OpenAI co-founder Andrej Karpathy posted on X on August 2, proposing an AI benchmarking idea: provide a model with the opening paragraph of The Lord of the Rings, then ask it to create the corresponding 3D world.
Previously, the industry commonly used SVG images generated from prompts such as “a pelican riding a bicycle” to test AI models, evaluating large language models’ actual capabilities in spatial reasoning, geometric logic, and complex code generation.
However, Karpathy pointed out that as models continue to improve, simple tasks are no longer a challenge. He therefore proposed a new AI benchmarking approach: have models build interactive 3D scenes based on literary texts, allowing their overall performance in code generation, scene design, and multimedia production to be assessed.
Karpathy asked Claude Opus 5 to use three.js to create a 3D world, with the opening paragraph of The Lord of the Rings as the input. The test’s maximum resource budget was 1 million tokens, at a cost of approximately 10 US dollars. Under this configuration, Claude Opus 5 took about 2 hours and produced approximately 5,500 lines of code.
The resulting work presents scenes related to the opening of The Lord of the Rings in 3D animation, including people attending Bilbo’s farewell party and a cave filled with treasure. Reports noted that the work could not compare with content created by humans over several hours, but considering that the AI generated it in about 2 hours, its level of completion was still relatively high.
