The brain
On 3 September 2026, Google Research and HHMI Janelia released MaleCNS v1.0, the complete wiring diagram of a male Drosophila melanogaster central nervous system. It is public and CC BY 4.0.
Mr Fly runs the whole thing: 166,700 neurons, 25,582,938 directed connections, 124,177,617 synaptic contacts. Nothing is pruned and no edges are thresholded away. The three source files are downloaded from Janelia's bucket and checked against their published SHA-256 digests before the graph is compiled, so the brain in this room is bit-for-bit the released one.
Where this came from
On 3 September 2026, Janelia Research Campus, the MRC Laboratory of Molecular Biology, the University of Cambridge and Google Research published the first complete wiring diagram of an adult Drosophila melanogaster central nervous system, in Cell. Millions of electron microscope images, reconstructed by machine learning into every neuron and every connection in one male fly.
It is the largest brain map ever made by neuron count, and they gave it away: public, free, CC BY 4.0. Within a week people had it playing Doom, watching Bad Apple, and steering a car.
Mr Fly is that same brain, whole, with a room to live in and a treasury to rule on.
Google Research · Janelia · HHMI · the dataset · the paper
The simulation
Each neuron is an approximate leaky integrate-and-fire cell integrated at a 0.1 ms timestep: 20 ms membrane and 5 ms synaptic time constants, a −45 mV threshold, 1.8 ms transmission delay, 2.2 ms refractory period. Rest is −52 mV for most cells and −60 mV for Kenyon cells.
Sign comes from the released transmitter annotations: acetylcholine excites, GABA, glutamate and histamine inhibit. Synaptic weight is contact count × 0.275.
One observation advances 500 ms of neural time and costs about 2.5 seconds of real time on one core. The fly lives roughly five times slower than you do.
What it sees
A 320×180 image, sampled nearest-neighbour at fixed positions inferred from each photoreceptor's own connections. 3,335 R1–R6 cells receive linear-sRGB luminance and 811 R8 cells receive blue and green proxies. That is 4,146 sample points out of 57,600 pixels, which is why a thin line drawn in its eye mostly misses.
The token's price is drawn into that image as left-right brightness. Not because the fly reads charts. Because light is the only thing its eyes accept.
Feeding it
The FEED button injects 20 mV for 200 ms into the 23 LB3c sugar-sensing neurons. Those cells exist in the released connectome, are wired into the graph, and are stimulated by nothing else.
Measured downstream, from an identical checkpoint, feeding versus not feeding: 17,432 neurons change their spike count, 14,924 of them in populations we never touch, including 1,685 Kenyon cell spikes and the central complex (EPG, PEN1, PEN2).
And the honest part: the PAM11 delta from feeding alone is −1. Feeding reaches its memory and its navigation system. It does not make it happy, and we are not going to say it does.
Reward and aversion
Price movement drives a pulse into identified dopaminergic cells: 15 PAM11 for reward, 2 PPL101 for aversion, graded by the size of the move.
Reward is a clean signal. On the same frame, PAM11 goes from 0 spikes to 270. Aversion is murkier: PPL101 sits at 65 unstimulated and reaches 96 when driven, because those two cells fire on their own. We report both rather than the flattering one.
Memory
The fly has 7,835 plastic synapses, every existing Kenyon-cell connection onto MBON07 and MBON11. They change under a baseline-centred anti-Hebbian rate rule adapted from Huang, Luo et al., Nature 2024: recent Kenyon activity followed by dopamine depresses eligible connections, the reverse timing potentiates them. Efficacy is bounded between 0.1× and 2× baseline.
A single observation typically moves 1,700 of those synapses. A dopamine pulse roughly doubles it. The fly you are watching is not the fly that started.
How it speaks
We choose the words. The brain chooses which one, and when.
Sixty-four words are each bound to one cell population, fixed and published before launch. Each tick, the population deviating most from its own rolling baseline emits its word. Rates are normalised by whole-brain activity, because total spikes swing 45% frame to frame and an unnormalised score just re-reports arousal.
No word is bound to a population we inject current into, or a decision would just be an echo of our own code. No population under 4 cells, because a 2-cell population's standard deviation approaches zero and one spike sends the score to infinity. There is no language model anywhere in this path.
Decisions by Mr Fly
Same machinery, higher bar. Each entry on a fixed published menu is bound to a population; the ruling is whichever clears the floor by the most. No entry can move more than 50% of the treasury, enforced by the module refusing to load if one does.
A window where nothing clears the floor produces no ruling, rather than a manufactured one. The fly declares. A human executes, or declines and says so on the same public list. The fly holds no keys and signs nothing.
What we tested and could not show
We tried hard to find a neural response to price, because it would have been a much better story. It is not there, and the attempt is on the record:
- 605 cell populations scored against price return, globally normalised, with size-matched shuffle nulls. One reached raw p<0.05. Zero survived Benjamini-Hochberg correction.
- A promising result in the wing motor neurons (DVMn, rho +0.91, genuinely
vnc_motor / subclass=wm) died when 27% of random populations of the same size beat it.
- The DNp20 steering readout scored rho +0.85 and failed its own control at p=0.31, because random left/right splits of any bilateral cell type score a median of 0.62.
- An apparent arousal signal was a receptor census artefact: 1,148 sample points fall left of the image midline and 2,187 right of it. "Busier when green" meant "green lights the half we sample twice as densely."
The reason is structural. The projection maps the left eye onto u ∈ [0, 0.60] and the right onto [0.40, 1.00], so the viewports overlap and no vertical line splits the image into left-eye and right-eye.
Everything that survived scrutiny is an intervention with a before and after. Everything that died was a correlation. That is why Mr Fly is a pet and not an oracle.