One brain for every task

Today we are opening flylabs, a lab built around one idea: stop designing a neural architecture per task, and start from the one architecture that has already been tested for a hundred million years.

In 2024 the FlyWire consortium published the complete wiring diagram of an adult fruit fly brain: 139,248 neurons, about 50 million synapses, each with a predicted neurotransmitter and therefore a sign. In 2026 two results changed what that diagram is good for. FlyGM showed that the connectome, used as the fixed structure of a graph network, learns whole-body locomotion in simulation faster than any unconstrained baseline. Eon Systems showed that the connectome with its biological weights drives a virtual body, walking and grooming, without a line of training data.

We take the second result as the proof and the first as the method.

What we ship

FlyCore is the engine: a sparse forward pass over the connectome, in PyTorch for training, in Rust for actuators and input devices, on WebGPU for the browser. Every weight set on this site runs on it.

Weight sets are the products. Cursor moves a mouse like a hand. Keys types with a pulse. Pad plays with thumbs. Gaze, Sono and Track see and hear. Dictate and Intent bring words in through a co-processor and let the brain decide when and where.

Adapters are the contract. Input adapters map your signal onto the 19,262 afferent neurons: screens onto the visual afferents, audio onto the antennae, embeddings onto the olfactory port. Output adapters read the 1,489 efferent neurons, or a decoder on the intrinsic population. Because the slot map is fixed, adapters and weights are swappable across tasks.

The brain is not the model. The brain is the substrate. The model is the brain plus its weights plus its adapters.

What is real today

Cursor is our first release and the one we are measuring against people: time to target, click accuracy and how closely its trajectories match recorded hands. Every other weight set is in beta or in development. Numbers on this site marked “target” are design targets for the first release, not measurements. We will replace them one by one, in public.

Why this matters

A network that starts with the right structure needs less data, less compute and fewer surprises. And a network that shares its structure across tasks can share its tooling, its adapters and its runtime. That is the bet.

If you want to build on it, the docs are open and the engine is on GitHub. If you want to train a weight set we have not thought of, talk to us.