Karpathy's New 'LOTR' Benchmark: Opus 5 Burns 1M Tokens to Build 3D World
量子位 · wechat · 2026-08-03
As AI models grow more capable, the once-popular 'pelican riding a bicycle' SVG test has hit its ceiling. AI luminary Andrej Karpathy recently introduced a tougher 'Lord of the Rings' benchmark to evaluate long-range planning and spatial understanding.
The new test requires models to read the opening of The Lord of the Rings and construct a real-time 3D world using Three.js code. Claude Opus 5 took 2 hours, consumed 1 million tokens, and generated 5,500 lines of code to build a rough but structurally coherent scene. Karpathy noted this exposes a current model limitation: they can write code to build scenes but cannot truly 'see' the video or play the games they make.
The benchmark sparked widespread community engagement and experiments:
- Workflow Vision: Karpathy suggested using procedural code for storyboarding, combined with video models (like Seedance) and audio APIs for high-fidelity rendering.
- Community Demos: Users built low-poly San Francisco using agents, created a 3D NYC model with real-time data, and generated an interactive Kanye West concert purely from code.
- Benchmark Debate: Critics questioned if this merely tests Three.js proficiency at a high compute cost. Supporters argued translating abstract text into 3D animations requires mastering spatial, physical, and mathematical reasoning, suggesting LLMs' general reasoning capabilities are naturally extending into the 3D realm.
Related event: Karpathy Tests LLM with 1M Tokens to Render 3D Lord of the Rings(6 posts)→
More from coding & agent
- From 'Works' to 'Verifiable': Trust and Audit in Enterprise AI Agents — pswider · 2026-08-03
- PullMD: Open-Source Tool to Convert Web URLs & Files to Clean Markdown — tom_doerr · 2026-08-03
- 10,000-Hour Programmers Embrace AI: Never Writing Manual Code Again — georgemillo · 2026-08-03
- AI Agents Self-Correct Causal Analysis Errors via Verifier and Cross-Review — hugobowne · 2026-08-03
- Open-Source Legal AI: 139 Agent Skills for GDPR and NDA Workflows — tom_doerr · 2026-08-03
- Prompt-Architecture as Literature: Writing Executable Text for LLMs — lnsip9reg · 2026-08-03