AI-powered games have been chasing the wrong dream—endless content generation. But more dialogue, more items, and more quests don’t make a game better. The real challenge is making every action meaningful, every choice connected, and every consequence real. At Atomic Dirt, we’re using LLMs not to flood games with generic content but to build true interactivity, where player decisions ripple through a world that actually reacts.

At Atomic Dirt, we’re pushing the boundaries of game design by integrating large language models (LLMs) to create more dynamic, player-driven experiences. But how do LLMs truly fit into gaming? Do they enhance player freedom, or are they just complex systems predicting text without real understanding? In this conversation, we explore the balance between constraints and creativity in game mechanics—and whether LLMs can actually follow the rules we set for them.

A Year of Discoveries at Atomic Dirt

We delve into the challenges we encountered while integrating AI-powered input methods into our game, Harker's Escape. We explore the inconsistencies that arise when players can use voice commands or typed instructions alongside traditional keyboard and mouse controls. For instance, players might "type" to pick up an object but then have to use the "W" key to move. This clash of input methods creates confusion and disrupts immersion. We believe that as AI opens up new possibilities for player interaction, we need to rethink game design to ensure a seamless and intuitive experience. Our goal is to leverage AI to enhance gameplay, not create unnecessary complexity.

A Year of Discoveries at Atomic Dirt

In this post, we delve into the strengths and weaknesses of Large Language Models (LLMs) like ChatGPT and Google Gemini, focusing on their potential for game development. We find that LLMs excel at understanding user intent, allowing for flexible and varied input commands, which opens up exciting possibilities for accessible game design. However, we also observe that LLMs struggle with maintaining long-term coherence and logic. This means they can misinterpret information or contradict themselves over extended interactions. While acknowledging the promising advancements in accessibility offered by LLMs, we conclude that they cannot yet independently handle the complex world-building and narrative consistency required for believable game experiences.

A Year of Discoveries at Atomic Dirt

In "Harker's Escape," players could break furniture and craft makeshift weapons like crucifixes, prompting dynamic reactions from Dracula's AI. However, replicating this generative AI in 3D for more complex actions, like choking Dracula with a crucifix attached to a wire, presents a significant challenge. Creating the necessary 3D models and animations for every possible player action would be an immense undertaking. This highlights a key limitation in current game development: while AI can generate text and simple visuals, realistic and dynamic 3D interactions driven by AI are still beyond our grasp.

A Year of Discoveries at Atomic Dirt

LLMs in Gaming: Are We Solving the Right Problems?

While Large Language Models (LLMs) hold exciting potential for game development, simply injecting them into existing frameworks doesn't automatically equal better gameplay. The real challenge lies in leveraging AI to unlock new gameplay possibilities, not just add cosmetic enhancements.

A Year of Discoveries at Atomic Dirt

Are LLMs a gimmick or game changer? A naughty wizard may have the answer…

Remember typing whatever you wanted into classic adventure games like King’s Quest? We're exploring how LLMs can finally fulfill the promise of those early games, letting you truly "try anything" to escape Dracula's castle.

A Year of Discoveries at Atomic Dirt

Many games focus so intently on sleek graphics and sheer volume of content they fail to make their worlds more immersive.

What if we could create games as dynamic and surprising as real-life? Enter Atomic Dirt. 

And enter Atomic Dirt’s blog, where I’ll share what we’ve learned in the past year.