FLY / DISTRICTA living advertising experiment
Preparing habitat
FLY DISTRICT 3.5 × 2.5 CM

Six places.
One curious fly.

A world of food, sport, screens and tiny distractions.
Put your brand in it. Watch the response.

Create a test campaign ↗TEST RELEASE · NO REAL PAYMENTS
Opening the habitat

Loading the anatomical model…

Preparing the model. Controls stay available while it loads.
Waking upGeneration 1
LIVE CONNECTOME · FLYWIRE V783

—of 139,255 neurons spiking

10 ms window · 15,091,983 edges

Spiking now · Dim points are inactive

  • Kenyon cells
  • MBON
  • Projection
  • Olfactory receptor
  • Lateral horn
  • Descending
  • Optic
  • Central
Cell inspectorClick a neuron

Pick any point in the brain to read its class, incoming synapses and firing rate.

Live signals

Neural activity · sensory inputs

Populations in the order shown, left hemisphere first within each. Brightness is the cell's firing rate in the last window.

Optic drive0.00L0.00R
Odor at the antennae · left | right

Plumes and the rendered retina are the only inputs the brain receives. Steering reads the brain, not these numbers.

Experiment desk

03 / Sense, intervene, compare

Rows are presynaptic, columns postsynaptic. Hover a cell for the summed signed weight and edge count.

Left and right are split by the medial plane of measured soma positions; bars are the mean rate of every labeled cell.

Last reward0.000Steering advantage0.000
Goal values Q + innate need
Steering weights 0 updates

First lifetime in progress. Death archives its outcomes and both learners.

No completed lifetimes to compare yet.

Experiences appear here as the fly interacts.

Lifetime archive

No completed generations

Each death archives its outcome and both learners. Records are currently stored in this browser; a shared 24/7 host is not connected yet.

The first lifetime is in progress. Its record will appear here when it ends.

What is real in this habitat?

Measured anatomy and wiring. The FlyWire v783 graph (139,255 neurons, 15,091,983 directed edges with signed weights) runs a leaky integrate-and-fire model on WebGPU, 1 ms steps, alpha synapses after Shiu et al. 2024. The 3D brain and the pixel matrix show every neuron, colored by live firing rate. The connectome matrix sums the measured weights between populations. The body is TuragaLab flybody's original meshes, joints, 78 actuators and MuJoCo contacts.

Authored senses. Each habitat zone emits an odor plume that decays with distance. Olfactory receptor neurons are split by hemisphere from their measured positions and assigned to five odor channels by index (the prepared data has no glomerulus labels). The antenna facing a source receives more drive. The rendered retina's left and right halves drive the left and right optic lobes.

Real reinforcement learning, two learners. Steering reads left−right activity asymmetries of six populations from the live brain and outputs a turn; it learns by REINFORCE from progress toward the goal. Goal choice is linear Q-learning over interoception, eight population rates and the sensed odor of each candidate. Innate need priorities, two reflexes (righting, escape) and a generation-decaying exploration rate coexist with what is learned. Rewards are algorithmic signals, not simulated dopamine.

Body. Turn and forward commands drive an authored tripod gait with velocity assistance. Navigation assist at 0% means the brain-derived turn alone steers; raising it blends in a coordinate bearing. This is not a measured brain-to-muscle mapping.

Sponsored world. Test-auction winners place their creative in a resource zone. Aroma boosts that zone’s authored odor plume, while signs and silent videos appear in the rendered retina. A bounded, need-dependent salience bonus also enters goal selection. Price never affects attraction. This is not a biologically validated advertising response or a measure of human effectiveness. Test profiles, bids, media and observations stay on this device; no real payments or shared advertiser accounts are connected.

Predator pressure. A periodic moving danger footprint applies proximity-based damage. The refuge is safe. Exposure adds a negative learning reward; this is an authored hazard, not a reconstructed predator or a validated fear circuit.

Life and inheritance. Hunger, thirst, rest, health, fertility, pollen delivery, eggs and a 240-habitat-second lifespan are simplified rules. On death both learners and the journal are archived and inherited by the next adult (an egg if one exists). Descendants can improve or regress; the lineage panel reports what actually happened.

Interventions. Disconnecting a cell zeroes its incoming and outgoing edges on the GPU; the synaptic-strength slider scales every measured weight. Both persist in the lineage so damaged-brain generations can be compared with intact ones. "Live drive" in the connectome matrix is presynaptic rate × summed weight, an estimate, not a spike-resolved current.

Sources: webgpu-fly · flybody · FlyWire · Shiu et al. model. Everything runs in this tab; no remote inference or hidden training service.