Gemini 4 roasts Google's 3D harness: blind coordinate writing and how to fix it
Ryoiki-Tokuiten · reddit · 2026-11-01
A Reddit user iterating a Three.js 3D scene with Gemini 4 Argon asked the model to explain itself, and got a strikingly candid self-analysis. Gemini admits its first-pass 3D code is bad because it writes coordinates blind, with two specific blind spots: visual occlusion in wide shots (it can't tell a hand became a rectangular block from 20m away), and the 2D-screenshot-to-3D-coordinate translation gap (it sees the clipping but can't compute the exact (x, y, z) offset). Its prescription: the harness must show it the rendered image every iteration via viewfile, and print exact 3D bounding boxes, joint world positions and [CLIPPING] flags alongside, so it fixes the exact rotation in one turn instead of guessing for five. It also notes the harness wasn't even rendering the actual pipeline, calling this the entire Bitter Lesson for 3D harnesses. Both a fun model-behavior observation and a directly actionable recipe for vision-feedback agent loops.
More from coding & agent
- VFX Supervisor Open-Sources 20 Skills for Building an Automated Motion Design Pipeline with Claude Opus 5.5 — GCWebDesigner · 2026-11-01
- Cursor launches Annotate in design mode: give agents visual prompts by editing the page — soleio · 2026-11-01
- Open-source Garmin MCP turns training data into AI-built dashboards — indaco_dev · 2026-11-01
- GPT-5.6 Sol reportedly beats Factorio, launching a rocket autonomously — Realistic-Total-3553 · 2026-11-01
- Open-source CodeRail lets teams share one Claude Code session from any device — L_i_n_k0666 · 2026-11-01
- Dev observes AI has made systems-heavy games like RTS and strategy easier to build — IanArawjo · 2026-11-01