I know there's a lot of push back in the game development community about generative AI, so I wanted to get the topic out of the way as soon as possible.  There's a few different ways I've been using gen AI in Rootless, and a few more ways I'm planning on using it.

Generative AI as a Development Tool

In my day job as a software developer I can say that generative AI has become a very powerful tool.  I often use it to automate tasks that follow established patterns that would otherwise take a lot of manual work, like writing unit and integration tests, utility classes, or small refactors.  In addition I use generative AI in combination with my own experience to learn new programming languages or libraries, combining asking the right questions with the language model's web search and ingested knowledge.

For game development I can confidently say Rootless wouldn't exist past an idea in my head if it wasn't for generative AI.  Using gen AI I've learned how to program in Unity, had discussions about the best approaches for certain features, taught myself about shaders and materials, and numerous other things.  Gen AI has been the perfect "interactive rubber duck" for discussing ideas and working through blockers, ensuring that I can stay excited and make progress on Rootless.

Generative AI as an Asset Generator

This is where things get a bit more dicey.  I've certainly had some success using services like Nano Banana and Meshy.AI to generate assets for my game, but none of them are quite "right".  They often have weird artifacts, spacing, highlighting, formatting, or other things that make it somewhat obvious that the asset has been generated by a model as opposed to hand-crafted by a human.

I fully support and intend to use human-crafted artworks on Rootless when the time comes to deliver a polished demo, but until then I feel generative AI assets as placeholders are a suitable substitute.  One of the biggest barriers to progress on my previous game projects was a lack of access to art; if I can use gen AI to overcome that barrier in the short term and be able to describe my vision easier, then Rootless will be that much closer to becoming a reality.

An example of this is this very blog's images and backgrounds, generated with Nano Banana.  Do they look great?  Not particularly.  Are they going to be the final product?  Likely not.  Are they better than having a blank grey screen and will they help draw others into checking out Rootless?  Heck yes!

Generative AI as a Feature

One of the defining features of Rootless will be an emergent narrative that responds to the user's actions and choices.  There are certainly games that have accomplished this without the use of generative AI - Wildermyth being a recent example - but I'm excited to pursue the possibility of using language models as the main narrative driver of this feature in Rootless.

Outside of my two years or so of experience working with language models through my day job, I've also created a "choose-your-own-adventure" game called Infinite Scribe to prove out this concept.  Infinite Scribe uses player input and game state to craft the next part of the story, allowing users to choose between three options (just like the old CYOA books) to see what happens.  The player's choice is combined with the game state of past events, the player's current health & stamina, effects, and inventory to craft a context-sensitive continuation of the story.

As an example, let's say your player character witnesses an NPC completely whiff on an attack or skill they were attempting.  Using the "say" action, your player character can call out to the NPC.  The game engine will generate two potential outcomes (maybe a Might -1 and a Resolve -1), and those outcomes will be fed into the language model to generate dialog options that would justify those outcomes.  This way the game logic won't be subject to generative error or hallucination, and we can provide a dynamic, contextual story option to justify the mechanics.

Generative AI as a Cost

Generative AI isn't free by any means.  An average "interaction" in Rootless right now is testing at about $0.006 per interaction.  While that doesn't seem like much at first, if you imagine a lot of players playing Rootless at the same time that could potentially add up.  For example, Final Fantasy Tactics: The Ivalice Chronicles, a recent popular and successful TRPG release, has seen a peak Steam player count of about 21,000 - if every one of those player did one interaction in Rootless, that would cost $126!

To broach this I'll be taking  a few measures from the implementation side to minimize unnecessary costs, such as rate limiting, caching, model selection, and others.  The goal is to provide a unique experience to everyone playing Rootless while minimizing the cost, power, and environmental footprint of generative AI use.