August 2, 2026
One pelican, infinite discourse
Karpathy’s Pelican
AI spent 2 hours making a Lord of the Rings scene — and the comments went feral
TLDR: Karpathy showed an AI spending hours turning a famous fantasy book opening into a rough 3D animated scene, hinting at custom-made game worlds in the future. Commenters split hard between “this is the future of personal entertainment” and “this is expensive corporate hype with extra pelican fatigue.”
Andrej Karpathy tossed an artificial intelligence model a chunk of The Lord of the Rings, gave it a giant budget and plenty of time, and watched it crank out a rough 3D animated scene in a web browser. The result was messy, weird, and undeniably impressive: thousands of lines of code, moving objects in a made-up fantasy world, and a glimpse of what happens when computers have the patience to do work no human would willingly sign up for. Karpathy’s big idea? Someday this could become ultra-personalized entertainment on demand — basically your own custom game world built from imagination alone.
But the real fireworks were in the replies. One camp was awestruck, with commenters saying this shows we may be only a few giant leaps away from entertainment made for one person at a time, not the masses. Another camp rolled its eyes so hard you could hear it through the screen. One user groaned that the old “pelican” image test needs to die already, while another sniped that this looked like a “perfect benchmark to burn more tokens” — translation: a flashy demo that conveniently spends more money.
And then came the drama. The hottest jab accused Karpathy of sliding from trusted AI realist into corporate hype man for Anthropic, suggesting his tone changed with his employer. Even the believers admitted there’s a big catch: these systems can make elaborate worlds, but they still struggle to properly check their own work, leading to plenty of glorious jank. So yes, the demo wowed people — but the comments turned it into a full-blown food fight over progress, propaganda, and whether anyone asked for a pelican sequel in the first place.
Key Points
- •Karpathy argues that older LLM tests such as simple SVG-generation prompts are becoming less representative of current model capabilities.
- •He tested Opus 5 by supplying the first paragraph of *The Lord of the Rings*, a 1 million-token budget, and a request to generate a Three.js render.
- •The model reportedly spent about two hours generating around 5,500 lines of code to procedurally render and animate the scene.
- •Karpathy says this kind of task illustrates how LLMs can make highly custom, low-priority creative work economically feasible.
- •He identifies a major limitation in game and world generation: LLMs still struggle to efficiently audit outputs because they lack strong native video perception and gameplay ability.