Karpathy proposes LOTR 3D scene as new AI vibe test

Andrej Karpathy fed Claude Opus 5 a single paragraph from The Lord of the Rings and asked it to build a 3D browser scene. Two hours and roughly $10 later, the model had written 5,500 lines of Three.js code, placed animated objects, and rendered a playable Tolkien diorama. Karpathy, who joined Anthropic's pre-training research team in May 2026 after co-founding OpenAI, argues this kind of stress test exposes capabilities that simple benchmarks miss.

Why pelicans and unicorns are no longer enough

Karpathy's premise: the existing quick tests for generative models are hitting their ceiling. Simon Willison's pelican test, which asks a model to draw an SVG of a pelican riding a bicycle, was designed to be absurd enough that the subject wouldn't already exist in training data. It worked as a rough spatial-reasoning check. Microsoft took a similar approach in 2023, prompting GPT-4 to draw a unicorn in TikZ for its "Sparks of AGI" paper.

Those prompts tested whether a text-only model could reason about shapes. But frontier models are now multimodal, trained on images and video as well as text. A pelican on a bike is no longer exotic. Karpathy wants a harder assignment: generate thousands of lines of executable 3D code, animate the scene, and do it cheaply enough that disposable custom worlds become practical.

What the demo actually produced

Karpathy gave Claude Opus 5 a one-million-token budget. The model rendered the scene in Three.js, a JavaScript library for WebGL graphics. Eleven Labs supplied the audio. Karpathy posted the source code publicly so others can fork or extend the scene.

He calls the result "rough but fun." Nobody would hand-author a world this way, he notes, but at roughly ten dollars per attempt, the economics shift. Teams could spin up bespoke game environments on demand, dropping users in as protagonists or background characters.

The limitation Karpathy flagged

Claude Opus 5 cannot watch its own video output. The model debugged by examining screenshots instead. That constraint introduced errors. Karpathy is clear: the demo is a vibe check, not a hard benchmark. It shows directional capability, not production reliability.

Other users have since pushed Opus 5 to write entire browser games, from first-person shooters to Minecraft clones. None of these outputs carry rigorous metrics. They serve as public probes of model limits, a form of crowdsourced stress testing that labs cannot easily replicate internally.

Karpathy's trajectory

Karpathy co-founded OpenAI and later led AI at Tesla, overseeing Autopilot's neural-network stack. He left Tesla in 2022 and launched Eureka Labs, an education startup. In May 2026 he joined Anthropic to work on pre-training research. The LOTR demo is not an official Anthropic project; Karpathy posted it as a personal experiment.

The question Karpathy leaves open: what happens when models can watch, not just screenshot, their own renders? That feedback loop would let them iterate without human prompting. For now, vibe tests remain exactly that, a gut check on whether the model can hold a complex task together long enough to ship something usable.

Source: The Decoder / Maximilian Schreiner