Gemini 4 Carbon tested: 100+ unattended passes rebuild a 3D fort in a raw engine
Prompt Engineering · youtube · 2026-11-11
Prompt Engineering (recently joined Google; testing predates that, with early access from DeepMind) got hands-on with Gemini 4 Carbon, Google's model for long engineering tasks. The video skips benchmarks and shows outputs:
- Three.js open world recreated from a video; 3D SDF dragon from one concept image
- A subagent team demo ("storm in a teacup"); a 3D YouTube studio in 15 passes; a 3D house from a floor plan
- The centerpiece: rebuilding the historic Qila Bala Hisar fort in 67.js—a 3D engine only one month old—across 100+ unattended passes. The model first researched 30 sources into a building history, then extended the engine's source code itself, using a self-scored review loop
- GPU experiments: galaxy collision on DGX Spark, training an ant to walk, and a neural network written by hand with no training
Failure points include reply limits and effort settings. The video closes with four prompting rules that mattered most, concluding that prompts matter more with this model than ever.
More from coding & agent
- dottxt launches Tool-call index: same open model fails tool calls up to 13x more on some providers — remilouf · 2026-11-11
- Artificial Analysis benchmarks 26 Search APIs across 14 providers in new agent-search leaderboard — ArtificialAnlys · 2026-11-11
- Keenable speed breakdown: 15s per task, only 2-3s of it search time — ArtificialAnlys · 2026-11-11
- Keenable cost breakdown: ~$0.040 per task, 80% from search, on the cost frontier — ArtificialAnlys · 2026-11-11
- Keenable Search 1.0 tops Artificial Analysis Search Index at 83, a third below median cost — ArtificialAnlys · 2026-11-11
- Dev calls on GitHub to ship push-based integration (webhooks, MCP) for agents — brianmichel · 2026-11-11