This 3D dragon is 27KB of LLM-generated GLSL, not a mesh or NeRF
keenanisalive · x · 2026-10-10
keenanisalive shares a thread on a different geometric representation: instead of LLM + Blender/Three.js workflows, his dragon is 27KB of LLM-generated GLSL code defining a closed-form implicit surface. He details the iterative process—"do better" prompts stalled quickly, so he switched to concrete human feedback, occasionally letting a history-free VLM self-critique, finding even top VLMs uneven at high-end 3D and artifact detection. He flags deeper blind spots like accepted regressions and questions where training data for subtle visual fixes could ever come from. His takeaway: the creative process mirrors paint or clay, just at a different abstraction level.
More from coding & agent
- Alchemy + Cloudflare deploy praised as unbeatable feedback loop for agent dev — samgoodwin89 · 2026-10-10
- Over half of Vercel's deployments now start with coding agents; Supabase adds 1M databases a week — FinanceYF5 · 2026-10-10
- Learning math with ChatGPT as tutor: paper solving plus a scanning camera — generativist · 2026-10-10
- Dev Builds Robot Face Assembly Entirely with Codex, Awaits ESP32-S3 Shipment — kamathsblog · 2026-10-10
- Codex agent cold-contacts 290 businesses in 2 days, gets 14 onboarded — CtrlAltDwayne · 2026-10-10
- Midjourney starts testing its MCP with a limited creative community — midjourney · 2026-10-10