Post Content
[[{“value”:”Gemini 4 Carbon is Google’s new model for long engineering tasks. I had early access, so this video skips the benchmarks and shows the outputs: a Three.js open world, a 3D studio, a house built from a floor plan, a historic fort rebuilt over 100+ unattended passes in a one-month-old engine, GPU experiments, and a neural network written by hand with no training. You’ll also get the four prompting rules that made the biggest difference.
Google’s announcement: https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon/
67.js, the 3D engine used for the fort and the jet engine: https://github.com/codepawl/67.js
The Three.js post the runner demo recreates: https://x.com/threejs/status/2119548992346533900
My video on weight-space programming: https://youtu.be/wXJDtImT_-0
Credits: dragon concept art by Inigo Quilez (CC0). House drawings: Historic American Buildings Survey, Library of Congress. Fort reference photos: NoahOmarY (CC BY-SA 4.0), Ibrahim Zauq (CC BY-SA 3.0), Marwatt (public domain), via Wikimedia Commons. Satellite image: Esri World Imagery.
Note: the fort footage labelled “pass 101” was rendered after I recorded the narration, which describes the run at pass 65.
Disclosure: I recently joined Google, and I started this testing before that. Google DeepMind gave me early access to this model. These are my own views.
Chapters
0:00 Gemini 4 Carbon
0:23 Two strengths: new research ideas and independent research
1:23 What this video covers and how the tests ran
2:11 Three.js open world recreated from a video
3:17 3D dragon from one concept image (SDF)
4:40 Storm in a teacup: a team of subagents
5:44 YouTube studio layout in 3D, 15 passes
7:02 From a floor plan to a 3D house
7:48 The long run: rebuilding Qila Bala Hisar fort
8:41 67.js, a one-month-old 3D engine
9:16 Research first: 30 sources and a building history
10:14 Carbon extends the engine’s source code
10:36 The review loop and its own scores
11:31 How to prompt a long unattended run
12:26 Jet engine cutaway in 67.js
12:56 GPU experiments: a galaxy collision on DGX Spark
13:40 Training an ant to walk
14:24 A neural network written by hand, no training
15:43 Where it fails: reply limits and effort settings
16:29 Four rules for prompting Gemini 4 Carbon
18:05 Verdict: prompts matter more with this model”}]] Read More Prompt Engineering
#Promptengineering #AI