comma’s seventh hackathon ran for two days in September. I spent them getting a fruit fly’s brain to drive a toy car toward a Wi-Fi access point.
It worked, in the honest sense of the word. The car crawled around a room, veered off walls before touching them, slammed into reverse when it hit one anyway, drifted up the signal gradient toward an access point named Flash, and buzzed through the comma’s speaker the whole time. It was also derpy: it meandered like something with better things to do, and the single most important control signal — the one that turns the wheel toward the AP — was the least neural part of the stack. I’ll get to that, because it’s the interesting part.
The premise is simple. Take a published connectome, simulate every neuron in it, then wire real sensors to real sensory neurons and real actuators to real motor neurons. Train nothing. Put no policy in the middle. Inject at the periphery, read at the periphery, and let the wiring do what the wiring does.
The hardware, which is the strange part
The compute is a comma 4 — a Qualcomm SDM845 dashcam-shaped Linux box with 3.5 GB of RAM — paired with an external AMD GPU that comma calls a chestnut. The chestnut is the part worth pausing on: a desktop-class AMD card connected over USB3 → PCIe through an ASM2464 bridge, driven entirely from userspace by tinygrad. There is no kernel PCI device, no DRM node, no amdgpu; /sys/bus/pci/devices is empty. tinygrad pokes the card’s registers over USB and schedules kernels on it directly.
It works about as well as that description suggests, which is to say it works: PCIe link up at LTSSM 0x78, USB negotiated at 5 Gbps, ~417 GFLOP/s FP32 steady-state on a 2048³ matmul. It also has opinions. Uploads have to be staged in 16 MB slices off a ~312 MB host-visible BAR window; a raw 4 MB PCIe zeroing pass took 1.164 s and blew through the transport’s 1-second timeout, which looked exactly like VRAM exhaustion until I measured it; and tinygrad’s BEAM autotuner is forbidden here, because its compile-time probing wedges the GPU.
The car is a Hiwonder MentorPi — an Ackermann-steered toy chassis, 1.2 m/s top speed, 29° of steering lock, 145 mm wheelbase — with a second comma 4 bolted on as the body. It has no eGPU of its own: the chestnut needs a 12 V supply the toy can’t carry.
What “running a fly brain” means here
The model is MaleCNS v1.0: 166,700 neurons and 24,469,412 signed synapses covering the brain and ventral nerve cord, released CC-BY by FlyEM/Janelia and collaborators. Each neuron is leaky integrate-and-fire with the Shiu et al. 2024 constants, dt = 0.1 ms, an 18-tick axonal delay, and weights that are signed synaptic contact counts scaled by 0.275 mV.
Most of the project went into one number: throughput. The final acceptance run completed 118.008 biological seconds in 59.7895 wall seconds with both cameras, Wi-Fi, IMU, and a browser viewer attached: 1.97× real time.
1.97×
Sustained simulation speed on the eGPU
19,737 completed ticks/s over a minute, full 166,700-neuron model, all sensors live.
Faster than real time, on a card talking over a USB cable, with a phone SoC generating the input. Getting there meant replacing dense Poisson input generation (403 ms per 36 ms biological window — a 0.09× ceiling by itself) with sparse event sampling at 3 ms, bit-packing the spike delay ring into uint32, and moving viewer JSON formatting into its own process after timing found 97 ms of display preparation in the critical path.
The guardrail was bit-exactness. This model is chaotic at the ulp level: one reference neuron crosses threshold by 7.6 × 10⁻⁶ mV after a hundred ticks of pure leak, and a single flipped spike diverges the whole network by 21% of total spike count within ten thousand ticks. So every optimization had to reproduce the reference second exactly — all 35,113 spikes over 9,990 ticks, every timestamp. It caught real bugs, including one where strict-float mode silently disabled tinygrad’s parallel compilation, so the spawned workers lost the no-FMA renderer patch.
What’s missing: the upstream constants were fitted on the FlyWire female brain and reused unchanged on MaleCNS, and the pack keeps only acetylcholine, GABA, and glutamate edges. That drops histamine — 7,891 neurons, including every photoreceptor — which is fast ionotropic chloride with a known sign and should have been kept. It’s also why the rig injects vision one synapse downstream, at the lamina, rather than at R1–R8.
Hardware in, neurons out
This is the part I came for. Every input drives specific MaleCNS neurons by body ID, not an abstraction of them.
| Population | Cells | Hardware source |
|---|---|---|
| L1 / L2 / L3 lamina | 5,327 | camera luma: ON contrast, OFF contrast, sustained brightness |
| ORN antennal lobe | 2,635 | Wi-Fi scan — SSID hashed to 1 of 53 glomeruli, RSSI sets the rate |
| JO Johnston’s organ | 672 | IMU accel/gyro, mic sound pressure, wheelspeed as wind |
| campaniform sensilla | 426 | motor strain: throttle commanded while the IMU reports no motion |
| PFN goal vector | 456 | Wi-Fi bearing, injected as a goal direction |
| TRN thermoreceptors | 25 | /sys/class/thermal — the SoC’s own board temperature |
| BM_Vib bristles | 22 | IMU jolt, as nociceptive drive |
| OCG ocelli | 22 | mean luma of the top camera rows, i.e. sky |
| sweet GRN | 17 | charging state — docking is eating |
Car hardware writes the fly's sensory periphery and is read back at its motor periphery. Camera luma pooled onto the ommatidial lattice drives the lamina's ON, OFF, and luminance channels; a Wi-Fi scan selects an antennal-lobe glomerulus and its signal strength sets the firing rate; the IMU, microphone, motor strain, battery, and board temperature drive Johnston's organ, sensilla, bristles, and insulin cells. Vision continues through the medulla and lobula loom detectors into the central brain, which is untouched connectome. Descending-neuron spike rates are read out as throttle, reverse, and an escape latch that drives the wheels, while ventral-nerve-cord wing motor neurons drive the car's speaker.
The retina is the mapping I’m happiest with. Camera luma is pooled onto the eye’s real ommatidial hex lattice, parsed out of the MaleCNS annotations, so each of ~530 columns per eye drives its own L1, L2, and L3 cell. L1 and L2 are transients that decay toward zero — the lamina’s actual band-pass behavior. Hold the camera still and the scene fades to its L3 floor; wave a hand and it blooms a bright trail that then vanishes.
A few of these do what the tissue does; most don’t, and the distinction is the credibility of the whole project, so I sorted every mapping into three labeled classes. Mechanism: the hardware performs the same transduction as the tissue. Homolog: the sensor measures the same physical quantity the organ measures — a camera and an eye both measure photons. Substitution: a different modality standing in for one the car doesn’t have.
Both hemifield eyes on the rig's real ommatidial sampling lattice, each shown on the three lamina channels. Drag the synthetic dark disc across the combined field: the sustained channel tracks brightness, the ON and OFF channels light only at the disc's moving edges, and one eye loads far harder than the other whenever the disc sits off centre.
View source values
| Channel | Driven by | In the figure |
|---|---|---|
| L3 · sustained luminance | absolute camera brightness, held while the scene is still | the disc as a standing dark patch; the only channel a motionless scene leaves lit |
| L1 · ON transient | positive frame-to-frame change, decaying toward zero | a bright leading edge ahead of the disc, and nothing at all once it stops |
| L2 · OFF transient | negative frame-to-frame change, decaying toward zero | a trailing edge behind the disc, mirroring L1 |
| Disc speed at zero | the two compared frames become identical, so the change term is exactly zero | both transient panels go silent while the sustained panels hold — the band-pass behaviour that makes a parked car nearly blind |
| Left eye lattice | 884 L1 neurons across 531 sampling cells, pooled 2x2 for display | 145 marks of measured sampling coverage; the sheared band is the hex lattice’s own shape, not an eye silhouette |
| Right eye lattice | 892 L1 neurons across 528 sampling cells, pooled 2x2 for display | 143 marks; one wide camera split left-half and right-half gives the two hemifields |
| Channel | Driven by | In the figure |
|---|---|---|
| L3 · sustained luminance | absolute camera brightness, held while the scene is still | the disc as a standing dark patch; the only channel a motionless scene leaves lit |
| L1 · ON transient | positive frame-to-frame change, decaying toward zero | a bright leading edge ahead of the disc, and nothing at all once it stops |
| L2 · OFF transient | negative frame-to-frame change, decaying toward zero | a trailing edge behind the disc, mirroring L1 |
| Disc speed at zero | the two compared frames become identical, so the change term is exactly zero | both transient panels go silent while the sustained panels hold — the band-pass behaviour that makes a parked car nearly blind |
| Left eye lattice | 884 L1 neurons across 531 sampling cells, pooled 2x2 for display | 145 marks of measured sampling coverage; the sheared band is the hex lattice’s own shape, not an eye silhouette |
| Right eye lattice | 892 L1 neurons across 528 sampling cells, pooled 2x2 for display | 143 marks; one wide camera split left-half and right-half gives the two hemifields |
Two things in that figure matter to the rest of the story. The first is why a still scene goes quiet: hold the disc anywhere and the ON and OFF panels fade while only the sustained channel survives, which is the lamina’s real band-pass behaviour and the reason a parked car sees almost nothing. The second is the asymmetry. One wide fisheye is split left-half and right-half into two eyes, so anything off centre loads one side harder than the other — and that difference, a few hundredths of a unit wide, is the entire signal a steering readout downstream ever gets.
Wi-Fi-as-smell is a substitution, and the fun one. The comma has no chemical nose, but it has a radio one. Each SSID hashes to one of the 53 real antennal-lobe glomeruli and its RSSI sets that glomerulus’s drive, so driving between access points shifts the smell of the room across the antennal lobe. A 2×2 directional antenna makes it a gradient: the two chains are the fly’s two antennae, and the delta between them is a bearing.
Reading the motor neurons back out
Nothing is fed back into the simulation — every output is a readout of descending-neuron spike rates, which keeps bit-exactness intact. Throttle is DNp09 (forward walking) minus MDN (the moonwalker, backward walking). Loom detectors LC4 and LPLC2 feed the DNp01 giant fibre, which latches a full-reverse escape. The bump reflex turned out to be nearly a dedicated line: driving the 22 BM_Vib vibration bristles at 200 Hz produces bilateral MDN bursts at 11.5 Hz with every other descending neuron silent, versus 2 Hz through the whole 672-neuron Johnston’s organ. A sharp accelerometer transient becomes a hard reverse ~300 ms later, and it fires reliably on the real car.
Plotting those populations on their real cell-body positions makes the shape of the thing obvious: everything the car does arrives through a few hundred cells at the bottom of a 166,700-neuron brain, and the wheel commands come through twelve of them.
Measured soma positions for the four populations the rig reads: the lobula detectors that report looming walls, the central complex steering cells, the twelve descending neurons carrying throttle and reverse and escape, and the wing motor neurons wired to the car speaker. Scroll moves attention down the funnel; every other soma is the untouched connectome.
View source values
| Population | Reads as | Plotted | In the rig |
|---|---|---|---|
| Lobula detectors — LC4, LPLC2, LC15 | objects and looming walls; feeds the swerve term and the escape latch | 437 | 437 |
| PFL3 — central complex steering output | heading against the Wi-Fi goal direction; the late, opt-in steering path | 24 | 24 |
| Descending neurons — DNa01, DNa02, DNp09, MDN, DNp01 | throttle from forward walking, reverse from the moonwalker, escape from the giant fibre | 12 | 12 |
| Wing motor neurons — DLM/DVM, b1–tp, TTMn, dMS2 | wingbeat and song; their missing partner is muscle, and the speaker is that partner | 78 | 80 |
| The rest of the measured atlas | untouched connectome, shown as a decimated background sample of every 23rd body ID | 6,057 | 139,473 |
| Sensory afferents — lamina and olfactory receptor neurons | not drawn: camera luma reaches 5,327 lamina cells and the Wi-Fi scan reaches 2,635 receptor neurons, but their somata sit outside the imaged volume | 0 | 7,962 |
| Population | Reads as | Plotted | In the rig |
|---|---|---|---|
| Lobula detectors — LC4, LPLC2, LC15 | objects and looming walls; feeds the swerve term and the escape latch | 437 | 437 |
| PFL3 — central complex steering output | heading against the Wi-Fi goal direction; the late, opt-in steering path | 24 | 24 |
| Descending neurons — DNa01, DNa02, DNp09, MDN, DNp01 | throttle from forward walking, reverse from the moonwalker, escape from the giant fibre | 12 | 12 |
| Wing motor neurons — DLM/DVM, b1–tp, TTMn, dMS2 | wingbeat and song; their missing partner is muscle, and the speaker is that partner | 78 | 80 |
| The rest of the measured atlas | untouched connectome, shown as a decimated background sample of every 23rd body ID | 6,057 | 139,473 |
| Sensory afferents — lamina and olfactory receptor neurons | not drawn: camera luma reaches 5,327 lamina cells and the Wi-Fi scan reaches 2,635 receptor neurons, but their somata sit outside the imaged volume | 0 | 7,962 |
Descending-neuron spike counts pass through a per-group exponential average into a control law producing throttle, steering, and an escape latch. That result is published at 10 Hz on an event stream, read over localhost by the control bridge, and sent as drive frames at 20 Hz to the car control server, which issues serial commands to the motor board and the wheels. Two independent stops protect the car: the bridge watchdog commands zero speed on a stalled stream, and the server halts the car when no frame arrives within 500 milliseconds.
The mapping I’d defend hardest is the speaker. motor_map carries the ventral-nerve-cord motor neurons themselves — the DLM/DVM wingbeat oscillators and the b1–tp steering muscles — and they drive the comma’s speaker directly. Fly song and fly buzz are wing vibration, so this one is mechanism rather than metaphor. A speaker converting wing motor neuron activity into sound performs the same transduction the wing does.
It measures as a real signal: 0.5-second renders go from 0.231 flight at rest to 0.698 under full-field flicker, about three semitones and +31% level. The probe also told me which circuit it came from — Mi1, T4, and T5 all read zero, so it doesn’t arrive via motion detection. It’s loom → LC4 → giant fibre → jump muscle plus wing power: the escape-takeoff program, reached by the correct route.
The split-brain rig
The fly needs the eGPU, the eGPU needs wall power, and the toy car can’t carry wall power. So the two roles live on two commas over Wi-Fi.
The car node carries a comma 4 on the toy chassis with the front fisheye camera, an eye streamer, a sensor streamer, a Wi-Fi scanner, the car control server, the motor board, the wheels, and the speaker. A separate antenna node polls a directional bearing. Both reach the compute node through a lossy access point. The compute node is a second comma 4 with the chestnut eGPU and holds the fly simulation, the GPU itself, the event stream, the control bridge, and the voice conductor, with both stream readers reading over localhost. Drive frames and sound commands travel back across the same access point to the car's control server and sound player.
Both event-stream readers live on the compute node rather than the car, because the AP loses enough packets that a long-lived car→compute read stalls while localhost never does. Every UDP path is latest-state only and every stream has a timeout, so a sense the fly has lost reads as silence, never as a frozen value.
Pain, and why the fly has a save file
The bit I didn’t plan and ended up liking most: the noxious channel has sensitization. Impact jolt, sustained wall-pushing strain, and board temperature past the TrpA1-equivalent threshold feed a phasic sensitivity that decays over 30 seconds, plus a chronic floor that ratchets up with injury history and never decays. A car that has been crashed repeatedly flinches harder and sooner than a fresh one, and because the sim divides pain by the neuropeptide-F tone, the same knock plays softer when the fly is doing well.
The chronic floor, the learned odor valences, and the biography — first boot, accumulated running seconds, lifetime injury count — persist to disk, and redeploys don’t reset them. There is deliberately no reset flag: killing this particular fly takes ssh and rm.
What was the brain, and what was me
Here’s the part a demo skips. Of 166,700 neurons, 11,753 receive hardware input or are read out to hardware — 7.05% of the brain.
7.05%
Neurons touching hardware
9,602 sensory in, 1,780 motor out, 371 instrumented but mute.
The other 93% is untouched connectome doing whatever the wiring makes it do. That’s the pitch, and I still like it. We inject at the sensory periphery and read at the motor periphery, and everything in between is the fly. It is also only true for throttle, reverse, and escape. Steering is where it falls down.
DNa02 is the real steering descending neuron, and it’s a pair: one cell per side. At the 0.2-second windows the control loop runs on, a “5 Hz” single-neuron readout is one spike, and a reproducible signed asymmetry per turn direction never materialized. Meanwhile LC4, the lateralized object detector that shows clean asymmetry under a synthetic bar, pegs at full deflection on a rich fisheye view of a real room and swamped everything else.
So the rig that drove around the office ran with visual fixation disabled and steered from a composite I wrote by hand: a two-sine wander term to explore, a graded turn-away-from-wall term gated on loom rate, and a direct RSSI-gradient bias. The descending neurons were live, published, and visible in the viewer — and not steering. That’s the derp: a control-loop problem dressed as a neuroscience result.
Late on the second day I got a real version working. The Wi-Fi bearing injects into the PFN goal-vector neurons, the central complex does its thing — EPG, PEN, hΔ, Δ7 — and steering comes out of PFL3, its actual steering output, with a slow baseline removed for the resting asymmetry. Signs were correct both ways: AP on the right gave +0.66, AP on the left −0.38. It shipped opt-in behind a flag — tested, but not the configuration that had been driving all afternoon.
Two more for the record. The reward system is instrumented but mute: the dopaminergic, octopaminergic, insulin, and neuropeptide-F populations collectively receive 212,836 synapses and have out-degree zero, because the pack only signs fast ionotropic transmitters. For a GPCR acting over hundreds of milliseconds that’s arguably correct, but it means the pleasure and pain visible in the viewer are carried by scalar laws beside the connectome, not by it. And there’s an open contradiction I never chased: in the last longer-window probe, loom measured 0.00 at every station including the giant fibre, against an earlier run that found it at 15–40 Hz. Loom is the stimulus that should drive that pathway hardest.
The one synapse a fly can change
There was a second, separate track, and it is the one place anything gets trained. The mushroom body is where a real fly’s plasticity lives: 4,064 Kenyon cells fan odors out into a sparse random code, 97 MBONs read that code as approach-or-avoid, and dopamine from PAM and PPL1 depresses exactly the KC→MBON synapses that were just active. So I trained gains on 1,797 of the 61,210 measured KC→MBON edges and nothing else — no edge added, no sign flipped, no zero made non-zero, nothing outside the mushroom body touched. On held-out arenas with a late, sparse radio signal, the starting policy managed 47 of 64 arrivals and the best trained run 62, 59 of them contact-free; against randomized radio lag it held 56 of 64, all contact-free. By the end of the hackathon that policy was also driving the car. It is still a population-rate surrogate rather than the spiking connectome, so it is a second model that learned to drive, not evidence that the connectome learned anything.
What two days bought
A real connectome, simulated whole and bit-exact, faster than real time, on an AMD GPU driven over USB from a dashcam, its sensory periphery wired to cameras and radios and its motor periphery wired to wheels and a speaker. It crawls, it flinches, it remembers which networks hurt it, and it does not shut up.
And on the connectome rig, the wheel that turns toward the Wi-Fi is turned by four lines of arithmetic I wrote, with 155,000 neurons of Drosophila melanogaster buzzing along beside them, having no say in the matter. Which is, if nothing else, an accurate model of a fly.