The long version — how it works → the brains and where they come from, the wave and the spiking model with their numbers, the harmonics, and every rule that turns the wiring into music.
Real · sampled · modeled. Real: the neurons, their positions and the connections between them come from the two fruit-fly connectomes, drawn in their real proportions; each connection is signed by its presynaptic neuron's predicted transmitter, and the 24 harmonics of each brain are computed from the wiring you are playing (his cells without a soma in the brain, mostly receptors and ascending neurons, sit at the center of their brain partners). Sampled: each brain shows a subset chosen by identity — 36,642 of her ~139,000 neurons and 36,724 of his ~146,000 brain neurons: every olfactory, auditory and pheromone receptor, the taste neurons (her sugar and water cells, all but his bitter ones), a slice of the visual receptors, the cells each dataset marks by sex (her sex-specific, dimorphic and fruitless/doublesex neurons; his 190 fruitless/doublesex ones, since his data has no sex-specific or dimorphic labels) and a stratified rest — with the strongest connections among them. Modeled: the firing, the inhibition, and every mapping from wiring to pitch, rhythm, and color are musical choices — this is a way to hear a wiring diagram, not a physiological simulation.
Two brains. On the left, the female fruit-fly brain from FlyWire (FAFB v783, 2024). On the right, the male brain from MaleCNS v1.0, released September 2026, with positions from neuPrint. Each has its own 24 connectome harmonics — the resonant modes of its wiring, eigenvectors of its graph Laplacian (stored at 8-bit precision). Tapping a mode chip rings mode k in both brains at once; they are different eigenvectors of different brains, and you can watch how similar (or not) the shapes are. Each chip's tooltip says what that harmonic divides — a patch of one eye from the rest, an eye from part of the central brain, or only fine structure — read off the eigenvector itself.
The duet. A spark excites a neuron and spreads along real synapses as a traveling wave — neurons fire, go refractory, the wave moves outward, inhibition pushes back. Duet picks one spot and sparks the nearest neuron to it in both brains at the same instant. The two waves then follow their own wiring. The brains differ — two animals, two reconstructions, two samples, with sex-specific cells among the differences — so the waves drift apart, and you hear that as two melodies pulling apart. It is not a measurement of sex differences. In The Band each brain plays two of the four voices (the first brain gold and rose, the second cyan and green); alone, a brain plays all four.
The instruments. Each of the four voices has an instrument you can change: nineteen built-in synths (harp, strings, electric piano, clarinet, organ, choir, flute, brass, nylon guitar, koto, marimba, music box, kalimba, steel drum, synth bass…) or any of the 128 sampled General MIDI programs (FluidR3, loaded on pick from a public CDN). Every species opens with its own band, and surprise me deals a new one. A genre (ambient, techno, jazz, orchestral) sets the tempo, the band, one of five drum kits, the chord bed, a bass line on the harmonic chord's root, note lengths, swing and the keys the conductor may visit. Pitch spans four octaves and each instrument sits in its own register, so a band spreads over about six, and the same choices are sent as program changes to MIDI live outputs and written into the exported file. MIDI file in (or a .mid dropped on the page) plays a piece into the brains: it sounds as written, with its own General MIDI instruments and drums; every note lights the neurons at its pitch's height (a new wave when a voice is free, otherwise it feeds a live one); the brain listens while it plays — no melody of its own, the chord bed following what lights up — and when the piece ends the brains answer with what they caught, in its key.
The music. Every wave is a voice: height in its brain sets the pitch — in a key chosen by what the brain is sensing (D Dorian by default; ripe fruit plays in F Lydian, motion in G Mixolydian, sugar in C Ionian, a mate in E Phrygian) and, when Play endlessly runs, changes every eight bars to a related key chosen by the song's seed, along with the spark pattern, tempo lean and number of voices — left–right position on screen sets the stereo pan (with headphones and binaural on, the wave's whole direction), and size sets both loudness and how long a note is held. Drums come from the combined firing bursts. The chord bed is the best-fitting harmonic of the busier brain in each of three bands. A song number brings back the same thresholds, band and opening; the waves themselves are stochastic, so no two plays are the same.
Singing. With the mic on, your voice's pitch is detected and a wave is sparked at that height in both brains, sized by how loud you sing (a soft note sparks a smaller wave). They answer in kind — their wiring turns your note into a phrase.
Stimuli. The five buttons drive the real receptor neurons — olfactory, visual, gustatory, auditory and pheromone — as a burst, and you hear the brain's response as music. Which neurons are receptors, and which are sex-specific, sexually dimorphic or express fruitless/doublesex, comes cell-for-cell from the datasets' own annotations. The two flies' receptor sets differ (for taste, her sugar and water cells against all but his bitter ones), so the same button is not an identical input. After a stimulus the readout says where the response went and how much reached the cells each dataset labels by sex; only her labels are complete (his data marks 190 fruitless/doublesex cells and none sex-specific or dimorphic), so the two brains' shares are not comparable yet.
The whole brain. With WebGPU, the full brain button loads every neuron with a position and every published connection — 136,984 ♀ and 143,768 ♂ neurons, 3.7 and 5.2 million connections, about 28 MB — and runs the wave and the drawing on your GPU, with the 24 harmonics of the whole graph (computed offline, like every brain's). The sample is the default because it runs anywhere.
The worm couple. The 🪱 worms button swaps in C. elegans: the hermaphrodite (♀ her, 300 neurons) and the male (♂ him, 360 neurons), every neuron-to-neuron synapse and gap junction from Cook et al.'s 2019 whole-animal connectomes of both sexes, with the same instrument, stimuli that drive its real olfactory, light-sensing, salt-tasting, touch and pheromone neurons, and its own 24 harmonics. Neuron positions are OpenWorm's; the 89 male-specific neurons (rays, spicule, CEM, MCM, CA/CP…) have no published coordinates and are placed by anatomy — head, ventral cord or tail — which is invented; the male pharynx, not wired in the male dataset, is left out.
The larva and the adult. The 🐛 larva button loads the whole brain of a first-instar Drosophila larva — all 2,952 neurons and 110,140 connections (351,828 synapses) Winding et al. mapped in 2023, the first complete insect brain — and puts it on one beat grid with the adult sample: one species, two ages, the same stimuli (smell, light, taste, touch, heat) driving each one's real sensory neurons. Signs come from the transmitter annotations in the reconstruction's CATMAID project (78 GABA/glutamate neurons); positions are the reconstructions' own root coordinates (somas; nerve-entry points for the sensory axons), except the 74 neurons rooted in the nerve cord, which sit at their brain partners' centroid.
A human brain. The 🧠 human button loads a human structural connectome: 400 cortical parcels and 14 subcortical regions, wired by diffusion-MRI tractography averaged over 207 people (Human Connectome Project, via the ENIGMA Toolbox), placed at their MNI centroids. This is the kind of graph connectome harmonics were first computed on (Atasoy et al. 2016). It is regions, not neurons, and tractography, not synapses; the same instrument plays it with the fly beside it, and its signals travel with the worm's conduction delays, slowed tens of times.
A mouse's cortex. The 🐭 mouse tier is a sample of the MICrONS cubic millimeter of mouse visual cortex (The MICrONS Consortium, Nature 640, 2025; CC BY 4.0): 29,727 neurons chosen by identity — every proofread cell first, then each cell class in its natural proportion — from V1 and the higher areas RL, AL and LM, wired by the strongest 322,698 of the 1.88 million connections the electron-microscope reconstruction found between them (1.16 of their 2.95 million synapses; pair weight saturating at 20 synapses), interneurons signed negative from the cell-type table, cell classes from the MICrONS metamodel (layer 2/3, 4, 5 and 6 pyramidal cells; basket, bipolar, Martinotti and neurogliaform interneurons). The stimuli drive the cortical layers: layer 4 is where the thalamus delivers what the eye saw. Two honest caveats: half of the neuron-to-neuron pairs have an inhibitory source because the automated reconstruction follows local interneuron axons far more completely than pyramidal ones (in vivo it is nearer a fifth), and the cortex here is a mammal's, so its harmonics are of a column of gray matter, not a whole brain — it plays on one beat with the fly for scale.
The spiking model. On the whole brain, the spiking model toggle swaps the wave for the leaky integrate-and-fire model Shiu et al. built on the female's wiring (Nature, 2024; here it also runs on the male's): rest −52 mV, threshold −45 mV, membrane time constant 20 ms, an exponential synapse (5 ms) fed 0.275 mV per synapse (inhibitory negative; counts above 50 are capped by the shipped weights), 2 ms refractory (2.2 in the paper), 1 ms steps, sixteen of them per engine step (the engine steps every third frame). Nothing fires on its own; hold a stimulus and its receptors spike at 150 Hz, and the readout names the circuits the real synapses recruit. Their model reproduces known behavior (sugar to feeding motor neurons, for instance); this is that model's arithmetic, not its full validation, with synapse counts recovered from the shipped log-scaled weights (the modulatory ones, which the wave stores at half weight, are doubled back first).
Find a cell type. Type a cell type or class — Kenyon cell, DNa02, LC10, olfactory projection neuron, or a word like clock or compass — and ignite them lights every neuron of that kind as one voice (in the spiking model it drives them at 150 Hz for 1.5 s). A second and a half later the readout names the cell types that answered and the neuropils the answer reached. A tap names the neuron it hit and, on the female brain, links it to its FlyWire Codex page (male neurons show their MaleCNS body id).
The look. The LOOK row is the visual layer, every piece a toggle: hot front & embers (a neuron that fired this step burns white, cools to its voice's color, then glows as an ember for its refractory period, so a wave has a leading edge and a scorch of where it has been), pulses (a bright dot runs along each synapse thread from the neuron that fired to the one it fired: the signal in flight), flow (at rest the strongest connections stream in the direction their signals travel), form (crowded interiors dimmed by local density, so the neuropils read as organs), film (vignette, grain, a filmic grade, a faint edge chromatic aberration), light shafts (a radial blur of the bloom from each live wave), breathing (at rest each harmonic displaces the neurons along their radius), trails (a fraction of the previous frame kept), focus (a fake depth of field), beads (the ribbons as chains of glowing beads with a bright head). A tap also lights the tapped neuron's own outgoing connections first. With the cinematic camera a stimulus pushes the camera toward the receptors and pulls back; left alone for forty-five seconds the camera starts to wander. Zoom has no floor: pinch or scroll zooms toward what is under your fingers or cursor, through the brain's surface and into its wiring, then flies on through it in the direction you point; drag looks around, two fingers (or shift-drag) slide, ⟲ view or R comes back out, and inside (I) flies straight to the middle of the brain facing down its length. Inside, the lens widens, the fog closes in, the glow and the lit wiring thin out and the organ's surface fades, so the waves stay readable around you. ✈ follow (F) is a camera drone on the signals: it finds the densest knot of firing in a wave, keeps it in the middle of the frame and trails it as it travels through the wiring, glides to a stronger knot when one flares up and moves on to the next wave when this one fades (and starts Play if nothing is traveling); drag to swing round it, scroll to come closer or stand back, tap to spark a wave for it to follow. On the worms, the larva and the human, whose cells are spread thin, it follows from a little way off and pulls back to the whole brain between waves. ⛶ theater (T) is the brain alone, full screen: touch the screen for a small bar with the view reset, inside, follow, the five stimuli and play; Esc or ✕ brings the controls back. Every species has its own resting palette. Phones start with the light shafts and the breathing off.
In VR. On a headset browser that supports WebXR, a VR button appears in the MIX row (experimental: not yet tried on a headset): the two brains hang in the room at 12% of their screen size, a controller's trigger sparks the neuron it points at (or fires a stimulus on empty space), and the binaural voices follow your head. The whole-brain tier renders through WebGPU, which has no VR path yet, so this works on the sample, the worms, the larva, the human and the mouse.
The gallery. save this song keeps a song — its seed, its band of instruments and its genre — in a shared gallery kept by the site's Worker, one list per kind of brain (the fly sample, the whole fly brain, the worms, the larva, the human, the mouse); songs people saved lists them — the most played, then the newest — and picking one plays it here. Titles are plain text, capped at 60 characters, one save per 20 seconds per address, 500 songs per list.
The internet jam. Join it and every tap and stimulus you fire is shared live with everyone else in the room; theirs arrive as violet rings and spark your brains too. It is one room for the whole internet, running on a small stateful worker at the edge — playful, not precise. Taps are shared as positions in the brain, so someone on the sample and someone on the whole brain spark the same spot; players on a different kind of brain don't reach each other.
What's invented. Positions (bar the placements named above), connections with their predicted signs, and each graph's modes are data. The sign of each mode is arbitrary; amplitudes are clipped at the 98th percentile. Firing thresholds, refractory time, the weight given to inhibition, the half weight given to the modulatory transmitters, and every pitch and rhythm mapping are musical choices. This is a way to hear two wirings differ — not a physiological simulation. The brains do not vibrate.