Building AI Tools for Roblox Creators

Welcome to the Naavik Gaming podcast, where it is our mission to explore the business and future of video games. We bring together the industry's brightest builders, investors, and thinkers to keep a pulse on current events, dissect emerging trends in games, share lessons learned, and have a great time. This podcast is also part of Naavik's growing ecosystem, which ranges from free and premium research to consulting and advisory services. For more information, visit www.naavik.co. This episode is brought to you by Heroic Labs, which powers some of the world's biggest games with its Heroic Game Stack. Its composable products are Nakama, the open-source backend game server, Hero, the game development framework for all your live service game features, including the complexity of managing offline mode, Satori, the

world-class live ops product for maximizing monetization and retention, and lastly, the Heroic Cloud to provide massive scalability. After nearly 10 years of operation, thousands of studios trust Heroic Labs to help them focus on their games and not worry about game tech or scaling for success. To learn more and reach out, make sure to visit heroiclabs.com or check out the link in the show notes. This episode is brought to you by Neon. For the first time in our industry, developers can finally have their own D2C channels to save on app store fees and build closer relationships with players. But, there's a catch. Selling direct means either building complex infrastructure yourself or choosing among legacy providers whose inexperience in games or questionable reputations pose real risk to your business. Luckily, there's Neon, the D2C platform built for games. As a merchant of record with customizable web shops

optimized for conversion and modern APIs that let you use your existing live ops tools, Neon is trusted by some of the biggest names in gaming to help you sell direct without the typical overhead. To start selling direct and finally take control over your full player experience, simply head to neonpay.com or check out the link in the show notes. And with that, let's jump into the episode. Hello everyone. I'm your host Devin Becker and today I'm delighted to be joined by Niklas Visoly, founder of Lemonade. Lemonade is an AI coding platform built specifically for Roblox game creation. And today we're going to discuss the intersection of AI and UGC development for the biggest platform, Roblox. So, just to start things off, uh could you give us a quick overview of your background and the origins of Lemonade, Niklas? Sure, and thanks for having me, uh Devin. Um my background comes from gaming since I was a kid. I did Minecraft servers very, very young.

I had the servers like true I I built servers mostly for my friends back when I was 12. And the servers that I built like started growing like getting spread between my school and then I quickly realized that I could start making money with these Minecraft servers that I had. And then that's when I got like red pilled into the UGC gaming industry and kept doing Minecraft servers for many, many years. Uh eventually I had one of the largest servers in Latin America. Uh we had a few thousand concurrent users. Um worked after that with Alex from Void X Games. They're now like the the largest Roblox startup. And before Roblox, they actually started with Minecraft servers as well. So, I helped out Alex for a few years and was building those games on Minecraft. And after that I I just move it to this big gaming company called Wildlife Studios. They I I joined it as a PM

intern. Uh worked at in a few games inside of the company. And then I moved it to Europe to work in a in a new UGC gaming app that was trying to actually compete with Roblox. It was called Soba. Uh raised a good seed round at that point. That was 2022. And and yeah, I just learned a lot about the platform angle. I was I was serving creators. And and was a creator myself as well for 10 years. So, that's when I started getting a lot of these ideas of what to do next. And that's how it connects to to the vision that we have right now for building on top of Roblox. Nice. So, I mean to get into a little bit of detail, how will how does Lemonade specifically work with Roblox? And like what does the the overall workflow look like? Yeah, so we are a coding agent connected to Roblox Studio. So, we know that like all this there's like a whole generation of kids that are just starting to get their feet wet.

They never could make games before. They they always had like these dream game ideas. They always wanted to do that. And making games was just incred- incredibly hard, right? And that's what we're helping right now. So, we're helping those kids make their first games. And that's the the very first product that we have. Uh so, why are you targeting specifically for Roblox as opposed to like the other parts of the the UGC game creation? Yeah. When you analyze the all the creative outlet that there is involved to making a game, you can see like you have to do in the past you you had to like mess up doing UI creation. So, you had to build the UI. You had to do game design. You had to do the map building, the 3D asset generation. And then you has you had also like this beast that was coding and it's just extremely complicated to learn and you had all this friction which you go through it. And yeah, when we start like thinking about what the piece inside of game

development we wanted to tackle first, so like being very narrow narrowly focused, making sure that we take a space and we really nail it, coding was very obvious. And I think like in in the very beginning of the company as well, we were exploring this idea maze on top of Roblox, we were realizing that a lot of those kids that make Roblox games that were making Roblox games after AI started like getting more important, they were using ChatGPT and they were going to ChatGPT and asking ChatGPT to like, "Hey, make me a script." They were copy and pasting this script inside of Roblox Studio manually and they were doing this process like hundreds of times. So like as we were interviewing these users, they we we could see like this pattern repeating and repeating over and over. And that was early 2025. And then it became very obvious for us that like, "Okay, coding is already a very evident use case. It's definitely one of the hardest things in the entire

like pipeline of making games." And we were just like, "Okay, let's let's jump in and let's try to nail this this space first." And right now we're starting to expand slowly starting to expand into other areas like UI development, UI generation. That's very interesting to us as well. And if if you look into Roblox, for example, 50% more or less of the of many games are purely UI. It's just like the 2D UI. You join the game, you see an inventory UI, you see a setting UI, a leaderboard. And using like AI to code these UIs right now programmatically, they do a good job. Like the visuals are not there yet. Like, it's quite sloppy, right? And that's like another space that we're slowly starting to expand from coding right now. It's like UI generation. Like, we're being very, very focused to nail those spaces first and expanding from there. That seems like a good expansion from there just cuz obviously the UI is what

drives a lot of the interactivity in these games and therefore like the actual code. So, having those two tied together seems like a good way to make sure that things make sense especially for a lot a lot of these game developers that are that are younger may not have a lot of experience doing UI. So, like that that that seems like a pretty natural expansion from there. Overall though, like how does Lemonade sort of compete with what Roblox is already providing? Yeah, so that comes precisely on the like the defensibility side that we're focusing right now. Like, we have a very intentional core audience that we're focusing. Like, for us it's really important to target this first-time developers. And you can imagine that Roblox is like this huge company even though they have these initiatives like to do the Roblox assistant, they're now launching the Roblox MCP integrations. They have to be very agnostic actually with the audience that they're serving. So, they're like trying to help kids

turn developers, trying to help like developers that already exist in the platform to build even better games. And I think for us like a lot of a lot of it comes from like being very intentional to the the audience that we're serving again. First-time developers, making sure that the installation pro- process is extremely easy for them. So, the way to install our coding agent right now in Roblox is just it's simply like a plugin inside of Roblox Studio. You go in the marketplace and you're like few clicks away. And then you use the interface in the web. It's a website. And then we have like all these initiatives focused on product development. And UI is one of them. Like we we recently built this feature that you can build like UI just by prompting. So, similar to like startups like lovable, V0, you could go to the website and prompt. You see the UI in the website. And you can interact with the UI from within the website.

And that's something we launched it recently. So, like we have all this like differentiability, like trying to differentiate ourselves by building like product that's very unique to us. And that's not going to be easily replicated using cloud code or code acts or any integration that you have. Um Yeah, there there's actually kind of interesting topic here. Like for us, we we've seen like this the paradigm of coding agents changing a lot over the last year. So, like it was relatively complicated to build a coding agent like beginning of last year, middle of last year. And right now it's actually fairly like fairly easy to like do an agent, do the loop, and play around with agents yourself. So, for us like we think coding agents are getting super commoditized. And a lot of the defensibility is actually going to come through like this product features, features that are very like that have this layer of like word of mouth, being

very intentional. Like how does how do we really surprise users after they use this feature, so they tell their friends, and we gain distribution after that. And also like focusing a lot on branding and this this modes that that we think are durable. But like coding agents actually um especially intelligence is not our focus. Definitely I'm happy to share more about this. I have a lot of thoughts, but like it's definitely not what we're thinking right now. Nice. Uh definitely definitely stuff you've thought about uh for the future there. I mean obviously Roblox is going to be trying to expand that themselves. So it sounds like you guys are maybe a little more focused on user experience than they are perhaps, which is
Yeah. which is good given the audience, right? Like but I I mean the kids are kids are very smart and and they're you know, their brains are still growing, but they also like may lack a lot of experience that game developers will have. So it's a a little bit different audience. Whereas Roblox's tools will be interesting to see if they gear a little bit more towards some of those game developers moving into the space as opposed to their their specific like

younger audience. So I'm I'm really curious to see how that expands. Uh but I'm also curious how are you making sure that Lemonade understands Roblox? Like that it stays current with the documents stuff like do you feed it the the docs that Roblox provides? Do you actually like give it extra bits of information that you know? Like how do you make sure that it understands Roblox? Yeah. We have all these layers of knowledge specific to Roblox embedded inside of our tool behind the scenes. So the developers and that's particularly good again for like this first time developers cuz they they don't have actually the time to go to cloud code and be building their own like skills of knowledge. So that's that takes time, right? So for us like we handle all of that and we had this like since last year. And one thing that's important to to realize as well like it's one of the the like realizations we had recently is like we you you don't want to be competing with intelligence when it

comes to this coding agent layer cuz like every single time you you have to like work on the model harness, you have to improve your tool, you're going to be working on like marginal improvements. Like maybe you improve the agent 5% here, 10% there. And in the end it's like it's marginal, right? And once the next model like the next release comes from the labs they usually like eat all these improvement that you had. So for us we're trying to be like very very intentional to actually leave our bottom act to being intelligence. So, yes, we do the basics like in terms of Roblox knowledge, we already covered that. Like, how does the models get even better? We don't do much work here to like improve the harness aside from the foundation we already built. We're much more more interesting again building this product experiences that are domain specific to Roblox that other coding agents are not going to deliver simply because they don't

focus on Roblox. Like, cloud code like Roblox is such a small market for them. And and through that we can like develop a better experience for the audience that we're targeting. So, like for first-time developers, how do we teach them what a core loop is, right? So, you imagine a kid joining the the platform like they don't even know what game design is. So, we have to like retain them. We have to make sure that they have guardrails around their first prompt. So, that what they're building actually connect and ties to to the pieces, the mechanics that they're doing sequentially. And all these like tiny details yeah, around the experience and not so much on the intelligence layer. I think that we we're Yeah, that's that's the operational level we have right now. Interesting. I I I'm curious too like both how sort of model agnostic you are and like whether or not you kind of have to lock it to specific models. Meaning

like, you know, hey, if a new model comes out, you got to test it first and stuff like that. But, also I'm kind of curious about what you said around these people maybe not understanding game design and stuff. And I'm thinking like, oh, well, like a new user, how are they going to prompt this? How do they know what to prompt? Do you provide any guidance or like how do you how do you help people get the prompt ball rolling, especially for kids that are new to this? Obviously, there's kids that are going to be like, "Oh, I use AI all the time for school and things like that." But, then there's people who are like, "I don't know how to prompt. It's It's like this art science kind of a you just blank white page sort of thing. Yeah. 100% This is one of the hardest and most like complex things we're having right now is like dealing with this user intent coming from kids, coming from young teenagers that don't even know how to articulate what they have in mind. So, we have like this whole layer of like agents and routers behind the scenes that like try to deconstruct what is the meaning of the user? Like when they when they say that they want to build a game,

let's say like they have their favorite games like Ro Garden for example. So, like when they prompt a mechanic that like, "Oh, I'm trying to build a farm mechanic." And we have this layer, we understand like that their favorite game is already Ro Garden because we do have an integration with Ro Blocks and we can pull the data to see what are what are the user favorite games. Then we we start to collect like this data around user intent that helps us understand what the user is trying to build even if they don't express with the right words what they're trying to to like exactly make, right? And yeah, we're being like very That That's one of the biggest challenges we have right now. It's just like decompressing and deconstructing user intent and making sure we collect a lot of data points that are external to the prompting layer. So, we help the user like get to to the dream game they want to build. And in terms of models like we're being 100% agnostic right now. So, we build our own eval pipeline and eval tests

under the hood. And every time like a new model drops like uh yesterday Opus 5.7 just launched like we we run the eval immediately like we measure we have like specific tasks for Roblox. So, we know like, "Oh, make me a punch mechanic." And then we have an agent that joins the Roblox studio, it tries the mechanic and we see like what are the mechanic works or not. So, we have like have many tests like that. And then we can measure like performance across different models and and very fast see which one is good for each kind of like different space. Like which is better for physics, UI. And then we just provide this to users. Like we we're very like agnostic to to the models we serve as long as they're good to to the user. Nice. Well, that makes sense considering like you never know these days. You have like for example what happened with open claw and clawed where suddenly clawed is like, "Hey, no. You can't use open claw anymore." And that was like the main platform, you know, model that I was using. So, obviously there's a situation

like, you know, that potentially they could just be like, "Sorry, like you can't use clawed anymore." And so, it's good that you have these other tests running cuz it does seem like building on top of models is always a little bit risky in in the current world and being agnostic definitely helps. So, having the evals seems like a smart way to like always be looking out. So, like for example if if Gemini suddenly was better or if codex was suddenly better, you'd be able to switch pretty quickly because you'd have the evals kind of telling you that. Yeah, exactly. We we can see high level like what are the models that are doing better each at each different time. This is actually like a strategy very similar to cursor. Like if if you're looking into the cursor, the way that they operate since the very beginning of the company, um it took them a long time to start like doing post training experiments and they were just like providing what was state of the art from the labs, right? So, they they were giving like sonar, opus when it launched, like Gemini. Like all these models, all these different models that the user could select

themselves. And only now they started like experimenting with post training and it's very it's funny cuz it's like very a very specific layer of models. It's like the cheaper version. Given that the AI labs are like they're leaning towards like pushing the frontier and not actually not shipping or dropping so many like expensive and super sorry at cheaper or models anymore. So for Cursor it's good for them to like post training a a cheap model to chat intelligent versus competing with like the frontier that the labs already provide very well. And for us it's the same thing like we're not experimenting with post training your own Roblox coding model yet, even though we have a lot of data and we like have the data label. And yeah, it's always interesting to think like when you start collecting this data how you can like make sure you have the entire trajectory of the data you're collecting cuz that's like this raw data

it's the most valuable piece of data you can have. And after that you can have like human humans annotating the data and you can do all these experiments on top of it. So yeah, that's that's the way we think right now is like not doing post training yet. We want to intentionally have intelligence as a bottleneck cuz the models are going to keep getting better here. And as long as we have a very strong foundation of evals under the hood, we're going to be good cuz then we can flip the models very fast and just make sure we're providing the best experience and like whatever works best right now for the users. On time with the users I I understand you could have quite a number of users already. You run game jams for this and things like that. What are some of the the more noteworthy results that some of the people have made like some of the things maybe from the game jams or anything that ideally stuff that surprised you that people were able to pull off or that helped maybe even shape the direction that you took.

Yeah. We started building this a year ago and the beginning was very like zero to one, doing things that don't scale. So, it was like very manual outreach and growing few hundred users every month. And we saw a spike coming a few months after in August. So, that's when we got like our first few thousand users. And after that, like because we were pretty much the first agent on top of Roblox to make Roblox games, the growth was vertical. So, like we have been growing very steadily. We Right now, we have um almost 100,000 monthly active creators, actually. And and yeah, like it's very fulfilling for me to see like all these kids turning creators, being able to like finally, finally have these dream games and turn them into reality. And for some of them, which is actually our mission in the company, like the reason we call ourselves Lemonade, is because we want to help teenagers

make their first dollar online. So, we're trying to be like the lemonade stand for the next generation. We think Roblox is the best place to get started right now. It's like massive to many users. And we do have users right now that are already making money with the games that they build with Lemonade. Like last month, we had a user that that that built a game and made over $20,000. So, like it was like a complete first-time developer. Like the the user never built a game before. It was completely done using AI. And and yeah, this shows like the potential. Of course, like this is a this is like the long tail. It's the extreme example, but it does show the the potential of focusing and like you you don't need to Yeah, like AI is explaining all this the the things that were complex before, like designing, doing UI, as I said before. Like you a single person can do that right now. So, if as long as you have the free time and you have the curiosity to explore,

there's a lot of fuel to to collect from from doing doing that on Roblox right now. I guess it's kind of helping people sort of accomplish what Roblox is sort of mission wants to enable kids to be able to sort of make money off of games and make games and things like that and now uh enabling that a little bit easier I think for some people that were kind of scared off by the coding side of things. Uh I'm kind of curious though like um you know, we're talking about zero to one and things like that sort of working from scratch. How well does this sort of platform work with like let's say I already had a game or let's say I just bought a game from someone. I'm like I don't want to read through all that code and try and understand it. Could I use this with an existing project? Is it going to be able to at least understand enough of the code to be able to make modifications, things like that? Yeah, even though we focus on again first-time developers, you're able to connect like any existing game from Roblox into Lamini. So like we uh you can do that for sure. Like you can do live ops. We have users

that have been working on their games for many weeks already. So that's possible. And and yeah, it's just like for us it's not the audience that we're focusing right now, but it's it is something that you can like you can have your code base and connect and like keep making a game from there. The live ops part sounds kind of interesting especially like I I just think you obviously that's a new concept for a lot of the the young developers and the idea of like hey, I want to keep this going long tail. I need to actually like update my game. Maybe they don't think of it as live ops. They're just like updates. You know, people want updates. So it's really interesting that it can help enable that because that's something that could be difficult especially for newer developers. So definitely interesting there. In general though like what are some different ways I mean live ops was a great example that you see AI fitting into like UGC development not not just in Roblox but just across game development cuz this is something that I think it's going to matter to this current generation coming up more so. Like obviously we've had our generations of UGC and mods and things like that. This definitely seems like a time when

especially with enable AI enabled stuff like you guys are building that this is like a big opportunity. So how do you see that across the whole sort of UGC game development ecosystem when it comes to AI? We see like AI coming pretty much in all the fronts in terms of like everything you can do in gaming. So you can do like the asset generation, the sound generation, the map building as well. One of the reasons like we decided on to touch map building early on was because we saw the Roblox actually doing an an extremely good job here. Like they have this own internal model that's open source called called Cube 3D. And yeah, they have been focusing on that doing super good job there. And it's coming out different fronts to be fair. I think like one of the one of the coolest ones we have been seeing is like come on Roblox specifically. If you open the their website and you go to see it the games in the home page, most of the thumbnails there are AI generated already. So like you see those

kids like it's at the very top of the funnel. It's like the very first impression that you have of a game. The thumbnail and even the the game name for sure. Like a lot of kids are using that to like refine their game name. Like making sure it has the best hook. So it's coming every single front. And yeah, it's super I'm actually like very curious about the things that are going to be explored from now on. Like there's a couple of areas that I feel are underserved still. Um one of them is like AI first games on top of Roblox. Like that's not super evident yet. So like that that does feel like an under explored space. Like we don't see that many games using AI chatbots or like any kind of AI visual generation on top of Roblox. And that's an area that I think like there's yeah, again, an opportunity to explore. And

something that I'm also like very curious and we think very actively about this for Lemonade is how do we solve user liquidity? So, in terms of when you think about the desire of that someone has to make a game, it's often one of the biggest desires is that like the social recognition, the social status that comes from from building your own creation. Roblox does that super well cuz they push the algorithm the discoverability is very sharp to get players in front of of your game if the game is minimally good. But for people that don't that perhaps want to make like a multiplayer experience and need like user liquidity, people online, um I think those like AI characters are they're going to start like becoming extremely good to the point like they they resemble actually a real human and you do have the same the same feeling of like you're playing a game with a an extremely intelligent agent that does give you the feeling that like he he's your buddy for real and he he knows the entire game like end to end. And

I know Roblox like is is exploring some some side of that internally, but like that's one things that that one of the spaces that we have on Lemonade that perhaps we can build games that come with bots like those intelligent characters attached to the game itself and then we're like helping users helping creators not only with the the creation part, but also with this feeling like they want to have have like liquidity, user liquidity. They want to have the social like character with you through through like the user journey. And yeah, that's that's another thing that's super interesting to me. Nice. Well, I think we'll definitely see a lot of that happen probably faster than we expect. Uh but I I am kind of curious since you've been doing this for a little bit in terms of building this out, how have you seen AI coding specifically for game development evolve during this time that you've been working on it? Cuz I think I mean I don't know exactly when you started, but

I imagine when you started it was pretty different than it how it is now just with how rapidly things have changed for coding especially. Like I I heard a lot of people, you know, saying that like before December last year coding was one way and then after it's like worlds different. I mean, have you seen like a lot of evolution? Like what have you seen kind of especially for game development cuz I know that's usually talking app development, but I think game development's a little bit different. Yeah. On our field like coding coding agents, which is again is like it's connected to game development because of the programming side, um I think we recoded our architecture over five times over the last year alone. So, we have been doing that all the time and we can like act we can very clearly see that the paradigm is changing like every 3 months and we have to keep adapting. So, one of the the things that's happening like we we were just discussing this internally this week is that this this like flow of like you plan your task first

and then you deconstruct the task into many smaller pieces. So, you have like a separate agent to planning and then you have a separate another agent to executing. This flow is actually starting to like vanish right now because like with with the model capabilities we that they can do like long horizon tasks better, they're more intelligent, you kind of like don't need a planner agent anymore. You're able to put like the planner inside of the executor and like actually share the entire context with a single agent. And it's funny to to realize this cuz like that that was literally like the paradigm for the past year and a half 2 years. Like everybody had plans, like the planner steps. And and one of the bets we have is like this is going to start to converge again. Like before you had your separate contacts, and now it's starting to converge back to like consolidating contacts in a single agent. And it's very like it's fun and at the same time chaotic to be working with this cuz we

have to keep like adding new layers on top of our technology. And in parallel, we have to be like keep sculpting and taking things out at the same time. So yeah, that's one of the observations we have. It's like very very chaotic to be working with it right now, but if you're in yeah, if you want to keep like building the agent the best agent possible, you have to keep like refining it all the time. And that's why we did like recoded up the the agent we have so many times already. Interesting. And and I mean, talking about this whole planning paradigm and things like that. I'm kind of curious, one of the things we've seen especially trend for a while is this idea of using markdown files to keep memory as opposed to say that database or something like that. Um you know, open claw was big about that. Cloud code is big about that. Is that something you guys are leveraging as well to kind of keep track of things, to remember stuff? Obviously like keeping things everything in context is kind of expensive to do. So like are you guys architecting similar to like two file markdown stuff or Yeah, we have a similar architecture. Um

this is actually an an area that we have want to put more time on right now. But yeah, we we do have like this memory bank that the the project of the user is like the constructed and we have all the information inside of like different areas of the project. And that agent can like pull this information selectively as as it it's doing the generation. So this is should be good to do and and again, with like context windows for for models for example getting much larger and more performatic, you don't have to be spreading like all your contacts within like dozens of files anymore as as happened in the past. Um we're seeing the same thing that I just said like we're seeing like the the context converging to fewer places and the agent is able to to pull the context from from fewer places right now. Um yeah. Cool. Well, I mean I would I imagine we'll see that evolve somehow there'll be a completely new paradigm or some other way of doing memory management and

context in the future as a especially since the context tends to lead to higher costs despite the window increases. So, be interesting to see. I'm curious though like uh most of the time when you see talk about vibe coding or AI coding, it's very app focused. I'm curious what you've seen different about game development specifically cuz you know a lot of people talk about coding game development, you know, very vaguely. You don't see a lot of details. What have you seen that's like really different about it? Yeah, I I think if you like looking at the fundamentals behind like vibe coding an app and vibe coding a game, I think like both of them have different purposes. At least that's that's the thing that comes top of mind. Like if you if you're trying to vibe code an app, it's very likely that the app has a specific utility that you want to get out of using it. It could be like a financial app that you just want to like check your savings for example. So, like you have this final goal that's very specific, very clear, and you just want

to get there. versus in contrast with gaming, I think the the experience is like the entertainment is is the goal, right? Like you're you're playing something and you're either like trying to escape like reality and just like get your time sucked inside of the game. You know that very well like with with the game design tricks. And or you're like trying to like socialize. Like how do you re-engage with your friends or how you meet new friends online. So, there's like these different purposes behind games. And also in in the actual as a creator the goal of games and that's something that we're doing internally like we we have this core principle that we don't try to one shot games like zero shot. We we see all these like AI game platforms that they try to go like from prompt to game in like 5 minutes and that for us is not fun actually like if you understand what a game actually is the the joy comes from the process right

like you're iteratively building mechanic by mechanic and the process is the fun is not like the final product. And and also like games you could have like this final piece kind of in your mind and you can imagine how the game is going to look like but the process of making a game is also very fractal. So like you as you build it new things start emerging new ideas start taking shape and I think that's the joy like for for us we're being very like if someone goes to lemonade right now and and try to prompt make me still a brain rot if they literally prompt prompt that we we don't we don't let the user generate anything like we return uh like some options to the user like hey these are mechanics that exist inside of still a brain rot which one do you want to start first right so we have these guard rails to make sure that we keep the process as the main thing as the the main joy and just allowing the user from to build from there.

Um that's something that that we think about a lot. Well then you definitely bring up something that's really important I think is the one shotting versus the iterative process right game development is extremely iterative as much as people would love to just like plan and and just produce straight through that's not really how games work right where as apps generally can a bit right you have your spec you generally you're producing it you can maybe iterate a bit but usually that's at the end when you're adding features as opposed to tuning tweaking iterating. How are you guys supporting that iterating process? Um like you mentioned a little bit of it. I'm I'm really curious how that that works, that iterative loop, because let's say I I you know I I tell it what I want, I get it, and I test it. Is Is the tool then like trying to be like, "Hey, how did that work? What do you want to change?" Or is it more like just whatever you want to do next? Uh do you want to add something? I'm curious how you're going to sort of guiding or supporting that iterative process uh differently than like something like Cloud Code would be doing. Yeah. High level for us like we're trying to understand all the abstractions that are

required to make a game. And then we're we have these agents that like build these abstractions like they they kind of like help the user get there. Uh again, our focus right now is just coding and UI, but we are going to expand to other areas. Like we are going to expand to asset generation, to sound generation. These things are definitely in the plans. And also we have like all these capabilities for a single person to build a a full game alone. Um then I think for us is like just becomes understanding the user intent, so like making sure we have the intent of the user like the final goal or like the the thing that they had in mind when they they started prompting. Kind of like have this in place and help placing guardrails through this journey so the user gets there. So they don't like derail and start making com- something completely different than what they wanted initially. And out of the sudden they have a Frankenstein. And when they see they cannot like even publish the game, right? So even though it could be a fun experience for for the

creator, I think for us it's really really important to help them finish and publish. So we have like specifically on the product we have a few guardrails to like when we build a mechanic, we send back to the user like, "Hey, are you happy with this mechanic? Or do you want to add it or refine it?" And then the user kind of like a human in the loop step that the user has to review what they just did. Um we also have like something that we're working internally now is like how do we have all this core loop architectures and how do we make sure that as the user builds the game iteratively, we have we can recommend mechanics coming next that tie to the core loop or perhaps like just the mechanic sequence that the user is is building so far. So we can finally like close a loop and help them build something that's playable. That's another thing that that we're doing to keep the guardrails and keep helping the user get to a final game.

Um yeah, that's that's partially like that's a few ways that we think about it. Interesting. Yeah, like it's just really interesting that that the harness parts that are not just the coding, right? The parts that are like helping users go through this game development process and sort of learn it. Eventually it becomes like a mentor as well, right? Like where you're almost like working with a a mentor to develop a game and go through the process. And the nice thing I suppose is you can also then implement like best practices and other things that sort of help them better do a good job. Like you said, trying to help them finish the game or or stop them from just going on all these tangents. Like that's that's a a scope creep, right? It's always a problem for game development. So anything you can do to help nip that in the bud early seems like a good idea.
Yeah. Something that we're really interested interested as well is how do we help those users monetize the games? Again, in the in the mission that we have as a company helping those creators earn their first dollar online. Roblox is a extremely healthy ecosystem

to do that, honestly. So as long as you have something fairly unique and you put like this granular taste, like you keep you you build a game in a very novel way, Roblox is going to like the algorithm is going to help you find players for your game for free. Like you don't have even have to spend anything and that's very very distinct from the gaming industry itself. You don't find this like super easy on Steam. Like if you want to crack growth in Steam on Steam, you have to do like all this like wish list growth hacks and it's like completely different. Even if if I come back to my time which was on Minecraft, you had to like find the server IP and then you had to like put that in the server list. So it discoverability was very a huge pain and that that layer like Roblox really really nails it. So with this solved, how do we help the users finish and actually make money with the games? And there's a lot of strategies here. Like

as as the user is finishing the game, how do we can how can we help them add perks to their games? How can they sell like in-app purchases, subscriptions and all these things that we can have an agent that is like there almost as a buddy for the user. And yeah, helping them like achieve these goals and ultimately just being very fulfilling for us. That's that's how we started like making money online through gaming and that's what we want to do with Mark Cuban kids right now. Awesome. Well, hopefully it has a that effect as a Roblox in general seems to have done quite a bit of that. I am curious though like uh just just recently even like you mentioned you know the the AI thumbnails and things like that. What do you think overall the impact has been on Roblox from AI? Uh not just you know Lemonade, but just AI in general, but also what you've seen from Lemonade. Like what is it what is it I mean is it taking over Roblox in general? Is it like like really growing the amount of games on there? Like what is what is the impact been?

On the on the Roblox side like what we observe externally is that it's very clear that one of the priorities of the company is safety. So that's one area that they're heavily using AI and they're they're fairly like doing well with the moderation, especially as the models are getting cheaper and like open source right now, which is like you can run your own model and Roblox is doing that. Um thumbnails, I think it's an area that I do want to double down like here. Again, connected to discoverability, the main piece to get your game discover is having a great thumbnail, is having like a great a great game description. So, that's how the algorithm can pick your game besides the game metrics. And yeah, if like by simply opening like the home page on Roblox right now, you're going to see like a bunch of AI-generated thumbnails or some flavor of it. Like I I I should think there's there's a mix that you need to put like taste of like a real human on top of the thumbnail. Like there's a lot of thumbnails that

that have like manual you draw everything manually. Um but thumbnails is definitely like a way that that's very present right now in the platform. And in terms of Lamin-aid, we have like dozens of thousands of games published using our tool already. And the thing with that we are like exploring right now is actually like back to the product parts that we were discussing. Like how do we make sure that the user finishes a core loop? How does the user has a gameplay that's a bit more connected, right? Like so, it it really gives a real chance for that game to work out. And that's like one of the things that we're working next on. I'm super excited about. Nice. Well, I'm I'm excited to see like the impact. I mean, obviously not everyone's a fan of AI in general, but I think if it can enable people to to join game development. Like I know we've had these sort of indie moments before in the past, but they've always been kind of difficult for people. It it was always a an audience that was very kind of small and really dedicated. And opening this

up, obviously it's not going to be good for everyone. We're going to end up with, you know, a mix of of quality of things here, but I think overall like generational benefit will be good. We'll see if Roblox still ends up being the platform of choice for that forever, but what do you think overall the sort of future for AI UGC game development looks like? And I mean even keeping in mind who knows if GTA 6 will have modding stuff or whatever like like the intersection of AI and UGC in general, what do you think that looks like over the next few years in terms of where we go from here? Yeah, our opinion is very I think like we we were very opinionated when we started doing what we have right now. For us, we think Roblox is going to keep being a part of online gaming. So, we think that we saw this last year already like all the success and like Roblox getting back to the spotlight. But I think they're doing like a few things that do feed more advantage to them.

Back to like this discoverability thing like being able to publish a game and have actual users trying it. That's rare. Like it's not super easy to to find distribution purely relying in the algorithm, so you don't have to spend much time like doing marketing and all of that. So, that's very rare and unique for Roblox. We like the way that they're like planning their road map and things like related to that as well. It's It's important that the platform has this possibility of success through making real money. So, they have the Robux, this economy inside of it. So, all these things like aligned it really make us believe that Roblox is going to keep consolidating and that's partially right why like we don't really look so much into other UGC gaming platforms. There's a There was a lot of like hype around UFN and I think they they should have so much to prove themselves. And and even like GTA 6 like let's see what UGC comes out of the out of the game. We

we are excited and we we want to see what they have, but right now like all of our eyes and like all of our focus is going to Roblox and we we are really betting that it's going to keep growing. And in terms of the technology like for AI that that we're really excited for game development, I think it comes back to the points I said before. So like user liquidity, if you understand why someone is making a game, uh there's a few like human desires connected to you you know like the the social recognition, social status status is definitely one of the the largest ones. So if someone is able to fix user liquidity inherently as the platform, that's a very interesting play. And AI first games again, it's something that is it's to an explore especially on Roblox. So that there is a lot of opportunities to exploring AI first content. And I think we're going to start have seen more of this come out through the

year. And just like the the fundamentals like nailing distribution um is something important for the platform. And yeah, I think like that that those those are some thoughts that come top of mind. So I'm curious then like having seen all this progress and the things that have worked out well and the the direction things are going, what do you think's missing or really really needs improvement uh just in game development using AI like that at least in from your perspective of what of what Lemonade does uh that maybe is either lacking in attention or just not quite there yet, especially in area like that maybe has just been problematic to deal with. It's funny you should think about this cuz it's like I think a lot of the areas are like literally improving so fast and like having these substantial improvements. So we're benefiting from like different tailwinds coming to us. Uh I think like the the thing that comes to my is like map building. I should think a lot of it's very hard to

it's very hard to build a map in granular detail through a text prompt. You imagine like you have to say the position in the axis like of which object you want to be and then you have to detail each object. So, I don't think like the chat interface is super fluid to doing map generation. And I think that's an area that that's going to be fun to watch. Like you how do you build like this complex maps and how do you keep the taste? It's possibly you're it's going to be the same as game development like you're going to have to keep iterating many many times until you got to the state that you want. It's not going to be something that you're zero shot and you're extremely happy and that was your dream thing. Um So, definitely I think like this positioning of objects within a 3D map / simulation that's very interesting and I I think that's one of the reasons like world models are in the spotlight more because they have

like a different procedure like the the the way that they generate everything is is different than than the the the LLMs per se. Um so, that's definitely an area that that I think there is a a stronger bottleneck that we're keeping closer look at right now. Yeah, it'd be really interesting to see if there ends up being like a sort of main like the the chat GPT equivalent of a world model that you can just use as a model to build things because it doesn't seem like there's you know, it's mostly these sort of demo tech demo kind of things coming out. There's not some sort of thing you can just use via the cloud or download to just build spatial things. So, that is that is an interesting problem. Yeah, building a game through a text is kind of a weird and yeah, although people tend to do that with game design documents in a way, right? But then once you actually have to rubber hits the road and and you actually have to have something that's playable in 3D that does seem like kind of a weird way to go about it. I have seen like some attempts for like Unity plugins and stuff to try and sort of populate things with like assets, but as you said, that fine grain detail, sort

of playability, all those things need a certain amount of spatial conception that maybe might not be like existent in uh those sorts of things. So, that's an interesting approach. Just a couple of last questions in this real quick to just finish things off. Uh Just uh two quick questions. Uh first off, favorite Roblox game? Hm. Um I'm I'm not going to go with the mainstream ones. There There's one that I play quite often. It's called Tank Game. It's like uh similar to Agar.io. You have to grind and like you have your character that gets bigger as you kill other players. Um I think that's a fun one and has gotten my attention for for quite a a long time. Plus. What's the uh biggest issue you've seen in Roblox game development? I think for us, connecting to our audience is still understanding what the audience means. As we said before, like they don't know how to prompt well, and we have to do like all this work behind the scenes to like understand user intent

when they prompt like, "Oh, make me a single game. No mistakes, right?" So, we have to do like the We have to understand again and like put them back in the guardrails and like, "Hey, the jar is not like one shotting this. Like, let's go here step by step, and we're actually going to help you finish, but you have to keep with us." So, I think like the the biggest problem is like, "How do you get the user through this finish line?" And at the same time, how does the user like doesn't derail and end up building like a Frankenstein and never finishing what they wanted initially? So, all these like connecting all these pieces and making sure the experience is very very tied during the process has been the the biggest challenge, but as well like the the most creative thing for us. Like, we we come as game developers, we love doing this and it's been a lot of fun working to to solve this problem for like teenagers to to build their first games. I'm sure it makes you think of game development differently when you have to kind of think of it from this

perspective of how do I explain game development to someone else as opposed to just doing it myself. So, it's an interesting interesting perspective, I imagine. Uh where can people find some of these games that have been made with lemonade? Uh is there is there on the website? Do people need to join the Discord? Uh is there something you can search for in Roblox? What's the best way for people to kind of experience some of the things that have been built out of lemonade? So, I think the best place for people to find us is definitely on Discord. We share all the games that have been built using our tool there. It's discord.gg/lemonade. And yeah, it's growing a lot over 50,000 members. And that's where we are most active as well. So, like I'm I definitely read all DMs there and probably the best place to contact me as well. Cool. Awesome. Well, uh thanks thanks so much for joining. I am really excited about this this project in general because I think it's it's really cool to see something that uh empowers people to at least experience game development even if they don't decide to do it for a career. I think it's a good way to get something hands-on and actually experience I I'm looking forward to the day when we have game development classes using this kind of stuff and whatever in schools and

whatever. So, everyone gets to get it give it a try. Maybe lemonade will be part of that. We'll see. Uh but thanks so much for taking the time today. I appreciate it, Nicholas. Thanks for having me, David. Cool. Well, thanks everyone for tuning in and and make sure to go check it out uh lemonade at uh lemonade.gg and of course the Discord he mentioned. Uh there's lots of cool stuff happening there. Always in motion, I imagine. And uh and make sure you're you're checking out of course Navigating the Metaverse and we also now have some new stuff around AI and gaming. Uh so, make sure you check that out. Just came out this last week. So, uh make sure that you're uh up on that on on Substack and everything else. And in the meantime, catch you guys on the next one. 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 want to reach out or provide feedback, shoot us a note at [email protected], or find us on Twitter and LinkedIn. Plus, if you want to learn more about what Navic has to offer, make sure to check out our website www.navic.co.

There, you can sign up for the number one games industry newsletter, Navic Digest, or contact us to learn about our wide-ranging consulting and advisory services. Again, that is www.navic.co. Thanks for listening, and we'll catch you in the next episode.