Karpathy Tests LLM with 1M Tokens to Render 3D Lord of the Rings

Andrej Karpathy recently pushed the limits of large language models in long-horizon task execution and coding. Providing the model with the opening text of The Lord of the Rings and a budget of roughly 1 million tokens (about $10), he tasked it with rendering the text into a 3D scene using Three.js. The model autonomously spent about 2 hours writing 5,500 lines of code, generating a playable, albeit rough, 3D game. This experiment highlights the astonishing potential of LLMs in complex spatial understanding and generation, while also exposing current technical bottlenecks.

Confirmed

• Test Conditions: The model was given the first paragraph of The Lord of the Rings and an execution budget of about 1 million tokens (approx. $10).
• Execution Results: The model worked continuously and autonomously for nearly 2 hours, writing 5,500 lines of Three.js code.
• Output Quality: It successfully rendered the text into an interactive 3D scene, demonstrating stunning spatial generation capabilities.

Unconfirmed

• Model Version: Due to conflicting reports from various sources, some claim the model used was Claude 3 Opus, while others suggest Opus 5. The exact model version used for the test remains debated.

Why It Matters

• Evolution of Evaluation: Karpathy noted that the industry is moving past simplistic static tests like "generating a pelican riding a bicycle SVG," and is beginning to explore autonomous task execution over long durations with high token consumption.
• Core Weakness Exposed: The experiment revealed a critical bottleneck in current LLMs: the lack of visual feedback mechanisms after code generation, preventing the model from effectively self-verifying and correcting errors.

2026-08-02 ~ 2026-08-03 · 6 related posts

• Karpathy Tests Claude Opus: Writes 5,500 Lines of Code in 2 Hours to 3D Render LOTR — andrewchen · 2026-08-02
• Karpathy Tests Opus: Writes 5,500 Lines of Three.js to Render LOTR in 2 Hours — TAbrodi · 2026-08-02
• Karpathy Tests Opus 3D World Generation: Spatial Skills Impress, But Lacks Visual Feedback — AI寒武纪 · 2026-08-02
• Karpathy Uses Opus 5 to Generate 3D LOTR, Exposes LLM Self-Review Limits — 新智元 · 2026-08-03
• Karpathy's New 'LOTR' Benchmark: Opus 5 Burns 1M Tokens to Build 3D World — 量子位 · 2026-08-03

1 near-duplicate retellings: TAbrodi