Field guide · 14 September 2026
Borrowed
brains
From mapped neurons to machines that move.
Evolution’s wiring diagrams are becoming engineering material.
An illustrated guide by Louis Lehmann
Evidence through 14 September 2026
The fly · a map to build from
A brain becomes something you can inspect.
A fly finds food, cleans its antennae and keeps its footing. Its nervous system has been solving these problems far longer than we have been building robots.
Now researchers can follow its neurons from one connection to the next. A connectome is that wiring diagram: who connects to whom. It gives model builders an unusual starting point, a circuit shaped by an animal’s life rather than designed from scratch.
Released in September 2026, the male fly map reaches beyond the brain into the nerve cord that helps control the body. This is an anatomical resource, not a newly switched-on digital animal. [1]
What was mapped, and which fly?
The male central nervous system map includes both brain and ventral nerve cord. The widely used 2024 FlyWire map is a different specimen: an adult female brain with 139,255 neurons and roughly 50 million chemical synapses. The different coverage matters when comparing the counts. [2]
The instrument · four layers of an animal
What can we map?
What can we reproduce?
Follow the path from a connection to a moving body. Switch the question, then select a layer. These are different kinds of evidence, not percentages of a brain “completed.”
From anatomy to action
How models make a digital fly move.
One route uses the map to explain biology. Shiu and colleagues built a simple brain-wide spiking model and tested its predictions in living flies. It helped identify circuits for feeding and antennal grooming. A useful model can answer a real question without reproducing everything a brain does. [3]
Another route borrows the wiring to build a controller. FlyGM turns the fly connectome into a trainable network, first learning from an existing controller, then improving through rewards. Its simulated locomotion results suggest that biological wiring can be a useful design choice. [4]
What the moving-fly demonstrations establish
FlyGM is a preprint, not evidence that a static map runs by itself. It uses imitation learning followed by reinforcement learning (PPO). The reported gains concern tested simulation tasks and comparisons; they do not establish broad resilience to blindness or injury. Each simulation step uses more computation and memory than the conventional neural network used for comparison.
Eon’s embodied fly combines a connectome-based brain model with hand-chosen links to trained motor controllers. Its own account says visual activity does not yet substantially influence behavior and full biological validation is unfinished. [5]
NeuroMechFly v2 and Flybody also demonstrate navigation or locomotion using designed or trained controllers. Those videos should not be read as the September 2026 male connectome coming to life. [6] [7]
The fish · wiring meets activity
A transparent window
into a vertebrate brain.
A larval zebrafish offers more than another neuron count. It is a vertebrate, with brain structures closer to ours, and it is transparent enough to watch many neurons at once. Researchers can record activity, then map the wiring of the same animal.
Fish Fire&Wire pairs structure and function in a six-day-old larva. Its draft wiring resource covers about 140,000 neurons. Researchers can then test whether a model built from these connections predicts those cells’ activity. [8]
What is available, and what is still unfinished?
Fish Fire&Wire is an ongoing, not fully proofread reconstruction. Research groups request access to versioned snapshots under publication and redistribution conditions. It is not an unrestricted download of a finished, runnable fish brain.
Other zebrafish projects map molecular cell types in a different specimen. They should not be merged into this animal’s measurements.
ZAPBench makes part of the challenge measurable: predict future activity from recordings of roughly 70,000 neurons, collected over two hours under nine kinds of visual stimulus. It tests forecasting, not whether a simulated fish can live a fish’s life. [9]
Calcium imaging is a readout of neural activity, not a complete record of every electrical event. Matching activity is a step toward explanation; changing a circuit and predicting the effect is a stronger test.
Branching futures
Three ways to borrow a brain.
These paths do not have to arrive together. Better scientific models can matter now. Useful machines may borrow only a few biological ideas. A faithful digital animal is a much larger claim.
Could this mean tiny, offline, exceptionally fast machines?
Compact offline controllers are a plausible engineering goal, not a demonstrated consequence of these maps. A processor’s clock rate divided by a neuron’s firing rate is not an emulation speed: updating synapses, neural state and body physics takes work. Performance, energy use and adaptation need head-to-head measurements on the target hardware.
The strongest alternative is simpler: ordinary trained networks may do the job better. Biology earns its place in a machine when the borrowed structure improves the result, not merely when the diagram looks like a brain.
Borrowed wiring still needs to prove itself against simpler controllers, in unfamiliar tasks and on the hardware where it will run.
Evidence shelf
Read the work behind the guide.
Research papers, project documentation and a clearly labeled preprint. Figures above are original explanatory drawings, not scientific data.