Andrej Karpathy tested the Opus 5 model for the autonomous design of 3D environments. In two hours of operation, the AI wrote over 5,500 lines of Three.js code, consuming approximately $10 in tokens. The experiment signals a shift from simple asset generation to complex cognitive engineering.
The era of testing AI capabilities through primitive SVG sketches—like the infamous pelican riding a bicycle—is officially over. Child’s play is being replaced by cognitive engineering backed by million-token budgets. Andrej Karpathy’s recent case study highlights this paradigm shift: the Opus 5 model was tasked with deploying a procedural 3D world using only the opening paragraph of "The Lord of the Rings" as a prompt and a 1-million-token limit.
After two hours of autonomous work and roughly $10 in direct costs, the model produced 5,500 lines of clean Three.js code. The system didn’t just render a static image; it orchestrated polygonal assets in 3D space and animated the narrative environment. This is no longer a mere creative exploration but a stress test of model endurance. We are witnessing a fundamental shift in content economics: tasks once discarded due to the staggering labor costs of human developers are moving into the "why not, it’s practically free" category.
This level of autonomy paves the way for ephemeral, hyper-customized worlds. Today, AI builds interactive backdrops for a single scene; tomorrow, it will generate personalized narratives where the user can exist as an NPC or the protagonist.
Strategic context
However, the ambitions of procedural design currently hit a technological bottleneck: the lack of native video and game-mechanic perception. Because models still cannot perceive dynamics in real-time, they are forced to audit their own work through a slow, agonizing series of screenshots.
The bottom line
This clunky feedback loop inevitably produces visual artifacts and the characteristic "jank" found in the final product's animations. Until AI learns to natively process gameplay and video streams, the gap between generating raw code and creating a seamless, high-fidelity world will remain. Nevertheless, for ten dollars, the result is more impressive than the output of an entire outsourced studio.