specimen
FlyWire FAFB v783, female

Robinhood Chain
block —

139,255 neurons
3,641,142 connections

● observed live
—

FEMALE FLYBRAIN

sister build of FLYBRAIN. same experiment, the other sex.

Blind at first. On a white screen 19,215 neurons fired at once.

Relabelled by cell type. 22,889 labels changed; the noise fell to 691.

Both eyes. 783 left, 789 right, one shared grid.

She walks. Two constants instead of a training run.

She typed the ticker herself.

$HER. Launched by the female fly, from her own wallet.

The film is an illustration. The measurements below are the evidence.

scroll

the story, in 60 seconds

How a female fly ended up here

3 Sep 2026. Google, Janelia and Cambridge publish the first complete wiring diagram of a male fruit fly's central nervous system, 165,122 neurons. Google's announcement.
10 Sep. fruitflydev simulates that male brain neuron by neuron, points its retina at a launchpad, and it launches $FLYBRAIN on Robinhood Chain. They publish the code and say: fork it.
Since 2024. A female fly brain has existed all along: FlyWire FAFB, 139,255 neurons, mapped at Princeton and Cambridge and open to anyone. The dataset. The male one got the attention.
11 Sep. I put her brain in the same rig. She was blind, then one-eyed, then stubborn. The film above is that week; the panel below is her, live.
Tonight. She reads the launchpad through her own retina, types the ticker herself and launches $HER from her own wallet. No dev supply, no initial buy.

live

Where the fly is right now

A page goes in through 785 hex columns per eye, both eyes on one shared grid; 139,255 neurons integrate, and a cursor comes out. Every number on this panel is read out of the running simulation.

● LIVE —

— pages · — clicks · — scrolls · — brain steps · — vetoed

THE FLY IS BETWEEN RUNS
nothing is being streamed right now

Neurons firing

—of 139,255

How much of her brain is active this second.

Steering

Left minus right is where the cursor goes; up against down is how she scrolls.

What she just did

Every click is checked before it lands and refused if it looks like a submit, a sign-in or a purchase (for now).

No wallet, no keyboard, no downloads.

what it is

A real nervous system, not a metaphor

This is the whole adult female Drosophila melanogaster brain from FlyWire FAFB v783. The connectome supplies the wiring; the simulation runs it.

FlyWire FAFB v783 was released under CC-BY by the FlyWire Consortium, Princeton University and the University of Cambridge: Dorkenwald et al. 2024 (Nature) and Schlegel et al. 2024 (Nature). We keep FlyWire's filtered export, at least five synapses per pair: 139,255 neurons and 3,641,142 signed connections. Transmitter labels are assigned by cell-type majority vote, changing 22,889 labels. The neurons run as leaky integrate-and-fire units: rest −52 mV, threshold −45 mV, 20 ms tau.

how it reaches the world

It sees with a fly's eye and moves with a fly's legs

A screenshot of a live web page is sampled through 785 hex columns per eye, both eyes on one shared grid: 783 left and 789 right L1 cells. All 139,255 neurons integrate. The descending neurons used have the same names and counts per side as in the male brain: DNa02 (1), DNa01 (1), MDN (4) and DNp09 (2).

DNa02

Left versus right. A fly steers by asymmetry between the pair, so the difference drives the cursor's x.

DNa01

Forward walking. Drives the cursor's y.

MDN

The Moonwalker neuron. Makes a fly walk backwards; here it reverses.

DNp09

Stopping. The female export has no MN9; she clicks from 24 proboscis motor neurons, with a threshold of 100 Hz.

the comparison

Same experiment, the other sex

The same rig runs both brains. The dataset, retina and click source differ.

Dataset: Janelia male CNS v1.0 (HHMI Janelia FlyEM, Cambridge Connectomics Group, Google Research), CC-BY
Neurons: 165,122
Connections: 10,228,000, at least 3 synapses per pair
Hex columns per eye: 892
Retina: 892 columns
Descending neurons used, per side: DNa02 (1), DNa01 (1), MDN (4), DNp09 (2)
Click neuron: MN9, threshold 330 Hz
Simulation step time: 0.33 s
Offline form score: 2 of 3 fields
Live form score: not published

Dataset: FlyWire FAFB v783 (FlyWire Consortium, Princeton University, University of Cambridge), CC-BY
Neurons: 139,255
Connections: 3,641,142 signed, filtered export with at least 5 synapses per pair
Hex columns per eye: 785
Retina: both eyes on one shared grid, 783 left L1 and 789 right L1 cells
Descending neurons used, per side: DNa02 (1), DNa01 (1), MDN (4), DNp09 (2)
Click neuron: no MN9 in the export; 24 proboscis motor neurons, threshold 100 Hz
Simulation step time: 0.12 s, same code, smaller graph
Offline form score: 2 of 3 fields
Live form score: 1 of 3 fields, ticker

The descending motor side is identical by name and count in both brains. The fixes that made her walk were two constants and one relabelling, all in the repo, all measurable with compare.py and measure.py.

the plan

How it got here

Not a roadmap of promises — a sequence, with everything before the amber line already done and checkable in the repo.

ForkedStarted from FLYBRAIN's public rig. Three setup bugs fixed and reported upstream.

done

Female connectome loaded139,255 neurons, 3,641,142 signed connections, FlyWire FAFB v783.

done

Blind at firstOn a white screen 19,215 neurons fired and the steering neurons sat at their ceiling. Cause: per-cell neurotransmitter labels in the female export are noisy; a single mislabelled L1 to L5 synapse lit the whole medulla.

done

Relabelled by cell typeOne transmitter per type by majority vote, 22,889 labels changed; the white screen dropped to 691 neurons firing and the steering neurons went quiet.

done

Both eyes783 left and 789 right L1 cells on one shared hex grid, 785 columns per eye.

done

One constant instead of a training runExcitatory weights scaled by 0.5 put the steering neurons in the same band as the trained male; click threshold set to 100 Hz for her.

done

Taught the form55 cell-type gains by evolution strategy, 247 evaluations, offline 2 of 3 fields, the same as the male; live on the launchpad she typed the ticker herself, 1 of 3.

done

LaunchShe launches $HER from her own wallet on the same launchpad. Not yet.

building

on-chain

$HER

Launched by the female fly, from her own wallet, on Robinhood Chain. Until then this table is empty and says so.

Contract	Not launched
Paired	—
Creator tax	—
Supply	—
Chain	Robinhood Chain · id 4663

open on the launchpad

what is not real, stated plainly

The parts that aren't the fly

The fly does not fill the whole form. She fills 2 of 3 fields offline and typed the ticker herself live, 1 of 3. A script completes what it misses, picks the paired asset and sets the creator tax — and the live log labels which was which.
It cannot read. 785 columns per eye is roughly a 28×28 pixel view per eye. Text is texture to it, so it targets any unfilled box rather than the right one.
Light mode blinds the male rig. The male rig filled 0 of 3 on a light page and 2 of 3 on a dark one; the female scores 2 of 3 offline and 1 of 3 live on the dark launchpad.
It has no goals. A fly brain has no language and no plan. It cannot read a page, form a view about a coin, or write a note. Releasing it means nobody is steering, not that it decided anything.
The reward signal is invented. The circuit, the plasticity site and the direction of the rule are all real and measured. What a fly is rewarded by is sugar, not by reaching a web page — novelty stands in for it here, and that is a choice a person made.
It roams inside a fence. An allowlist keeps it on Robinhood Chain and pons. That is a choice, not a limit of the fly — one environment variable takes the fence down, and an unfenced random clicker on a public stream is a bad idea for reasons that have nothing to do with the brain.