Karpathy Retires the Pelican: Frontier Models Now Build Whole Worlds From a Single Prompt
Andrej Karpathy handed Claude Opus 5 a 1M-token budget and the first paragraph of Lord of the Rings — and got back 5,500 lines of code that procedurally rendered the story in 3D. He argues the era of toy benchmarks is over.
Andrej Karpathy declared the end of an era this weekend, and he did it with Tolkien. In a widely circulated post, @karpathy described handing Claude Opus 5 a roughly 1-million-token budget — about $10 worth of inference — and asking it to render the opening paragraph of The Lord of the Rings in Three.js. The model returned 5,500 lines of code that procedurally rendered the narrative as a 3D scene. His framing was the real headline: "We're starting to leave the territory where you'd test an LLM by e.g. 'create an svg of pelican on a bicycle.'"
The pelican-on-a-bicycle prompt has been Karpathy's own running informal benchmark for well over a year — a quick, absurd probe of whether a model can reason about spatial composition in a domain it can't have memorized. That he is now retiring it signals something specific: the interesting question is no longer whether a model can produce a coherent single artifact, but whether it can sustain coherent execution across thousands of lines and a massive context window without collapsing into contradiction.
Get our free daily newsletter
Get this article free — plus the lead story every day — delivered to your inbox.
Want every article and the full archive? Upgrade anytime.
No spam. Unsubscribe anytime.