RoleplAI
summary
A prototype blends classic procgen rules with transformer output, producing thousands of JSON universe outlines offline on a Mac Mini using LM Studio and a local 12GB model; generation uses multipass prompting, strict JSON schemas and validators to keep outputs on-rails. Backend uses Deno and Hono to serve precomputed worlds; dialogue renders as branching trees with sentiment scoring. Results: thousands of playable seeds exist, per-query generation still costs seconds and occasional nonsense, but the hybrid system works as a low-cost, controllable way to enrich procedural games.
AI-written summary · about
chapters (17)
- 0:00RoleplAI and D&D inspiration
- 3:24Procedural generation history
- 8:23Generative AI vs procedural generation
- 11:36Structured design with generative AI
- 14:26Validation for generative AI
- 17:21Multi-pass worldbuilding
- 19:53Seed data and prompt engineering
- 23:27Rich world and character generation
- 26:38AI cost and environmental impact
- 28:47LM Studio local models
- 32:06Latency and real-time challenges
- 33:10Precomputing content
- 35:20Game architecture and API
- 36:45Dialogue trees and player choice
- 38:29Live demo
- 41:02Limitations and future
- 42:07Conclusion and Q&A
related talks
44mBringing stories to life with AI, data streaming and generative agents
02h 48mCompose & Dragons: The Role-Playing Game of Agents Fed on Tiny Language Models
46mThe Future of Role-Playing Games IS... A.I. | But How?
34mCan AI Make Your Game Better?
25mTurning Data Into Generative Chaos
09m#BuildPropulsionLab at #DevoxxUK: @jbaruch with @charlyouki
49mAI Apocalypse Survival Kit: coding agents & Tiny Language Models
09mRethinking AI: Building What Matters, Not Just What's Possible