Andrej Karpathy Proposes 3D Scene Generation as a New LLM Benchmark
OpenAI co-founder Andrej Karpathy has proposed a new AI benchmarking method that tasks LLMs with generating interactive 3D scenes from literary text using three.js. In a test, Claude generated a "Lord of the Rings" scene within a 1-million-token budget, producing 5,500 lines of code in about two hours. As LLMs quickly master simpler tasks like generating 2D SVG images, this benchmark raises the bar by testing complex code generation, spatial reasoning, and scene design. It highlights the growing capability of AI to handle long-context, multi-step creative engineering tasks. The generated 3D animation depicted Bilbo's farewell party and a treasure cave, and while it does not match human-professional quality, its level of completion is impressive for a two-hour automated run costing around $10. Currently, this remains an informal proposal rather than a standardized, automated benchmark.
## BACKGROUND
Previously, developers used informal tests like asking LLMs to generate SVG code for a "pelican riding a bicycle" to evaluate their spatial reasoning and coding abilities. Three.js is a popular open-source JavaScript library used to create and display animated 3D computer graphics in web browsers using WebGL.