Karpathy Tests Claude Opus 5 with 1M Token LoTR Render
Original: Karpathy’s Pelican
Why This Matters
Demonstrates LLMs' emerging ability to generate complex interactive 3D worlds autonomously, while exposing critical gaps in multimodal self-auditing.
Andrej Karpathy shared on X that Claude Opus 5, given a 1M token budget (~$10) and the first paragraph of Lord of the Rings, autonomously spent ~2 hours writing 5,500 lines of Three.js code to procedurally render the story in 3D.
AI researcher Andrej Karpathy posted on X that he tasked Anthropic's Claude Opus 5 with generating a Three.js 3D render of the opening paragraph of Lord of the Rings, providing a 1 million token context budget costing approximately $10. The model worked autonomously for roughly 2 hours and produced 5,500 lines of code that procedurally placed and animated polygon assets in (x,y,z) coordinates. Karpathy described the result as 'janky but fun,' and made the output playable and forkable at karpathy.ai/lotr-movie/. He noted the experiment illustrates a new class of value: tasks 'no one would ever do' due to effort now become trivial with LLMs. He envisions hyper-custom interactive worlds — such as an on-demand GTA-style Hobbiton — where users participate as NPCs or characters. However, Karpathy also highlighted a key weakness: LLMs cannot efficiently perceive video or play games natively, forcing Opus 5 to take slow, painstaking screenshots to audit its own work, leading to several errors. He identified native multimodal gameplay perception as a capability still significantly lacking in current models.