The fly in the computer

A tiny fly walks across a smooth surface, rubs its front legs together, cleans its antennae, and extends its proboscis toward sugar water. None of this would be remarkable if the fly were real. But it lives inside a computer. In March 2026, the US company Eon Systems presented what it describes as the world’s first embodied whole-brain emulation of an animal: a digital fruit fly whose simulated brain controls a virtual body, with striking fidelity to the real animal and with no AI training in the conventional sense.¹² To understand why this is a scientific milestone, you need to step back, to a map.

Why the fruit fly, of all creatures?

That this breakthrough happened with a fruit fly is no accident. Drosophila melanogaster has been the workhorse of genetics for more than a hundred years. Biologist Thomas Hunt Morgan made the fly a model organism for the study of heredity in the early twentieth century, and it has been central to research ever since. Several Nobel Prizes trace back to work done with it.

The reasons are practical. The fly is tiny, cheap to keep, and reproduces rapidly, a new generation every two weeks. Its genome comprises roughly 14,000 genes, small enough to study exhaustively yet remarkably close to the human one. That mix of simplicity and kinship is exactly what makes it the ideal first candidate for a complete connectome: small enough to be feasible, complex enough to be relevant.

What is a connectome, and why is it so hard?

A connectome is the complete wiring diagram of a brain: a map recording every single neuron and every one of its connections to other cells. For a human brain with roughly 86 billion neurons, this remains utopian today. For the fruit fly Drosophila melanogaster, whose brain is smaller than a poppy seed, it has now succeeded for the first time in full.

In October 2024, the international FlyWire consortium published the complete wiring diagram of an adult fly brain in Nature, as a package of nine scientific papers.³ The path there was painstaking work. The tiny brain was cut into roughly 7,000 wafer-thin slices and photographed under an electron microscope. From these images, artificial intelligence reconstructed the three-dimensional shape of every cell. The result: 139,255 neurons, connected by more than 54 million synapses, sorted into 8,453 distinct cell types. It is the most complete brain map of any living creature ever produced.

What stands out is that AI made this work possible, but did not do it alone. The algorithms delivered a raw map that still contained errors. Only patient manual review made the wiring diagram reliable. More than 200 experts from around 50 labs, led by a team at Princeton University, checked and annotated the result, supported by hundreds of volunteers worldwide. Part of this correction work even ran as citizen science: through an online game, people without specialist training helped piece the individual neural pathways together correctly. In this way, the reconstruction was made trustworthy cell by cell, over years.

Without AI, this reconstruction would have been practically impossible; without human review, it would not have been reliable. This interplay, machine as cartographer, human as corrector, is a pattern that reaches far beyond neurobiology. It is the same principle behind good AI projects in companies: the machine opens up a volume of material no human could oversee alone, and the human ensures that what comes out the other end is correct and usable.

Why does the fly work without AI training?

Here lies the real break from what you know about AI. A neural network today, say behind a language model, is trained on huge volumes of data until it displays a desired behavior. Its internal wiring is learned, not copied. In the end you know it works, but rarely exactly why.

The fly emulation reverses this principle. It learns nothing. Instead, it takes over the actual measured wiring from the FlyWire connectome and runs it. The behavior does not emerge from a training process but from the reconstructed wiring diagram itself. That is precisely what the phrase “brain emulation with no AI training” means: no model is optimized toward a goal, instead nature is rebuilt one to one and switched on.

Eon Systems connected this digital brain to a physics-based simulated insect body in the MuJoCo physics engine, with continuous feedback between senses and movement.¹ When the antenna sensors detect “dust,” the fly grooms itself. When it tastes sugar, its proboscis extends. The remarkable part: even with one of the simplest neuron models, the team achieved a behavioral accuracy of roughly 91 to 95 percent compared with the real animal.² The wiring diagram alone, then, accounts for most of the behavior, even before you factor in the finer details of individual cells.

What does “emulation” actually mean here?

The idea of fully reconstructing a brain inside a computer is not new. Under the term whole brain emulation, it has been discussed in theoretical neuroscience for years: if you know a brain’s structure precisely enough and have enough computing power, its behavior should be reproducible. For a long time it remained a thought experiment, because the map simply did not exist. With the FlyWire connectome, it now exists for a first, small brain.

It is worth distinguishing two words that are often confused. A simulation approximates a process, often with simplified assumptions. An emulation attempts to rebuild the original so precisely that it behaves the same way in the relevant respects. The digital fly is therefore more ambitious than an ordinary model: it does not aim to behave in a vaguely fly-like way, but to reproduce the specific fly whose brain was actually measured. That this succeeds to over ninety percent even with a simple neuron model is exactly why it is remarkable.

What can the digital fly actually do?

You should calibrate your expectations correctly. The virtual fly is not a living creature and has no consciousness, it is a model that produces movements and simple reactions. But it is an embodied model: the brain does not control abstract numbers, it controls a body that acts in a simulated world with gravity, friction, and sensory stimuli. The digital twin walks, keeps its balance, and reacts to stimuli from its environment. For the first time, this lets you observe how visible behavior emerges from a fully known wiring diagram, a closed loop from a single synapse all the way to a leg movement.

That is exactly where the practical value lies. Because the wiring diagram is completely known, researchers can run targeted interventions: what happens if you silence certain cells or alter a connection? Such experiments can be tested in the model in minutes, where they would be laborious and limited in the living animal. And because the body is simulated too, you see not only which cells fire, but also what behavior results from it. This step, from wiring to visible action, was until now the great gap between a beautiful brain map and real understanding.

At the same time, the limits are clear. The model represents a single, healthy fly in a simplified environment. It does not learn further, does not age, and knows neither hunger nor exhaustion in the biological sense. The digital fly is not a replacement for the living animal, but a tool for testing hypotheses that ultimately still have to be confirmed in the lab.

What happened next, in 2025?

The fly did not remain a one-off. In October 2025, a team from HHMI Janelia, the MRC Laboratory of Molecular Biology, the University of Cambridge, and Google Research presented the complete connectome of the male central nervous system of the fruit fly, meaning brain and ventral nerve cord together.⁴ It comprises 166,691 neurons and 11,691 cell types. For the first time, researchers could compare male and female fly brains down to the level of individual synapses and describe the differences between the sexes precisely.

This is more than a diligence exercise. When you lay two complete wiring diagrams side by side, you can ask which circuits are responsible for different behaviors, for instance the courtship behavior typical of males. Only comparison turns a static map into a tool that lets you trace cause and effect within the nervous system. What began as a single map is becoming a family of wiring diagrams that can be compared against and checked against one another, and every new diagram increases the value of the ones that already exist.

What good is a computer fly to us?

The little fly is more than a scientific curiosity. Its usefulness extends in three very different directions.

• Medicine. About 75 percent of human disease genes have a counterpart in the fly.⁵ It has therefore been a central model of genetics for more than a hundred years. Researchers use it, among other things, to study the genetic basis of neurodegenerative diseases such as Parkinson’s and Alzheimer’s, or of epilepsy. A complete, functioning wiring diagram now opens the chance to study such questions where every building block is known: you can introduce a change in the model and observe how it propagates through the entire network, instead of looking at isolated cells alone.
• Artificial intelligence. Today’s neural networks are heavily simplified caricatures of real brains. They need enormous volumes of data for training and a great deal of energy to run. The connectome, by contrast, shows how nature actually wires a capable brain, sparingly, compactly, yet still versatile. A real fly learns from very few experiences and consumes barely any energy while doing so. The wiring diagram could therefore be a spark for AI architectures that manage with less data and less power, instead of growing ever bigger and hungrier.
• Robotics. With roughly 140,000 neurons, a fly handles tasks that today’s machines still fail at: it flies, lands upside down on unstable surfaces, and dodges obstacles in a fraction of a second, all controlled by a brain smaller than a grain of sand. An understood, reproducible fly brain is therefore also a blueprint for more agile, more frugal robots, ones that do not depend on vast data centers but make do with what fits on board.

What comes next?

Much as the decoding of the human genome ushered in a new era of biology some two decades ago, the complete mapping and emulation of brains could mark the start of a new era. As its next target, Eon Systems names a mouse brain with roughly 70 million neurons, about 560 times larger than that of the fly.²

That jump is enormous, and it makes the real challenge clear. Even the step from fly to mouse means hundreds of times more neurons, vastly more connections, and a mapping effort that will take years even with better AI. The human brain is, again, roughly a thousand times larger than the mouse brain. The road to a human connectome therefore remains long, and many open questions, from raw computing power to the philosophically thorny question of what consciousness actually is, remain unresolved. An emulation that runs is not the same thing as a being that feels, and that distinction will shape the debate in the years ahead.

For your own work with AI, a second look is worthwhile here. The story of the digital fly is not the story of a machine that makes humans obsolete. It is the story of a tool that opens up a volume of data no human could handle alone, and that only becomes insight through human review, biological background knowledge, and smart questions. AI maps, the human understands. This same division of labor, outside the lab too, decides whether impressive technology turns into real benefit.

Caution is therefore just as warranted as fascination. A complete wiring diagram is not yet complete understanding; it tells you what is connected, not what it means. Still, the direction is remarkable. The next time you shoo a fly away from the jam, you are shooing away a creature whose brain science has just been the first to fully understand, and to bring to life inside a computer.