How AI Increased Game Engagement By 500%

the integration of AI and gaming has been one of the most discussed topics in game development over the past year while many studios are experimenting with AI for asset generation or development workflows few are building truly AI native games experiences that would be impossible without AI at their core today's conversation is with CMAC founder of campfire a company building both AI native games and the tools to enable them what makes their story particularly interesting though is how they discovered their Direction not through careful planning but through an unexpected Insight when they added an AI companion to solve the simple problem of empty lobbies they found users spending five times longer talking to the AI character than playing the actual game this observation led them to develop cozy friends an AI native social game where players build relationships with AI characters as well as Sprites doai a platform that helps other Developers

create engaging AI characters and experiences in our discussion we'll explore how they built and optimized their AI character system the challenges of scaling AI interactions while controlling costs their vision for the future of AI in gaming and practical advice for Studios looking to integrate AI let's dive in right now and so I wanted to first ask you about kind of the origin story of what you guys are working on and then if you can go into specifics about exactly what you're building and how AI is a central component of what you're building yeah absolutely campfire is my second company my first company was in the AR VR space so I that was my segue into realtime 3D gaming space around 2021 I started the Venture into full gaming with the precursor to campfire we started building game

the co lockdowns were happening so there was a need for something to connect people I kind of pit it as Roblox for adults then a lot of our customers were people with remote teams or like distributed teams who use this platform to get together and hang out and that platform did great during that time and we were working on it adding features over time and one of the things we added with the first wave of kind of the large language model based new wave of AI kind of started swooping this was before for Chad GPT but something was in the air we did some internal hackathons and added as you said like a concierge feature like a player in the lobby that mimicked kind of what it's like to join a full Lobby um that had a personality like talked in a social chat initiated games and we thought it was just more a way to solve the cold start problem for lobbies and player counts and geographies where we didn't have player liquidity but people actually just talked to that AI agent essentially for a longer period of

time than they actually played the games on the platform so that was the key Insight like the engagement time just talking to the AI character was a 5x playing the games for that week's cohort of users so it was kind of like staring at us in the face there's inside here there's some new modality some new player kind of like interaction Paradigm is happening that's possible with jni so we Dove like head first to try and figure out like Hey how do we make this insight into a kind of company and product right and so CAC did you do this you said you started this product during covid so this is before chat GPT so were you working on large language models or other kinds of AI Tech before chat GPT and is that what you used to enable this this kind of user engagement feature this was when the kind of apis were first becoming publicly available okay uh but before the kind of Chad gpts switched to GPT 3.5 so if you remember

this was like summer of 2023 with the release of GPT 3.5 and the kind of capabilities of Chad GPT going through a big like leap something shifted where like they really captured the zge with possibilities we were playing around with their apis like before that kind of like big shift happened so this was a little bit earlier in the year in 2023 I've done some classical AI with natural language processing and things like that but nothing that directly kind of translated into foundational work on large language models we are an applied AI company and so what you observed is a lot of high engagement with a sort of like a companion or somebody that people could talk to and then maybe you could walk us to what you have now what are you working on now and how did you take this insight and then transform what you were working on to something of like a game and and if you could walk us through the thinking process and as well

as kind of describing what you're building now that would be great yeah so the process was very organic we had this community on our Discord and people who had tried our earlier games so we had a kind of audience to test this stuff with we did a bunch of prototypes kind of like a best buddy that followed you around in games the murder mystery go and talk to people and do detective work and a few other prototypes and we tested all this with this audience and they're kind of like the most interesting patterns were people just liked talking to the characters and going off script talking about personal stuff talking about what they had going on in their real life so we kind of like started trending to a game archetype which is like the life Sim that pretty much really allows for that like you know you go on do your daily chores in the village and Animal Crossing for example but a lot of it is centered around like talking to different characters checking

in keeping your Island um active and keeping all the villagers happy so we thought that's a fantastic play pattern that fits really well with this capability that AI now enables we can just let players now open in Social Sandbox and go around and talk to all these characters about whatever they want we have basically gameplay that's powered by that AI but we don't try to get in the way if you just want to go and do your own role play Adventure in the Social Sandbox as well so that's kind of what we settle on which is a Sandbox model open-ended social interactions there's some linear kind of story content again powered by the AI but the most engagement is just people logging on going around talking to characters going like fishing with them and that's their kind of PlayStation so just so I understand better the specific gameplay model or the player experience so is this more like Animal Crossing or is it more like a second life or more like a GT RP server what what does that

experience specifically look like for the players yeah so um that's a great question and are like full vision of what's possible is really something like you can imagine the most kind of well-known examples from like popular fiction like ready player one or something like that where it's almost like things become interchangeable between like what's an agent and what's a player like in a fully kind of developed version of a simul powered by this technology those lines can get blured right but with the current product and current capabilities of AI and our team size and things like that these are like multiplayer lobbies where you can see other humans and interact with them and there are also AI agents roaming the same lobbies but there's a clear kind of distinction between these human characters look like humans and the agents are all the kind of anthropomorphic animals the loop is similar to Animal Crossing you log in

you have some chores you're upgrading the island you're getting more kind of animal characters to move to your Island the way the progression works is through developing your relationship with the AI characters and getting to know them better leveling up your relationship and that's not so mechanical kind of just giving them the gift of the day or just chatting with them Etc you really have to actually invest time and get to know them personally to progress that relationship and that's the core Loop of the game center around developing friendships with AI agents and the presence of other player characters in a Lobby is more of a novelty at this one there really isn't like meaningful multiplayer interactions between players but that's something we we plan on adding oh so that's interesting so it sounds like the way that you are characterizing the gameplay and kind of the findings of the product that you have launched is really around if I'm understanding this correctly that the players really are engaging more in ter terms of the interactions with the

anthropomorphic animals and so I would assume that each of these animals has a certain personality has a certain kind of way that the character has been established and so could you talk a little bit more about that how do you think about those personalities how do you develop those personalities how much depth do they have and maybe you could kind of tell us how the AI is creating these personalities how different are they how deep are they it'd be great to hear a little bit more about that side of it yeah absolutely so you hit it right on the head it is about um people want to interact with these characters and personalities they're talking style specific activities like conversational activities or conversational flows take an offer to the user like hey I'm just a good friend I can hear you out or one of them is like uh sassy and just will make fun of you some of them are like very empathetic and some of them have very deep Back stories where you get to learn and dig deep with them and to give you a

little bit of context right so at campfire we look at ourselves as we're building tools that enable generative AI native games and consumer apps basically and then we make the apps and games that are only possible with these tools so we kind of like look at ourselves in a mold of some like epic games where they have the Unreal Engine and then they can make the best content with Unreal Engine as spell right to enable these Personalities in cozy friends we had to build so much tooling and Technology on top of basically large language models that are available out there that we realize this is a tool set right so we offer that tool set under the banner of Sprites which is our like AI character engine basically so you can set up these personalities give them a tone of speech optimize the type of conversation you want to derive with the given character

how long the conversation should be do you want to optimize for engagement do you want to optimize for specific outcomes so the way we're able to fine-tune these personalities to be so engaging for people to log in day after day and just talk to them was with our own tool set basically and that was a process that took us like almost a year of R&D to nail yeah it's definitely not something you can just build it's not really something you can just write a basic system prompt and then start interacting with the AI and expect your users to fall in love with a character it takes a lot more work than that right and I I do think in terms of some companies out there that are actually specifically working on this problem that you're talking about and so like there's inworld and convey and these companies actually specialize in these NPC behaviors these kind of interactions with real real people and and how are you differentiated if the the approach that you're taking versus another company that might try to Leverage and enroll or convey is is there a specific difference

in terms of the approach or the way that they're doing things versus how you're doing things yeah I think so so our tools organically came out of us making basically AI native AI products as opposed to us building tools in a vacuum and then trying to sell it to companies later so we have like an end to end solution for you if you want to build a game that has NPCs that do stuff like the proof isn't a pudding we have a game we've launched like several smaller apps with the same tool set and it works like the numbers kind of speak for themselves as opposed to I think like companies that started from the tools it's not so much the same thing as like let's say Unity I don't know if they started in 2004 2005 something like that like it was really well understood what it takes to make a modern game you need like 3D rendering you need a sound engine you need a physics and gen it's not really the same case for what it takes to make a gen AI native game so I think there's

a lot of guess work that has gone into these products and at a very large scale and that wasn't the case for us like we built everything to solve very specific problems along the way for a like AI native game basically okay and M maybe I could ask like where are you guys specifically as far as you know development phase roll out and so maybe you could speak to us a little bit in terms of the current progress and road map yeah so um basically basically cozy friends is a desolation of um what is possible with Gen in in games and and US showcasing that and we plan to basically continue uh building that out but it's also our uh kind of like R&D uh playground where we just like keep pushing the AI and making it better so it gives us a direct line with enough consumers like end users to make our AI tooling the best in the business while

still having a direct relationship with our community and we get to push the boundaries of what's possible with AI more and more in that game so for any other game studios that want to get involved and use our tooling like the game speaks for itself and we've solved every problem along the way we are a seed Stage Company so with this kind of split product Focus where we have to keep improving the AI tooling and continue building content for the game the pace of releases or improvements on the game aren't going to be the same as like if you were a large scale gaming studio so that's kind of kind the Market's going to decide like how big the game gets to be and whether it's just like a kind of Showcase of AI game features and it doesn't get to scale beyond that or if people love it we're going to keep basically adding to it and we have a public road map out there our first experience of Dipping our toes into hey what does the world think about this was with steam nextfest this year

before that we were only doing private play test on our own Discord and we put out the game on nextfest we didn't really do a lot of promo for it and it did really well so I shared the number a couple weeks ago on info world where where we had an article about our learnings from the nextfest kind of I don't know Tech test public beta whatever you want to call it and we did a 1.7 million chats served basically summary for that so Nexus did really well for us so we feel pretty good about how uh people perceive the game and relate to the characters and if you check on our Discord the the reception was really positive specifically to the not just like the mechanics of the AI output but also the kind of quality of the characters and the IP itself so yeah I think we're going to keep pushing along both ends make the tooling as good as possible and that translates to the game getting better and then as we reach the next mil St for both of those

products so we have D customers we're going to go public with soon on the spr AI side which engine and the game getes more traction we reach that next level of basically a financing Milestone I think that's going to give us a bigger team and more resources to attack the opportunity with more aggression got it and for potentially some of the early stage game studios or gaming entrepreneurs in the audience you mentioned that you are currently seed funded I I think there was an announcement that you guys raised like 3.95 million in seed funding and you you've just talked about you know potential additional financing plans as well as growth plans but can you go more specifically in terms of how you're thinking about a funding and growth so what would that look like in terms of like you know how much additional Capital you're looking for for this seed what was the specific Milestone that you were looking to achieve was that achieved and for the next funding round how do you think about that in terms of

where you how much you want to raise and what you want to achieve with that additional funding yeah that's a great question so I think we're a little bit different than a typical game Studio because we are really a kind of technology company at our core right our journey to make products like cozy friends and our other entertainment chat experiences that we'll be releasing more of soon is really all centered around the platform that's going to enable other people to build similar things so when we approach investors like the way our company gets valued and the potential upside would be more like a company like epic games or Unity or something like that right so we position ourselves more as hey the tooling is the heart of the business and then we build these kind of showcase contents and maybe one of those reaches product Market fit on their own but we're really a technology company at our heart so with that positioning that kind of

changes the conversations where we are in a cycle where there is a paradigm shift with Gen Ai and a lot of people have like strong thesises of how is that going to affect various Industries and if you're basically making a solid case for why you could be one of the winners in this paradigm shift that changes how an investor would view you versus if you are in a pretty established Market kind of like maybe pitching only a content company or something else so um I don't know if I have like broad advice for gaming entrepreneurs I've been on that side before with my previous company which was in arvr as I mentioned more of a Content play and Dad you really need to like show exceptional product Market fit potential or like really something special with how people are receiving a product either metrics or something qualitative but for us our secret sauce is is that we build amazing tooling on

top of gen models and we can translate that to Real World products and we have other customers doing the same and having success and because this is at a kind of inflection point and paradigm shift in the industry the threshold for what's considered as a kind of series a milestone might be more generous than a typical Tech cycle if that makes sense so just to underscore some of what you're saying and for clarification so the company that you're really that you would describe yourself as most like is more like epic because you're making cozy friends but to your point that's more of like a technology showcase to show what's possible and so really the focus of your company is to build just a lot of great tools to enable other people to make games like cozy friends or other games is that correct yeah exactly and we think the best way and really only way to do that because the tech ology isn't established no one

knows what to do with it is to solve the problem end to end which is like what do you even do with AI in games what does unit economics look like whatever we'll solve all that for our customers so like by the time we give you the tool set you just need to think of the content right so building stuff like cozy friends is how we get to build an end to-end solution for right uh people who want to build in this space and something like cozy friends this could be our like unreal or this is our Gears War or this is our fortnite we don't know that the market gets to decide that right we don't want to just build it for the SE we do want to make profit on those efforts so to some degree your most direct competitor would not be another gaming studio it would be more like an enroll or convey is that correct I think we look at ourselves not so much in competition with those guys more so the overlap is going to be people building platforms to build conversational AI

agents or autonomous AI agents for these simulations some people are going to come at it from the gaming side and some people are going to come at it from more traditional entertainment or consumer apps stuff like character. a I think because we're in this interesting inflection point I don't even view people building pure gaming tooling as necessar L competitors I kind of look at people like character. a maybe 11 Labs like some of these other broader platforms who aren't going at it gaming first we're going at it kind of like gaming and consumer app first uh so yeah that's how would I would think of it so in that case could we maybe dive a little bit deeper in terms of your tools and kind of the capabilities that you're enabling and so I understand you guys have like what you call the Sprites platform um that basically allows developers to create you know AI NPCs or agents with some notion of memory emotions can you walk us through how

does that work and then how would that be different from what other potential competitors in the marketplace would enable yes so Sprites doai is our Flagship AI tool set to build deploy and optimize conversational AI agents the set of actions those agents can take is going to expand over time right in the context of a game um information driven from the convers can actually inform what actions they take in the simulation for example and in other context that can mean maybe opening a link to a website or something like that so they can also take actions it's not just conversation and agent part just means like they're in a loop to make decisions based on some user input it's not a kind of scripted linear like do a do B do C or like rule base if this then that it's a kind of stochastic or probabilistic blackbox some set of rules some set of instructions and then they do something

with the user input and then you kind of have to monitor like how that goes so you give it a set of instructions like hey be an empathetic companion to the player use like memories from your previous interactions to enrich the conversation start the conversation by following up about a previous interaction like however you set up your agent in practice that can go like off ra so you also want to like monitor how that's going and once you kind of like set up what you think is a good user experience you want to verify that that is the case right so for a traditional product where you have like a UI or like a gameplay interaction you can kind of test that you can do some play test get a sense for it and be like hey is this fun or not is this a satisfying user interaction or not with AI you it's really not as easy because the kind of space of possibilities of what users are going to say or ask or

try to push the conversation especially a few turns in so like let's say 20 messages have been exchanged my conversation with the character is going to look so different than yours that it's impossible to predict ahead of time whether this is going to be a fun conversation or not so then you kind of need it's a framework to monitor that and be able to qualitatively analyze like is this agent that have built actually fun for the user is this engaging is this keeping them hooked that's the optimization piece of sprite so we basically simplify building the agents for you where we we're opinionated enough about the use case where you don't have to learn primitive AI building blocks like hey what's a Chain of Thought do I need to do rag like these Concepts were like hey we'll do magical work for you so you just focus on your user experience but once you have your user experience prototype or something that passes the vibe check

you really need to monitor how those real user interactions go with your AI and do they like the AI output and do they like the AI output in message one or message 80 and what does that look like and what you need to change to satisfy both of those you really need a kind of full endtoend platform that monitors allows you to iterate improve and basically optimize the output of your AI so Sprites kind of serves all of that and each piece of that we had to build for cozy friends so we had to build an abstraction to make creating personalities prototyping they easy and then we're like I love this character let's ship it and then people hate it in practice because like 10 messages deep it starts getting super aggressive or hallucinating making up nonsense so we need to monitor the full user journey and across all our users and have a kind of qualitative way to score the quality of and the engagement of these

conversations and then like okay we're at production and people love it and we're exchanging like a thousand messages per user over the kind of average user lifetime but that's costing us a lot of money so let's build a kind of distillation framework where people can fine-tune smaller models to save money once their AI agent is working Sprite solves like a toz a geni native game or consumer app with AI loop with users and I don't think any other platform can claim that same thing right and maybe just to dive even a little bit more deeper in terms of what you're describing as far as a Sprites AI platform I think the way if I understood what you spoke about it seems like we could kind of characterize the Sprites AI platform in three ways input model and output and so maybe we could talk about each of those three things in turn but starting with the input and if you look at some other AI platforms that enable AI NPC behaviors and interactions

the way that the input generally works is that there could be a set of knobs right that Define personality or the the the character's personality and so whether it's outgoing or introverted or this or that different companies will the knobs are different then there might also be a knob in terms of how NSFW the the model is or not depending on your specific application and then there is generally like an input text field where you just write the backstory you know this is if you're making RPG game you know this is a elf Ranger with blah blah blah blah or you know if it's an animal you are a turtle that is uh reclusive and bl blah blah but there there's a backstory that you would input and so on the input side for some of these types of products you generally have the personality knobs and sfw sort of knob and then a character input backstory is that how your model also works or is

there something different in the way that you guys approach the input as well yes so that is basically a stop set of what Sprites offer so in practice we found that level of granularity of control not being enough you really need more to make meaningful user interactions possible so you can start with basically picking your model and just writing a super simple description as you set which is like hey you're you're friendly you're outgoing and then things like sliders on NSFW content that is really kind of like a piece of abstraction over like what instructions get added to the AI ultimately when a request to the model for a completion is made right so you can say like hey be very safe for work or be very not safe work and that's all that slider is doing is translating into words to put in there in instead of limiting you to that sort of an

interface we offer you those types of interfaces via our templates which is like hey if you want to build a simple companion we've solved it all for you behind the scenes memories work out a box you can do it based on sliders on sentiment or personality traits but we also offer you like the raw power of how those things work and with that you can do things like I'll give you an example build a Chain of Thought before and after the AI speaks so that the experience for the user is enriched one example of doing something if you think of it like a series of AI tasks that kind of culminate in AI speaking and then maybe enriching what you do with that AI output in your game or app with some additional metadata so before the AI speaks you can set up like monitors like hey if the user talks about something like self harm we can just let the AI speak right so you can set up like a control chain on on on your AI output or

let's say you want to build like custom content or triggers like if the user mentions like Cosmetics you want to push the AI agent to like upsell them certain content or even promote like whatever your your new season of quest content that you have you can set up these like triggers in conversation and push the conversation to a very specific point or outcome that you wish so it's not just a personality it's what is the full exchange with this agent is going to get to what output what deterministic point do I want to push the agent to get to with the user right so you can do that with these chains of thought which means like do work before the AI speaks to kind of like push the AI to answer in a specific Way Beyond just a personality kind of description for things where after the AI speaks you can enrich the user experience for example um basically derive the sentiment of the AI

agent after the fact when they've spoken by running a simple analysis like hey what's the sentiment of the ai's message and then you can map that for example one to one to animations in your game so you can drive the animations of the character with what the AI thinks is the closest kind of mapping of one of your animation sets to what the AI has spoken so you basically get a way richer interaction with the user that way essentially so we start at giving you that basic input of like hey personality sliders what you can do with our platform is zoom one level out which is like what do this agent supposed to achieve with the user in the context of my product and we give you the tools to basically build that full chain which is like is this agent supposed to upsell push to a quest just be a companion keep the conversation going for 80 turns whatever Beyond just

their talking style what are they supposed to do in my product you can control that now with these chains of thoughts okay so going back to like that simplified model input model output in terms of the model so I I guess this is where you you guys are probably adding a lot of your value so basically by abstracting the technology so that games game developers who are likely not experts in AI can leverage the technology in an easy way can you speak to where have you guys put a lot of your effort energy in terms of like providing value for developers or simplifying things for developers in terms of like whether it's a the language model or whether it's around Chain of Thought or whether you know what what are the key things that you guys have done on the model side um what are maybe some of the hard problems that you you guys have had to tackle yeah so the Insight that we have to make the agents stateful right like

they can't just be a personality prompt and be let loose with the user came from us trying to make Quest content right so we're making some Quest content at a specific point of the story agent is supposed to have a state of being like angry or unhappy or unsatisfied and the way to do this is allow the quest designer to be able to Define key moments like hey just change the character's personality so a lot of these insights added up to what abstraction can we build on top of basically what gets sent to the large language model for a completion task so that you are happy with the output so I can say there's whatever like a thousand pieces of magic happening between what you set up in our UI and the kind of response you get and allowing people to get

custom data or set custom data with their own in the context of their own app so the agents can have emotions and emotions are happy angry whatever developers get to Define their own State for the agents and then that can mean different things so you can add Fields like the location of the agent in your game world or like exact Quest they're doing or exact conversation flow they're engaged in and have those things mean mean uh very specific things with the instructions that get sent to a large language model so you can tie a set of instructions conditionally to very concrete moments in your game and those are things that uh get very hard to build yourself on large language models because you want to constantly experiment iterate and change the behavior of the agent and this all gets locked into your code base like it gets locked into code files and some like a quest designer can't go in there and change the state of the AI agent at a

specific key moment and pulling all that stuff out and putting your more creative product owner or content designer to be fully in charge of how does the agent behave at any given moment that's what really opened us up to be able to make some awesome content like our most popular characters were prompted by the character artist on our team for example and that's not something you would be able to do if you kind like keep that locked away or locked more on the technical side right and maybe the most interesting thing that you've talked about in which I have not seen by some of your potential competitors at least they have not discussed it at all is the output part right so input model output in terms of the output when you said that you actually have some whether it's an approach or technology or features that allow the optimization of that personality based upon the engage engagement and interactions of the NPC

with the player and I think that's super interesting because it's you know for from a game developer perspective that's potentially even something more Dynamic like a multi-arm bandit model of some kind so that you're able to dynamically optimize this character which I think is so important and so critical so could you talk about that what what is this approach your this optimization approach where you know players are interacting with a character of some kind and then you are somehow optimizing the personality or those interactions it'd be great to hear how you guys are thinking about that what is the specific approach you're taking yes so it's great that you picked up on that Joseph because as you said super valuable yes in consumer apps and free playay games and now increasingly even AAA life service games it's all about you do your split tests you do your ab tests everything from when upsell happens to the size of UI buttons and even the

content exposed to users and if you can't do that with AI you're going to lose it's just not going to be interesting you can't just slap something together and put it in front of user and thinking that it's because it's AI it's going to be fun or engaging putting the user in loop with AI which is like character. a is the biggest one they have companies like Taki chai these guys all built their own experimentation framework where they put maybe tens of variant of an agent into production and then judge user retention length of conversation qualitative stuff like thumbs up thumbs down like human feedback you see on stuff like tragedy with your character those get directly fed into their decision making and for the more advanced companies for example chai has this approach you can even

train an AI model to predict basically how well your users are going to interact or like their attention is going to be impacted with the AI output they've been exposed to right and this is tooling you're going to have to need if you're going to succeed in this space even if your IP is fantastic people love the early conversation you're going to lose out to people who have this tooling so we built all of this where you can basically set up as many variants as you want for an agent and you can change what large language model it uses for example so if you want to test like hey maybe this open-source model like myth Max or Quin is going to work better for me for cost or performance or because it allows like more borderline not safe for work content versus like no I want to stick to CLA I want to stick to GPT 40 because my game is whatever very important and the AI is very intelligent more so than just entertaining chat you can ab test the model level you can ab

test at the prompt level you can make the character very empathetic you can make the character very sarcastic whatever your parameters for that are so you can ab test along just what the core personality is and you can ab test that whole like Chain of Thought concept that I mentioned so you can simplify that workflow expand it and see like hey if the AI takes like five steps to figure out what to say is that going to work better or maybe it's even just better to do it one shot right and we basically aggregate all the performance data from that which is retention conversation length like hert score so basically rate this allow your users to rate their interactions with AI users seem to like it actually on a side note it's a good ux in in certain cases where actually increases retention to allow them to give that feedback but all that gets aggregated and then you're just always pushing most of traffic to the best variant of an agent and that's how

that's the only way to succeed with something as random and stochastic as output from AI model you're exposing like millions of your users to basically right and I actually do want to touch upon the feature of AI or how AI impacts gaming in a second um one thing I want to touch upon then I want to just ask you a question. business model but one question with respect to what we're talking about in the potential future AI is and when we talk about optimization and the output of what you're doing is if you think about a potential super intelligence type of AI future in that kind of a scenario you could imagine that you're not just using Ai and creating different personalities and you're creating different kind of variants for testing that is given to a user to and for the user to give feedback but you could potentially imagine with more of a super intelligence type of scenario that the

AI could actually be the proxy for a user understands a user so well that they can give the feedback in which case you could then potentially scale millions of experiments with the feedback being given by the AI so that you can create an incredible game in like an instant right and so I wanted I I was wondering if you could touch upon that what is that potential for the AI to act as a proxy to the player that's a great question so like I mentioned there's ways to approximate that a little bit today so if you have a lot of existing data like whatever your output is and and I use conversational examples because that's the most well-developed kind of gameplay or kind of AI exchange Paradigm today but tomorrow it could be imag in or like whatever like runtime procedural content generation could be anything and you

have some random output from Ai and you want to evaluate like hey are people going to like it or not and if you have enough of that data you can basically build that predictive agent in a narrow sense even today but you need to have a lot of data so you need to have like a lot of people thumbs up thumbs down AI output for you and then do the work to basically prepare that data chain your model and it's not something that every company's going to have the capability bandwidth or the resources to do hopefully with a platform like Sprites for certain types of features and products is a oneclick thing where we're like hey you have enough data now do you want to train your own predictive uh model of like hey is my next agent is my next variant is my next quest line going to perform well and we're just going to compare it against other historic data and and give you a prediction yeah or even something like if you could Define certain player personas and then you can say okay who

does this game resonate with in terms of player even to that level I think would be very valuable yeah exactly like hey we have the people who like the argumentative characters we like the people who like the empathetic characters and they're going to have different like spending profiles and kind of like session links you're going to want to get out of and maybe you even optimize for them in in parallel right and like even pin them to specific variants over time and the the super intelligence piece though is really interesting because yeah if if you can just in an instant expose the AI to some content and have it predict like yeah the bar for content and and um creative efforts is going to be like zero right you can like build at this speed of thought and immediately also get a sense of whether other people are going to like it so that sort of a future is like someone's going to own that infinite

pleasure box I think that's going to be infinite pleasure box like every user or every user cohort to their perfect taste is going to be engaged like we know exactly what makes people take and we can generate that in an instant and show it to them and have it be spoken to them and Etc yeah I mean but you can imagine with that kind of capability there is no chance for humans to be able to create the best music the best movies the best anything anymore if you have an AI that can instantly find what will resonate and do a billion iterations like just like that I still think like as humans this is more of an opinion for me and I I do feel strong about it I think we have such a strong bias for like the intention behind a piece of work that's why things like sometimes don't resonate when they seem like too like corporate or commercial even like some games you're like this is this looks like AI generated the game already right some of the kind of very copy and paste like

life service stuff you see sometimes doesn't resonate people because they don't respect intention or Judgment of the artist or something like that so I think there's always going to be a space for hey this is an intentional thing that I created but I think the process for that is going to be like I just had a thought and I told it to the super intelligent AI content machine and then it just made the best version of it possible and still having that link to a human intention makes end content way more interesting in Resident for end users as opposed to it came from a kind of a nameless box right so I do want to hit like a feature of AI and gaming but before that the one question I do want to ask just for the audience because you know you have mentioned that your primary business model is going to be around offering developer tools to game studios to build their own kind of NPCs and and a AI integrating AI in in into those NPCs so what is like the current

thoughts in terms of the business model as far as your relationship with game developers that would be using your tools absolutely so I'll I'll just add a kind of amendment that it's beyond NPCs if you want to integrate user facing generative AI in your game or game like consumer app Sprites basically will satisfy your needs so even if you're doing things like runtime image generation for users it's conversational agents is kind of what we're pushing the most but the same tooling can be used because it's just like Chain of Thought stateful agent and controlling the output of the model in the end and the model can be a multimodal model can be just language only model agents can be voiced we have people using it to build like just their kind of like onboarding or intake kind of questionnaire and then personalize the content to the user you can do things like personalize

specific messages or greeting messages to the user so if you want to just at runtime have ai make stuff that your users see and interact with Sprites is going to help you the most kind of visible and not worthy example as NPCs but you can do way more with it sorry just quick question in an NPC context what would be the advantage of using just a llm versus a multimodal model for it like a NPC in a game I think in for a mobile game if you have gaml that is powered by like real world content or interactions where like the user can upload a picture of themselves or some scenery or whatever to personalize their experience or have some content generated for them or have the NPC comment on it engage with them that's possible for like more traditional PC and console games just like we don't have that input layer like as like you don't really interact with games like uploading files or giving

them access to photo gallery or anything like that so I think it's going to be more limited for the like PC and console side of things but on the free to playay mobile or just mobile side of things I think can think of as even net new types of gameplay being possible like something like Pokemon go was only possible once basically we had um like GPS available right and using that map data in your game what's going to be possible if you can interact with the game when you can yeah input or expect output of images right so right that's up to game designers hopefully more creative than me to figure out right so something like if you wanted a character and then it's got access to camera it's like hey you look tired today or I noticed you're wearing glasses today and you didn't you weren't wearing glasses before or something like that so that would be more on like the my outfit like I cook this thing what do you think like for the companion use case it's very much like stuff you would send a friend on like whatever Instagram

WhatsApp and and be like hey check this out comment on it but it could also be used to like personalize the game content for you in a way if you're like hey I saw this thing like can you like talk about it or like make up a quest for this or like whatever right or you have in your cozy friends you might have a a raccoon that's a painter and say then you ask a player hey I'm going to draw a picture for you what do you want me to draw exactly I saw this outfit like you just like upload a screenshot from Pinterest or whatever and have that cosmetic be generated for you at runtime right so yeah that's a very exciting feature right and so great the last question oh sorry ask to your business model question it's traditional software as a service for onboarding onto Sprites and basically from there we add a kind of small processing fee to like every AI request per interaction pricing model and I think you know I did uh host a

podcast with inworld I I believe that's a model they use as well so is it per interaction you charge pointx cents for example is that how you work so I would more liken it to like credit card processing or something like that like okay way less kind of egregious on the terms but we will basically route requests uh we will basically give you the interface to prepare your agent and Define how it's going to behave and for that that's just a software as a service then if you want to serve requests once you build that agent if you want you can just take the specs and like go and you have your engineers coded for you if you don't want to run your traffic through our platform we don't lock you into like hey you only can serve traffic on spr you can use spry at AI as your like prototyping ground or your like figma for the AI experience almost and then like going if you're like Hey we're

locking this in we're going to run our own data center where we have ai Engineers cool take that with you as the blueprint of the user experience and codify it and optimize for cost if you guys have that capability otherwise we are basically routing traffic on your behalf to commercial model providers right so whether that's open AI or anthropic or whoever it might be in the future or if it's open source models we use hosted providers like fireworks or t. a on your behalf again and then whatever they charge for that computation which is I giv you on a per token basis let's say it took a dollar to get like one million requests done for you we just add like three cents to that dollar basically for for our infrastructure that had to be involved in the middle but we don't have a flat like per token it's if you want to use a very expensive model that's up to you if you want to use a much cheaper model that's up to you we just add a flat kind

of percentage to the tokens that went through our system okay would it be possible to do some tasks on device for end user at Edge I think if within this generation we have like great product Market fit with AI features it can be a thing where like there's even a dedicated like inference chip getting added to these devices in future Generations but right now the models that that run on device with the latency that's user facing kind of like you can't do let's say like coherent dialogue or you can't do really good speech to text text to speech for sure not that has to be on a cloud right now image generation takes a long time but you can do things like autocomplete like press tab to autocomplete like on your email client where like you're just predicting like two words so in a kind of game maybe that's something like personalizing tiny bits of copy or like

tiny bits of pre-written dialogue you just want to personalize it a little bit like adjust a couple of warts things like that you can definitely do those workflows on device but if you want the user to like speak into a microphone or type a response and basically without getting bored or shutting the game off getting a response that's high quality enough into time right now that's not a paradigm that can happen right okay so I'd love to talk to you about the future of AI and gaming I think that the way that I would characterize things right now is is probably in the TR like if you think about the technology hype cycle it feels like the games industry right now is in the trough of disillusionment Right Where You know people got super excited about AI they kind of tried it and either you know just messing around with chat GPT or some of the generative AI stuff and then there they're like oh well you know this it doesn't work it's

too far out but what they're not seeing is the extreme rate of improvement right and so I feel like there are too many game studios now who are oversold on AI when they should really be leaning in a lot more and from your perspective and as somebody who's also you know building a game I think it'd be very interesting to understand where do you think the future is going how will AI impact games um and maybe you could also think about because the one thing I really think about is this notion of having an AI native game meaning a game that fundamentally depends on AI and so if you have any thoughts on what some of those games might look like that the prototypical example that I mention is Dungeons and Dragons if the AI can be the dungeon master that's an experience that you can't have without the AI and that therefore becomes a lot more scalable and better so what is the future of AI cross gaming the earliest St because no one wants to miss out

right and then a lot of those bets don't don't pay off because they were like very speculative and they came from a place of like hey we can't lose right like we have to if there is an opportunity we have to capture it but the kind of like getting actual value out of AI in the gaming industry that's going to be a long tale by the way everyone I know like uses stuff like Chad gbt Journey like on the daily like even very established companies like we did so many customer Discovery calls even people who take great pride in their work like they don't want to put AI generated Assets in front of users but they use mid Journey as like a creative co-pilot or like just to visualize small bits of like some concept they're working on so it's already like infiltrated workflows to some extent I kind of view it as three themes of how how AI is impacting the industry same here one is like asset gen and workflows

right so that's like the most obvious one that's what all the like big incumbents like EA and Ubisoft and whatever they've been doing like narrow AI for years for this stuff inside and now it's like boom now you got this like magic box that can do like maybe 70% of some of the workflows more automated and that's the holy grail for incomers they just want to make better content cheaper higher Fidelity I think this year a lot of that stuff will mature like I think we're going to have like legitimate kind of new streamlined workflows and some of that kind of like is signaled by how the incumbents are positioning themselves like in one of the EA earning calls Andrew Wilson the CEO just said like Hey we're we're an AI company and use the example of their college football game where it just wouldn't have been possible to have have the full roster at the Fidelity of that comparable to their

NFL products if it wasn't for their workflows and I think that's going to continue that's going to bleed into a lot more products Netflix has put like a gen workflow Zar on their gaming side I think he was a VP of gaming now he's a VP of like gen content creation at Netflix so I think the incumbent see the workflow thing they're going to invest in there it is possible it's just going to be very custom like I think like a global solution is going to be tough I think like everyone's going to build bits and pieces of it themselves but one kind of theme that might be a global type of product is like co-pilots for these like creative software everyone uses like Unity onreal blender these have like really intricate uis a lot of repetitive processes where like you're dragging and dropping things like Arcane kind of with days day work in which like okay you got to do this first then you got to create a material then you got to

apply texture then that's a prefab and today you could have a co-pilot which is like AI in the loop between the end user and the software do a lot of that work for them so if you want to set up a prototype scene in unity or set up your avatar prefab and now you have 80 more cosmetic variants that you got to create prefabs for you could just have basically robotic process automation type of co-pilots powered by large language models or driven with a chat interface do that so I think this year we'll have like a legitimate like game engine or like cat software co-pilot at the level of a a GitHub co-pilot or cursor that's already pretty widely adopted on the coding side of things yeah I mean the way that I think about AI in gaming is also three buckets but somewhat different and the way that I think about it is ai ai leverage AI

integration and AI native and what I mean by a Leverage is kind of what you're talking about as like a helper right so you've got a bunch of spreadsheets you just kind of throw it into chat GPT to do your data analysis for you or you've conducted a bunch of play tests and you've got videos of the play test and then you transcribe it and then you upload it into like Jupiter llm and then you have it characterize the biggest problems that players are seeing things like that but there's that you can leverage it or building concept art that you then paint over or you able to see a lot of variations so that's leveraging AI the second bucket is like AI integration meaning I think we're getting closer I I don't think it's fully there of having AI take over a complete pipeline so whether it's a 2D art pipeline or even something like a feature like battle pass and because you've run so many battle passes before that the AI can actually take over optimizing you know the battle pass figuring out what assets to offer in

terms of rewards it can actually take over that entire Pipeline and the third is this notion of AI native right which is like how do you build a game that really natively like at the core of the game experience is the AI but that's kind of how I think about AI in games and hopefully we'll be able to see more and more of the leverage in the AI native type of stuff in the future yeah I think if you look at the last 12 to 18 months we have these little things bubbling under the surface and I'll put cozy friends in there but there is stuff like suck up that kind of captured the Z guys for a moment I don't know if you saw it is this awesome game it has it's like it's the type of game where it was impossible before Ai and I think that's the kind of bar like hey is this a game or gameplay modality that is possible without gen or is it only possible with J and suck up was only possible with J where like you're a vampire and you go

into like this town and you can it's like kind of like Rog likee so you can do it again and again which has a cool loop and then you just go door too and convince people to let them into your home and it's like an open-ended conversation very fun I think did something like tens of millions of like YouTube hits so it was great kind of like streaming pop but again like small Indie example then there was like yandere AI girlfriend simulator like that one is huge on YouTube YouTube shorts Tik Tok where it's kind of like escape room where where it's kind of like a joke game but it actually has like millions of players you are like in your apartment and the kind of like girlfriend character isn't allowing you to leave and you have to like within a very confined sandbox of an apartment have a open-ended interaction with this AI it's like as if it was like alien isolation or something but a convers ational uh agent who's like your your girlfriend in this game to just like

escape the apartment and we have these like little pops of like oh these gen games you actually really well for content and it's I think only a matter of time before something like character. a which is like hey these are like tech people who were really good at AI just made such awesome AI output that people get hooked on and then chat with these characters like hours a day and then more creative stuff like suck up meet in the middle and have like a pop where you have like the first maybe freet to-play AI native and I think it's going to happen on mobile more likely you you get into that like double digit million monthly actives and stuff like character. AI is like under monetized because like their background isn't in gaming I think like there will be a gaming company that cracks this basically first I think it's going to be on mobile great and so actually maybe one final question I I can ask you CAC

is for some game studios out there that are interested in Ai and want to learn more or start to experiment with integrating AI into their games in some way what advice would you give some of those people um I think it's just really fun to like play with it so we started with just doing a onewe hackathon or like a game jam and you can just make a lot of little Concepts and test it out because um it's it's very open-ended and and people are are um not really sure like hey how can we utilize this like exactly and and everyone has different like points of strength right so the the DND example you mentioned like hey if you're really good at like World building and building text based RPGs or whatever that's a slim dunk because you know exactly what to do with that but if you're making like multiplayer party games you're like how can we like make something with AI

maybe it's a jackbox Style game with like a AI narrator in the loop or something like that so I think like starting with small prototypes game Jam hackathon mindset and then like finding a core loop with AI that works uh is the best place to go if you're planning to build your entire content around AI you really have to find that new paradigm um if you're in in coment and you want to like add AI features I think that's like really straightforward like you can extend your um basically session lengths uh by significant percentage just like just personalize the dialogue just let them talk open-ended like that already you're going to get so much of a lift so um I think it's yeah it's a little bit different if you if if you're trying to like yeah put a little thing in in an existing product or like build that AI native game okay CAC thank you so much

for all of your insights telling us more about your company I think that pretty much wraps us up and for those in the audience that may be interested I I'll put links to your company and and to you but would it be okay for people to reach out to you and what would be the best way if they wanted to uh contact you yeah absolutely so you can email me directly uh CAC campfire doto so that's s i a m K campfire doto if you want to work for us if you're working on the content side or more traditional software engineer want to work on the tooling side uh we're hiring across both of those uh products uh if you're a Game Dev who just wants to kind of try out Sprites or get some advice uh feel free to reach out um or just talk about anything I love I love meeting other Founders um or people in the industry and yeah reach don't don't hesitate to reach out all right thanks so much simac

and for audience we will catch you next time thanks thank you Joseph [Music] [Music]