Vote for what the flies sense next
How the brains work
Each of the six voices you hear starts from a real map of a fly's nervous system. Four brains are built from the Male CNS, a complete wiring diagram of a male fly's brain and nerve cord published in 2026. Two are built from FlyWire, an earlier wiring diagram of a female fly's brain alone. Both maps show where every neuron sits and how it connects, and that wiring never changes during a performance. What we add is a simulation of activity: each neuron firing, moment to moment, as smell, wind, warmth and light move through the circuit the way biology would carry them. All six run at once on a computer in a room in San Rafael, California, driven by the town's live weather, sun and air, and their activity becomes one piece of music that never ends. What you see is every neuron of each brain lit by its own activity, and nothing here is recorded or written by hand: you are hearing what the brains did about twenty seconds ago. Every ten minutes you can vote for what the flies sense next; if nobody votes, the box picks at random.
The wiring is real
Each brain's connectome has about 166,000 points for the male brain and about 139,000 for the female, joined by roughly 125 million and 34 million connections. Every point is a real neuron, traced from thin slices of an actual fly brain under a microscope, and every connection is a real synapse counted the same way. Each neuron also carries a label for the chemical it most likely releases onto its targets, some exciting the neurons downstream and some quieting them. That label is itself a prediction, not a direct measurement, and most of the wiring is left exactly as predicted, including real disagreements between the male and female maps; only a handful of cases where the prediction is flatly contradicted by decades of lab work, such as a well-studied calming cell type mislabelled as exciting, are corrected first.
A pulse, not a flow
Inside the simulation every neuron is a small leaky battery. It collects charge from the neurons wired onto it, loses some of that charge every instant, and fires a brief pulse once it crosses a threshold. After firing it goes quiet for about two thousandths of a second before it can fire again, the way a real neuron's channels need to reset. The whole brain is stepped forward a thousandth of a second at a time, so a pulse reaches the next neuron only after the short delay a real connection has. This is a simplification of a real neuron, which is far more complex, but it keeps the wiring exact and the timing in order.
From a smell to a signal
A stimulus starts as a written profile, not a live sniff of the real object. For a pile of horse manure, papers on dung chemistry tell us which odor receptors respond to its main compounds, phenols and ammonia, and that becomes activity on those receptors' neurons, read overall as mildly unpleasant. A human hand held near the box is mostly not a smell: it is warmth on the fly's heat sensors, carbon dioxide from breath, and a looming shape its visual system treats as a threat. A gust of wind carries no odor at all; it bends the tiny hairs and antennal joints that sense moving air, and it cools and dries the antenna as it passes. Every profile is a careful estimate from the literature, not a measurement of that particular object.
What the composer listens to
Nothing composes directly from voltages; the piece reads how many neurons of a kind are firing and how fast. How excited the whole brain is, averaged over thousands of neurons, sets loudness and how busy the texture feels, not how fast the beat runs, which stays fixed within a stretch of music. A separate set of neurons, the ones a fly's brain uses to grade a moment as worth approaching or worth avoiding, decides whether a voice leans toward the stimulus or away from it. Habituation, a brain's own way of getting bored, runs on two clocks: a fast fade over a second or two models a receptor going numb, and a slower fade over tens of seconds models the whole circuit losing interest, so a smell that lingers keeps fading toward a quieter, steady hum rather than staying at full strength.
Six brains, one clock
The six brains do not send signals to each other today: each one only senses its own version of the world, and nothing crosses from one brain's neurons into another's. What makes them sound like a group is a shared clock and a layer above the simulations that compares their states about ten times a second. When the brains broadly agree, excited or calm together, that agreement sets the key and mood everyone plays in. When they disagree, one aroused and one bored, that split becomes counterpoint, a held note against a moving line rather than one blurred chord. Whichever brain is responding most differently from its own recent past gets handed the solo for a while.
The male and female difference
The two sexes are not the same brain repainted pink or blue; they are separate reconstructions of separate flies, and the male one carries clusters of neurons the female one does not have, built for courtship and, in some of the same cells, for aggression. A cluster called P1, about 150 neurons wired to hear another fly's song, see it, and taste it on contact, exists in the male connectome with no counterpart in the female's at all. So the same stimulus lands on genuinely different circuitry in the two sexes, and their activity drifts apart on its own, with no rule written to make it happen. The piece hears that drift as two voices answering each other in different registers, which is real biology, not an effect added afterward.
How the sound is made
Every voice you hear is a fruit fly brain. Six of them are running at once: four built from the wiring diagram of a male fly and two from a female, every neuron and every connection taken from the real maps that scientists traced from electron microscope images. What reaches each brain is what reaches a fly in the room: the local weather as a steady background, and now and then a smell or a flash, voted by whoever is watching, or drawn at random from a library of things a fly can sense when nobody votes, from rotting fruit to a thunderstorm. The neurons respond the way the wiring says they must. How excited the whole brain is sets the loudness and the pace, its learning center sets whether the music leans toward the stimulus or away from it, and how used to a smell it has become decides when a voice tires and drops out. The music is composed from that activity; nobody wrote the notes.
The tone of each voice comes from the fly's anatomy too. A fly sorts smells in a part of its brain made of 55 small stations, one for each family of smell receptors on its antennae. We turned those 55 stations into 55 resonators, tuned as the partials of one note, and we wired the resonators to each other exactly as the stations are wired in the fly: a connection that excites in the brain excites in the instrument, and one that inhibits damps. Play a note and the whole network rings and settles the way the fly's own circuit would. The smell the brain is sensing then decides which of the 55 keep ringing and which fade, so the same note sounds different when the brain smells vinegar than when it smells rain. Some of the brains play a second instrument built the same way from the fly's compass, the ring of eight cell groups that tracks which way the fly is facing, which gives a smaller, purer tone; and each brain strikes, bows, plucks or breathes into its instrument in its own way, so no two of the six sound alike.
What to listen for: the female brains, drawn in pink on the display, sing against the male brains in blue, and they differ because their wiring differs. When a new smell arrives you will hear a fly's courtship song for a couple of bars, the pulsing wing song a male fly sings to a female, and the voices bloom and brighten as the brains get excited, then darken as they grow used to it. Under long notes you may hear a faint sympathetic ringing: that is a model of the fly's own ear, the hearing organ in its antenna, listening to the music and answering.
In more detail: the 55 resonators and the neurons behind them
The 55 and the wiring between them
The antennal lobe is where a fly sorts smell. It is made of small stations called glomeruli. Each receives one family of smell receptors from the antennae and hands the signal on through a handful of cells called projection neurons. The 55 are those stations, picked by one rule: every projection neuron type whose name carries a station label. That rule gives the identical 55 labels on both the male and the female maps, 51 for smell plus 4 for temperature and humidity. Behind each label sit real cells: 2 to 15 projection neurons per station, 284 in the male brain and 275 in the female, plus about 420 local interneurons per brain that connect the stations to each other.
The coupling between the 55 comes straight from the synapse tables. Two kinds of path count: a projection neuron of station B synapsing directly on one of station A, and the two-step path where B's projection neurons excite a local interneuron that then synapses on A's. That second path is the fly's own way of balancing one smell against another. Deliberately left out is the path from the smell receptors through the interneurons, because the receptors are the outside drive, and the live simulation supplies that separately. Every synapse carries a sign, excite or inhibit, from the predicted transmitter of the cell that sends it, and each input is counted as a fraction of the receiving cell's total input, so a large station cannot dominate by size alone. The result is averaged over each station's cells into one table of 55 by 55 numbers. The entry for A and B reads: the signed share of a station-A projection neuron's input that comes from station B, directly or through one interneuron.
| male | female | |
|---|---|---|
| couplings that are nonzero, of 3025 | 3025 | 2971 |
| spectral radius (how strongly the network feeds back) | 0.25 | 0.27 |
| share of the coupling that inhibits | 0.3 percent | 2.2 percent |
Two nuances worth knowing. After the fractions are taken, the coupling is mostly excitatory in both sexes, and the two datasets disagree substantially about the interneurons' transmitters, so the female's stronger inhibition is partly an artefact of the data. Both facts are left as they are: the project's rule is that the biology is never tuned to make the music nicer.
How a station becomes a resonator
Each station is one sine partial at the played note's frequency times a fixed ratio, with an optional narrow band of noise around it, the breath. What the wiring controls is the partial's loudness, through a small network that runs a thousand times a second. Every station holds an excitation value. Each step it moves toward a target made of four parts: the strike from the note-on, shaped by how hard the note was played; the held drive while the note is down; any push from the fly organ layers; and the coupling term, the sum over all other stations of their excitation times the signed coupling, with the exciting and inhibiting parts scaled separately. A negative target counts as zero. The time constant is about 60 milliseconds.
So a loud partial pumps the partials its station excites and damps the ones it inhibits, and the whole spectrum settles the way the circuit would. The exciting gain is capped so that gain times the network's feedback strength never exceeds 0.95, which is why the network rings but never runs away, and after the note is released the loop gain halves so every note decays. The partial levels then pass through smoothing with a 15 millisecond attack floor, a gentle tilt toward the bass, a shelf above 2.5 kHz, a gate 40 dB under the loudest partial, and a cap of 32 audible partials per voice.
Which ratio each station gets is the musical half of the design, and it changed once. The first design put seven scale degrees into every group of seven stations, and the owner heard it as dissonant. The current default places stations on a consonant set of harmonics, 1, 2, 3, 4, 5, 6, 8, 10 and 12, with several stations sharing each harmonic and 3 to 12 cents of detune between them, which gives a slow chorus. The order still comes from biology: stations sorted by how attractive their smell is to a fly, the attractive ones on the low warm harmonics and the repellent ones high; within a group, the well-connected generalists get the simplest ratios and the specialists the color notes. How broadly a station's receptors respond sets its breath: broad channels are airy, and the two channels for the fly's own pheromones are close to pure sine. One station doubles the fundamental so a chord is never rootless.
What moves it while it plays
The live link is the neural bus. About ten times a second, every station's activity from the whole-brain simulation becomes that station's held drive, blended with a fixed pattern, with a compensation that keeps a sparse smell from going quiet. How excited the brain is scales that drive; reward and punishment signals from its learning center tilt the balance between exciting and inhibiting; and a slow neuromodulation gesture shifts the same balance so a held note re-blooms without a new attack. Pressure adds drive, and with the current instrument, color.
On top of that, with the current instrument. Excitement above a brain's own recent baseline opens and brightens the tone, habituation darkens it, and a new stimulus makes it bloom. Each brain strikes, bows, plucks or breathes into its network in its own way. And some brains play the fly's compass instead: eight stations, the four head-direction cell types on each side of the brain, with the same recipe applied to that circuit. The compass has no smell channels, so the smell does not steer it; the composed notes alone play it, and it gives a smaller, purer tone.
The male and female instruments share the same 55 labels and the same recipe. They ring differently only because the wiring differs, which is the point.