decision. And then a lot of the calculus comes down to how well can we proxy the data that we get from this to commercial data, right? So like I said, and that was always the issue with soft launches. It was like, well, okay, we've got day seven retention or whatever and we see some acquisition costs. How well can we proxy that out to like if we're spending a million a month or two million a month or five million a month, whatever that becomes the challenge of game launches, which is like an analytical exercise, but it just allows you to do that and collect actual real data much faster. - Yeah. And when you soft launch, the game is complete. - Right. - That's the issue. Say soft launch, not looking very good. Some people do some alterations, some alterations are possible. Let's say your economy is all balanced. Like, you know, the IP is too expensive or the free gems are way too cheap that people don't need to buy, right? There's mechanics that you can actually tune to find balance the game. That's possible. But what if the core loop itself, like people don't find it right, you know? At that point, that's the game's done. So it's hard to undo done. The idea of this iterative playable testing is the game is not done and you get the feedback from the people that matter most instead of like people saying, like, this is fun is, this is what fun is. Like the best gesture of fun is the people that are going to play your game. So you get that signal. - Right. I mean, to your point, you've made this point a much time. But like the MVP for a soft launch candidate is 90% of the way to the finished product. I mean, you have to build all the systems. You have to build the core engines in place. You've got all the systems in place. You've got the economy in place. You've got, I mean, you have a lot of things in place. It's not like throwing up a, you know, if you're launching like a website or something, you've just got a landing page. I mean, this is a serious undertaking to get to that point. But even then, it's a lot. It's like a fully baked product and now you're just doing tweaks, right? So there's, there's, when you get just like a vertical slice as you described it with the playable, the baked in decisions are reduced, right? - Yep, that's right. - Mobile game developers no longer need to pay up to 30% in major App Store fees. With XOLA Web Shop, you can create a direct storefront, cut fees down to as low as 5% and keep players engaged with bundles, rewards, and analytics. Start today at xola.com. That's xsolla.com. For use the link in the episode show notes. - We hear a lot about like AI and game development and all the different ways that developers will be able to avail themselves of these tools. I just had Yoast Van Droin in on the podcast. We got a whole episode about this. How does Unity support developers as AI penetrates further into the development process and becomes kind of like core to making games? How does Unity support developers with that? - Unity at the inception of our core value and product design is open. We have a vibrant plugin ecosystem in which like many of the tooling you wanna call it that is not really made with Unity, right? It's made by another third party. Unity is mainly used as the assembly point where you basically check that all those tools are all used in the right context of building this game that you really would like to ship and then you ship it with Unity. It's built with Unity, Unity runs with it. In fact, that's a fun fact. We ship with a game. An old game engine in the past, they basically spit another executable, but in Unity's case, the reason why when you actually look at click play in the editor when you're making the game, that is Unity right there. That is actually your game running. Like it's just happened to be in the editor window. So when you ship, Unity ships with you. So for us, we are the assembly point, the destination point, the endpoint of a game because we ship with it. The tooling, it could be a tooling made by a first party. It could be tooling made by third party. That's the pre-AI age. So we don't feel differently when it comes to AI. We have Unity AI system, which is a first party AI that we build. Are we gonna open up with third party? Yeah, private beta. We have MCP gateway that we have launched with a couple of partners. We're also planning to actually find a way to open this up with more people that actually use other workflows outside of Unity through MCP server. We never shy away from opening us up from third party. We believe Unity at the end of the day is the destination point where people actually create and bring their ideas to life. Whether the utility of some aspect of creation is using another tool that is outside of us or our own tool, we are indifferent. Now, back to our AI product. One reason why we invest in a first party AI product is because we believe we understand games. That's basically the nature of it. There's a lot of models out there where the model's really large, right? It's a large language model. It is called LLMS, I guess, for a reason. Sometimes when you use a large language model that is generally smart about many things and you're trying to use it for remedial tasks like creating an NPC character, it's a bit of an overkill for two reasons. One, it's expensive. The inference cost alone is very expensive, but secondly, they're not specialized. They're not grounded. They're not trained. They're not tuned for the object of making an NPC, which is a very specific task. Layered out with the fact that creating NPC has dependency, right? Like you want to create animation rigging. OK, first, you need to make the mesh. OK, before that, you need to make the texture. OK, before that, you need to make the concept design. There's a sequence of operations to make an NPC. Why we believe Unity can really excel in this area in our own agent is because we understand the Unity context. We can make a model that is actually very cost effective, very smart, and making video games. We don't care about having being smart in doing anything else. That's not our job. But we understand developers. Developers don't want to spend too much in burning tokens to create games. They want to make games by consuming token as an expression of proffetivity and an intersection of themselves. What we want to invest in our first party AI is to solve that problem. We believe we will have an AI agent that really understand game development in the Unity context environment. We feel we have a really strong advantage in that area. But at the same time, if you want to use some stuff outside, we're going to be open. So we are agnostic to what you use us for outside. I do think that we're going to see-- I don't know what's called like a reckoning, but there's going to be a recognition at some point that the way that a lot of people have been interfacing with some of these development tools does not scale. I see people-- they're kind of like brag about it. It's like, OK, I built some product or some tool with whatever a cloud code. And I needed to add a new feature. And so I basically just rebuilt the entire thing with that feature in it. Oh, yes. Yes, you just hit-- you got it. It's like, well, you're referring to, I can one-shot a game, right? Right. I don't want to do anything. Yeah, you can one-shot anything. But the reality of video game is it's an enterprise great entertainment that you need to think about, not just the visual fidelity of it, the mechanics, the services you use, what monetization do you want to make, what economy design is, all of these interactions-- oh, your game is multiplayer. How do you balance that, right? All of these things require thought. Now, the thing that you mentioned that actually is good. The reason why we're thinking about this thing differently, you don't want to basically reprom and make from scratch. That's also another thing. The idea of game development with AI is human in the loop. You have context of what you've done before. You want to build based on that so you don't regenerate. Yeah, we got that too. You want to change the small little bit. Yes, we got that too. You want to change the small little bit, but don't touch anything else outside. That requires a deep understanding of how human and AI in RAC and that's our environment. The other thing also to note is, yes, there's a lot of smart AI stuff going on outside. The reality of a professional great game development is you can't want to shot a game with a professional grade. Secondly, most of the time when you multiple prompts a game, you still need human in the loop where you need to make final layer review or edits. That's why our role is to make sure that we can create an environment where that is possible with the idea of not just humans doing it, but human in the loop with AI agents. Right. Just being mindful of these token, I think people look at tokens as just like immaterial or rounding down to zero and any given prompt. For some tiny prompt, they might put these accumulate into, at some point, real costs here. It's not just free to you. I think people don't recognize that the view of Genie I is just like, well, I'll just continue to reprompt or try to one shot things and I can rewrite the whole code base. I don't need to refactor. All those core design principles that were developed in software engineering over the course of how many decades, they're important for a reason and those are going to be important whether you're using Genie I to write the code or whether you're writing the code by hand. In fact, they might even be more important in the Genie I case because you could have all these agents just burning tokens with no sort of discernible or scalable or extensible output. Yes. I agree. All right. What are you most optimistic about? What are you most optimistic about with mobile gaming? Well, at its core, gaming has never been about having the best graphics or the most intricate complex design or economy design and all that stuff. It's about fun. Mobile, I believe, continues to be probably the best place for fun to show up. Smartphones are more than powerful enough to display rich graphics and experiences. They're portable enough that anyone can access that content and a massive scale. That's exactly where Unity sits. Unity is mindful that games is not just about graphical fidelity, but although ongoing services and the design of the game itself, that makes it such a good entertainment experience. What I'm most optimistic about mobile, I don't know if it will continue to become a portable platform, robust, allows fun to manifest. Mobile will also benefit with this AI revolution that is happening right now. Mobile has the most device penetration. It has the most volume. It has a working discovery dynamics. It has a flexible monetization models, IP ads, paid and all these stuff. It also have support of graphics at any level complexity that Unity extends. We believe like we sit in the answer section of allowing mobile to happen. Mobile has propagation and device. It has a performance differences in generation and device. Unity sits to enable this propagation device to exist and it can thrive and with the content generation with AI becoming more and more accessible, more and more devices will get entertainment content. I have high hopes for how mobile will go in this AI era. Felix, this was great. This is very informative. Thank you very much for taking the time. I just really appreciate that we got the chance to have this conversation. Same here, Eric. We should do this more often. [BLANK_AUDIO]