Anthropic首席科学家卡帕西推出了一项新的大模型评测基准，通过让Claude Opus 5模型根据《指环王》开篇文本生成完整的3D中土世界场景来考察AI能力[1]。该实验中，模型用约两小时生成了5500行代码，消耗100万tokens，成本约10美元，最终成功渲染出可交互的3D游戏场景[2][3]。这一实验在社交媒体上引发广泛关注，在X平台获得超过280万次浏览[3]。
卡帕西表示这项新测试将接替已饱和的"鹈鹕骑自行车"SVG基准，用来更全面地考察模型的空间理解、代码生成和长期规划能力[1]。该项目已由卡帕西开源，使用Three.js库实现3D场景生成[1]。然而卡帕西同时指出了当前大模型的关键局限：虽然代码生成能力强大，但模型无法像真人测试员一样进入自己生成的游戏进行连续体验和问题反馈[3]。卡帕西指出"模型生成内容的能力正在快于验证内容的能力"[3]。网友基于此方案进一步创作了"Earth Online"、纽约3D数字模型、AI演唱会等应用[1]。
Andrej Karpathy, chief scientist at Anthropic, has introduced a new evaluation benchmark for large language models centered on generating a three-dimensional Middle-earth scene from the opening text of The Lord of the Rings [1][2][3]. Using Claude Opus 5, the model produced approximately 5,500 lines of code over roughly two hours with a reasoning budget of 1 million tokens, costing about $10 [2][3]. The interactive 3D game generated through this process has garnered over 2.8 million views on social media [3].
The benchmark, built using the Three.js library [1], is designed to replace the saturated "pelican on a bicycle" SVG test [1]. Karpathy framed the achievement as both a demonstration of model capability and a revelation of its limitations. While he noted that "no reasonable person would spend time writing such highly customized code, but LLMs have all the patience in the world," [2] he emphasized that the experiment exposed a critical weakness: large language models cannot natively perceive video or enter the environments they create [2]. "The model's ability to generate content is now outpacing its ability to verify content," Karpathy observed [3]. The project has been open-sourced by Karpathy [1], and the community has already built upon the concept to create applications including "Earth Online," a New York 3D digital model, and AI concerts [1].
