LLMs: gimmick or game changer? A naughty wizard may have the answer…
A Year of Discoveries at Atomic Dirt
Excuse me, your ChatGPT is showing
For the better part of 2023 you couldn’t step out your front door without hearing, “AI will change the world.” In gaming this mindset opened the floodgates for companies to offer generative tools to create assets and artwork, enhance(?) NPCs and even let non-coders program games themselves. So much potential! But let’s look at the reality: currently many AI implementations in games are little more than OpenAI API integrations. If you look close enough, you can see ChatGPT behind the curtain injecting variety into the copy or giving the player the illusion of deeper gameplay. But ultimately what’s happening is a gimmick. And that’s understandable, because Generative AI is a big, clunky beast of a technology and finding out its best use cases takes time.
“Raised by a naughty wizard!”
Atomic Dirt is still very much a young company, but after one year of planning, discovery, and production, we are well on the way to creating a game that’s purpose-built to leverage LLMs to power an experience that simply couldn’t exist without AI. But it took a considerable amount of blood, sweat, and testing for us to get here. So I thought I’d share what we’ve learned in the trenches, and by trenches I mean the ghastly corridors of Dracula’s castle.
Let’s look back at the beginnings of our first prototype – Harker’s Escape, as well as the technology we developed to power it, and how this evolved and informed our thinking on how AI (combined with other technologies like XR) will shape how we play games.
One of my earliest gaming memories is of playing King’s Quest III on my friend’s Tandy computer. Something about entering any text I wanted to try and escape the clutches of an evil wizard felt so free. I could do anything! In reality that wasn’t true. Most commands didn’t work, and so we ended up typing “fart” into the command prompt. That worked! Many, many times.
A poor upbringing
Why did that style of game never mature past the early days of keyword groupings? Maybe it needed the magic of LLMs to help make the experience what it could be? Instead of getting “I can’t do that” responses to every prompt, what if the game engine would let you say what you wanted in your own words, understand that and then let your character attempt it i.e. let you try just about anything to escape the castle?
Our first prototype “Harker’s Escape” was born of this very concept – we put the player in Dracula’s castle as Jonathan Harker, where they needed to escape before nightfall when Dracula would arrive to…not drink wine.
The idea is simple: you wake-up in Dracula’s castle and have to use your wits to escape. You can try anything ex. Crafting a crucifix, picking locks, making a poison or you can even try to convince Dracula to let you go. And while it was a small sandbox and a straightforward concept, it provided no shortage of AI-powered gameplay possibilities and conundrums to sink our teeth into. I will not apologize for vampire puns.
Next week: We’ll dig into the lessons we learned making Harker’s Escape.