Solo dev builds local LLM-powered GUI where every pixel is vector code generated by the model
Wix86 · reddit · 2026-09-24
A solo researcher demoed a GUI harness (Minecraft-inspired) where the LLM generates code using just 7 vector functions (Rectangle, Circle, Line, Image, Text, List, Div) — every pixel on screen is model-generated vector graphics.
Notable engineering:
- Written in Go as a single 40MB exe (CPU version), 70–100MB RAM, no browser; bundles llama.cpp, whisper.cpp, omnivoice.cpp, a code interpreter, and custom text/layout engines
- Storage built from scratch with full change history: revert user or agent mistakes, keep only the last hour/day/week
- Minimalist brush interaction: hold CTRL to select layout elements and talk; the agent never sees the screen, it edits code and a knowledge base
- Testing gemma-31b, Qwen-27b, DeepSeek-V4-flash at fp8; optional OpenRouter for SOTA models
Author says context engineering is 50% done and waiting for agents is the biggest UX problem.
More from coding & agent
- Running models 24/7: we're in the harness phase, not autonomous builders yet — BLUECOW009 · 2026-09-24
- HeyGen guide: wiring Meta's Muse Agent to MCP for automated avatar videos — HeyGen · 2026-09-24
- Dev runs npm publishes and GitHub chores from a Tesla via Grok Bot — Baconbrix · 2026-09-24
- Grok Bot adds voice calls, 1Password and Slack drafts in a big feature week — mark_k · 2026-09-24
- Proof and Superfluid launch agent wallet verifying a human behind every transaction — csuwildcat · 2026-09-24
- Turning LLM hard classifiers into tunable soft classifiers with logprobs — JnBrymn · 2026-09-24