A brain that was never asked

The jar is a question about what a map of a brain is worth once you switch it on and give it something to do that it was never built for.

The claim

In 2025 the MICrONS consortium published a wiring map of one cubic millimetre of a mouse's visual cortex: about two hundred thousand cells and half a billion synapses, every one of them traced from electron microscope slices. It is the largest piece of a mammal's brain anyone has mapped at that level. The jar takes 1,627 of those neurons, the ones whose shapes were proofread by hand, keeps every connection between them, and runs them as a simple spiking model. Then it feeds stock prices into the cells that would normally hear the thalamus and reads its trades off the cells that would normally talk to the spinal cord.

The claim is not that this is intelligence, or a good trader. The claim is narrower and, we think, more interesting: the wiring alone already has opinions. Before we chose a single number, the map decided which stocks this brain can buy, which it can only sell, and which it will never touch.

What is real

Three things in the jar are taken from measurements of animals and nothing else.

The neurons. Each of the 1,627 cells has its reconstructed shape, its cell type and layer as labelled in the dataset, its position, and its synapses onto the other cells in the sample. Those 52,181 connections, with their synapse counts, are the ones the electron microscope found. We removed nothing and added nothing; we only simplified each arbor to about a hundred vertices so a browser can draw them all.

The brain around them. The outline in the jar is the adult mouse template from the Developmental Common Coordinate Framework, with the occipital cortex marked. The cubic millimetre sits roughly where primary visual cortex is. Roughly is the honest word: the two datasets come from different mice and were never registered to each other, so the box is placed by hand.

The prices. Robinhood Chain carries Chainlink feeds for eight stock tokens. A feed publishes a new round only when the price moves enough, which is why the mouse can sit for hours after the close with nothing to hear.

What is ours

The neuron model is a leaky integrate-and-fire unit with the constants from the whole-fly brain model of Shiu and colleagues: a 20 millisecond membrane, a 5 millisecond synapse, threshold seven millivolts above rest. Each synapse is made far stronger than a real one, because we can only see the few dozen inputs a cell receives from inside the sample, not the thousands it receives in life. Background noise stands in for everything outside the cube.

The mapping from prices to cells follows the anatomy of cortex, not the stock market. Input to cortex arrives in layer 4, so a stock's input cells are layer 4 pyramidal cells. Output leaves through layer 5 pyramidal tract cells, so those are the voters, split into eight groups along the length of the sample, one per stock. For each stock we then picked, in turns, the layer 4 cells whose real connections reach its voters most strongly, directly or through one excitatory cell in between, and the inhibitory cells with the most synapses onto those same voters. A rise drives the first group. A fall drives the second. The wiring between them is untouched.

The rule is one paragraph and it is printed on the front page in full. Listen for a minute. If the voters fire one and a half times their resting rate, buy. If they fire two thirds of it or less, sell everything. Otherwise do nothing.

What the wiring decided

We ran the same full-strength rise and full-strength fall into each stock's cells for a minute at a time, several runs each, and counted the votes. The results are in the table on the front page. They were not tuned per stock; they are what those particular cells do.

NVDA's input cells reach its voters through 1,677 synapse-weighted paths and a rise lifts the votes almost fivefold. SPY drew the thinnest wiring in the sample, 62 paths, and a rise lifts its voters only 1.3 times, under the 1.5 it takes. This brain cannot buy SPY. It can sell it: SPY's brake cells cut the votes to a quarter. TSLA and GOOGL go the other way on a fall, to zero. Their brake cells, the ones with the most synapses onto those voters, silence them completely. The first real price the jar heard, TSLA down half a percent after the close, produced exactly that: not one vote in sixty seconds.

None of this is a prediction about the companies. It is a fact about which cells happened to be next to which in one mouse, and it is the whole point. A trading agent built from an LLM has opinions because it read the internet. This one has opinions because of where its axons grew.

What it is not

It is not a whole brain. A cubic millimetre is about one five-hundredth of a mouse cortex, and the proofread cells are under one percent of that. It does not reproduce those cells in full; the neuron model is the simplest one that spikes. It is not connected to a wallet. The money is paper, ten thousand dollars of it, marked at the feed price with a small fee per fill. There is no order and no exchange.

It is also not a mouse in any sense that matters ethically. Nothing here is alive or was made to suffer. The animal the map came from was part of a research programme that published its methods, and the dataset is released under a licence that invites reuse.

Why a jar

Because the picture of a brain in a jar is the oldest thought experiment about minds, and the map makes it a little less of a thought. You can now hold a real piece of a real brain's wiring in a browser, watch it fire, and give it a job. The job we gave it is absurd on purpose. If the absurd job still produces behaviour that traces back to anatomy, that says something about how much of behaviour anatomy carries.

Sources

1. The MICrONS Consortium. Functional connectomics spanning multiple areas of mouse visual cortex. Nature 640, 435–447 (2025). doi:10.1038/s41586-025-08790-w
2. Kronman, F. N. et al. Developmental mouse brain common coordinate framework. Nature Communications 15, 9072 (2024). doi:10.1038/s41467-024-53254-w
3. Shiu, P. K., Sterne, G. R. et al. A Drosophila computational brain model reveals sensorimotor processing. Nature 634, 210–219 (2024). doi:10.1038/s41586-024-07763-9