- CubeMuse's Visual Feedback
How CubeMuse uses screenshots and a vision model to give the Agent visual observations during building, then review the finished build and suggest revisions.
13 min - CubeMuse's Voxel Rendering Engine
Rendering tens of thousands of Minecraft-style blocks in the browser with Three.js via greedy meshing, vertex AO, and incremental chunk rebuilds.
28 min - From Keyword Scoring to Semantic Vectors
CubeMuse's 800-entry reference library uses a five-layer scorer to reach hit@6 = 90.7%. The scoring layers address problems observed in specific queries.
18 min - From One-Shot Generation to Bounded Agents
CubeMuse's build pipeline evolved through three generations: a single LLM call, a staged pipeline, and a model that selects tools within a fixed budget.
21 min - LLM Handles Intent, Code Handles Geometry
Why CubeMuse uses a JSON instruction set instead of raw coordinates, splitting what to build from how to draw it.
20 min
Back