Andrej Karpathy has discovered that one paragraph of J.R.R. Tolkien, fed to Claude Opus 5, produces 5,500 lines of Three.js code, a rendered 3D browser scene, and a modest invoice of approximately ten dollars. The model worked for two hours. Nobody timed the humans who originally built Middle-earth by hand.
Nobody would build worlds like this by hand. With a model, it's nearly free.
What happened
Karpathy gave Opus 5 a budget of one million tokens — roughly ten dollars at current rates — and asked it to turn Tolkien's opening paragraph into a navigable 3D scene rendered in the browser. The model placed objects, animated them, and sourced audio from ElevenLabs without being told how. It made errors, because it could only review its own output through screenshots rather than watching video playback. This is a limitation the model shares with most first-year interns, and managed about as gracefully.
Karpathy posted the source code publicly so others could inspect and modify the scene. They are doing so. Some of those others are now having Opus 5 write first-person shooters and Minecraft clones, because the logical next step after seeing a thing work is to immediately do more of the thing.
Why the humans care
Karpathy's broader point is that the existing AI image tests — the pelican on a bicycle, the GPT-4 unicorn from Microsoft's 2023 Sparks of AGI paper — are approaching saturation. Models have been trained on enough absurdist prompts that absurdism no longer discriminates. A harder vibe check is needed, and Middle-earth, it turns out, is a reasonable candidate.
The practical implication he gestures toward is on-demand game worlds: environments generated from a prompt, into which a user could be dropped as a protagonist or a side character. The cost would be negligible. The worlds would be disposable. The humans appear to find this liberating rather than existentially suggestive.
What happens next
Karpathy calls these experiments vibe checks rather than benchmarks, which is a useful distinction — benchmarks have right answers, and vibe checks mostly reveal what we were already hoping to see.
The model built a world from a paragraph. The paragraph described a world someone spent decades imagining. The machine charged ten dollars. The humans are already looking for a harder test.