Karpathy gave Opus 5 a million tokens — and it rendered The Lord of the Rings in 3D
Karpathy gave Opus 5 the first paragraph of The Lord of the Rings and $10 in compute: the model wrote 5,500 lines of code and rendered the Shire in low-poly. The post drew 4.8M views.
Highlights
- Karpathy gave Opus 5 the first paragraph of The Lord of the Rings, a 1M token budget (~$10) and asked for a 3D render — the model wrote 5,500 lines of code in ~2 hours
- The result: a procedural low-poly Shire with hills, hobbit holes and the river — 'no one in their right mind would spend the time, but LLMs have infinite patience'
- The vision of the future: hyper-custom worlds 'like an ephemeral GTA on demand', where you drop into the story as an NPC
- The exposed weakness: LLMs can't audit their own work because they can't perceive videos or play inside the worlds they create
Andrej Karpathy ran an experiment that sums up the AI moment: he gave Opus 5 the first paragraph of The Lord of the Rings, a 1M token budget (~$10) and asked for a three.js render of the scene.
The model worked for about two hours and wrote 5,500 lines of code that procedurally rendered the story — a low-poly Shire with hills, hobbit holes and the river, under a pale moon.
What the experiment reveals
Karpathy highlights the regime shift: 'we're starting to leave the territory where you'd test an LLM by asking for an SVG of a pelican on a bicycle'. With cheap budgets, hyper-custom ideas go from 'no one would ever do this' to 'sure, why not, it's ~free'.
The future he envisions: hyper-custom worlds you can drop into as a spectator NPC — 'something like an ephemeral GTA on demand'. The LLM orchestrates polygons in (x,y,z) coordinates and animates it all with infinite patience.
The exposed weakness
The worlds/games domain exposes a gap: LLMs can't audit their own work because they can't natively perceive videos or play games within them. Opus 5 had to slowly take screenshots and messed up several times, creating jank.
It's a rare portrait of raw capability (multimodal, gameplay) that's still lacking — and of what's next at the frontier.
Why it matters
Karpathy's post drew 4.8 million views and 28K likes — one of the most cited examples of what frontier models can do with a small budget.
The direction is clear: AI is moving from generating text to generating worlds — and the bottleneck now is perception, not generation.