Andrej Karpathy, co-founder of OpenAI and head of AI at Tesla, recently tested the capabilities of Claude Opus 5 by converting a single paragraph from 'Lord of the Rings' into a 3D browser scene. The model generated 5,500 lines of code in about two hours, with a budget of one million tokens, roughly $10. The result, rendered using Three.js, demonstrated the potential for creating custom game worlds at a low cost. Karpathy shared the source code online, allowing others to play back and modify the scene. The test, he said, is more of a 'vibe check' than a hard benchmark, highlighting the creative possibilities of AI-generated content.
The experiment involved placing and animating objects autonomously, though the model made mistakes due to its inability to view its own video output in real-time, relying instead on screenshots. Karpathy noted that while the outcome was rough, it was still fun and showcased the model's ability to generate complex scenes. Users have since used Opus 5 to create entire browser games, including first-person shooters and Minecraft clones, further illustrating the model's versatility. These projects, however, do not provide hard benchmarks but rather serve as creative explorations of AI's capabilities.
Karpathy, who has worked on pre-training research at Anthropic since May 2026, mentioned that the pelican test, a popular benchmark for assessing AI's spatial reasoning, is nearing its limits. He referenced earlier tests, such as Microsoft's GPT-4 drawing a unicorn in TikZ in 2023, as examples of how AI models can be used to explore new creative frontiers. The focus remains on demonstrating the potential for on-demand game worlds, where users could be immersed as characters without manual development.
Source: thedecoder
