What's this about?
Andrej Karpathy gave Claude Opus 5 the first paragraph of The Lord of the Rings, a 1-million-token budget (about $10), and asked for a three.js rendering of it. The model worked for about two hours and produced 5,500 lines of code that procedurally render the story — placing polygon assets in 3D coordinates, orchestrating scenes, and animating it all. His verdict: "kind of janky but fun," and mind-boggling that it works at all.
What exactly is "three.js / procedural rendering"?
three.js is a JavaScript library for drawing 3D graphics in a web browser. "Procedural" means nothing is hand-modeled or downloaded — the code itself generates every hill, tower, and character out of raw geometric shapes, like building a movie set entirely out of LEGO instructions the model wrote itself. The model can't see what it's building while coding it, which is why getting a coherent animated scene is impressive — it has to hold the whole 3D world in its head as coordinates.
What's the community saying?
- Karpathy's own framing drove the discussion: simple one-shot tests ("SVG of a pelican on a bicycle") are saturating, so evaluations must move to long-horizon, budgeted, open-ended projects.
- The economics got attention: two hours of frontier-model labor and 5,500 lines of code for ~$10.
- It lands the same weekend as the "Habsburg frog" benchmark thread on HN — the community is actively hunting for the next unsaturated eval.