LLM Generates Procedural 3D World from Text

The author describes an experiment pushing the boundaries of large language models beyond simple image generation. Using Opus 5, the LLM was tasked with procedurally rendering a 3D world based on the opening paragraph of "Lord of the Rings," consuming a significant token budget. Over two hours, the model produced 5500 lines of JavaScript code, orchestrating various polygon assets in 3D space to animate the story. This showcases an impressive capability for complex, custom code generation that a human would rarely undertake due to time constraints. The author envisions a future where LLMs can create "hyper custom worlds" or "ephemeral GTAs" on demand, allowing users to interact with stories as NPCs or characters. However, the experiment also exposed a key weakness: LLMs' inability to efficiently audit their visual output. Opus 5 had to slowly take screenshots to check its work, resulting in some "janky" elements, indicating a current deficiency in multimodal perception and self-correction within generated environments.