Fly Brain · WebGPU Connectome
Real FlyWire (female) and Janelia MaleCNS (male) Drosophila wiring running as spiking neurons on your GPU. Stimulate pathways, read and rewrite synaptic weights, train it with reinforcement learning, and pit it against the language model.
Wiring from FlyWire FAFB v783 (female) — 134,181 reconstructed neurons and synapses across 78 neuropils. Dorkenwald et al., Whole-brain annotation and multi-connectome cell typing of Drosophila (FlyWire FAFB v783), Nature 2024. Segmentation by Google Research; hosted by Princeton Neuroscience Institute.
AL_L
0.000
AL_R
0.000
AME_L
0.000
AME_R
0.000
AMMC_L
0.000
AMMC_R
0.000
AOTU_L
0.000
AOTU_R
0.000
ATL_L
0.000
Train Behavior
Describe any behaviour. Astra builds the 3D environment, the connectome learns it with reinforcement learning, and the language model maps the result onto neurons.
Sugar source approach
Fly to the sugar drop while avoiding the spider's web.
Sensors → neuropils
odour left (goal_left) → AL_L
odour right (goal_right) → AL_R
looming (obstacle_ahead) → ME_L, ME_R
danger (hazard_proximity) → LO_L, LO_R
Actions
turn left turn -1 · throttle 0.5
forward turn 0 · throttle 1
turn right turn +1 · throttle 0.5
reward · approach 10 · reach 2 · collision -1 · hazard -0.6
Edits modulate the 787-neuron simulation on the GPU immediately; the pristine published matrix is kept underneath, so gain changes are always reversible.
This brain is registered on the platform driver bus as flybrain.stimulate, flybrain.train, flybrain.evaluate and flybrain.state — any other engine, agent, or chat skill can drive it and read its firing state — and as flybrain.weights, flybrain.edit_weights, flybrain.activations, flybrain.gains and flybrain.arena, so agents can read and rewrite the synaptic weights themselves and run the reinforcement arena against the language model. flybrain.design_behavior, flybrain.train_behavior and flybrain.map_behavior expose the prompt-driven behaviour trainer.