We’ll explore the promise and challenges of using generative techniques to create new experiences in established worlds. Large language models are a fraught technology, including in their lack of regard for the boundaries of intellectual property. At Hidden Door, we’ve been combining them in constrained ways with procedural generation, templating, and “classical” natural language techniques to allow stories that feel meaningfully different every time, but which respect the world in which they occur.
We’ll speak to the kind of input that creative partners (authors, film producers, and other creators) want, and discuss the risks and pitfalls of generative technology (oh hi, bias) and how we mitigate those. We can also share how we’ve approached these challenges in our game, with examples like:
- Using language models to choose which templated story beat happens next, instead of generating them from scratch.
- Treating language as a unique affordance for player characters (my strong character is different than your strong character).
- Creating lots of content using procgen and language models together, with a game designer’s curation.
- When we take the reigns off and let worlds get weird (cybernetic mech skeleton in the Wizard of Oz?).
- How to do this ethically, from a technology, data, and business perspective.
Takeaways:
- A grounded perspective on the use of large language models: both boons and pitfalls.
- Some approaches to mitigating bias when using those models.
- Examples of non-LLM machine learning techniques and their use in generative stories.