Karpathy’s Pelican

	Karpathy’s Pelican(twitter.com)
	613 points by delichon 73 days ago \| 432 comments
	tl;dr: Karpathy tested Claude Opus 5 by giving it the opening paragraph of Lord of the Rings and a $10/1M-token budget to build a Three.js render, which produced 5,500 lines of procedural code over two hours—janky but functional. He argues this points to a future of hyper-custom, on-demand generated worlds (e.g., an "ephemeral GTA") that no human would bother building manually. However, it also exposed an LLM weakness: they can't natively perceive video or gameplay, so Opus had to iterate awkwardly via screenshots to audit its own output.
	HN Discussion: ↑The janky output is the point; this is a valuable new benchmark for measuring model understanding of the physical world ↓Three.js results are unreliable signals because Anthropic specifically trained models to be good at it ↓The model failed at text interpretation (e.g., Bilbo's disappearance), undermining claims of understanding •Sharing personal experiences of using LLMs for similar 3D/scene generation tasks ~AI exposure has lowered our quality expectations; we shouldn't declare pelican/rendering problems solved based on janky output