The Most Honest AI Adoption Story in Gaming

Innovames adopted AI because the alternative was killing a game. That's the most honest AI adoption story in gaming right now and it's one of that nobody's telling because everyone else is either overselling the tech, hiding their usage, or simply has no understanding of the technology they're talking about. Today I get to talk to the guy who built the system, Thomas Lair. Welcome to the podcast.
Hi Mishka, thanks for having me. Happy to be here today. Games Industry Biz wrote an amazing article of how AI saved Sunrise Village, which was making about $160 to $190,000 a month in revenue, had a 25 person team. So the first question is like when did you know that the game was dead? First of all, the game was never dead. So it has a very active community and a lot of players loving this game. And for sure we always want to continue a game

for long for for years for 10 years or decades like we did with a lot of other games we have at games. But yeah in the end if you look at the numbers you you start to to grow to a certain degree uh you have a business case in mind you have a scale in mind and you can't reach it. Um even if you try it with a lot of iterations over years at some point you need to realize okay this is not growing any further but you have the cost of the team and um then you need to to yeah change something in the evaluation and uh the numbers you um just told are not absolutely correct they are I think from sensor tower or um you know what what other other tool um so it's missing some part of our revenues direct to consumer. We are also on browser so it's not seeing everything but uh yeah it's more about the total um level we were at. So it was not sustainable for us to run it with that

big team uh for many more years. So we need to take a decision and uh luckily we had another option than killing it. So there are different options you always can discuss. Yeah. um like can you scale down the team and uh run it with a skeleton team in some way. Um but this would not be sustainable either because we cannot create the content we needed for this game. But luckily we experimented a lot with this with AI um and could bring another option to the table which was creating content with the help of AI. Who's in the room when you're making a decision of like, hey guys, let's let's think outside the box. We can't sustain the this large of a team on a game that makes this amount of revenue. Was it an engineeringled decision or was it a leadership made decision or was it somewhere in between?
It was a joint decision between management and leadership and

engineering side of things. Um so we had exper experiments successful experiments um with proof of concepts on how AI can help um but we never proved it in a production. Um so this is completely different thing. So um you really need to um yeah prove it that it's reliably working with a with a reliable outcome you can really use. Um but once we presented this um management was on board and said okay this is something nice we we should try so we cannot lose so we can always kill a game in the end or shut shut it down we don't want that and we give it a try with AI if it turns out to be helpful let's use it if not think about other options
and when you said other options when you were originally putting these different options on the table you said skeleton crew is one of the examples but that's basically just milking the cow until it runs dry. Uh were there other options as

well like spinning off the game and giving somebody like Deca Games or these other type of you know companies that essentially do life support uh perhaps working with you know cheaper outsourcing partners you know Eastern from Germany. Um yeah, those the skeleton crew I already already talked about is not really the solution as we cannot really create the content we need to keep the players in the game. So we will slowly um decrease the amount of players and then slowly die. Um outsourcing is something we don't do actually. So it's um several reasons. So um lot of lot of management overhead for example but also you just shift to costs somehow also ownership of the game um handing it over um so a lot of overhead in there. So this is not our our option we usually choose um yeah and then we we um did approach with

with AI. So this is um something we we bet on. So hopefully onboarding the agents is is easier than onboarding an outsourcing studio.
Yeah. Okay. Well, I'm I bet a lot of outsourcing studios will now chasing you telling you that it's not that difficult but [laughter] we work with the best co-development as they say. Yeah,
absolutely. Different approaches work. It's just not a model we are doing. So we don't have much experience in or the experiences we had didn't work out very very good. So that's not something we
we strive for.
All right. Uh last thing before we jump in into the uh the uh the more details of of this this story. In the article, it said that you know the team that worked on this game, they got moved to new projects. How many of those 25 people are still at Inoame today?
I would say all of them. So um didn't

count them through but uh um no one left the company because of this decision. Also because I had the opportunity to work on a new game. So many people like getting to something new and getting very creative again, start something up. So this option um yeah was was there and the majority of the old team just shifted to a new game and some people then stayed on the project with high ownership then with a lot of expertise and knowledge about the game and then continued the game and uh it's also um always important to mention that there are still humans really holding the whole game um and uh putting a lot of effort still into this to really um yeah continue um continue as a

to translate this to a human language from engineering language. No humans were sacrificed on the altar of AI and the game is not run purely with agents, right?
Absolutely not. So important. [laughter] So, it's it's it's heavily AI supported, but it's not AI not run by AI.
Got it. That that's a that that's the distinction I was looking for because I know if there's AI skeptics, they're trying to listen to things and if we're not mentioning something specifically, then there's somebody's trying to hide the evil parts about AI. So, let's uh let's move into practicalities, how you built the system. Uh can you talk about the game team right now in terms of how it is manned and are the revenues in the same level as they were when this was purely a human-led operation and not AI and humanled operation? Saturi, it's the Japanese Zen Buddhist term for sudden enlightenment. That flash where everything clicks. We're told that Heroic Labs team had theirs

when they realized that most studios run live ops like a horoscope. vague, kind of generic, same advice for everyone. So they built Sattorii, real time experimentation, player level personalization, events that actually respond to player behavior, save to your old live ops stack, and Konichiwa to Sattorii from heroic labs. So the the title was a bit of a of incline already. So when we started with a with AI approach but since we implemented it it's stabilized. So um yeah sure um we we cannot do marketing at a scale we want to do it. So it's not not growing but it's it's at least not declining very quickly. So we see that with a new pipeline with a new content we create we can keep the people engaged. How many

people are working on the project right now?
Currently, three people working on the project.
Nice. So, we went from 25 to three and the revenue declined but stopped the decline after the transition was made.
Yes. At least slowed it drastically down. Yeah. I'm curious what did week one with operating with just three people versus 25 look like versus let's say week 8 or week nine when those same three people and agents were operating the game.
Oh first of all um when we started this whole transition to AI we first really need to break down the whole game the whole production of the content into what actually consumes time. So in the beginning it was really breaking everything down where actually goes the time. So what are the big blocks and and then this was this discussing and observing. So an

engineer um an experienced engineer with AI. So we have an AI task force at Inoame and one of one member of this AI task force was joining the team to observe and discuss what are actually the parts in the production pipeline and what are the high the big blocks and what are the real detailed elements to then think about okay can AI even help. It's not that you can switch AI on and then it's working in some way. You cannot give an agent a task say okay run this game throw out the content we need it's not working um and if you have it very detailed decomposed you can think about okay what are the first steps where AI can directly help don't go for this big full thing go for the small parts there were for example um we were noticing that a game designer

was entering all the game design definitions in the wiki into confluence also providing some kind of ids in there structure in there so it's not meant for that but it was a practice so they were starting with a with a wiki then they were transporting it to Excel and then they were exporting it from Excel to some kind of CSV or JSON and then importing that to to Unity a very cumbersome very errorprone annoying but very time consuming work. So you directly see okay this could be a thing where AI can be be a lot of help and this was taking days before for all these game design definitions getting them right in the right structure and exporting them importing them into Unity and then um luckily with the help of AI we were able to build a tool super simple tool but that was doing all of this heavy lifting into getting this a

this game design into Unity directly um um accessible and yeah was cutting down the the time drastically. Just a simple thing simple observation direct win if you tackle it and then you continue with the next thing and um from then um lot of things get compound in the end so that you are getting faster by solving stepbystep things in the pipeline. So wait a minute, you broke down the workflows, analyzed every step, and then started implementing different tools or proprietary tools to improve each of those steps and make them more efficient. You didn't start with an AI lab or other type of innovation function or subscribing to 15 different tools and then throwing them on the team. So this is already like one and a half years ago. because it was in the early early stages of how you can really use AI for

for engineering and coding. So um the tools are now way more advanced but already then you were able to create custom tools as you need them like in an hour. So this effort has not been done prior to this to AI because it was just too timeconuming. Oh I might not going to need it very often. So I do it by hand or in the classic way. Um I can't um get an engineer to provide me a tool because it was taking weeks. Now it was like I can do this in the morning. I can provide you a tool like an MVP tool for for um um for something like I prove it. It's working and then iterate on that. Um, this tooling, creating custom tools is one of the real strengths of using AI. It's not necessarily you use AI like it's generating content for you like creative AI in creating a story also helpful, but

this is really creating tools with AI that then you integrate in the pipeline that help you drastically. I was being facicious because I' i've talked to enough of different companies and consulted with many of them and the the ones that are getting true ROI from their AI projects have been the ones who have been almost boringly structural with their AI implementation where they actually figured out how they're working where they actually invested into the people understanding their own processes and choosing the right tools and the right workflows that improve their own work rather than the ones who are subscribing to the newest tools to the ones that are straight out of, you know, mega funding or winning some kind of a a weird trophy for being the best tool for whatever. And and it's no wonder that that this type of a sort of a structural

uh methodical approach has worked for you guys. But I wanted to kind of dive deeper into one of the things that you mentioned in the article and that was the content for Sunrise Village. Um, and you talked about this a little bit. You talked about the uh the sixstep AI stage designer that you have done in the game. So what are the stages of this pipeline and what was the hardest and what was maybe the easiest step out of these ones? If Apple owns your checkout, Apple owns your player, their receipt, their refund flow, their win back email, leaving you as a tenant in your own economy. App charge hands you the keys to the castle [music] you've built. A worldclass web store and a direct line to your highest spenders the [music] platforms have been quietly keeping from you for a decade. The 30% you save is [music] the boring part. Owning the player's relationship

is the actual unlock app charge provides. I think important to say is that um there is no real asset creation in the sense of um of graphics or like um art assets created. We luckily had a library of like 10,000 assets already um in the game created handcrafted by real artists and we worked we use these and um um um in the in the new stages we then create. So it's mainly about creating the story. It's a storydriven game which works with um with with chapters composed of stages. So you have an exploration stages you go through, you unveil areas and you collect um whatever objects and um but the art assets already there. So it's not about asset creation, it's about level creation, using these assets, creating a

story, creating quests and and the like. So the first step in the pipeline is creating the story. This is where LMS are very very good in creating text. Important to say this text is in the maximum mediocre. So not a super high quality usually, but it's a it's a start. But that's where the crafting part already starts with a game designer. I will I will always emphasize this here. So it's always a human in the loop really crafting it to the end. But begin with a story. Put in a rough idea into the prompt. Like it's all inside Unity. So all the whole tooling is inside Unity. So you open Unity, you have your editor tools ready and then you start with step one. It's really step one of the stage designer and then you have a rough idea of the story and AI works it out and then you iterate on it either by steering it manually more by prompting

it more in some direction. Um, you can even say, "Okay, this is all crap. I put in my own text is better however you created it." But then it's a great assistant in getting the story uh created in a lot shorter time. So, it cuts the time to actually have the story ready. Once you have the story ready, you get to next step, the quests for the story. It's it's quest lines we have in the game. Mhm.
So um these are created derived from the story created with a knowledge of all the game mechanics. So you don't need to handwire them. So they are created as a draft and then you yeah you work on it tweak it a bit here and there. You can always just just use it as is. So you can create a stage in like an hour but the stage would not be fun but it would would be there technically it would be there. You could you can play it but there are a lot of

issues with it. Um but to really to get it right and to get the fun in there and really get it consistent you need to always have this human in the loop and also someone really experienced and knowledgeable also knowing the game very well. This is something AI can never give you or from perspective at least not give you really the soul, the touch, the humor maybe in there. So, but you you're starting with a like a 70% um status. So, maybe just 50%, but depends. Um but you have a story, then you have quests, then you um get to the balancing. Balancing is tedious. This is where AI also had has its issues. But you get already a rough balancing and then just tweak it. It's very different and very shorter very much shorter to just tweak things than to um create things in the first place. All this wiring in in in Unity, all these objects, all this the code connections and so on. This is all predefined by the

tooling. Then you go to the maps. similar thing this we also need to to um differentiate between um L&M's creating being creative creating the story and quests and procedural generation of levels. So this is something we already do in the industry for a long long time that you have some kind of program creating you um maps or level structures that is also what we're using heavily. So you can create bases this procedural generation also by AI again. Um but this is not by definition not an AI creating the level by a tool which has been created. You can create the levels and then you orchestrate with a AI level higher for example. Um but this is works hand in hand procedural generation plus L&M on top plus a human on top and the human is always involved

in all these sequence steps. So you don't chain them and just run play. We always stop between the steps because if you get a 70% result in the first with the first step you get a 70% in the second step you already just by 50%. So
you will never get to good good quality in the end. But breaking it down using all these techniques um this is working out. Then there's object placements and then there's cinematics even and this was the most surprising part that it's also working as a as a um as a director for creating the cinematics. So you just put in okay roughly I want in this in this 5 to 10 seconds cinematic um some characters having a dialogue running around on the wall a bit interacting with with each other. You just phrase it as a prompt and then the L&M is doing the the work of wiring it all up and thinking about

okay where do I need to place a camera? How's the camera movement? Where are all these key frames whatsoever? This is done by the AI and this was a very tedious work before. So a lot of time was spent into the cutscenes manually before and this was was surprisingly easy to let the AI do it.
Not not super easy by the sense okay you don't need to think about what what you want to do but it was very good in the end like in two weeks we had a very good working um proof that it can do this. This was of course for a live game with a lot of assets with a lot of data for for how things are done with workflows that are already set. How difficult would this be to replicate on new game? Especially since you were talking about that you get about 70% right with each step and you kind of have to verify between each steps in sickness and in health in soft launch

and in live ops through every meta shift and every platform policy [music] change that's the kind of loyalty vibes built between your game and your players AI powered targeting a [music] real metalayer rorowass as optimization that actually compounds. That's what makes Vibes the fastest growing loyalty platform in mobile games till churn do us a part. Vibes vybs.co. [music] So it depends on how you create a new game. So um common approach is and that what what we are also doing is to to try to prove the fun first somehow doesn't need to be the final assets and creating non-final assets AI also is very very good image creation asset creation in general if you don't want to have it in the final product you can you can start with with with doing all that so you create a game have these temporary

assets in there and if the game works out it's fun you believe in it then you can we can continue in uh investing more. Um so we have now approaches where we want to think this more radical. So if you start fresh okay how does a process look like if you start AI first. This is an experiment. So we we this is not not a proven thing we have in the industry and just can use it as a template. So you need to find out for yourself. What we we're doing here is automating things, automating workflows. Um, put a layer on top what we already have, creating something fresh needs different approaches, different okay the team also very small I would say in the beginning um but really think everything through how how I friendly can I even make this this process um that LM can really help.

It's a bit different problem we need to solve with a new game than with this existing game. Has AI made new game development easier for you? And I don't mean in the sense of passing the sort of a pre-production gate where now you're getting into production phase and it's technically new game development, but it's not really. You're building the game that you already decided to build. What I'm curious about is has AI helped you to find the fun faster.
Absolutely. So um so the game designers are now really able so they are iterating on the ideas. They're brainstorming. They are discussing ideas and now they can actually really build themselves like the prototype. Is it fun to play? So is what I have in as mechanics as a core loop whatever in mind is it actually working out? They did it before with some kind of uh custom tooling and uh
some did it even with PowerPoint and animations and stuff to somehow somehow

transport the idea. But now they can really transport it in like a a game format. And this is helping a lot to find out is this way a a good approach. Should I should I continue? Should I try out five different steps? And then you always use just the best or the best two. But you have a lot broader space you can you can you can pick off pick from. One of the interesting thing in the piece was that you stuck with chatbt for model and you didn't switch to any new models. You didn't get, you know, fomoed into claude or any other LLM that would have been on top of the charts at that moment. So what was the reason that you didn't go for a switch? I'm sure your UI team is good. The question is whether the [music] data feeding their decisions is keeping up with them. Pathmatics from Sensor Tower will give them [music] what your internal dashboards can't. Historical

spend across every major platform. Competitor creative rollouts, share of voice trends, and seasonal patterns you'd never spot otherwise. It's the difference [music] between playing with fog of war and having the whole map revealed. Pathmatics from Sensor [music] Tower because you want your UA team to be OP. It was um GPT 40 the model behind TEPT for a long time I think. Oh wow. Well from my
perspective this was a breakthrough model. So there was TBT started with 3.5 um then 4.0 was another breakthrough model
and there we saw the abilities uh the intelligence actually that this approach with village was even possible. It wasn't possible before but then the intelligence get very good already and you get also you can somehow rely on

the structures um the LNM puts out or you can even define okay I need this structure if you can't create it retry so you get some kind of a contract of what actually the output is absolutely necessary to to chain it in some kind of um pipeline and programmatic um yeah workflows. But as this worked out and we get to the results we wanted um we didn't see the necessity to switch also because every model behaves differently across vendors but also if you have one provider like let's say open AAI from model to model updates nuances are differences very different or it's behaving drastically and we didn't take the risk of um of touching this running um pipeline. So we actually achieved that

we the stage creation was taking months before like 3 months and we get down to like 3 weeks. We even can create it in one week but then we didn't reach the quality but um with a lot of iterations three weeks um some crafting on top we get very um good quality in three weeks three months before. So um further gain was not directly visible. Okay. Should I get should we get to to two month two weeks or something? Um so it wasn't really the model was a problem. The whole workflow you need to improve here and there. Um at some point we were forced to to change it because 40 wasn't supported anymore and was shut off. So we needed to switch. Um it's now on um 5.4. Um it was also working quite well. the switch, but you need to adjust a few things and you need to really um test the whole workflow again again and again and again because you don't know how reliable is it now. It was quite

reliable before and you don't know you need to test it.
Yeah. Yeah. I like how you're focused on the workflow and not the uh the tools all the time.
When we're talking about the models um so for this summer build pipeline we we stuck with for for a long time. So this model was working very well. that was solving the problem. On other aspects or in other fields like on engineering, we always went for the latest models especially then funamentic which turned out to be a very well um product for um doing engineering coding there. We always um as soon on the first day it arrived directly switched to these new models because the capabilities the intelligence gets so much stronger or often very much stronger with next model update and that you want to have in this field
and talking about the costs of using all these models like what is the AI API

cost to run Sunrise Valley?
Sunrise Village. Yeah. Sunrise, sorry, Sunrise Village. Sunrise Valley is the sequel that I that I was not supposed to mention, but uh but yeah, Sunrise Village.
Um so the cost of running the pipeline is really um not relevant even. So it's it's a few hundred uh months. I don't even have the exact number in mind because it's just not relevant. The way getting there creating all these tools and stuff that was relevant costs but also not um um like something fourdigit something um where it gets expensive is when you use agents now in the whole engineering process and coding there it's getting uh a lot more expensive and this is where the where the tokens burn through. in the same article. So this was on games industry biz and they did a survey on games industry biz. This is like a industry publication for professionals and they had this survey

where 84% of the respondents said that you shouldn't use generative AI in any stage of game development. So you mentioned internal push back uh in in the article as well. What was the uh the sort of like most memorable conversations that you had internally without of course na naming anybody?
I think we need to clar clarify a bit. As far as I know, it wasn't about in any stage of the production.
It's more about uh in production.
Yeah,
that's how I interpreted it. I'm not 100% sure. Maybe you know better. Oh,
but as far as I know from this from this survey, it was more about okay uh you don't accept it in the final
if today's studios are not using it in any aspect of the production pipeline. So,
okay, hold on. I'm going to read the exact quote. I have it open here. Let

let me let me actually let me let me do a let me do a screen share so this is no misinformation.
Let's be let's be correct here. Yes, make sense. Let's let's let's be 100% correct here. So, uh we'll go in here. Boom. Screen share. This is game industry biz uh game industry biz survey of people working in the industry. Uh 84.2% of respondents said that they thought no amount of AI generated content was acceptable at any point in the development of the game.
Okay. Yeah, sounds like in any point of the production process.
You're right. [laughter]
Yeah. Yeah. Just just making sure. So, um yeah, [clears throat] this is this is the sentiment at least according to this publication. I don't know how widely spread the sentiment is outside game industry biz, but what were the sort of like memorable conversation given that at least this is one of the data points

or how people feel about the usage [clears throat] of AI? So there's a lot of truth in there. So there are a lot of aspects where people are afraid of or fear um have fear in theirs. Um the most prominent thing um we always talk about is quality. So creating slop um yeah this is this is a real risk. So um you can always create something arbitrary, something random. This is super easy. It's just a click. Everyone can do this. And if you are not taking the time into really investing time and this is a real decision that we want to do this, you can always take shortcuts and then you the quality is lacking. And this is um very often what I hear. So this is not a quality I can stand behind especially

the artists. So um engineers are more used to pick something from Stack Overflow and here and there some kind of snippet. Um but artists really had issues with uh with um yeah it's not crafted by them. It's it's something without soul. It's coming from somewhere. We don't know where it's coming from. So some artist is behind behind that. Maybe it's trained from that artist. A lot of concerns around this and this is absolutely understandable.
Mhm.
Other aspects. Yeah. Always job security. Yeah. Uncertainty. What happening to us? It's a drastic change, big transformation. Where does it lead to? And you cannot give a clear answer to this. Also me um dealing with AI every day like ingesting like every news I can find around this um I don't know where this goes in like three to five years. I can somehow foresee the next month. This is this is somewhat

possible.
Um but it's super clear that you need to use this powerful technology and ideally shape it for you. But you need to define as a company where do I want to go with the technology? What are our quality targets? Um for example and then also you can somehow mitigate this issue in the company. So okay we are not aiming for slop. Pretty sure in the next years well next year the next weeks or like like always from now on we tons of games are thrown out like every day. We know this already for the past past years but it's super easy to copy something. It looks even good but this is not a not a new game. This is nothing you really need on the market. So what are we aiming for? We are aiming for something more unique, something really interesting to play, something crafted with the new tooling.

But the tooling is super strong and um there's no way around it despite the push backs we also receive. Um maybe something I really feel personal is um this kind of AI fatigue in the sense of it is draining a lot of energy. So I'm using it like every day. I'm coding with it like every day. But as you get a lot faster with the tooling, you need to take a lot more decisions. You're producing just a lot more. So, um, big parts, complex things, but you need to take decisions a lot. And while one agent is running, the other one is running too doing something else. So, a lot of context switches on top. And this is something we really need to address that you are very exhausted by the end of the day. You had a lot of AI support but you were feeling more exhausted than before and this is a real real thing I experienced myself. But you need to shape it. You need to find new new rules how we do we deal

with it. But actively shaping is something we um we we are doing and we openly discuss it. The article also quoted a senior lead artist at Inogame. He this person said that he is not feeling threatened by the AI, but he also said that it's legitimate to be scared. If you kind of listen to those two statements, which one is closer to what you hear now in the hallways, 7 years, that's not a sponsorship. That's a marriage. And that's exactly how long we've been partnering with Apps. [music] And the reason it has lasted is simple. Every growth team I talk to on this podcast, every operator who comes and [music] breaks down what is actually working for them, eventually the conversation lands on measurement. And in gaming, we refer to reliable measurement with a single word, [music]

absor back this platform, too. [music] Which means the people competing hardest for your budget all agree where the single source of [music] truth lies. I hear both a lot. Um so the first thing is people felt threatened but as they practice it they see they have a lot of shortcomings. They are not perfect and your craft is still needed and then they are somehow relieved. Okay, it cannot replace me. I'm getting more powerful. I can do more, but it's not getting my craft is not getting replaced. My skill set is still valid, most of it at least. And it was kind of a relief, but this was the first threat. Um, and then you have these this fear about the future. Um, you don't know what happens. You cannot foresee it. This is um yeah, this is a this is something um I also hear very often. Um

the transformations pushed on them. It's not that that they didn't they did have a choice if if they want to they are some are pushed they need to adopt to this new technology which they don't know where it goes and another thing that was mentioned in the article and something that you said was that AI is a learning accelerator. Can you give an example where somebody at Inoame genuinely got better because of AI in their craft. So it it can be used as an accelerator. It doesn't necessarily necessarily um um accelerate your learning. You need to use it right. So if you're lazy and just take shortcuts, you you lose a lot of skills. there is skill erosion if you don't don't use it um in the way you remember things you think things through um but you can use it differently you

can ask AI a lot um okay why how what other ways would be be able to do the same thing um this is something juniors can do um not all are doing this for sure but they can do it they can use AI as as a senior at your hand asking tons of question using as the pepper crema. Um so this um um um juniors are using this but I can also give you an example on on a game designer. Um um so what I meant already before they are not just can prepare the ideas they can bring it in kind of a game. This is what I say is a is a is a is a improvement in your craft getting to the next step getting adjacent skills to to what you already can do. Um I also see this with uh with developers. So we had the distinction between back-end developers, front- end developers. What about a software engineer with the help

of AI? I can get better and also as a backend developer can do front- end stuff. I can learn it a lot easier easier for me. For example, I'm also an engineer. Um I never did Rust programming, Rust low-level high performance programming. I never did that before. Now I can because I understand the concepts, but I can't write Rust code myself. I can read it and I have the concepts. But I would say I'm now I have a new ability and others see this as well.
How how do you work around that? You mentioned skill erosion and that is a very important element because it's so fast to speed through the most difficult parts of your work with the help of AI. uh let's say you know in my case a lot of my work is writing and presentation surprisingly and speeding through those kind of takes away from that important thinking process and it backloads the work to the

uh to the QA part of the process where now I'm QAing what the AI did and I'm kind of like taking a mediocre answer and making it okayish from like you know errorfilled mediocre to like yeah this All right. Versus spending a lot of that time that I'm spending in QA to actually producing the work and then it's actually good and requires less QA on the other side. So I'm kind of trying to, you know, maybe that's a bad example, but it feels like we're moving the block of work from execution to QA,
judging things, steering things. Yeah, that's what skill skill is. Yeah. Um and I'm I'm honest here. I think you we'll lose skills. Um I think some example have been presented also in some of your past podcast episodes about the comparison like or the analogy with uh with with Google Maps or something. Um the same thing. So you you

is it's kind of this is right. It's it's who is able to really navigate through the city? I can't. And the same thing is you will pretty sure as an engineer you will lose the skill to write specific program language code by yourself. how the syntax is, how specifics are in the language, you will lose this.
But important to me is to really is it? I don't know.
Is it I travel all the time. Sometimes when you're in a different country, your internet doesn't just work as well as it should
and suddenly the Google maps is not showing me the points on different restaurants where I'm going to stop while I'm walking through. Is this a 4.2 or 4.8 sandwich spot? there's a there's a highly important information for me to [laughter] to to make decisions and then suddenly it's really weird to be in this world where you have to trust other senses than what Google or whatever platform you're using for navigation and

telling you on different spots. And I would imagine that let's say Google like there's some kind of a thing and and the navigator doesn't work at all. Can we even you know travel from one city to another without without a navigator these days?
Yeah, you will f find your way. There are still some signs out there which
some direction
but I felt it myself a few times already. Um like I was I was driving in the opposite direction. I didn't start the the navigation. I I thought, okay, I know I know where to go. And then suddenly I I just drove in the wrong direction for like half an hour or so.
Yeah. Would you be able to make it from Hamburg to Potam? I don't know. I don't know.
Roughly. Yeah. Not not in the direct way maybe.
But to answer it, it's um why it's from my perspective not a real problem. So sure always if you lose a skill, you can't do it anymore. But is it really

relevant is is is a key question. And as a software engineer, is it relevant for me to really write or understand? Yes, but really write the syntax of of Python or of of Rust or is it important to know this the engineering concepts behind and create something? So I need to compose something. I need to to architect something. This is the engineering skill, not writing the code actually. What at least this leads to if you are released from some building work, you can likely create more complex solutions. You can spend the time thinking through the whole architecture. Um, compose something more complex, create a create more parts of a game, um, create better systems. Um so so you can't you can you save spending time into one

thing which is not really necessary anymore and use it for something else.
Mhm.
Um you don't need to be laying bricks to build a house anymore or do it with the electrician. Someone is doing it for you. The agent is doing it for you. So you don't need to build uh this house in every aspect on yourself. But you can design the whole house. You can architect the whole house. Maybe you can even architect the whole city. This is what you what you now empowered to do. So not the single thing, the basic thing. Um but yeah, some people are very proud of these simple things and it's even not simple. So it's very hard work to to write Python or Rust code specifically. Uh so super um hardly earned skill over years. Absolutely. So absolute respect on that. But the reality is it is not really necessary anymore to do this. You need to really think about what is your core

skill you need to keep and what what other skills I might be allowing to yeah not completely vanish but to to degrade or erode in some way. And this is you need to accept hard to do so. But um the agents or the the LMS can write a lot more code in better quality than you you likely can do. When thinking about this 84.2% of the game industry biz report um audience who answered that geni shouldn't be used at any point of development. How many do you think are influenced by the fact that they spend decade or maybe even more to learn these skills that now people can kind of inst like not even learn just start using what they've learned immediately almost is that what pays the effect of being

anti- AI
it's a part of it for sure there a lot of reasons being anti-I so we have also prop variety in our company. Um but one thing is uh is losing skills. Yeah. Um just I can just tell okay from my perspective um where does it go and um being ahead of it or being uh in the in the in the being somehow in the driver's seat or together in the the front seat somehow and not just a passenger somewhere. Um I think this is important. So um make up your mind okay what is really necessary in my profession and this is a good part of it. Your skills to a high degree are needed. A lot of what you have learned and experience you have is super crucial. The judgment the steering is super important.

Luckily, we are the ones need it and we are the ones um really driving it to something great. No one is interested in slop or something mediocre or or you name it. It's it's yeah some some beginner can do maybe uh write some code I need to spend learning five years but they cannot construct something complex out of it. They cannot maintain it. They cannot make it make it great. And I fully believe that we are getting flooded with tons of slop. It's already everywhere. And u the ones really creating something great. this will be recognized and you can also use this technology and iterate craft something just with a higher leverage with more power um with broader skills and um I'm

absolutely optimistic that uh we can we will see also great games products heavily inspired or created by AI but with a lot of crafting on top
just Final final notion on this 84%. I'm I'm still like just amazed by this number. But if this number reflects the western game development, what's your honest look on the outlook of the western game development? If such a large portion of developers don't want to even engage in these tools, if it stays this way, they will get into trouble for sure. So um if you use it, you experience how po powerful these tools are. You cannot keep up. That's a reality and the outcome is clear. If you don't use it, you cannot compete. And this

leads to various effects. All right, I'm going to take us out with a direct playbook. Let's not talk about Inoame. Let's talk about Hypothetical Studio. So, I am a CEO of Hypothetical Studio. I have a team of 15 people. We got a game. It's making about 500K a month. It's tying up all my resources and it's not making enough money where I can start something new and the revenue might be declining. So, what's the step one? Where do I start? I listened to this podcast. I read the the story on Game Industry Biz. I'm calling you. Thomas, help me out. We're stuck. We're going to just die if if we continue this path in like next three years. Imagine I would just have one person able to continue this game. Yeah. What how what do I need to adjust? Don't know if it's possible, but it

makes sense to think about this. what tooling needs to be in place that what was this one single person gets enabled to run the whole game. So if if it if you need to create tons of uh assets um yeah might be difficult but you need to think about okay how do I get this these assets for example but it is about okay decomposing finding actually where the effort is spent and can I um um yeah create tooling can I use AI um in supporting me in this um but it's it's about a drastic scale down like a t 10 x scale down of the team or maybe just half it or it's drastic and and bring it to the max like one person how would the tooling look like how what does this person need and then you

automatically come to the point okay or this is a big complex thing to do this is a big complex thing to do you need to decompose further what is complex about that and where actually if you have it completely decomposed what steps can be taken over by a workflow by automation by AI by procedural generation whatsoever creating this tooling will take time so you need to invest first um but if you if you and this This is my real advice here. If you if you start with the things maybe the simple things which eat time and you have the first nice results in speeding up some kind of sub process, you already have a win and then you continue with this. So um there is quick return I would say if you start with a um big blocks and with some kind of obvious things to which you can automate. We were wondering about when we observing how how people work even

they work for for the title for years already and they did a lot of optimizations still when you do observing people in the sense of okay discussing with them and looking over the shoulder what are you doing here and there you see a lot of things which are good candidates to to um to automate and with a today's technology so it is super easy to automate something you can even ask AI how can I automate this ask me questions uh about the context. So, and then you you you interact with the AI and it proposes you something and you get more specific um on the way. Um if you can define it, if you can specify what you want in the end clearly specify what you want, what's the input? What's the out what's the output? What is expected? If you can be clear about this, if you work this out, I have a good chance that AI will provide you a tool doing exactly this for you. I assume that the communication to the team is primarily important because if I

start using AI and looking over people's shoulders of how they're doing their work so that I can automate their work that sounds quite dystopian.
Yes. If you just use this this is frame. Yeah, for sure.
Yeah. Um but you releasing them from
Thomas is only seeing how you do your work so that he can make computer do it.
A lot of this work is not fun. So it's not about you have a creative part you're working you like working on these are the the problematic fields but most of it is just work. It's not it's not nice work. It's just there. You need to repetitively do um repeating doing it and and this is not you you even relieved if someone is taking away this repetitive work from me.
Yes. Yeah.
But sure communication is key here. So what happens then if all is automated you need to have a plan. What's what

what is next? Okay. I help you with automating all this but obviously we need less people in here. Yeah. Okay. But then if you have a plan what to do next um and if it's just about cost cutting and not finding crows somewhere else yeah okay then you are somehow screwed you cut the costs you're not needed anymore this is not how most businesses or not the business I am with are working they are growor oriented like we we want to we want to find this next big thing the next great business case the next new game the next best idea we can work on if we have more chances, more bets in the field. Um, we can distill them the best ideas from um we have a higher chance of succeeding with a very nice game and freeing up people um yeah bound to some kind of tedious work in uh and and and freeing them up

for a bigger leverage for bigger opportunities. I think then you have an easy story to tell and this is how we see it.
Yeah, I understand. Still dystopian but I understand. Uh, can a studio without without a strong internal tooling even attempt this type of a?
Yes, because you can create this tooling quickly. If you can solve it logically and specify as I said, you can easily create this tooling. But sure, it helps if you have already like CI/CD in place, if you have um a lot of workflow tools in place. Yeah. But um I have a start startup background. So um I did a startup 10 years. I would have loved to have this this technology at hand. We were building a lot up from scratch. I was I I needed to solve problems uh foundational problems uh um

tooling problems over months. And now I would have be able to do this in days. Um so startups should this should be very startup friendly technology. So um they are very dynamic uh super agile and they will see um they can build up things very quickly where big uh companies uh needed a long time to do so.
And final one on this is like the the vendor lock in that people might be a little bit afraid of. You say that the uh the model is that you have on Sunrise Village, not Valley, Sunrise Village [laughter] that is model agnostic. Um how much the model swap would cost in practice in terms of time? Is it days? Is it weeks?

It's weeks. Depends. So you can try it out and just it's it's super easy to switch a model. It's just a like a like a label and then the the connector is connected another LM.
Um but yeah the the devil is in detail and it will not just work out as as before. Um but with testing and stuff it will take uh like one to two weeks. So not not not forever. More problematic it's um and we invest a lot into our engineering tooling. Um so with Entropic for example um if you build up this ecosystem or if you use this ecosystem for building up all the your your your tools your workflows the harnesses harness engineering are the big thing at the moment. So really have your
harness what's a harness engineering
it's wiring all the tools the skills the tools the connectors up. So you have um

you have an agent connected to your in-house knowledge um have created custom skills for doing something um having prepared prompts sub agents a lot of things you wire a lot of things up so we have basic intelligence of the of the of the model so it's important to have a good model there but more important is getting how I use this model, how I connect it to different things. Um, and if you if you are um on one ecosystem like Inotropic or OpenAI, you more and more use these um abilities from that ecosystem to define all these skills. The skills would be okay, create a PowerPoint presentation for me. This is input. This is a text. Now create a presentation and you can define a skill for doing this. You have another

connector for your internal slack. You have a connector for your bit bucket for your gitlab um for confluence whatsoever. Um you have tons of connectors. You have defined when they trigger uh you have defined quality gates. This is important thing by the way. Um feedback loops and quality gates. So an agent um will create something. Can I find hard criterias? What good quality means for me? If I can do this, um, the agent gets feedback from the hard criterias and can selfcorrect. So, you reduce the amount of human in the loop steps, the interrupts. You only want to do really substantial decisions, not um, oh, it's not working, please continue or please or please fix it or whatever. And um this is this is also part of the of the whole flow and and and the harness to to set these quality

gates to maybe get set sub agents that do a review um or feedback something on this is a lot lot of puzzle pieces you can put together and there's a big puzzle and doing this um is is um is prompt engineering was was a start. So how do I prompt my my model? Yeah,
tons of techniques. Then there was this context engineering. So prompts nice, I have good prompts, but now then the the thinking models say more or less uh makes it obsolete to write craft handcraft prompts. But then it was about okay, what do I have in context? Models don't know about your specific problem about your specific product. They don't even know about Sunrise Village. They maybe have some information from from like last year um some wiki entry about some village but you have a lot of internal company knowledge about the game. You have the game design documents and all that and you need to um to wire that up and um if the agent is able to

ingest on demand the relevant information about the problem about the product then it has a high chance to succeed. So this context engineering and now comes the next thing and this is a harness engineering. So having orchestrating the agents, having multiple agents, um having workflows defined, um quality gates, constraints, feedback loops. This is where a lot of um the engine neon now goes to. It's more important now as as the models. Why is Ino games talking about this? Why are you talking about this openly? because there's not a lot of companies that talk openly about how they use AI.
Yeah. So, um so we set ourself um this um this um we are AI friendly. We decided we want we are optimistic on technology. We want to use it. We want

to craft it with it and we want to bit um influence on how negative that is also seen. Um so um there are a lot of aspects to it but um it is a real for me super amazing technology super interesting and um we we we use this technology we will create will create games with it and we also want to somehow influence a few on this and um yeah this is um also also This is a bigger strategy also. Um we have also push backs in in in the company as well. So um we want to be very openly we are using AI we want to leverage it. We want to build great game with it and um by talking openly about this. I think um it is it is very clear and um

well Thomas I want to thank you Ninoames for talking openly about how use AI giving concrete examples running publications coming on podcast explaining how you'd use different tools how do you design your workflows what were the learnings what were the mistakes and what were the success cases because I think this type of a conversation helps to lower the bar of usage of AI and takes it away from the hype talk of the newest tool and the newest model into much more grounded conversation. It's not AI labs. It's not uh the latest startup rounds. It's not the super big corporations running their latest video models. It's actual work, workflow designs, and coming up with tools that make our work more efficient, that allow us to do produce more, better, at lower cost, and at the higher quality.
Thanks having the chance being here on your podcast.

And
folks, you can reach out to Thomas on LinkedIn link below in the description, ask him more questions about how they are using AI and use AI smartly. On that note, thank you Thomas for for sharing the uh
sharing the learnings and thank you everybody for listening.
Thanks Mishka. See you.
You did it. [music] You made it to the end of the episode. As a fan of the show, it would help us out if you'd subscribe and leave [music] us a review on the podcast service of your choice. Those stars and likes, while a small action on your part, mean a lot to us and help to grow the show. More importantly, are you a member of the Deconstructor of Fund Slack group? If you have 5 years or more of games industry experience, go to deconstructorofund.com/slack [music] and apply to join. Join the games industry best professional community filled with peers always willing to lend a hand or [music] subscribe to our newsletter to get all

the latest insights from the Deconstructor of Fun content creators. Thanks for listening. [music]