Fly Lab / a simulated nervous system
The fly, live from inside.
Follow a simulated fly's movement, sampled neural activity and visual inputs. Choose a published source, then switch between its body, brain and vision.
Connecting to the live body view…
Connecting to the live full-CNS activity…
Waiting for the next neural activity sample.
Drag or use arrow keys to rotate. Scroll or press + / − to zoom; press 0 to reset the fitted view.
Sampled soma anchors from the public presentation layout; activity is the real per-neuron count reported for the interval. Sampled connections are not exported by this viewer.
Make something from the signal
Play the nervous system.
Capture published neural activity, shape it into a musical phrase, and take it into your music software. Reopen a saved project to keep creating without running another fly.
A local project up to 40 KiB. Nothing is uploaded. Imported source identities are declared, not independently verified.
No signals captured. Sound is off.
Capture first, then arrange up to 24 musical beats. Musical repeats do not duplicate neural measurements.
Capture uses the same read-only feed as this brain view. It sends no commands to the fly and starts no extra simulation.
What you are hearing
This is an instrument driven by observed interval counts, not a recording of the sound of neurons. Mean spike rate selects pitch, the active-neuron fraction controls intensity, and the published group-count distribution changes timbre. Glass and Reed are browser synthesis settings, not neural models.
Each selected interval becomes one musical beat. Trimming, repeats and transposition change only the arrangement. Playback remaps time and does not represent biological real time or individual spike timestamps. Missing intervals are reported, not filled in. Up to twelve samples are captured; capture stops after twenty seconds or when the source or view changes.
WAV exports the playback samples as PCM16. MIDI carries the same arranged notes and tempo; your music software supplies its timbre. The editable project retains the original source, graph, mode, declared cadence and counts, separately from sound settings. It is a local record, not a signed experiment receipt or proof of musical understanding. Reopened records never become live measurements.
| Sequence | Neural end, ms | Interval, ms | Spikes | Active neurons |
|---|
Connecting to the live vision feed…
No vision source is currently bound.
Live measurements & source identity
- Source
- Not bound
- Stream identity
- —
- Graph identity
- —
- Mode
- —
- Execution
- —
- Sequence
- —
- Measured RTF
- —
- Sample rate
- —
- Sample age
- —
- Active neurons
- —
- Spikes / interval
- —
- Neural compute
- —
- Site build
28fd17f5b5e5
How to read this view
The simulation uses a complete retained fly nervous system to drive a simulated body. The source selector distinguishes the female BANC and male MaleCNS networks; a source stays unavailable until it is published.
The body panel shows real articulated poses from the live FlyGym physics simulation, drawn as bounded part proxies and the measured joint skeleton. The brain panel samples soma anchors from the viewer's public presentation layout and colours them with the real per-neuron counts reported for that interval; it is not a spike-timestamp recording and no activity is generated in the browser. The vision panel shows the fly's rendered-eye or consumed-input images when the running viewer build serves them, with a source-produced reconstruction when available.
Real-time factor (RTF) is the measured simulation time divided by wall time reported by the live viewer. A full whole-CNS interval is computationally heavy, so this page labels a low RTF honestly as best-effort live rather than claiming real-time playback. When the running viewer declares its physics/controller cadence, the execution row labels a decimated run as fast approximate; the page never infers a mode from timing. Samples expire; the last verified sample is shown until its freshness window closes.
Read-only by construction: the public site worker stores bounded snapshots published outbound by the lab. There is no inbound route, no command channel and no operator data in the public payload.
Compare with Doom AI and Minecraft AI.