How much of a fly is in its connectome · Lobeworks/17

A dissection of the fruit-fly connectome read as a problem of observation. It follows one brain from the fixative to the simulation: the chemistry that keeps its structure and erases its state, the diamond knife and the ion beam that cut it, the electron optics and cameras that photograph it, the software that sews twenty-one million images into one volume, the networks that trace each neuron and guess the sign of each synapse, the cell types that turn a graph into an index, and the leaky integrate-and-fire model that runs it. At each step it separates what is measured, what is inferred, what is borrowed from other experiments and what is absent, sets out the words that separate a model from an emulation, and reads the virtual flies of 2026 against them. It stops at the fly, with the human brain in the margin.

How much of a fly is in its connectome. How a fruit fly's brain is fixed, cut, photographed, traced and run as a model, and what of the living animal each step keeps

In March 2026 a company called Eon Systems posted a video of a fruit fly that had never been alive. It stood on a patch of ground inside a physics engine, walked on six legs in the alternating tripods real flies use, stopped to rub its forelegs over its head, and lowered its proboscis to something sweet, and the post that came with it was titled We've uploaded a fruit fly1. Six months later the wiring of a second fly, a male, with the nerve cord that runs its legs still joined to its brain, was published under an open licence on a Thursday, and on the Friday a graduate student at Georgia Tech had it steering a fly through Minecraft; within days other people had it playing Doom and Super Mario 642. Both videos stand on the same object, a connectome, the map of every neuron in a nervous system and every chemical synapse between them, and both invite the same reading: the map is complete, so the animal must be in it. The question this dissection opens is how much of a fly a connectome actually holds, and it is a question my own trade has a word for. In control engineering a system is observable when its internal state can be worked out from what its sensors record over time, and the useful form of the idea is narrower and harder: given these measurements, which parts of the system can be recovered, which can only be guessed with some confidence, and which no computation will ever bring back. A connectome is the most complete measurement ever made of a brain's structure, and it is taken by a method that destroys almost everything else. Each section below is one stage of how it is made or used, from the fixative to the simulation.

1Eon Systems, We've uploaded a fruit fly, March 2026. The post lists the ingredients as the wiring graph, synapse counts as weights, a map of which neurons excite and which inhibit, and leaky integrate-and-fire neurons, and says the body was not scanned. 2Janelia Research Campus, Male CNS connectome, released under CC BY on 3 September 2026; Gizmodo, Google mapped a fruit fly's brain. Now it's playing Doom and Super Mario 64, September 2026.

What the fixative keeps

A brain holds at least three kinds of information, and they behave very differently under a microscope. The first is structure: which neuron touches which, through how many synapses, along what path. The second is the set of parameters that turn that structure into behaviour: how strongly each synapse pushes its target, which receptors sit on the far side of it, how fast each cell forgets its input, where its threshold lies. The third is state: what the cells are doing at a given moment, how much of each modulating chemical is in the tissue, whether the animal is hungry, asleep or has just been frightened, and which of its synapses were changed by the last hour of its life.

Layer What it is What a connectome records

Structure Cells, their shapes, the contacts between them and how many All of it, at a few nanometres

Parameters Synaptic strength, receptors, time constants, thresholds Indirect traces at best: synapse counts, the look of a vesicle

State Electrical activity, modulators, hunger, recent learning Nothing

The table is set by chemistry. A fly brain is dissected out of the head in a few minutes and dropped into glutaraldehyde, the fixative, a small molecule with an aldehyde group at each end that hooks one protein to its neighbour wherever it touches, until every protein in the cell is welded into one cross-linked mesh1. The pumps that hold sodium out and potassium in stop with everything else, and the voltage across every membrane, which is where a neuron keeps its present tense, runs down to nothing. Osmium tetroxide then binds the fatty chains of the membranes and leaves a heavy metal atom wherever there was a lipid, so that under the electron beam every membrane will show as a dark line and the inside of every cell as pale ground. Finally the water is replaced by an epoxy resin cured in an oven, and what had been a soft organ a quarter of a millimetre deep becomes a speck of amber plastic hard enough to be cut like a block of wood. Each step keeps the shape of things, and none can keep what they were doing, because doing is a flow of ions and molecules and the first step stops every flow.

1Zheng et al., A complete electron microscopy volume of the brain of adult Drosophila melanogaster, Cell 174, 2018, which fixed the brain in 2% glutaraldehyde, post-fixed it in 1% osmium tetroxide, stained it en bloc with uranyl acetate and embedded it in EmBed 812 resin.

The destruction is the price of the resolution, and nothing yet avoids it: freezing keeps one instant at best, and light microscopy can watch living neurons but cannot resolve a synapse. The route the field has taken is to measure the same animal twice: record what its neurons do while it is alive, then fix it and map its wiring. MICrONS did this for a cubic millimetre of mouse visual cortex, recording the responses of about 75,000 neurons with calcium imaging before the tissue was cut1, so that activity and structure belong to one individual. It still records only a few hours of state, and none of the parameters directly.

1MICrONS Consortium, Functional connectomics spanning multiple areas of mouse visual cortex, Nature 640, 2025.

The comparison that helps me most here is a forensic one. Imagine an engineer handed a computer that was switched off and then cast in resin, with the instruction to work out what it was running. Sectioned finely enough, the board gives up every trace and every pin, and the engineer can draw the complete schematic. The contents of memory and the registers of the processor at the moment of shutdown are gone, and so is the configuration of each chip, which on this board was set by the chemistry of each part rather than printed on it. The likeness fails at one point, in the computer's favour. A computer's state is just as physical as a brain's (charge sitting in transistors), and an engineer can read it out before cutting the power, because the machine was designed to be dumped; a brain keeps its state in the same chemistry that has to be stopped to see it. The likeness also points at the three ways around the loss that the field actually uses, and they recur below: read the state before fixing, as MICrONS did; borrow the parameters from other animals measured alive, which works only if individuals are built alike; and infer what is missing from what the structure itself shows, which is what the networks of the later sections do. What can be known about a fixed brain is what its structure records, plus what other experiments can lend it, and every claim about a simulated fly should say which of the two it rests on.

Forty nanometres at a time

The plastic block has to become images, and the electrons of a transmission electron microscope only pass through material a few tens of nanometres thick. For the fly brain of 2018, the full adult brain of one seven-day-old female imaged at Janelia and called FAFB, the block was cut into 7,062 serial sections about 40 nm thick, some two thousand times thinner than a human hair1. The machine that does this is an ultramicrotome. The block is clamped on an arm that swings it down past the edge of a diamond knife, and between strokes the arm advances by the thickness of the next section, a step made by heating a metal rod so that its own expansion pushes the block forward, or by a piezoelectric crystal that lengthens under a voltage. Behind the edge the knife carries a small trough of water, and each section slides off the diamond onto it and floats there, still joined to the one before, so the cut brain leaves the knife as a ribbon of grey plastic whose interference colour (grey, then silver, then gold as sections thicken) tells the operator how thick each piece is.

1Zheng et al., 2018, cited above.

A ribbon of serial sections floating on the water of an ultramicrotome's knife boat, each section joined to the one cut before it. Photograph by euphras, 2004, CC BY-SA 3.0, Wikimedia Commons.

From the water the sections are lifted onto grids, metal discs about three millimetres across with a slot covered by a plastic film on which the sections rest. FAFB used about 2,400 of them, three sections each, every grid marked with a barcode so that the order of the brain survives being handled. Picking sections off water by hand is where a cut brain most easily loses a slice, and a later system, GridTape, replaced the grids with a long slotted tape that collects more than 4,000 sections a day, over ten times the rate by hand1.

1Phelps et al., Reconstruction of motor control circuits in adult Drosophila using automated transmission electron microscopy, Cell 184, 2021.

Slot grids from the FAFB brain: a grid with its support film carrying sections, and a custom grid with barcodes and a serial number so that the order of the brain survives being handled. Adapted from Zheng et al., Cell, 2018, figure S3, panels D and H, CC BY 4.0.

The other way to slice a brain is never to cut a section at all. In a FIB-SEM, borrowed from the semiconductor industry where it is used to open chips, a focused beam of gallium ions sweeps the face of the block and blasts away a few nanometres of material, like a sandblaster working one atom deep, and a scanning electron microscope photographs the fresh face; the cycle repeats tens of thousands of times, and the block is consumed as it is imaged. The milling can be finer than any knife, so the voxels (the three-dimensional pixels of a volume) are 8 nm on every side. Past about a hundred micrometres of depth the milling streaks and the images degrade1, so the brain is first cut by a second, heated knife into slabs 20 µm thick, each milled on a different machine and stitched back together at the end. The hemibrain of 2020 was built this way2, and so was the male central nervous system of 2026, imaged by seven of these microscopes running for thirteen months3.

1Hayworth et al., Ultrastructurally smooth thick partitioning and volume stitching for large-scale connectomics, Nature Methods 12, 2015. 2Scheffer et al., A connectome and analysis of the adult Drosophila central brain, eLife 9, 2020. 3Berg et al., Sexual dimorphism in the complete connectome of the Drosophila male central nervous system, bioRxiv, October 2025, published in Cell in September 2026.

The choice between the two methods is a choice of voxel shape. The transmission images of FAFB have pixels of 4 by 4 nm, finer than FIB-SEM in the plane of the section, but the third dimension is the thickness of the section, 40 nm, so each voxel is a brick ten times taller than it is wide. A neurite that runs straight through the sections is drawn crisply by either method. One that runs almost parallel to the cut is smeared, because each section averages it over 40 nm of depth, and the fine branches of an insect brain, often narrower than 100 nm, run in every direction.

Serial-section TEMpixels 4 nm across, sections 40 nm deepFIB-SEMvoxels 8 nm on every side100 nmdashed: where the two neurites really are The same two neurites, 60 nm thick, sampled by each method. Each voxel is shaded by the share of it the neurites fill. The one crossing the sections square is crisp in both; the one running at 12° to the cut is smeared across several bricks in the TEM volume and stays a clean tube in the FIB-SEM one.

The same block of fly neuropil, 600 nm on a side, with cubic voxels of 4 nm (top) and with the 4 by 4 by 40 nm voxels of serial-section TEM (bottom). Seen in the plane of the section the two agree; seen from the side, the bricks blur into stripes. The scale bar is 100 nm. Adapted from Xu et al., eLife, 2017, figure 2, CC BY 4.0.

The instrument decides which errors the data will contain before anyone looks at it, and that will hold at every stage that follows.

Why it takes electrons

To resolve two points is to see them as two rather than one blur, and every instrument that forms an image with waves has a smallest separation it can resolve, set by the wavelength. Ernst Abbe wrote the limit down in 1873: about the wavelength divided by twice the numerical aperture, a number for the angle of light a lens collects that reaches about 1.4 for the best lenses. Green light, at 550 nm, gives about 200 nm. A wave cannot carry detail finer than its own ripples, in the same way that ocean swells pass a wooden post and close behind it as if it were not there, while they break on a breakwater as wide as they are long. The parts of a brain a connectome needs are mostly finer than that: the thinnest processes of a fly neuron, the synaptic cleft (the gap of about 20 nm across which one cell signals another) and the vesicles of about 40 nm that carry the signal across it.

Electrons solve this because they are also waves, as Louis de Broglie argued in 1924, with a wavelength that shrinks as their momentum grows. An electron accelerated through 120,000 volts, the voltage of the microscopes that imaged FAFB, has a wavelength of about 3.35 picometres, some 160,000 times shorter than green light1. No glass can bend such a beam, so the lenses are coils of wire whose magnetic field curves the electrons' paths toward a focus, and their aberrations hold the resolution of a biological microscope to around a nanometre. For a connectome that is ample.

1The wavelength follows from de Broglie's relation with the relativistic correction for the electron's speed at 120 kV, about 59% of the speed of light.

The microscope that photographed FAFB was a standard one rebuilt for throughput. Electrons that pass through a section are scattered where they meet the osmium of a membrane and pass straight through pale cytoplasm, and below the section they strike a scintillator, a screen that flashes with visible light where each electron lands. In TEMCA2 that glowing screen was filmed by four scientific CMOS cameras at once, each taking a quarter of the field, about 50 megapixels a second, forty times a conventional microscope, while an automated stage swapped grids by itself so imaging could run for days1. The camera at the end is a cousin of the sensor in a phone. The whole brain came out as about 21 million images and 106 TB on disk.

1Zheng et al., 2018, cited above.

The camera array of TEMCA2: below the specimen, a scintillator turns the electron image into light, and four cameras each film one quarter of it at once. Adapted from Zheng et al., Cell, 2018, figure S3, panel E, CC BY 4.0.

What one of those images contains is a field of grey. A section through the neuropil of a fly (the dense tangle where neurons make their contacts) looks like the cut end of a bundle of spaghetti of a hundred different thicknesses, each strand outlined in dark membrane, some pale, some filled with the small circles of vesicles, some with the dark sausages of mitochondria.

Fly optic lobe neuropil imaged by FIB-SEM: hundreds of neurites cut across, outlined in dark membrane, with red arrows on synapses; the field is about 7 µm across. Adapted from Xu et al., eLife, 2017, figure 3c, CC BY 4.0.

Where a synapse is, the presynaptic side carries a dark density shaped like a T, with vesicles clustered around it, and across the cleft sit several receiving profiles at once. Fly synapses are polyadic: in the hemibrain there were about 9.5 million of these T-bars and 64 million postsynaptic densities1, close to seven partners per release site. One packet of transmitter is heard by several cells at once, a broadcast at the smallest scale that a model built on pairs of neurons quietly splits into independent wires.

1Scheffer et al., 2020, cited above.

Polyadic synapses in the hemibrain: each red arrow marks a presynaptic T-bar, and the white triangles around it mark the several postsynaptic cells that face it. SVs are synaptic vesicles, M mitochondria; the scale bar is 0.5 µm. Adapted from Scheffer et al., eLife, 2020, figure 4A, CC BY 4.0.

What the image does not contain matters as much. There is no colour, no name of any cell, no trace of activity, and no direct sign of which molecule a synapse releases or which receptor waits on the other side. Gap junctions, the channels that join two neurons' interiors directly and pass current without any transmitter, are a few nanometres across and are not identified reliably at this resolution, so the fly connectomes contain chemical synapses only.

Twenty-one million photographs

Electron microscopy has photographed cells since the 1950s, and for most of that time a study meant a handful of sections examined by a person who knew what a mitochondrion looks like. One section of a fly brain needs thousands of overlapping tiles, and the brain needs seven thousand sections. Every tool in the chain exists because the step before it produced more data than the next step could handle by hand: grids gave way to tape, one camera to four, tracing by eye to networks, a single proofreader to a versioned database. Within each section the tiles are stitched like a phone's panorama; across sections each slice was cut, floated, stretched and laid down at some angle, so bringing one onto the next takes a field of small local warps. Janelia's software for this, Render, keeps the original tiles once and beside them the transformations each tile needs, and draws the aligned image only when asked1. An error of a few tens of nanometres here joins two neurites that never touched, and every later stage inherits it.

1Mahalingam et al., A scalable and modular automated pipeline for stitching of large electron microscopy datasets, eLife 11, 2022, describes a pipeline built on Render and the reasons for keeping transformations apart from pixels.

The scale of these numbers is an argument in its own right, and it cuts both ways. H01, the largest piece of human brain reconstructed so far, is one cubic millimetre of temporal cortex removed during surgery for epilepsy to reach a deeper lesion: 5,019 sections imaged over 326 days, 1.4 petabytes, about 57,000 cells and 150 million synapses1. A human brain is about 1.2 million cubic millimetres. At H01's rate that is about 1.7 zettabytes of raw images and about a million years of one microscope's time.

1Shapson-Coe et al., A petavoxel fragment of human cerebral cortex reconstructed at nanoscale resolution, Science 384, 2024.

Neurons in the nervous system10²10³10⁴10⁵10⁶10⁷10⁸10⁹10¹⁰10¹¹worm, 1986fly larva, 2023fly brain, 2024fly brain and nerve cord, 2026mouse brainhuman brainRaw images of one reconstruction1 TB1 PB1 EB1 ZBfly brain, TEM, 106 TBH01, 1 mm³ of human cortex, 1.4 PBa mouse brain at that resolutiona human brainfilled: measured · hollow: H01’s bytes per mm³ times the volume of the organ Neurons per nervous system and raw image data per reconstruction, both on logarithmic axes. The hollow dots multiply H01's 1.4 PB per cubic millimetre by the volume of a mouse brain (about half a cubic centimetre) and of a human brain (about 1.2 litres).

I read those numbers in two ways at once. They put a human connectome out of reach of any machine that exists, and they say the obstacle is throughput, the kind engineering has worn down before: a whole mouse brain, some 2,400 times smaller than a human one, is the stated target of a programme the NIH began funding in 20231. The manifesto's missing Common Crawl of neurons is a shortage of recordings from living people; this is a shortage of storage, imaging time and human attention, and it scales differently. The human brain adds obstacles a fly does not have: the fixative has to soak through an organ of 1.2 kilograms before its middle decays, which is why every piece of human cortex imaged at this resolution so far was cut from a living patient in surgery and fixed at once, and a brain has to come from someone who gave it.

1NINDS, NIH BRAIN Initiative launches projects to develop innovative technologies to map the brain in incredible detail, 2023.

Following one neurite through seven thousand sections

An aligned volume is still only grey. To become a connectome, every voxel has to be given the identity of the cell it belongs to, a job called segmentation, which in the picture of the spaghetti means colouring each strand its own colour along its whole length through thousands of sections.

The first person to do it for a whole nervous system was Eileen Southgate. She had been a technician at the Laboratory of Molecular Biology in Cambridge since 1956, and had worked on the structure of haemoglobin with Max Perutz and John Kendrew, when she joined John White's reconstruction of the worm C. elegans, begun in 19691. She took the micrographs of the sections Nichol Thomson had cut, printing them at twelve by sixteen inches, and following each neuron from one print to the next with Rotring pens, drawing its outline and giving it a number so that she could find it again on the next sheet2. The result, published in 1986 with her name second on it, filled 340 pages: 302 neurons in 118 classes, some 5,000 chemical synapses and 600 gap junctions, which the worm's micrographs did show3. The errors the authors listed are the errors of every connectome since: in long bundles of featureless processes it was easy to jump to a neighbour, and processes running along the cut were hard to follow. Thirty-eight years separate her 302 neurons from FlyWire's 139,255, and the idea did not change in between. What changed is that every link of her work was handed to a machine.

1MRC Laboratory of Molecular Biology, The Eileen Southgate Staff Prize. 2Emmons, The beginning of connectomics: a commentary on White et al. (1986), Philosophical Transactions of the Royal Society B 370, 2015. 3White et al., The structure of the nervous system of the nematode Caenorhabditis elegans, Philosophical Transactions of the Royal Society B 314, 1986.

The automated version faces two errors, and they do not cost the same. A split leaves one neuron in two pieces, cheap to see and cheap to repair. A merge joins two neurons into one; every synapse of one is then credited to the other, and the graph is wired wrong in a way that can look plausible. The classical pipeline marked each pixel as membrane or not and then grouped the regions enclosed by membrane, flooding outward the way water fills the basins of a landscape, and a membrane that failed to take up the stain for a few nanometres is a hole in a dam: the flood pours through it from one cell into the next.

The method that changed this came from Google in 2018, which is why the company's name is attached to two of the fly volumes. A flood-filling network starts where a paint bucket in an image editor starts, from one seed point inside one cell, and grows outward. In the paint bucket the rule for growing is a fixed threshold on colour. Here it is a three-dimensional convolutional network that looks at a small cube of the image and, as a second input, at its own current guess of which voxels in that cube belong to the cell it is tracing; it returns an improved guess, which is fed back to it, and when the guess reaches a face of the cube with high confidence the cube moves that way. The second input is what makes it work. Asked whether a voxel belongs to the object whose shape it has been following, the network treats a gap in a membrane the way a walker treats a gap in a wall beside the path.

Boundaries first, then groupingone gap in a membrane, and two cells become oneFlood fillingthe network keeps to the object it is tracingunstained membrane1234567dot: the seed · squares: the field of view, moved on where the object reaches its edge Two neurites touch at one point where a stretch of membrane failed to stain. Grouping the regions between detected boundaries joins them through the gap. A flood-filling network, seeded in one, moves its field of view along the object it is following and leaves the neighbour alone. Drawn by hand to show the mechanism.

On its first test, a volume of zebra finch brain, it traced an average of 1.1 mm of neurite before the first error and made four merges in 97 mm of test path, an order of magnitude better than earlier methods and much more expensive, because each object is filled on its own with fields of view that pass over the same voxels many times1. Google segmented the hemibrain and the male nervous system this way; the female brain of FlyWire went through a different pipeline, from Sebastian Seung's group at Princeton2. The metric Google chose carries the theme of this whole piece: the distance a reader can follow a neuron before the data starts to lie.

1Januszewski et al., High-precision automated reconstruction of neurons with flood-filling networks, Nature Methods 15, 2018. 2Dorkenwald et al., FlyWire: online community for whole-brain connectomics, Nature Methods 19, 2022.

Segmentation in one section of fly medulla: the raw image, and the same image with each neuron filled in its own colour. Adapted from Sheridan et al., Nature Methods, 2023, figure 3o, CC BY 4.0.

Every pipeline still needed people. Correcting a machine's segmentation is called proofreading, and it took about 33 person-years for FlyWire and an estimated 44 for the male nervous system1. A proofreader does not inspect voxels: in a viewer that shows the sections beside the three-dimensional shape of each cell, they look for anatomy that cannot exist. A neuron with two cell bodies is almost always a merge, and a branch that stops dead in the middle of tissue is almost always a split; then they drop into the sections at that spot and cut or join. The edits are kept the way version control keeps a codebase, so an analysis can be rerun on the brain as it stood on any day2. The female brain that came out of it has 139,255 neurons and about 54.5 million synapses, and it stops at the neck: 1,303 descending neurons carry its commands toward the legs, and none of their targets are in the map.

1Dorkenwald et al., Neuronal wiring diagram of an adult brain, Nature 634, 2024; Berg et al., 2025, cited above. 2Dorkenwald et al., CAVE: Connectome Annotation Versioning Engine, Nature Methods 22, 2025.

The sign of every connection

A wiring diagram says which cell can influence which, and not in which direction, and a model cannot run without that. The message across a chemical synapse is a molecule, a neurotransmitter released from vesicles into the cleft, and its effect is decided on the far side by the receptor that catches it. If the receptor is a channel that lets sodium or calcium in, the receiving cell moves toward its threshold, which is excitation; if it lets chloride in or potassium out, the cell is held down, which is inhibition. Potassium leaving a cell is how a neuron ends its own spike; the signal between cells is the transmitter. So the sign belongs to the pair of molecule and receptor, and one molecule can excite at one synapse and inhibit at another.

Papers still speak of excitatory and inhibitory neurons, and the habit rests on Dale's principle: a neuron releases the same transmitter, or the same mixture, at all of its synapses, so a cell can be labelled by its transmitter, and in the fly a rough rule turns the label into a sign. Acetylcholine, the main fast transmitter of insects, excites; GABA inhibits; and glutamate, which excites in most of the vertebrate brain, inhibits in the fly's olfactory centre, where it opens a chloride channel1. Whole-brain models apply that rule everywhere, a useful approximation of a fact shown in particular circuits.

1Liu and Wilson, Glutamate is an inhibitory neurotransmitter in the Drosophila olfactory system, PNAS 110, 2013.

The transmitter was known, before 2024, for about 700 of the roughly 7,000 cell types of the fly's central brain. The rest were filled in by a network that learned to read it from the picture of the synapse itself1: shown a small cube of micrograph around a synapse, it chose among six transmitters and was right for 87% of single synapses and 94% of neurons, once each neuron's synapses voted. It is a convolutional network because that kind slides the same small filters over the whole image, which builds in two facts true of micrographs: what tells transmitters apart lies in a small neighbourhood (the size and texture of the vesicles), and a vesicle looks like a vesicle wherever it sits in the frame. A connectome's sign is a prediction stacked on an approximation, and every model downstream inherits both.

1Eckstein et al., Neurotransmitter classification from electron microscopy images at synaptic sites in Drosophila melanogaster, Cell 187, 2024.

One fly for all flies

Every neuron of the FlyWire brain, each in its own colour, seen from the front. Adapted from Dorkenwald et al., Nature, 2024, figure 1a, CC BY 4.0; the image is mirrored relative to the brain itself.

A graph of 139,255 nodes is unreadable as it stands, and FlyWire's companion paper sorted its neurons into 8,453 cell types, sets of neurons with the same shape, position and partners1. It would be easy to read the typing as bookkeeping, and I think it is the step that makes the rest usable. Types make the graph readable, and they match one animal to another: 3,643 of FlyWire's had first been proposed in the hemibrain, a different fly imaged by a different method, much as a function can be recognised in two builds of the same program after every memory address has moved. And types open the door to experiments, because fly geneticists keep lines of flies in which one cell type, and only that type, can be made to glow, switched on with light or silenced. The name of a type is the key that joins the map to the laboratory.

1Schlegel et al., Whole-brain annotation and multi-connectome cell typing of Drosophila, Nature 634, 2024.

The comparison between the two flies also answered whether one fly can stand for all of them. Within one brain, the number of cells of a type differed between the two sides by less than one on average, and any connection carried by more than ten synapses was found again, more than nine times in ten, in the other hemispheres compared1. Insect brains are built from a genetic programme that repeats itself closely from animal to animal, and that stereotypy is, to my mind, what pays for the whole enterprise: a map of one fly is a map of the species, to the precision of its strong connections. It is also where the human case diverges most: a human cortex is shaped far more by what happens to its owner, so a generic human map would say much less about anyone than the fly's says about any fly.

1Schlegel et al., 2024, cited above.

An index without the datasheet

Three words name different objects here. The connectome is the graph of single neurons and their synapses. A projectome is the coarser map of which regions send fibres to which, derivable from a connectome by summing over its cells. The effectome is something else in kind: a causal map of how much activating one neuron changes another, and with what sign, which a group at Princeton has begun to estimate in the fly from experiments that switch neurons on with light, using the connectome as a starting guess1. Two neurons joined by fifty synapses may barely affect each other, and two with no contact may be strongly coupled through others. To someone who writes software the difference is familiar: the connectome is a call graph, every function with the functions it calls; the effectome is what a profiler with causal tracing would report, how much each call changes the result.

1Pospisil et al., The fly connectome reveals a path to the effectome, Nature 634, 2024.

The comparison that fits best, for me, is a board of unknown purpose with every pin labelled. Take a circuit board on which every chip, pin and trace has been mapped, with a part number on each chip so that two chips of the same model can be recognised, and a guess, for each output pin, of whether it pulls the line high or low. That is roughly what a fly connectome with types and transmitters provides. What it lacks is the datasheet. For a fly cell type it would hold the stimulus it responds to, its timing, what the animal does when the type is switched on or off, the modulators it answers to and its row of the effectome, and each entry is written by an experiment on living flies, often by switching the type on with optogenetics. The connectome is the index of the fly's manual, and the manual is still being written one cell type at a time. For the human brain neither exists: there is no complete index, and the genetic access that lets a fly geneticist switch one cell type on is, for people, essentially closed.

A bucket with a hole in it

To turn the map into a moving model, each neuron needs a rule for how it responds to its inputs, and the rule used for the whole fly brain, the leaky integrate-and-fire neuron, is the oldest and simplest one that still produces spikes.

A leaky integrate-and-fire neuron reduces a cell to one number, the voltage across its membrane.

The membrane is a thin insulating film between two salt solutions, so it stores charge like a capacitor, and the charge stored is the voltage. Channels that are always slightly open let that charge seep away, which acts as a resistor, the leak. Synapses inject current, positive from excitatory ones and negative from inhibitory ones. The voltage therefore integrates its input, drifts back toward rest when the input stops, and, when it crosses a threshold, the cell emits a spike and the voltage is set back to a reset value. Think of a bucket with a hole in the bottom: pour slowly and the hole empties it as fast as it fills; pour hard and it reaches the brim, tips over, and starts empty again. Inhibition is a cup taking water out.

Written as an equation, the voltage VVV of each neuron obeys

τmdVdt=(Vrest−V)+R Isyn(t),and if V≥Vth: spike, then V←Vreset\htmlData{sym=1}{\tau_m} \frac{d\htmlData{sym=0}{V}}{dt} = (\htmlData{sym=2}{V_{\text{rest}}} - \htmlData{sym=0}{V}) + \htmlData{sym=3}{R\,I_{\text{syn}}(t)}, \qquad \text{and if } \htmlData{sym=0}{V} \ge \htmlData{sym=4}{V_{\text{th}}}: \text{ spike, then } \htmlData{sym=0}{V} \leftarrow \htmlData{sym=5}{V_{\text{reset}}}τm​dtdV​=(Vrest​−V)+RIsyn​(t),and if V≥Vth​: spike, then V←Vreset​

VVVthe voltage across the membrane τm\tau_mτm​the membrane time constant: how long the cell takes to forget an input VrestV_{\text{rest}}Vrest​the voltage the leak pulls back to R Isyn(t)R\,I_{\text{syn}}(t)RIsyn​(t)the synaptic input, a current times the membrane's resistance VthV_{\text{th}}Vth​the threshold at which the cell spikes VresetV_{\text{reset}}Vreset​where the voltage starts again after a spike

The threshold and the reset are a rule added by hand, because the equation has no spike in it; a real spike is a burst of opening and closing sodium and potassium channels that this model replaces with a line in the code. Philip Shiu and his colleagues at Berkeley gave every one of the 127,400 proofread neurons of FlyWire the same numbers: a time constant of 11 ms, rest and reset at −52 mV, threshold at −45 mV, a refractory pause of 2.2 ms1. The weight of each connection was its synapse count, positive if the sending cell was predicted to release acetylcholine and negative for GABA or glutamate, scaled by one free number chosen so that driving the sugar-sensing neurons at 100 spikes a second brought the motor neuron that extends the proboscis to about 80% of its top rate.

1Shiu et al., A Drosophila computational brain model reveals sensorimotor processing, Nature 634, 2024.

-55-50-45-40membrane voltage, mVthresholdrest and resetin050100150200time, mssparse input: the leak winsdense input: it fireswith inhibition: silent A leaky integrate-and-fire neuron with the parameters Shiu et al. gave every cell, driven by illustrative inputs drawn in the strip below: sparse excitation leaks away before it accumulates, a dense train carries it over threshold again and again, and the same train with inhibition mixed in leaves it silent.

Every neuron in that model is the same neuron. Everything that distinguishes one cell from another in Shiu's model comes from the connectome: whom it listens to, through how many synapses, and with what sign. That makes the model an experiment: whatever it gets right, the wiring got right, because nothing else was allowed to vary. It got a surprising amount right. Activating the sugar-sensing or water-sensing neurons in the simulation predicted which neurons downstream would respond and which would be needed to start feeding, and of 164 predictions that could be checked on real flies, 91% matched; of eleven cell types the model said would drive the proboscis, ten made real flies extend it when switched on with light. The model does nothing below the level of a whole spike: no vesicles released, no cycle of the SNARE proteins that fuse them with the membrane, no calcium. All of that is folded into one number per connection.

The authors listed what the model leaves out, and the list is the third layer of the first section seen from the other side: gap junctions, neurons that never spike, internal state, neuropeptides that act at a distance, with the warning that circuits run by neuromodulation would be modelled poorly. Two of those absences should be held apart: neuromodulation changes how a fixed circuit responds, for seconds to hours, and synaptic plasticity changes the circuit itself with use, which is the physical form of learning. Neither appears in a fixed connectome beyond the residue of past plasticity in a synapse count, and neither is in the model. Its neurons cannot learn, and they cannot be in a mood.

That is why the simulated fly cannot be hungry. Hunger in a fly is a state of the body that reaches the brain as chemistry: hormones rise or fall with the sugar in the blood, and dopamine released onto the sugar-sensing neurons of a starved fly makes them respond more strongly to the same sugar1. The wires that carry the taste to the proboscis are all in the connectome. The knob that sets their gain is chemical, it is turned by a body the connectome does not include, and the model has no knob, so a hungry fly and a full one, simulated, respond identically to the same drop of sugar. What the model shows is the circuit at one fixed setting of every knob it lacks, which is still worth having: it says where to look.

1Inagaki et al., Visualizing neuromodulation in vivo: TANGO-mapping of dopamine signaling reveals appetite control of sugar sensing, Cell 148, 2012.

The other answer to the missing parameters is to learn them. Janne Lappalainen and his colleagues took the wiring of the fly's motion-vision circuits, tiled it into a network of 45,669 neurons of 64 types, kept every synapse count and sign fixed, and let a training procedure of the kind used for deep networks choose the unknowns (one time constant and one resting voltage per cell type, one strength per pair of types) by asking the whole network to estimate motion in natural video1. Nothing in the training mentioned any recorded neuron, and the trained networks agreed with measurements from 26 independent studies. Machine learning enters fly models in three places: Shiu's model learns nothing, Lappalainen's learns what the wiring leaves open, and the virtual bodies of the next section learn to walk by reinforcement learning, because the brain model does not reach the legs.

1Lappalainen et al., Connectome-constrained networks predict neural activity across the fly visual system, Nature 634, 2024.

Where it comes from What it supplies How far to trust it

Measured in the volume Neurons, their shapes, synapses and their counts High, after proofreading

Inferred from the images Transmitter per neuron, cell type 94% for transmitters by neuron; types matched across flies

Assigned by a rule Sign from transmitter An approximation shown in particular circuits

Fitted or trained Time constants, scale of synaptic strength Only as good as the task or the behaviour fitted

Borrowed or assumed Thresholds and time constants equal for all cells Chosen for the experiment, never measured

Absent Gap junctions, modulators, peptides, plasticity, state, the body None

Simulation, emulation, upload

Most of the argument about these flies is about words, and the words have fairly precise meanings. A model is a mathematical description of a system that leaves things out on purpose. A connectome-constrained model is one whose architecture is fixed by a measured connectome, as Shiu's and Lappalainen's are. A simulation is a model run forward in time. A validated simulation is one whose predictions have been checked against measurements, and for the fly that has happened circuit by circuit (feeding, grooming, motion vision), never for the brain as a whole. An emulation, in the sense Sandberg and Bostrom gave the word in 2008, reproduces a system at the level of its causal parts closely enough that it responds as the original would to any input1. An upload is an emulation of one particular individual, with what that individual had learned. The fly has the first four and neither of the last two, and the distance between the fourth and the fifth is the whole of the third layer of the first section.

1Sandberg and Bostrom, Whole brain emulation: a roadmap, Future of Humanity Institute, technical report 2008-3.

What Eon built in March was a chain of existing parts and the glue between them, which its own post and a critique published days later describe the same way1.

1Zeleznikow-Johnston, No, we haven't uploaded a fly yet, March 2026.

1The FlyWire connectomefemale, brain only2Shiu's leaky integrate-and-fire model, run in Brian23Chosen sensory neurons driven, the firing of descending neurons read out4Those rates mapped to commands for walking, turning, grooming and feeding5NeuroMechFly's controllers executing the commandsa body in the MuJoCo physics engine

Each link is real work, only the last two are new, and the last one is where the body's movement comes from. NeuroMechFly is a model of the fly's body with its joints and muscles, whose legs are moved by controllers written or trained for it1. In a real fly the brain's descending neurons talk to the ventral nerve cord, the fly's equivalent of a spinal cord, where the rhythm of the legs is generated, and FlyWire's brain stops at the neck. The critique put it in one sentence: the connectome selects the behaviour and NeuroMechFly's controllers execute it. It also named what the work was, the first closed loop of connectome-constrained brain and body models, a description nobody in the field would have objected to. The post's figure of 91% behavioural accuracy is not defined in it; the published number closest to it is the share of Shiu's tested predictions that held. The September games are the same construction with a different brain: the most-watched one runs the male connectome as leaky integrate-and-fire neurons, turns events in the game into the activity of sensory neurons, and reads hand-chosen output neurons that select and modulate scripted movements of the fly's body2. Every fly in every video of 2026 is a real wiring diagram with uniform neurons choosing among movements that something else performs.

1Wang-Chen et al., NeuroMechFly v2: simulating embodied sensorimotor control in adult Drosophila, Nature Methods 21, 2024. 2Evan Sinclair Smith, NeuroCraft Fly, GitHub, September 2026, whose description of the pipeline says the readouts select and modulate scripted body programs.

Image volumeWiring diagramActivity modelBodyJoined up19862015201820202023202420252026Worm sectionsserial TEM, printedand traced by penC. elegans302 neurons, withelectrical synapsesOptic lobe partsFIB-SEM: 7 columns in2015, a lobe in 2019FAFBfemale brain, TEMCA24×4×40 nm, 106 TBHemibrain volumeFIB-SEM, 8 nm cubeshot-knife slabsHemibrainabout 25,000 neuronsof the central brainLarva3,016 neuronsserial TEMFlyWire139,255 neuronsbrain only, no cordShiu et al.whole brain as LIFnothing trainedNeuroMechFly v2body in MuJoCo, legsrun by controllersLappalainen et al.motion vision, timeconstants trainedMale CNS volumeFIB-SEM, 8 nm, sevenmachines, 13 monthsMaleCNS166,691 neurons, withcord and neck intactflybodya body that walksand flies, by RLEon, MarchFlyWire run as LIFover NeuroMechFlyGames, SeptemberMaleCNS run as LIF,scripted bodiesnew measurements of a nervous systembuilt from earlier worka line ends where the work was used The projects behind the simulated flies, by what each produced and the year it was first released. Solid boxes are new measurements of a nervous system; dashed ones were built from earlier work. Only the two columns on the left hold new data about a fly; the two integrations of 2026 added none.

Each column depends on the one before, the columns were filled by different teams, and nothing in the last column measured a fly. Why the games, and not Eon, became the larger wave is a question about attention rather than data, so my answer is a reading I trust only moderately. The male data is the first with the brain and the nerve cord joined through an intact neck (It basically lets us get from eyes to legs in one go, as Greg Jefferis of the MRC Laboratory of Molecular Biology put it1), it came under a licence that asks only for credit, every part of the chain already existed as public code, and AI coding assistants had cut the cost of joining them to a day. Half a year earlier the same integration had needed a company; in September it needed a student and a laptop. A dataset spreads when the cost of trying something with it collapses, well before its quality catches up with the claims made for it. One detail is easy to miss: the nerve cord that made the male data complete is the part the games did not use, since their bodies are scripted.

1Janelia Research Campus, Researchers reveal connectome of the male fruit fly central nervous system, October 2025.

A connectome holds a fly's structure completely, to a few nanometres, for one animal that stands well for its species; it holds the signs of its connections by prediction and its cell types by a classification that matches across animals; it holds the time constants and strengths of its cells only where an experiment or a training task has supplied them; and it holds nothing of its state, its modulators, its learning or its body. That is enough to make predictions about the feeding circuit that held nine times in ten, and too little to make a fly hungry. The virtual fly that walked in March was a real wiring diagram choosing among movements a body model knew how to make, and the test that would show a connectome doing more than choosing would be a model of the male's brain and nerve cord, with no hand-written controllers, that produces the stepping rhythm of a walking fly and loses it when the same synapses are shuffled among the same neurons.