Building an AI game studio: what we've learned so far

I think a part of the reason why I&#x27;m kind of dismissive of AI when it comes to creative endeavors is that I haven&#x27;t seen it do something that isn&#x27;t already possible using a &quot;not-intelligent&quot; solution.<p>For example, AI website builders. It&#x27;s impressive as a tech demo that you can prompt an AI to make a website for you, but otherwise I don&#x27;t think this is going to be disruptive to the web design industry since it&#x27;s already pretty easy to make a website using a drag-and-drop builder like Squarespace or Webflow. Granted, it can get complex when you add certain special features on these platforms, but same goes for an AI-generated website. AI doesn&#x27;t eliminate the complexity, it just hides it. At best it&#x27;ll get you maybe 80% of where you want to go and you can get to the remaining 20% by hand. At worst, the complexity is so thoroughly hidden that addressing it isn&#x27;t an option. It simply does not allow you to manually edit the outputs, or makes it unreasonably difficult to do so.<p>In that same vein, an AI game maker is pretty impressive as a tech demo, but I feel like it isn&#x27;t a huge improvement over a code-light game maker. You&#x27;re not eliminating the complexity of making a game, just hiding it.<p>Nonetheless, it&#x27;s a pretty interesting tech demo.

Spot the fun logic bugs! Here&#x27;s a great example of where AI requires some careful attention to detail:<p><pre><code> const tentsFacingNorth = getAllUnitIds() .filter(unitId =&gt; getUnitString(unitId, &quot;unitStrings&#x2F;tags&quot;)?.includes(&quot;tent&quot;)) .filter(unitId =&gt; Math.abs(getUnitRotation(unitId) % (2 * Math.PI)) &lt; 0.01).length; </code></pre> This code only works if getUnitRotation() returns values between -pi and pi, with north pointing up. This isn&#x27;t an unheard of convention, but it would surprise me -- if you asked the AI to generate movement code I&#x27;m sure it would do const moveX = Math.cos(); for instance.<p>I would more traditionally expect a range of 0 to 2pi, but then this code wouldn&#x27;t work for anything close to 2pi on the other side. And if the range was -pi to pi, why have the modulo use 2pi :)<p>Additionally, &quot;facing north&quot; probably implies a bigger epsilon than 0.01 radians. That&#x27;s only half a degree! Even if it was super tight, I would expect maybe 10 degrees slop in either direction, which would be 0.17 radians.<p>The gameplay programmer way to do this would be to get the facing vector and dot that with the vector for north, and compare the results. Dealing with angles isn&#x27;t worth it.<p>Oh, also, I&#x27;m guessing getUnitString().includes(&quot;tent&quot;) would work for anything with substring tent, so no adding tentacles to your game.

Hard facts: Steam has over 73k games - and is growing at over 10k games per year and increasing. The median game on Steam earns less than $800 - down from $4k in 2020.<p>Making a game is not hard - marketing a game is.

I think this would have more chances of success if it focused on just one aspect of game creation, instead of trying to solve engine development, gameplay development and content development all at once.<p>I.e. if an programmer wants to create a game, then getting the artwork in a consistent style is a big problem, and solving that alone (i.e. as a plugin for Unity or Unreal or Godot or whatever) would be both powerful and difficult.<p>Conversely, if someone without programming skills wants to create a game, then it would be best to focus solely on the AI assistance for gameplay design with a fixed, well-known engine and an integrated well-populated market for content, and assume that they will get the content packs there, leaving that problem out of your scope but making it easy to integrate the AI-generated logic with the pre-made content packs.

They mention using a service called Meshy for generating models, and I noticed that they have a free plan where the generated assets have a &quot;license&quot; of CC-BY-4.0[1]. Any lawyers in here that can comment on that? They&#x27;re claiming copyright ownership over models generated from customer prompts, and I doubt they trained their model exclusively on stuff they own&#x2F;licensed.<p>1: <a href="https:&#x2F;&#x2F;www.meshy.ai&#x2F;pricing" rel="nofollow">https:&#x2F;&#x2F;www.meshy.ai&#x2F;pricing</a>

I know this is a small team and experimenting, but in my mind, this engine is already limiting the potential of games it can output.<p>I don’t know, but Unity is already pretty much point and click game building. I spent a few weeks building a card game with it during lockdown, and without a lot of code, I had an ugly, and basic, but totally playable game.<p>The hardest part to a game is building something fun yes, but it also has to meet some level of quality bar and work on any platform.

How will you ensure the content created for a game is free from copyright violations? Or is that up to the user to determine?

Not a bad idea. But what&#x27;s next?<p>For example, I am indeed interested, but will anyone really play the game I make with AI?<p>If I spend the same time watching YouTube and learning Godot or Unity, will it be more likely to make games and even make money, and I can also use Meshy to produce my assets. Unity also has Muse (though it&#x27;s hard to use).<p>If the games I make don’t make money, what is your business model?

Hi HN. We want to make it possible for anyone to build games without first learning to code or make art, by giving them access to an “AI game studio”. It’s early days, but here’s what we’ve learned so far building this product. Would love to hear what everyone thinks, and see if we can find some alpha testers for this!

This is genius! I currently work on attention loop games (the kinds you see ads for on your social newsfeed). A big bottleneck is the number of new games we are able to produce per week. New games generally have a better ROI and get more installs and clicks than older games, which age out.<p>Having the ability to programmatically generate new games would be amazing. It would be great to have the game completely generated by AI, and then pipe the screenshots into a video AI tool to create the ads.<p>Will it support slot based games and gambling card games? My head is just buzzing at the possibilities and the opportunity for profits from this!

This is very impressive. If i&#x27;m understanding the post they basically created an inrernal api and just have the llm use that api. Seems like a great way to workaround llms struggling to keep track of large codebases.

What a neat idea!<p>Next: a strategy game where you give orders in plain text, instead of hundreds of clicks per minutes

Any plans for LLM based agents :) ? NPCs will be amazing. AI Dungeon is probably a good example for character setup and story snippets to weave in (so you don&#x27;t get all generic &quot;helpful assistants&quot;) :).

I&#x27;m very into the idea of using AI to change the types of games we make, but I think that &quot;prompting the game to do stuff&quot; is barking up the wrong tree. As the top comment says here, you can develop non-intelligent solutions to do what is going on here in probably the same if not less time. This is more like &quot;end of chain&quot; AI work and doesn&#x27;t center AI as necessary for the game itself. If anything this is using LLMs as a shortcut for keyboard shortcuts or something, which is fine, and likely necessary work to get to something new and interesting, but already makes the category error at the top of constraining what is possible based on the need to fit the LLM into the rules box.

This is really cool thanks for sharing. I have built a generic app builder that follows the same idea. I am dealing with the same problems!<p>It is an iOS app called Burning Idea, it lets you build small apps by talking to it.

Wonder how well this will do given that, going by reactions to even fairly anodyne AI usage in games, gamers really <i>really</i> loathe gen AI used to make games and game assets.

Question: the demo uses star wars imagery, including X wings and tie fighters, going as far as naming BB-8. How are you not getting sued to kingdom come?

How robust is the system to fix bugs? Can you just point at a thing and say &quot;fix this bug&quot;? How does that side of it work?

Common remark about game making (also here) is that “there are already a lot of games, building them is not the limiting factor“, and yet I find myself without anything to play despite these 70k games on steam. I like playing city building games, all the good ones I already beat and there is nothing new that seems any good!

First thing I noticed in the demo is the lasers casting shadows on the ground. Which doesn’t make any sense

I find the &quot;threads&quot; approach to prompting in a complex app very intriguing. Did you experiment with threading at different levels and how they interact? If so what can you share about this?

This looks like the perfect tool for rapid prototyping. I am not yet sold on this being a viable method to produce commercial games, but I do think you guys are on to something. Very cool!

That&#x27;s sweet! Like the UI and threads approach.<p>How does it generate 3d models?<p>Btw there is also <a href="https:&#x2F;&#x2F;frvr.ai" rel="nofollow">https:&#x2F;&#x2F;frvr.ai</a> for more casual games

Could you leave the tapping and chatting to the AI? This way, I can let it work on some basic guidelines and let it surprise me with a game every day.

I wonder why their studio is named after a way to cheat on exams?

This doesn&#x27;t happen to be the AI game startup at which the host of the wonderful <i>Last Week in AI</i> podcast works? <a href="https:&#x2F;&#x2F;lastweekin.ai&#x2F;" rel="nofollow">https:&#x2F;&#x2F;lastweekin.ai&#x2F;</a>

This is incredible. Thanks for the great write-up!

we are in need of a good 2D strategy game such as Advance Wars or Fire Emblem that can be played with infinite variations

AI is so creatively bankrupt and hated by nearly every professional in the industry that isn&#x27;t a programmer that using it feels like a real liability.

I think most people don&#x27;t actually want to make games.<p>Most people want to take an existing basic game, and tweek it a bit, incrementally, adding new art, ideas and levels, and removing the stuff they don&#x27;t like.<p>They definitely do not want to write code.<p>They most certainly do not want to debug generated code and I think you&#x27;ve hit the nail on the head with &#x27;generating actual code sucks, is hard and basically doesn&#x27;t work&#x27; with current models.<p>What I see in the demos looks different though; it looks like you have a <i>configuration</i> driven set of basic game elements and let the AI tweek the structured and easily-to-validate-schema config.<p>Which I think is fantastic.<p>&quot;Make this unit bigger&quot;, &quot;Add new type of fighter&quot;, &quot;Get rid of this unit&quot;, &quot;make the terrain here smoother and add a lake&quot;; I feel like this kind of structured tool use is currently possible, if you write a framework of constrained actions (which is what it appears you&#x27;ve done).<p>So, props to you. This feels like something that is both actually novel and interesting.<p>...but; I feel there is going to be a difficult path ahead.<p>The more <i>generic</i> you make the engine, the more different types of games you can make with it; however, in doing so, the tools you have (eg. &quot;modify unit config&quot;, &quot;generate 2d sprite for unit&quot;, etc.) will have to become more generic; and the more generic they become, the less effective the agents will be at using them.<p>Our current batch of GPT4-ish models seems like they&#x27;re relatively well suited for using <i>specific tools</i> with <i>specifically constrained</i> inputs to plan and achieve goals; however, as the tools become generic, the solution space balloons out widely and you start getting random crap instead of actual real solutions to tasks.<p>Specifically, I&#x27;m skeptical about your generic `createOrUpdateRule` and that you can effectively scale it to complex interactive behaviors; it just feels like the agents will never express complex ideas with that kind of coding; it&#x27;ll only ever be an array of trivial behaviors.<p>Maybe that&#x27;s all you need in some cases; but I don&#x27;t think you can build actual games that way.<p>So, practically speaking, this might actually only ever really work as like a &#x27;super modding tool&#x27; that takes basically a <i>fully working game</i> and lets you mod it in a very specific set of ways; but that would mean creating a &#x27;template game&#x27; for various different game types, and different &#x27;modify game&#x27; tools for each one. Otherwise, the final product is never going to really be beyond &#x27;unit moves back and forth randomly&#x27;.<p>...but, that would still be really fantastic.<p>I guess, I hope you don&#x27;t get lost trying too hard to build a generic Unity clone using AI that is kind of so-so (ie. generates random code that doesn&#x27;t work and you have to constantly debug), instead of an amazing thing that is slightly less generic, but lets you do amazing things <i>within a specific set of restrictions</i>.

The models look awful. If only there was a place where you could go to download human-created 3d models, either for free or a nominal fee. &#x2F;s

So ... does every single AI demo put copyright and intellectual property infringement on the box?<p>Just asking? How on earth can an entire ecosystem be so damn tone-deaf and lack self-preservation?