01→
Give it a view.
Game pixels become a high-contrast visual stimulus.
OBSERVATIONAN OPEN NEUROSCIENCE ARCADE
TINY BRAIN.
BIG FLAP.
166,700 mapped neurons. One very difficult game. An experiment in turning a fly’s neural activity into a well-timed flap.
Full map verified. First action decoders trained.
SPECIMEN 001D. MELANOGASTER
166,700RETAINED NEURONS
25.58MWEIGHTED CONNECTIONS
1BUTTON TO FIGURE OUT
Full map. Standard CPU.See the benchmark ↗
01 / THE ARCADE
NO COINS REQUIREDSTAY CURIOUS.
STAY AIRBORNE.
Same gravity. Same gaps.
Let’s see how you do.
FLIGHT DECK / 01SCRIPTED PREVIEW
SEED 0019
P TO PAUSE
BUILT FOR CURIOSITY. OPTIMIZED FOR “ONE MORE TRY.”30 FPS / DETERMINISTIC PHYSICS
02 / LEARNING IN THE OPEN
GOOD SCIENCE KEEPS THE MISSES
EVERY FLIGHT
TEACHES US SOMETHING.
The learning pilot trains an action decoder from full-network activity. The mapped connection weights stay fixed.
DECODER PILOTRECORDED
Mean pipes cleared on three unseen layouts.
| CONTROLLER | PIPES | FRAMES |
|---|---|---|
| Results are loading. Raw reports are on GitHub. | ||
The scripted teacher sees privileged game state. Three layouts are an early check; they don’t establish a biological advantage.
Inspect the complete report
THE DECODER’S FIRST TEST FLIGHT RECORDED
Loading recording…
0 / 0
Actual evaluation, re-rendered in the arcade style. Always the first test seed, whatever the score.
03 / UNDER THE EXOSKELETON
BIOLOGY → COMPUTATION → FLAPREAL WIRING.
NEW QUESTIONS.
A connectome is a map of connections. We use the retained MaleCNS network inside an approximate simulator, then teach a small decoder to act on its activity.
02→
Meet the retina.
A documented adapter drives mapped photoreceptors.
MODELED INTERFACE
03→
Run the full map.
166,700 retained neurons. Mapped weights held fixed.
SIMULATED NETWORK
04↗
Learn the flap.
A small action decoder learns when to flap or wait.
TRAINABLE READOUTThe current pilots use imitation learning. Reliable play and a benefit over simpler networks are still open questions. Read the method ↗
04 / SHOW YOUR WORK
LESS HYPE.
MORE DATA.
The complete retained map ran on a standard GitHub CPU runner. These are measured responses to six saved game frames.
0.91 GiBPEAK NEURAL-PROCESS RAM
27.4 msPER 10 MS NEURAL WINDOW
September 11, 2026 · six samples after 100 ms warmup · learning disabled for this runtime benchmark
RUNTIME / SIX RECORDED SAMPLESVERIFIED
Wall time per neural windowMS
- Wall time
- 24.4 ms
- Spikes
- 6,052
- Active neurons
- 6,028
Choose a bar to inspect its exact model input and recorded response. These measurements are not live game telemetry.
THE LAB DOOR IS OPEN
MAKE THE
NEXT DISCOVERY.
Try the game. Question the method. Reproduce a run. There’s a lot left to figure out.
Get your hands on the code
~/flappy-fly
git clone https://github.com/jackspiece/flappy-fly.git
cd flappy-fly
python -m http.server 8000 --directory docs
Then open localhost:8000.
CURIOUS? GOOD.
A FEW
FLY QUESTIONS.
Wait. Is this a real fly-brain map?
Yes. The wiring comes from MaleCNS v1.0, released by HHMI Janelia, Google Research, and collaborators. We retain 166,700 neurons under the upstream inclusion policy. The simulator and its interfaces are approximate models; the map alone does not recreate a living fly.
What’s actually learning?
A small action decoder learns from the full network’s activity. The mapped connections stay fixed. Every pilot records the trained parameters, evaluation seeds, baseline scores, and a checkpoint.
Is the fly controlling the game on this page?
The arcade lets you play or watch a scripted demo. The separate recorded flight shows a real decoder evaluation. The full simulator runs in Python and C++ on a GitHub runner.
Is fly wiring better than regular AI?
We don’t know. Establishing an advantage needs reliable play, many more test layouts, and matched comparisons with simpler and shuffled networks. That’s part of what makes the experiment interesting.