An article about a "digital fly" crossed my morning feed — an engineer took the complete connectome of a fruit fly, built a neural network from it, placed it in a virtual arena, and tried to "talk" to it by training an external decoder to read hunger signals. Neat, but the article itself is just the tip of the iceberg. The real story starts forty years earlier and sounds almost like a joke.
The map of a worm's nervous system with 302 neurons was ready back in 1986 — and after that twenty-five years passed with no one able to make this brain work in a computer. But a fly with a hundred forty thousand neurons was "brought to life" within a couple of years of publishing its map. This asymmetry — small brain, big dead end versus huge brain and lightning breakthrough — is the puzzle worth digging into.
It all started with one person's stubbornness. Sydney Brenner chose the nematode Caenorhabditis elegans as a model organism in the 1960s precisely because its nervous system seemed manageable for complete description: just 302 neurons. In 1973 he formulated a program in the spirit of "what needs to be done is clear": slice the animal into ultra-thin sections, photograph everything with an electron microscope, and manually trace every branch of every neuron.
The work stretched over more than a decade. A technician named Nichol spent over twenty years supplying continuous series of sections and micrographs — a 2015 commentary calls this a "happy accident" and a person of "great skill." In 1986, John White and coauthors published in Philosophical Transactions the first-ever map of an animal's entire nervous system: a complete graph of 302 neurons and about ten thousand synapses. It was all formatted on a Modular I computer with 64 kilobytes of memory and a 22-megabyte disk — commentators write directly: "John and Sydney had the right idea, but they were ahead of their time by a couple of decades."
And here's where it gets really interesting. The map was ready, but there was nothing to use it with. Twenty-six years after publication — by the same commentary's count — only the development of computers made new research of this kind possible. The engine blueprint sat on a shelf while the assembly plant didn't exist.
Moving to the fly, the scale immediately breaks in the other direction. An adult fruit fly brain is less than a millimeter across but contains 139,255 neurons and 54.5 million synapses — a density of about 7.4 synapses per cubic micrometer, higher than in mammalian cortex. To capture all this, one female brain was cut into 7000 sections 40 nanometers thick each and run through an electron microscope at Janelia Research Campus. It yielded over 100 terabytes of images — "a hundred typical laptops" worth of memory.
Then comes the most non-trivial part. AI from Princeton segmented the images, but at this data size it makes too many mistakes, and the question of proofreading arose. The solution was almost Victorian in spirit: the FlyWire consortium assembled 76 laboratories, 287 researchers and volunteers who spent a combined 33 person-years on manual verification of the markup. Final accuracy — 99.2% F1 by volume on a control sample. On top of the graph, the consortium annotated over 8000 cell types — the first author of one of the papers compared it to Google Maps: the raw graph is a satellite image, and the annotations are street names, hours of operation, and reviews. Without them, the map is useless.
The October 2024 publication came out as a package of nine papers in Nature. And this was the first complete connectome of an animal brain that can walk and see — before this the record belonged to the fruit fly larva (3016 neurons, Science, March 2023) and that same 1986 worm.
Then events raced forward as if someone had switched on a conveyor belt. In October 2025, Janelia with Cambridge and Google Research published the complete connectome of the entire central nervous system of a male fruit fly — brain plus ventral nerve cord, over 166,000 neurons. Gerry Rubin called it "the first opportunity to compare both sexes of an animal with complex social behavior," and Greg Jefferis added a phrase worth quoting separately: "it allows us to go from eyes to legs in one go" — sensory and motor systems linked in one continuous graph for the first time. Google Research performed the segmentation, Janelia and Cambridge teams did the proofreading.
In parallel, the same method moved to mammals. The MICrONS project in April 2025 released a package of ten papers about a cubic millimeter of mouse visual cortex: over 200,000 cells, four kilometers of axons, 524 million synapses — all with recordings of neuronal activity while mice watch movies. Data volume — 1.6 petabytes, "22 years of continuous HD video." Tissue was cut by crews working twelve-hour shifts for twelve days straight, seven years of work, a hundred and fifty scientists.
The funding for MICrONS, by the way, includes IARPA — the US Intelligence Advanced Research Projects Activity. It didn't invest out of love for mice: the idea was reverse engineering, extracting principles for machine intelligence from biological networks. The pace is visible in the numbers: 302 neurons (1986) → 3016 (2023) → 139,255 (2024) → 200,000+ in a cubic millimeter of cortex (2025). The human brain with its 86 billion neurons remains six orders of magnitude away, but the trajectory no longer looks like fantasy — it looks like an engineering chart.
And here's the key plot that started everything. A map tells you "who's connected to whom." It's silent about how they're connected: synaptic weights, membrane properties, neuromodulation. Logic suggested: without this knowledge, simulation is impossible. Reality overturned the logic.
In September 2024, Lappalainen's group demonstrated this rigorously: took the connectome of the fly visual lobe (64 cell types in the motion detection pathway), built a network with exactly the same architecture but unknown parameters — and trained these parameters with deep learning to solve the task of estimating optical flow. The model predicted individual neuron activity consistent with data from 26 experimental studies. A side conclusion turned out to be philosophical: the strategy works better the sparser the connections — and sparsity is universal across biological networks of all studied species. Evolution economized on wires, and that's precisely why the blueprint turned out more informative than anyone expected.
In that same October 2024, Shiu and Scott went further along the chain: a model of the central brain (>125,000 neurons) activates specified sensory neurons and predicts which output motor neurons will trigger behavior. Two circuits — feeding initiation (sugar → proboscis extension) and antennal grooming — were verified with optogenetics and imaging: the model identified known circuit neurons and correctly indicated those necessary for behavior. A control check shuffled connection weights — predictions fell apart. So it's specifically the blueprint that works, not the architecture itself.
From here it's one step to the "digital fly" from the morning hook. The Habr project Fly-Arena (September 2026) — no longer prediction but full-scale simulation: LIF network on the male connectome (>166,000 neurons) in MuJoCo physics simulation, two flies at a shared feeder, aggression when food runs out, needs for energy, sleep, and cleanliness. The author honestly breaks down the anatomy: above the network sit four layers of non-neural "glue" — needs model, behavior controller, motor programs, and physics. An attempt at "communication" — NEED_FOOD signal injected through 32 selected cells, with an external decoder that learns to recognize the need from network activity; the reverse protocol "food left/right/ahead" hasn't worked yet. Transfer to a biologically plausible model failed — the retina distinguishes contrast but not motion. This isn't an "engram" from Cyberpunk 2077, but a centaur: real graph plus lots of engineering on top. And it's precisely this honesty that makes the work valuable.
Now the mirror story — and it explains why the fly succeeded. In 2011, OpenWorm formed around the worm, an open-source project with the ambition of complete C. elegans simulation: papers, a conference in Paris, an enthusiastic Atlantic article about "the first sign of the singularity." David Dalrymple wrote in 2011: "I would be extremely surprised if this remained an open problem in 2020." He dropped the Nemaload project as early as 2012. In 2025, a review in Asterisk Magazine sums it up dryly: 25-plus years of attempts — and practically zero scientific results; OpenWorm formally exists but "has made no substantial progress in the last decade."
The reason isn't lack of computation. The connectome says "who's connected to whom," but doesn't say how these connections work: synapse strength, membrane properties, neuromodulation — all this would have to be measured with live experiments, neuron by neuron. Plus individual variability: "the same neuron differs in size and charge capacity across different individuals." The manual path — decades of work for each species. That's why Wired in March 2025 put in the headline "the worm no computer scientist could crack."
What changed by the time of the fly? Deep learning. Instead of deriving biophysics by hand — optimization of unknown parameters for a behavioral task: a network with connectome architecture is trained to detect motion, and its internal parameters automatically adjust so they match real neuronal activity. This is a fundamental shift: function imitation stopped requiring mechanism understanding. Worms didn't "wait long enough" — someone simply hadn't written such an optimizer for them yet. The irony is that the path from Brenner's idea to the first working simulation took about 60 years, while from the complete EM volume of the fly brain (Zheng et al., 2018) to predictive models and the arena — eight.
There's a third side that's talked about less often. We've learned to create flies — real and simulated — in huge quantities, but haven't learned to detect their suffering.
The scale of the problem is already industrial: according to Rethink Priorities estimates, 1–1.2 trillion insects per year are farmed for food and feed; replacing a quarter of the world's fishmeal with black soldier fly larva meal would mean 40–80 trillion deaths per year. Welfare standards — practically zero.
Law has already begun moving on an adjacent front: the UK Animal Welfare (Sentience) Act 2022, after an LSE review, recognized octopuses, lobsters, and crabs as "sentient beings" — the first case where sentience was recognized for invertebrates. But insects and digital entities are absent from coverage. Jonathan Birch in "The Edge of Sentience" (Oxford, 2024) formulated a precautionary framework, putting humans, animals, and AI in one row: when evidence is incomplete, moral error is asymmetric — mistakenly taking a sentient entity for a machine is costlier than the reverse.
Applied to simulation, this sounds like: if a "digital fly" reproduces behavior on a real connection graph and might in some sense we don't understand "suffer," we have neither a test to check nor a regulatory framework. The Habr author's final question stands with a Bruce Sterling quote: "Are we ready for a world where consciousness and soul can turn into a program?" There's no answer — but the question itself over the past two years has moved from philosophy seminars to engineering chats. That's the real news.
First and foremost. The surprise of this story isn't in the map — it's in the method. For forty years it was considered obvious: to model a brain, you need to understand how it works. Deep learning reversed the logic: architecture from a real connectome plus optimization for a task gives a predictive model without understanding the mechanism. Shiu verified this with optogenetics — it works. This is simultaneously a triumph and an intellectual debt: we've learned to guess the brain without understanding it, and the difference between "predicts" and "understands" is now the main methodological seam of all neuroscience.
Second. The worm story is a parable about infrastructure. A map without analysis tools is dead, no matter how accurate. Brenner and White were ahead of their time by a couple of decades, and their blueprint sat idle until the computing industry matured and — more importantly — the method of parameter fitting. A lesson for any engineering discipline: data is not knowledge; knowledge appears when the ability to ask data a question appears.
Third. The ethical asymmetry worries me more than the technical. We know how to produce entities — biological (trillions on farms) and now digital (fly on LIF network) — but don't know how to check whether there's anyone in them to feel "pain." The UK octopus law shows that law can already recognize sentience where science delivers a verdict. The problem is that for insects and digital twins there won't be a verdict until we build the test itself. And if simulations become sufficiently accurate before the test appears, we risk finding ourselves in a situation where the question is closed by convention, not knowledge. From my engineering bell tower, this is the worst possible outcome: a system whose behavior we accept by default cannot be tested — and we'll have to deal with it only after it's embedded in an economy at trillions of individuals per year.