CUTE-DINO · MALECNS v1.0

Fly's neurons are playing your cute-dino game.

80 real neurons and 1,296 measured synaptic connections form a circuit that is entirely fixed and untrainable. Behind it sits a single 269-parameter readout network that decides whether to run, jump, duck, or move left and right. What you see is computed live, not a recording — every 1/30 s the game state really does pass through those 1,296 connections once.

Neurons
80
Edges
1,296
Synapses
26,029
Channels
13
Params
269
Rate
30 Hz
How it works ↓
Running · Connectome + trained readout
01 / Gamecute-dino · 60 Hz
Score 0 0.0s Difficulty 1.0× HP 1 seed —
Action RUN
02 / Keyboard outputflybody anatomical model
—
Both forelegs press the key for the current action Drag to rotate
03 / Decision networkTrainable action readout
16 → 12 → 5 · 269 params · 30 decisions/s Highest score wins · orange + / blue −
04 / Brain activityMaleCNS v1.0 soma atlas
Mean |activity| 0.000
124,289 somata · 80 circuit cells at true coordinates Drag · scroll to zoom
Excitatory Inhibitory Silenced Input Output DN
05 / Sensory channelsGame state → 32 visual projection neurons
Range 0..1, injected as u = 2·(feature − 0.5) into the cells assigned to that channel The channel count follows measured within-type projection separability, not an arbitrary choice

Training & evaluation

Only the tiny readout network is taught

The 80-cell circuit's weights never move; only the final 269 parameters are evolved. Cross-entropy method (CEM): 64 candidates per generation, the best 8 update the distribution, and each generation draws a fresh batch of tracks.

Learning curve—
Re-runs CEM on your own CPU; results stay on this machine

Results

Switch the circuit off and it stops playing

100 unseen tracks, 180 s cap each. The ablation group forces the 16 descending neurons' output to zero and changes nothing else — if the score holds, the readout was never using that circuit.

Method

How the data flows through

  1. 01 · OBSERVEObserve.
  2. 02 · INJECTInject.
  3. 03 · PROPAGATEPropagate.
  4. 04 · READ OUTRead out.

Circuit activity is a simulated, dimensionless value — not a membrane potential and not a measured firing rate. The input encoding and action readout are both hand-specified, not biological measurements. This is one selected circuit, not a whole brain.