Gemini 4 Carbon Wrote 20,000 Neural Net Weights by Hand and Scored 217

Google's Gemini 4 Carbon hand-wrote a neural network of roughly 20,000 weights for the game Breakout in a single attempt, and the resulting net averaged 217 points per game. The best trained network in an MIT research post averaged 25. That's the headline number from a battery of tests reported by Prompt Engineering — and, as always with these demonstrations, the caveats sit close behind.

The weight-writing trick worked for simple games only. When testers moved Carbon to a harder task, called Sequest, the model hit reply limits and failed to produce a functional network at all. Notably, only about 1 in 10 of those Breakout weights contributed anything; the rest sat at zero. So the impressive result rests on a sparse, lucky construction, not a general method.

What can it do with 3D reconstruction?

Carbon ran an autonomous reconstruction of a historic port in a lesser-known eastern city over 21 hours, using a newly released 3D tool — not Blender, not 3JS — despite thin online documentation.

A second project is more revealing. Carbon replicated Kilabalahisar, a fort in Peshawar, Pakistan, by analyzing 30 sources: colonial records, aerial photographs, and satellite imagery. It identified the visible walls as British-period bricks laid over an older mud core, and — to its credit — labeled undocumented sections as "inferred" in the 3D model rather than inventing them.

That project has run 65 iterations over about 20 hours. A review loop brings in a fresh sub-agent every 10 passes to compare renders against real photographs, reducing bias from earlier iterations. The replica's accuracy score has climbed from 3 to 5 out of 10. Photorealism starts at 8. During the project, Carbon also modified the open source 3D engine it used — six changes, including two bug fixes.

How does it handle self-correction?

When Carbon recreated a small open world from a 3JS video, its first output had jittery animations. Rather than patching errors directly, it wrote a program to audit the animation frame by frame, flagged problems, and corrected the walk cycle. It built the scene from simple shapes. Human reviewers still had to step in to catch the jitter in the first place — a point worth keeping in mind.

A separate YouTube studio model took 15 passes: render, review the images, fix visible flaws, repeat. The final interactive model lets users switch camera angles, adjust lighting, and receive equipment suggestions for camera, key light, desk placement, and seating distance from the back wall.

Can it coordinate sub-agents?

Given the phrase "a storm in a teacup," Carbon divided the work across sub-agents. It built a shared plan and timeline to synchronize elements like lightning strikes, and assigned specific components — porcelain, tea surface, clouds — to separate agents. The output ran to nearly 13,000 lines of code, with each sub-agent managing its own files. Modularity came out clean.

Engineering tests push further. On a DGX Spark cluster, Carbon simulated two colliding galaxies with over 500,000 stars, verifying stability with a single-galaxy test first. It also trained a reinforcement learning algorithm that taught a four-legged robot to run; when the robot sprinted off the floor's edge, Carbon extended the floor and re-rendered the video.

A jet engine project produced a museum-style cutaway: users can slice the engine at any angle, disassemble it into modules, and export sections as 3D files.

What breaks it?

High effort settings — where the model thinks longer before responding — actively hurt. In one case, Carbon spent its entire reply limit on reasoning and delivered no network. Long runs exceed the context window, so the model relies on a state file tracking the current pass plus the 10 biggest unresolved problems, reloading it after memory resets.

The pattern across all tests is consistent: Carbon shines on long, iterative, heavily prompted work and disappoints on simple single-shot tasks. Human feedback remains essential — testers needed it to catch animation jitter and to refine outputs throughout. The original claim that Carbon beat Claude Opus 5.5 in 16 of 18 tests appears in the headline but nowhere in the test descriptions themselves; readers should treat that comparison as unverified until the methodology is published.