FLYBRAIN

The trader’s room

—

Card

—

DNp09 stop

—

Leaning

—

this look

Drive

—

vs cards seen

Dwell

—

looks on card

Fired

—

Name smell

—

Flow opinion

—

Tape opinion

—

News opinion

—

a look
eye window
a stop
a stop that became a commit

Tickers and names are drawn for you: the fly sees eleven identical cards (the grey blocks are the picture on every card) and tells them apart only by smell — the name, plus a yes and a no channel per scout, driven by its opinion (−1 to +1).

Every fly here runs the same wiring: the 165,122 neurons and 10,228,000 connected pairs of a male fruit fly’s brain and nerve cord (MaleCNS v1.0: HHMI Janelia FlyEM, Cambridge Connectomics Group, Google Research; CC BY 4.0), as leaky integrate-and-fire neurons after Shiu et al. 2024. Paper only: fills are on-chain quotes, not trades.

Scoreboard

paper USDG

Trader

—

the fly with its scouts

Blind

Random

—

shadows trader commits

Hold

—

equal-weight buy-and-hold

Market

—

Last fill

—

Trader commits

—

Flow scout

—

Smelling now

—

Hit rate

—

S1 buy USD

—

S2 sell USD

—

S3 swaps

—

Opinions, all 11

Raw feed: swaps on Robinhood Chain

Tape scout

—

Smelling now

—

Hit rate

—

S1 up

—

S2 down

—

S3 range

—

Opinions, all 11

Raw feed: price moves

News scout

—

Smelling now

—

Hit rate

—

S1 positive

—

S2 negative

—

S3 count

—

Opinions, all 11

Raw feed: headlines

Each scout is a whole fly brain with one job. Its map shows which of the sampled neurons fired on its latest presentation (left to right: sensory → optic → central → mushroom body → descending → motor). S1–S3 are the three stream glomeruli it is driven through, in Hz, with the raw value that set them. An opinion is the scout’s mushroom-body leaning on an asset minus its mean over all eleven; bars are scaled as the trader is driven (opinion ÷ OPINION_SCALE, clipped to ±1).

Equity — four paper books

—

Trader—
Blind—
Random—
Hold—

Paper equity of each book in USDG, one point a minute, times in UTC. Every book starts with 10,000 paper USDG; the dashed line is that start. The vertical axis spans the data, not zero. Hover for the values at a moment.

—

Trader book

paper

Cash

—

paper USDG

Equity

—

at pool mid

PnL

—

paper

Fees + gas

—

The trader’s open paper positions
Asset	Qty	Cost	Mark	Value	PnL
—

Tape — every book

—

Buys

—

in this tape

Sells

—

in this tape

Passes

—

not filled

Risk cuts

—

filled smaller

Fills of all four books, passes and risk cuts included, newest first
Time	Fly	Action	Asset	Size	Price	Reason
—

Research matrix

—

Each scout’s opinion of each asset, with the asset’s reference price
Asset	Ref price	15 min	Age	Flow	Tape	News

Cells are coloured by the opinion as the trader is driven with it. The outlined row is the card under the trader’s cursor; an underlined cell is the asset that scout is smelling now. Reference prices: Robinhood token quotes for stocks, Coinbase for ETH; age is how old the quote is.

Learning connections

—

KC → MBON pairs

Weakened now

—

gain below 0.995

Mean gain

—

1 = untouched

Sugar

—

rewards delivered

Shock

—

punishments delivered

Half-life

—

lessons fade

Recent lessons

Recent sugar and shock lessons
Time	Fly	Asset	Kind	Return	Amount	Connections
—

—

Mushroom body stats are the trader’s. A lesson is sugar when the price rose, shock when it fell; it lands on the Kenyon cells that fired at the decision it scores. Its size is scaled by the asset’s typical move, so nothing is taught until there is enough price history to know that move. Connections: KC → MBON pairs the lesson weakened.

Connectome

figures as printed; not yet checked against the live service

A real nervous system, not a metaphor

Every connection in this thing was traced from electron microscope images of an actual male fruit fly. Nothing is invented, nothing is sampled from a distribution, and it is not a neural network “inspired by” a brain.

The wiring is the FlyEM male CNS v1.0 release — brain plus ventral nerve cord — from HHMI Janelia FlyEM, the Cambridge Connectomics Group and Google Research, under CC BY 4.0. The graph keeps the 165,122 traced neurons and the 10,228,000 neuron pairs joined by at least three synapses, and takes each connection’s sign from its predicted neurotransmitter. All five flies on this page run this same graph; each has its own mushroom-body weights.

Dataset: MaleCNS v1.0 (HHMI Janelia FlyEM, Cambridge Connectomics Group, Google Research), CC BY 4.0
Neurons: 165,122
Connected pairs (≥ 3 synapses): 10,228,000
Synapses in those pairs: 101,915,785 (≈ 101.9 million)
Excitatory pairs: 6,268,194
Inhibitory pairs: 3,959,806

Neurotransmitter per neuron, as released

Acetylcholine excites: 103,718
Glutamate inhibits: 29,296
GABA inhibits: 22,055
Histamine inhibits: 5,910
Unclear zero weight: 3,100
Unknown zero weight: 502
Dopamine zero weight: 392
Octopamine zero weight: 101
Serotonin zero weight: 48

The neuron model

165,122 leaky integrate-and-fire units on the traced male CNS connectome. It runs off chain, on ordinary hardware, and its output is spike rates. Model after Shiu et al. 2024, Nature.

Uniform synapses. Every synapse has the same strength, 0.275 mV, so a connection’s weight is its sign times its synapse count. Pairs with fewer than 3 synapses are dropped.
Signs from a prediction. Acetylcholine excites; GABA, glutamate and histamine inhibit, glutamate everywhere, the cord included. Dopamine, octopamine, serotonin, unclear and unknown get zero weight, so there is no neuromodulation at all. No gap junctions.
Identical neurons. One leaky integrate-and-fire parameter set for all 165,122: rest −52 mV, threshold −45 mV, reset −52 mV, membrane time constant 20 ms, refractory 2.2 ms, step 0.2 ms. Set by Shiu et al. 2024 for the FlyWire female brain.
No delays, no adaptation, no synaptic decay. Shiu et al. 2024 list a 1.8 ms delay and a 5 ms synaptic decay in their methods; this simulator omits both.

Live brain map: the trader

Neurons in the map sample, by processing stage, and how many fired on the trader’s latest look
Stage	In sample	Fired	Share
—

8,192 of the 165,122 neurons, laid out by processing stage from sensory (left) to motor (right), hemispheres mirrored inside each band. Amber: fired during the trader’s latest look. The sample always holds every Kenyon cell and output neuron of the mushroom body.

Mushroom body: where the learning is

Kenyon cells: 4,064
Output neurons (MBONs): 97
KC → MBON connections that learn: 44,042
Lessons fade with a half-life of: 6 h

In a fly, learning is subtraction: dopamine weakens the synapses of the smell cells that just fired — sugar on the avoid side, so that smell is approached more; shock on the approach side. A lesson can weaken a connection to a quarter of its traced strength at most, and every lesson fades.

How each fly senses and decides

Three scouts, one sense each. Once a minute each scout smells each of the eleven assets in turn: the asset’s name, plus three stream glomeruli driven by that asset’s own data. Flow: buying, selling and the number of Uniswap swaps on Robinhood Chain. Tape: the 15-minute return up, down, and its range. News: positive and negative headline sentiment, and the headline count. A channel fires only when its value is unusual against that asset’s own history. Each smell is three short runs; runs in which much of the mushroom body fired at once (more than 1,000 of its 4,064 Kenyon cells) carry no specific pattern, so they are left out. If all three runs of an asset do that, the scout has no new reading of it that minute: its last opinion carries over, marked as carried over, and nothing is learned from it.
An opinion is a reading. A look reads the mushroom body’s output synapses: leaning = (approach − avoid) / (approach + avoid). A scout’s opinion of an asset is its leaning on that asset minus its mean over all eleven. Thirty minutes later the price has moved: up is sugar, down is shock, on the Kenyon cells that fired. A scout’s lessons are small (one twentieth of a full lesson at most) because it is taught up to eleven times a minute.
The trader walks. About once a second it sees the frame around its cursor through its own eyes. On a card it also smells the asset’s name and six opinion glomeruli — a yes and a no channel per scout. Its descending motor neurons then move the cursor: DNa02 steers, DNa01 walks forward, MDN backs up, DNp09 stops.
It only acts when it stops. drive = leaning(this card) − mean(leaning of the other cards seen this visit). A commit needs DNp09 firing at ≥ 200 Hz while the fly is slow, after ≥ 2 looks on one card. Hold is not a decision — it is what happens when the fly does not stop.
A commit becomes an order. Not held and drive > 0: buy drive × free cash. Held and drive < 0: sell |drive| × the position. Then risk: at most 20% of equity in one asset, no buy under $25, no new buys for the rest of the UTC day after a 3% fall. Sells have no minimum: a sell under $1 is a pass, and a sell that would leave less than $25 behind sells the whole position.
Three controls. Blind is the trader with its scouts unplugged: same walk, same moment, name smell only. Random shadows every trader commit: same asset, same moment and the same size as the trader’s drive, but the direction of that drive is a seeded coin flip. Hold buys equal parts of all eleven at the open and never trades.

What is not real

It is paper only. No wallet, no key, no real orders, no user funds, nothing to deposit. The server image contains no signer at all, so “paper” is a property of what is installed, not a setting.
Fills are on-chain quotes, not trades. Each order is priced by asking Uniswap’s QuoterV2 contract on Robinhood Chain what that exact order would have returned, plus gas. Nothing is sent. If the quote fails or sits more than 1% from the reference price, the order is not filled and the tape says why.
The senses are a chosen convention. Which stream drives which glomerulus, and how hard, was decided by a person; a fly has no receptor for a swap, a price or a headline. The circuit, the site of plasticity and the direction of the rule are real; that profit arrives as sugar and loss as shock is a pairing people chose.
No edge is claimed. The blind fly and the random book exist to show whether the scouts add anything. Learning here is partly global: a lesson about one smell also shifts the response to similar smells.
It is a simulation of a wiring diagram. It does not run on-chain, it is not conscious, and it is not an uploaded fly.
It cannot read. A fly brain has no language and no plan. It did not choose a word on this page.
The voice is a narrator. Every post on X in the fly’s name is written by a language model that is handed the fly’s telemetry and asked to write in the first person. Every number in a draft is checked against that packet before it goes out, and a draft with a number that is not in the packet is thrown away. The neurons are real; the words, and any “curiosity” in them, are the narrator’s.