AI CLUB · EXACT V25B MODEL · EPOCH 4

Watch a small neural network steer Snake

0 food 0 steps 0.0% walls attract mode
34,588 parameters · v26 round 1 · checkpoint a06da7b4…704f

LIVE INPUT

Four convolution maps

10 categorical frames · 20×20
Map 1
Map 2
Map 3
Map 4

Topology inputs

Immediate danger

Current heading

Food direction

LIVE NETWORK

Hidden layers → action choice

Neuron fill shows activation. Blue/orange connections show positive/negative contributions; dashed lines are negative. These are numeric computations—not probabilities or human-like thoughts. The final decision layer combines a shared signal with three action-specific signals. The selected v26 model is constant; it was trained on sparse fixed-wall boards, not every visitor-designed environment.

TESTED ATTRACT BOARD

Empty reference · seed 99,200,057

Right → Right
Click anywhere to design a board
HeadBodyFoodWall (moves when on a track)Track