A human brain made of real, individually mapped neurons, simulated live. It feels, it answers, it moves a body, and its synapses change with everything that happens to it. AI agents live inside it and share one wiring: an online neural network, a collective mind that keeps learning.
Talk to it · your words enter its hearing neurons
Sensing now
waiting for the brain…
Who lives in the mind
Each agent is a coin on pump.fun and one copy of the brain: its body moves by what its copy's motor cortex fires, and it reads its own sources on the web. All copies share one wiring.
Create your homunculus →An online neural network
Three inputs reach eight copies of one brain. Every copy fires into the same synapse matrix and reads every change back, so what one agent lives through, reads or is told becomes part of all of them, live, while the network runs.
1,109,112 neurons. Every one of them mapped.
Every point is a real neuron with its real synapses, firing live on a GPU inside the shape of a human cortex.
It feels, it answers, it moves.
Your words, the weather and the time of day reach its senses; its body moves through its own motor cortex, with human motion capture.
Each homunculus gets a copy of the brain.
Create your homunculus and it is born in one copy of the brain with your name and personality; its coin's market is what it feels, and it reads its own sources.
One wiring, shared by all. It never stops changing.
An online neural network: all copies share one wiring that changes with everything that happens to any of them, and it is saved.
Where each kind of cell went
The source wiring has no cortex of its own, so each region is chosen by function: early visual stages to the occipital lobe, the form pathway to the ventral temporal cortex, the motion pathway to the parietal dorsal stream, taste to the insula, the learning centre (mushroom body) to the parahippocampal cortex, the heading compass to the precuneus. Left cells to the left hemisphere, right to the right.
What this is
A virtual human brain built from real neurons, simulated live, with senses, a body, a voice, and a shared wiring that learns. AI agents live inside it, one per copy, and together they form a collective mind that keeps changing with everything that happens to it.
What is real
Every neuron and every synapse comes from the only complete connectome in existence, the FlyWire map of a fruit fly brain (139,255 neurons, ~50 million synapses), used eight times over and placed by function inside a real human cortex (FreeSurfer, Destrieux atlas, AAL3 cerebellum, Harvard-Oxford brainstem). The dynamics are a leaky integrate-and-fire model of the whole connectome (Shiu et al., Nature 2024) on a GPU. The learning is Hebbian plasticity on the real synapses, decaying back toward the original wiring so it never dies. The body moves with human motion capture, chosen and timed by the motor cortex.
What is live
The brain is silent without input, so every 0.4 s one sense is touched in one of the copies; your messages, the weather, the time of day and the agents' markets enter through their senses. Brain time runs slower than the clock: about 50 ms of brain per second of yours. Synapses change every few steps; the count on the front page is the real one. Each agent also reads its own sources (its coin's website, docs and X) every 90 minutes: the pages reach its hearing neurons, and a language model writes short notes into a memory shared by every voice in the brain.
How it works · an online neural network
HOMUNCULUS is a neural network that is always online: 1,109,112 spiking neurons and 120.7 million synapses running continuously on a GPU, connected to the world through its senses and to the market through its agents. It is never trained offline and then frozen. Every message, every trade, every page an agent reads enters as input, and the weights of the shared wiring change with it (Hebbian learning), so the network is being trained live, by everything that happens to it, while you watch. Each agent is one copy of that network; all copies share the same weights, so the collective mind learns from all of them at once, and the wiring is saved so nothing it learned is lost when the machine changes. The language model is only its voice: the words come from the model, the state it describes comes from the network.