The brain
for every task.
flylabs is a connectome-based AI lab. We do not build an architecture per task. We take the one architecture evolution already tested, and train it for yours.
An intelligence layer with a lineage
A fruit fly has 139,248 neurons and about 50 million synapses. With them it flies, lands on a ceiling, finds food, courts and learns. Since 2024 we know every one of those connections: the FlyWire consortium reconstructed the whole brain from electron-microscope sections, every synapse with its sign.
The connectome is an architecture that survived a hundred million years of selection. We treat it that way: a fixed network whose wiring is not negotiable, but whose weights are trained. Adapters bring your signals to the 19,262 afferent neurons and read the answer off the 1,489 efferent ones. Co-processors from classical AI take over what a fly brain never could: words.
Why fixed wiring wins
Randomly initialised networks have to learn their structure first. The connectome brings it along: motion detectors, target tracking, turn-taking, gait patterns are already wired. In 2026 FlyGM showed that a connectome-based network learns to walk and fly in simulation faster than any baseline. Eon Systems showed the brain drives a body with fixed weights, without a line of training data.
One engine
FlyCore runs the connectome as a sparse forward pass, in PyTorch, Rust and WebGPU. Same code, same answers.
A fixed slot map
Which neurons get which modality is a contract. That is why adapters and weights are swappable across tasks.
Co-processors where needed
Words and class labels come from classical networks and enter through the embedding port. The brain decides.
Who we are
flylabs is being built by a small team in Germany. More here as soon as there is more to say.