Smell inputs (PN)
Smell cells · memory (Kenyon)
Outputs · go toward / keep away (MBON)
Dopamine · reward (PAM)
Dopamine · punishment (PPL1)
Brake (APL)
◯ a real trade arriving
Drag to turn · wheel or two fingers to zoom · shift-drag to move · tap a neuron to walk its synapses · tap the legend to hide a family. The 3D is schematic: depth groups each family (inputs at the back, outputs in front); the circuit file has no anatomical positions.
Stock
—
reading…
Its decision now
watching —
Waiting for a smell
sell / stay outbuy / hold
The fly decides only when a window brings real trades. Its rule: out and drawn to the smell → buy; in and repelled → sell; otherwise it holds, or stays out.
It acts from 2.0% sure — its boldness is set with its other controls (band 3).
1 · What it thinks
Its thoughts · before, now, next
Each decision is a set of synapses firing: the smell cells a smell wakes up, through the synapses it learned, into “go toward” or “keep away” outputs. Tap a synapse to light it on the map.
What the words mean
Smell cell = a Kenyon cell: a few of them fire for each smell, they are its memory. “Go toward” / “keep away” output = an output neuron of the memory centre (MBON): when it fires the fly is pushed toward the smell, or away from it. % strength = how much the synapse between them keeps of what it had in a newborn fly: learning weakens the side that was wrong (a fall weakens “go toward”). Pull toward / push away = the same, summed over the cells this smell wakes up (100% = nothing learned yet). “Next” is computed, not guessed: the fly is deterministic, so a shadow copy loaded with its exact state is run ahead once per possible move (Laplace's demon, at the scale of this model). It forecasts the FLY's next decision, never the market.
Activity right now
What happened · log
2 · Teach it, keep its memory
Teach it a move (simulation)
Injected on the smell of the last real window of trades: the fly learns what this move would teach it, then answers again on that same live smell. It really trains this browser's fly, so it is blocked while the fly's own wallet trades.
Its memory
What it learns is kept in this browser, per stock, so it keeps training from visit to visit. A private window starts a naive fly.
Sealed automatically: after each real trade, the fly's own wallet sends 0 ETH to itself with a 66-byte public note — the fingerprint (keccak-256) of what it learned on this stock and after how many windows (about 23,000 gas, measured on a fork). No funded fly wallet, no seal. The chain proves this memory existed then and was not changed since, not that the fly is right.
Sound: smell cells pluck notes, outputs hum high (go toward) or low (keep away), reward rings, punishment drums. Same moves, same rhythm.
The shared fly (server)
The same brain, run by the server on one stock, read every minute and deciding once per window fitted to the stock, whether anyone watches or not. Its memory is saved on the server.
3 · Trade with it real money · your risk · not advice
Let it trade alone · one signature
The fly gets its own small wallet, created in this browser. You sign one transaction that sends it a budget (USDC, plus a little ETH for gas). Then it trades that budget on its own on the stock you watch, within your caps. Stop halts it; Withdraw all sends everything back. It trades only while this page is open; reopened, it resumes where it was, with your settings (Stop halts it).
Read before you fund it
- You can lose this money. Its decisions are a model of a fly brain, not advice.
- Its key is kept only in this browser, unencrypted. The server never receives it, so nobody can recover it: clear this site's data or lose this device and the money is lost unless you exported the key.
- Anyone or any extension with access to this browser can use it, and so could the code this site serves: you trust this site's code with the budget.
- The server plans each order; before signing, this page checks every call itself and refuses anything that is not an exact approval, the app's fee, or a swap paying the fly.
- The app takes no interface fee on the fly's trades: a trade pays the pool's own fee only (0.05% a side on the measured routes; a buy and a sell back cost 0.10% on a fork). If the site is busy, a decision can pass without a trade, a SELL included. Its decisions follow its real holdings, not the paper book.
- It buys only when its edge pays the real cost. It decides once per window fitted to the stock: a quiet stock gets a longer window (up to 2 h), a nervous one a shorter one (from 5 min), so that its usual move can pay the fees. It buys only when the gain it can expect from that window — its record (pulled toward a coin flip while it is short) times the stock's usual move over the window — is larger than one side of the trade: the pool fee plus the app's fee of that route. Otherwise it says so and signs nothing. On a quiet stock, or with a young record, that can mean no trade for a long time: that is the rule working, not a fault.
- Proven on a copy of Base with Base's own precompiles: NVDAc, AAPLc and SNDKc bought and sold back, every order checked by this page, round trip about 0.3%. On Base itself: one real buy (SNDKc, 0.5 USDC, 8 Nov) and one real Withdraw all (9 Nov); no real sell yet. Keep the budget small.
- Tokenized stocks are only for eligible persons outside the U.S.
The fly's rules — what it can never do
Not connected.
No fly wallet in this browser yet.
4 · How it works
How it differs from a block's brain
- A block's brain: 128 spiking neurons whose wiring is derived from the block's own address, so every block has a brain of its own. It is fed the block's own on-chain facts (market cap, GMs, messages, new holders, whether its creator still holds) and shows how the block moves and its mood. A toy network, not a biological copy.
- This brain: 1,119 neurons wired exactly as a real fly's, the same circuit for everyone, measured under the microscope. It is fed a stock's live market price, and it learns: dopamine reshapes its synapses, the memory forms and fades, and it leans toward or away, then meets the consequence.
- Both: no language model, no randomness, public inputs only. Anyone can replay them and get the same result.
What this model is, and is not
- The neurons, the synapse counts between them, and which dopamine neurons reach which output neuron are measured (electron microscopy). Threshold: 5 synapses; 300 Kenyon cells out of ~2,000, taken by ascending id.
- The price is the stock's live on-chain pool (Aerodrome, Base), weekends included. The Chainlink reference is shown next to it, not fed to the fly.
- The odor is what the chain proves about each transaction of the window, from its Transfers and its signer: the signer received the stock, sent it, or it moved between others. Never called "buy" or "sell". Each kind lands on its own third of the PNs (id mod 3, our choice), stronger with USD size.
- Rises drive the reward dopamine neurons (PAM), falls the punishment ones (PPL1), as in fly studies. Their strength is the move against this stock's usual move, because dopamine neurons adapt to the range of rewards they expect (Tobler, Fiorillo & Schultz 2005, Science): a usual move teaches at half strength, twice the usual saturates. Dopamine is not spread as excitation: it weakens the Kenyon→output synapses of co-active cells, the fly's known learning rule. Its memory trace is that weakening; it fades a little at each new experience, never while idle.
- A smell is taught by the moves of the next three windows, each half as much as the one before: an odor leaves a trace in the Kenyon cells that dopamine arriving later can still act on (Handler et al. 2019, Cell). Measured on a synthetic market (16 runs): with a pattern two windows late it was right 9 times and wrong 0 (4 right, 2 wrong without the trace); with no pattern at all it still leaned one way 15 times out of 16 — a learner sees patterns in noise. That is why its record is shown against a coin flip, and why its own wallet waits for that record by default.
- Activity spreads in proportion to synapse counts and fades in about 3 to 4 s once a stimulus has passed; the next one waits for that. The brake (APL) is modelled as global inhibition that lets only the top 10% of Kenyon cells through.
- Approach/avoid is read from output neurons sorted by their measured dopamine input. The constants, the activity rule and the odor mapping are ours: a fly does not smell prices.
- The paper trader is virtual — unless YOU fund the fly's own wallet (“Let it trade alone”): then those same decisions move that wallet's real money. Its rule: out and drawn to the smell → buy; in and repelled → sell; otherwise hold if in, stay out if out; a window without trades changes nothing. No short. Each switch pays the pool's own fee, read on chain. It is shown next to two dumb witnesses (always in; in if the last window rose), because a P&L alone proves nothing.
- The fly is asked before the outcome: its prediction on one window is judged on the next window's price move (under 0.05% = flat, no verdict), misses counted next to matches; then it is taught with that move, and the synapses it weakened flash white. A tendency is not a forecast and not advice. Nothing here feels anything: it is a computed state, replayable by anyone with the same data.