📖 Table of Contents
1. TL;DR — What Happened
Andrej Karpathy — ex-Tesla AI director, founder of Eureka Labs — posted an experiment that took over Hacker News this week (414 points). Using Opus 5 with a 1M-token budget (roughly $10), he asked the model to procedurally render the opening passage of The Lord of the Rings into a 3D scene. The model generated 5,500 lines of code that ran for 2 hours to produce an explorable, programmatically-rendered world.
Why it matters: Not because a book passage became 3D — but because of the economics. "Nobody would ever do this" content — bespoke, one-off, deeply customized — just became ~free to attempt. That's a category change, not an incremental improvement.
2. The Experiment: 1M Tokens, 5,500 Lines, 2 Hours
| Parameter | Value |
|---|---|
| Model | Opus 5 (frontier reasoning model) |
| Context budget | 1M tokens |
| Token cost | ~$10 |
| Input | Opening passage of The Lord of the Rings |
| Output | Procedural 3D scene generator |
| Code generated | ~5,500 lines |
| Runtime | ~2 hours to render |
The method is the interesting part: instead of asking for a video or a pre-baked animation, Karpathy had the model write a procedural rendering program — code that builds the world (terrain, trees, paths, atmosphere) algorithmically from the text's descriptions. The 1M-token context lets the model hold the entire book excerpt, the scene graph, and the rendering code in view at once, iterating coherently rather than generating disjointed clips.
This is the pattern that matters: long context → generate an engine, not an artifact. A video is a fixed artifact; a program is a generative system you can re-run, tweak, and extend.
3. Why "Pelican"? The Nobody-Would-Ever-Do-This Principle
The name comes from the classic LLM demo of asking for "a pelican" — a trivial, silly request that showcases capability without pretense. Pelican extends that spirit to a grander scale: asking a model to do something so bespoke and so uncommercial that no human would ever have paid to build it.
Karpathy's framing: there is a vast territory of "nobody would ever do this" content — a 3D rendering of one specific book paragraph, a custom tool for one person's workflow, a game about your own neighborhood. Before LLMs, all of it was economically dead on arrival. The fixed cost of bespoke creation was too high, and there was no market to amortize it.
The shift: when marginal cost of bespoke generation approaches zero, the "would anyone ever pay for this?" filter stops being the gatekeeper of what gets made. The filter becomes "would anyone ever enjoy this?" — which is a vastly larger space.
4. What Karpathy Says LLMs Still Can't Do
The experiment is valuable partly because Karpathy was explicit about the boundaries he hit. Two limitations stood out:
- Models can't natively perceive video. The model couldn't "watch" its own output and react to visual problems. It works from code and text descriptions — so visual bugs have to be caught by a human watching the render, then fed back as text. The loop works, but it's clunky and slow.
- Models can't audit their own output well. With 5,500 lines of generated code, verifying that the render matches the book's description ("the Party Tree", "the mill by the Water") requires human reading of both the source text and the output. Self-verification remains shallow.
Takeaway: the ceiling right now isn't generation — it's perception and verification. Generation is cheap and getting cheaper; the human-in-the-loop for quality control is still the bottleneck. Expect this gap to narrow as multimodal models get better at watching their own outputs.
5. What This Means for AI App Developers
For anyone building on top of LLMs, the Pelican experiment is a useful calibration of where long-context capability actually is in mid-2026:
- Long context is now a creative medium. 1M tokens at ~$10 means "hold an entire book + scene graph + codebase in view" is a practical operating mode, not a research demo. Architect your products around that, not around 32K-token windows.
- Program generation beats artifact generation for interactive content. Games, 3D scenes, and tools want engines, not clips. Ask the model to write the generator, then iterate on the generator.
- Design for the verification gap. Whatever you generate, assume the model can't fully check its own work. Build lightweight human-review loops — or secondary models that act as auditors.
- "Bespoke" is a product category now. One-off personalized content — a custom dashboard, a neighborhood game, a personalized story world — is now buildable at near-zero marginal cost. The moat moves to taste and iteration speed.
Bottom line for builders: the expensive part of bespoke software used to be the first copy. Now the first copy is ~free, and the expensive part is knowing what to make. That inverts the classic indie-hacker calculus — distribution and taste beat implementation.
6. FAQ
Q: Is the rendered scene actually good?
By game-studio standards, no — it's a procedural approximation of a book passage, with the rough edges you'd expect from 5,500 lines of model-generated renderer code. But that's precisely the point: it doesn't need to be polished to be valuable, because the cost of attempting it was ~$10.
Q: What exactly is "procedural rendering"?
Instead of modeling every object by hand, you write code that generates the scene from rules and parameters — terrain from noise functions, trees from growth algorithms, lighting from a sun position. The model wrote this code from the book's textual descriptions of the Shire.
Q: Could this scale to a full book or a playable game?
Incrementally, yes — that's the direction. The 1M-token budget already covers a substantial excerpt; multi-pass workflows (scene by scene, then assembled) could extend it further. The constraint isn't context size anymore, it's the perception/verification gap Karpathy flagged.
Q: How does this relate to AI video generation?
They're complementary. Video models generate the artifact (the clip); the Pelican approach generates the engine that produces the artifact. For interactive or explorable content — where a fixed video isn't enough — the program-generation route is currently the stronger play. See our AI Video Generation 2026 comparison for where the artifact side stands.
Q: Where can I follow the discussion?
- Karpathy's original post on X/Twitter (search "Karpathy Pelican")
- Hacker News thread (~414 points)
- MiniMax H3 Review — the other big open-model release this week