Where's Fly-do?

Ten fruit flies, wired from a real fly's connectome, learn to find the striped guy. Then you race one.

▶ Play it in your browser (phones too)

Each fly sees through an eye with a real fly's facet count and learns with mushroom bodies (the fly brain's learning centers) wired with that fly's real connections, taken from its complete wiring diagram. Land on the striped guy: sugar. Land on anyone else: a zap. Flies that can't learn find him about 1 time in 8. Flies that learn pass 8 in 10 within about ten rounds and settle at 9 in 10. You can also make any fly helpless and watch it give up, or race one: can you find the striped guy faster than a fruit fly's brain?

In 2026, a team from HHMI Janelia, Google Research, and the University of Cambridge published something that had never existed before: a complete wiring diagram of a fruit fly's central nervous system. That's every neuron in its brain, both optic lobes, and the nerve cord that runs its body: about 166,000 neurons and roughly 125 million connections between them.

The way they made it is wild. A single fly's nervous system was cut into thin slabs. Then an ion beam shaved each slab away 8 nanometers at a time, for 13 months, while an electron microscope photographed every newly exposed layer. That produced 160 trillion voxels of images. No team of people could trace every neuron through all of that by hand. Google's part was the AI that does the tracing: flood-filling networks that follow each neuron through the 3D images, plus Neuroglancer, the open-source viewer researchers use to explore these maps.

Maps like this have grown fast:

I've always been drawn to psychology: why we learn, what we fear, what we want, why we do what we do. For most of its history, psychology has had to study the mind from the outside, by watching behavior and working backward. Freud and Jung built whole theories of the inner world with no way to look inside a brain. A wiring diagram is the first real look inside: the physical circuitry that behavior actually runs on.

And a fly isn't as far from us as it sounds. It learns from reward and punishment. It forms memories, and it gets used to things. It courts, fights, and makes decisions. Its learning runs on dopamine, the same chemical that carries reward signals in human brains. The complete map has already let researchers compare male and female brain wiring cell by cell, to see how the circuits behind courtship and aggression differ. And when researchers turned the earlier whole-brain map into a working simulation, it predicted which neurons drive feeding, and experiments confirmed it.

This project is a small version of that idea. Pavlov trained dogs with food. Fly researchers have been training flies with electric shocks since the 1970s and with sugar since the 1980s. Where's Fly-do does the same thing, with the real cells, and the real wiring, that do the learning.

I believe this is the start of something much bigger. Wiring diagrams have gone from 302 neurons to about 166,000 in forty years, and the AI that reads them keeps getting better. A mouse brain has about 70 million neurons, and people are already working out how to map one. A human brain has about 86 billion. That's still a long way off, and a wiring diagram isn't the whole mind: chemistry, timing, and experience matter too. But I think maps like this are how we'll one day understand the brain, and maybe the mind, as fully as it can be understood. The fly is where it starts.

flowchart TD
    eye["Compound eye<br/>886 + 893 facets,<br/>nearly blind to red"]
    vpn["253 real visual neurons<br/>of 143 types"]
    apl["APL neuron"]
    zap["Landed on anyone else?<br/>Zap: PPL1 dopamine<br/>weakens the approach<br/>synapses that just fired"]
    kc["335 visual Kenyon cells<br/>163 left, 172 right"]
    sugar["Found the striped guy?<br/>Sugar: PAM dopamine<br/>weakens the avoid<br/>synapses that just fired"]
    approach["Output neurons that<br/>push toward approach"]
    avoid["Output neurons that<br/>push toward avoid"]
    land{"Land here?"}

    eye --> vpn
    vpn -->|"17,890 real synapses"| kc
    apl -.->|"lets about 5% fire"| kc
    zap -.-> approach
    kc -->|"10,816 real synapses"| approach
    kc -->|"27,458 real synapses"| avoid
    sugar -.-> avoid
    approach --> land
    avoid --> land

    classDef zap fill:#e6eaff,stroke:#3b5bff,color:#16181d
    classDef sugar fill:#fdf1d8,stroke:#e9a21b,color:#16181d
    class zap zap
    class sugar sugar

Each real visual neuron reports brightness, darkness or stripes from one patch of its eye. Each turn, a fly glances at everyone nearby, flies to whoever looks most like past sugar, looks closely, and lands if that person looks better than what it's used to.

All from the Male CNS connectome, pulled by scripts/extract_counts.py and scripts/extract_wiring.py:

Their wiring: 253 visual neurons of 143 types feed the visual Kenyon cells through 17,890 synapses (counting connections of 3 or more synapses). The median cell has 4 visual inputs, and 5 cells get none at all, so they never fire. The cells send 10,816 synapses to output neurons that push toward approaching and 27,458 to output neurons that push toward avoiding.

The whole fly has 166,700 annotated neurons, 4,064 Kenyon cells, 316 PAM and 16 PPL1 dopamine neurons, and 97 mushroom body output neurons.

Nobody hand-labeled the output neurons. Each one is sorted by the dopamine neurons that connect to it: mostly punishment dopamine (PPL1) means a zap retrains it, and mostly reward dopamine (PAM) means sugar does. That rule alone reproduces the known layout of the mushroom body, with punishment dopamine in the γ1, γ2, α2 and α3 compartments and reward dopamine in the rest, and it cleanly sorts 95% of the visual Kenyon cells' output synapses.

• The striped guy: red-and-white stripes, a striped bobble hat, and glasses.
• Mimes: black-and-white stripes. Flies barely see red, so to a fly a mime looks almost exactly like the striped guy. Only the hat gives him away.
• Sailors: blue-and-white stripes, which look faint to a fly.
• Referees: vertical stripes, so the fly has to notice which way the stripes run.
• Everyone else: solid shirts, including plenty of red ones that look dark to a fly.

Yes, and you can check it in seconds. node test/learn.test.js runs 10 flies that learn against 10 identical flies whose synapses never change, 100 rounds each, and scores the last 50:

10 flies each, 100 rounds, scored on the last 50
  learning flies:        90% found, 43.0 visits per round
  flies that can't learn: 12% found, 114.1 visits per round
PASS  learning flies find him far more often
PASS  learning flies need less than half the visits
PASS  the worst learning fly (82%) beats the best fly that can't learn (24%)

A fly gives up on a round after checking 120 people, so the flies that can't learn usually run out of tries. Every run is seeded, so you'll get the same numbers.

Some of these ideas are built into how the flies work. Others are ways of looking at what they do.

Real: the cell counts above; which real visual neurons connect to each visual Kenyon cell, and with how many synapses; each Kenyon cell's real connections onto approach and avoid output neurons; flies' weak red vision; sparse Kenyon cell firing; and dopamine weakening whichever Kenyon cell connections were just active (sugar weakens the pull to avoid, punishment weakens the pull to approach). Learning is fast, which fits real flies: they can form a lasting memory from a single training session with sugar.

Simplified: the fly looks straight down at a flat page. The optic lobes' processing is replaced by four simple measures per patch, and what each real input neuron reports is picked per neuron: one of those measures, from one small patch of its eye. (Giving neurons wide fields of view blurred the brain's codes, and the flies found him 66% of the time instead of about 90%.) Output neurons are pooled into approach and avoid instead of being simulated one by one.

• Play: open index.html. There's no build step.
• Prove the flies learn: node test/learn.test.js runs 10 learning flies against 10 identical flies that can't learn.
• Redraw the learning curve: node scripts/learning-curve.js runs the same flies and writes assets/learning-curve.svg.
• Regenerate the connectome numbers: pip install pandas pyarrow, then python scripts/extract_counts.py.
• Regenerate the wiring: python scripts/extract_wiring.py. The first run downloads the connectome's full connection list (about 1 GB).
• Redraw the link preview image: scripts/social-card.html draws social-card.png from the game's own crowd and eye; the Chrome command is at the top of that file.

index.html, style.css   the page
src/eye.js              compound eye: 1,779 facets, 12 patches × 4 measures per eye
src/wiring.js           real visual neuron → Kenyon cell synapses (generated)
src/connectome.js       cell counts from the connectome (generated)
src/brain.js            mushroom bodies: Kenyon cells, APL, approach and avoid, dopamine
src/game.js             one fly's search: glance, choose, look, land, learn
src/scene.js            the crowd
src/race.js             racing a fly
src/ui.js               the lab, the panels and the chart
src/sprite.js           the pixel fly
src/util.js             seeded random numbers
test/learn.test.js      learning flies vs. identical flies that can't learn
scripts/                connectome extraction, the learning curve, the preview image

• Connectome data: Male CNS connectome v1.0 by HHMI Janelia FlyEM, Google Research, and the MRC Laboratory of Molecular Biology and University of Cambridge. CC-BY 4.0. male-cns.janelia.org
• Sexual dimorphism in the complete Drosophila male central nervous system connectome, Cell (2026)
• A connectomics milestone: mapping the complete male fruit fly brain, Google Research
• Ten years of neuroscience at Google yields maps of human brain, Google Research
• A Drosophila computational brain model reveals sensorimotor processing, Nature (2024)
• Visual input into the Drosophila melanogaster mushroom body, Cell Reports (2020)
• The neuronal architecture of the mushroom body provides a logic for associative learning, eLife (2014)

Not affiliated with Where's Waldo or its owners.

MIT