Karpathy’s Pelican(twitter.com)
581 points by delichon 1 day ago | 395 comments
tl;dr: Karpathy tested Opus 5 by giving it the first paragraph of Lord of the Rings, a 1M token budget (~$10), and asking for a Three.js rendering—it spent ~2 hours writing 5,500 lines of procedural animation code. He sees this as a shift from "no one would bother" to "why not, it's free," opening possibilities for hyper-custom on-demand worlds like an ephemeral GTA. The exercise also exposed a key LLM weakness: they can't natively perceive video or play games, forcing Opus to laboriously screenshot to audit its own work.
HN Discussion:
  • The jankiness is the point; this exercise is a valuable new benchmark for physical world understanding
  • Sharing personal experiences building similar LLM-generated 3D scenes from text descriptions
  • Three.js results aren't meaningful because Anthropic specifically trained models on it
  • The output shows poor text comprehension and modeling quality, undermining the demo's impressiveness
  • The demo isn't reproducible without the prompt, unlike Simon Willison's pelican benchmark