Local LLM now watches its own Three.js demos and rewrites them, open-sourced
jacobpederson · reddit · 2026-09-06
Developer RowanUnderwood updated autodemoscener, an open-source "endless demoscene machine" where a local LLM writes Three.js visual effects — and now watches its own output to improve it.
- Loop: LLM writes effects → code streams into a retro mock-OS editor → runs fullscreen → crashes/blank frames auto-detected → model asked to fix → working demos archived.
- New visual feedback: with Ninfer's video check, it captures 30s at 2fps and sends the video plus HTML source back to Qwen for a visual improvement pass; on a single 5090 it hits 110-220 tok/s decode with MTP speculative decoding.
- Qwen 3.8 still makes dumb mistakes (quarter-screen layouts, backwards maze walkers) that the video validation catches.
- Adds a thinking-effort selector, retains LM Studio support; needs ffmpeg. Project, setup guide and a browsable demo gallery are on GitHub.
- The author also suggests using it as a model benchmark since it tracks failures and deletions.
More from coding & agent
- Astra builds a polyphonic desktop app in just 9 hours — RileyRalmuto · 2026-09-06
- Local 27B + Pydantic AI writes the code that computes your real ledger — samuelcolvin · 2026-09-06
- Andreax Lists 157 AI Agent Tools with x402 Micropayments, No Signup — AncientLawfulness926 · 2026-09-06
- hostely: open-source Apple-native CLI self-hosts containers, Metal LLMs and HTTPS in one tool — ayo_ham · 2026-09-06
- Mollick: You Can't Really Turn Off Memory in Local AI Agents Like Fable and Astra — emollick · 2026-09-06
- PaperLens MCP Verifies Papers Are Actually Implemented in Code, Now Open Source — CraftyScore4348 · 2026-09-06