117 → 1.00
Average Sunset color error, in RGB levels, on 4,096 held-out samples. A small color network trained for 1,500 steps while the 3D geometry stayed frozen.
Glacier: 122 → 1.56. Targets were authored synthetic color curves, not photographs.AN EXPERIMENT IN NEURAL COMPUTATION
Type a world.
Watch it change.
Snow. Sunset. A rainy neon night.
Turn words into weather and color while you play.
Trained text understanding. Neural pixels. On your device.
PRIVATE · LOCAL INFERENCE
Loading the text model…
Mix snow, rain or fog with sunset, night, neon, monochrome or icy colors. Use “no snow” to remove an effect. Unmentioned settings stay as they are.
What can this model change?
It learns instructions for these weather and color controls, including short combinations and negation. Requests it rejects leave the world unchanged. It cannot create arbitrary objects or rewrite the game rules. Snow geometry and motion are authored neural circuits; text intent and color response are trained. Short explicit phrases work best; some paraphrases are missed.
TRAINING LABSee what changed ↘
Loading experiment…
THE WORLD / LIVE INFERENCE
Tap the frame to flySCORE 00 BEST 00
ROTATE VIEWQ / E
WHAT ACTUALLY LEARNED
Small experiments. Clear results.
Gravity learned: 2.00
Starting at 1, the fit recovered gravity 2 and jump impulse 26 from example trajectories. The exported values were projected to integers to preserve exact rules.
Same pilot, same start: 248 → 193 ticks to finish. Faster completion was a side effect, not the training target.Lower loss ≠ valid rules
Updating every simulation weight reduced prediction loss by about 49%, but 21.5% of held-out transitions had invalid Boolean flags in PyTorch.
The exported JavaScript checkpoint also failed: 16.2%. That unstable version is documented, not used in a playable cartridge.All runs use a fixed seed. Training ran locally in PyTorch; your phone loads the finished weights. Full metrics & learning curves ↗
THE IDEA, MADE PLAYABLE
Every rule.
Every game pixel.
The 3D cartridge supplies a neural simulator and a neural ray renderer. Convex solid intersections, nearest-surface depth, facet lighting, and the entire HUD become ReLU weights. The same runtime also plays Flappy and Pong.
Read the technical notes ↗Download engine source ↓
01 / INPUTState + buttonsFixed-step recurrent state
→02 / FORWARD PASSHand-set + learned weightsWeighted sums → max(0, x)
→03 / NEURAL RENDERERPixel coordinates → RGBThe network paints the frame
← THE STATE LOOPS BACK. THE RULES STAY IN THE WEIGHTS. ←
01
Solid, in three dimensions. Walk in X/Z and jump in Y. Orthographic rays intersect 3D solids; neural depth comparisons decide what is visible. Rotate the camera through four views.
02
A deliberately small wrapper. The browser collects inputs, evaluates the networks, and presents their RGB output. The surrounding controls and inspector remain ordinary interface elements.