Can a poem move a fly?

A fly listens. We inspect the traces.

The recording becomes sound data. A published model of antennal mechanics transforms that sound; its vibration envelope stimulates the same 138 hearing neurons used by fly.ai. The MaleCNS connectome stays fixed. No words, meaning, or language model enter this process.

The antenna-to-neuron connection is an engineering assumption. This is a literature-informed hearing model coupled to a simulated nervous system, not a validated prediction of a real fly listening to speech.

Selected nerve cells supply walking, turning, and flight commands. Separate, pretrained flybody controllers coordinate the legs and wings in MuJoCo physics. Our adapter, takeoff and landing rules, and arena steering are additional modeling choices. The fly can move in silence too. These movements are model outputs, not a score for the poem or a prediction of a living fly’s behaviour.

The microphone, cable and speaker stage the listening experiment. The box shows the existing 20 × 16 × 10 cm movement bounds; the fly keeps its body-model size. Sound is imposed equally wherever it moves. Distance, glass reflections and body feedback are not part of the hearing model. The line on the floor and the map show its recorded path from above. Close views hide the glass for visibility.

RESPONSE means what the simulation did. Any interpretation is an inference from those measurements; it does not reveal feelings or understanding.

Reading the display

Reduce visual motion
Use the 2D map instead

Amber points mark recorded spikes in a fixed sample of 12,000 neurons, at their supplied coordinates. For the examples, that spatial view shows one run; the traces show the average and range of four runs. Visitor results use one run and an equal-duration silence control with the same starting state and random seed. That silence is calculated in advance, then checked and shortened to match your recording, so your device only needs to calculate the response to your voice.

The grey trace is the same model run with no sound, for the same duration and starting conditions. The two recordings have different lengths. Their levels are matched with linear gain; pauses and dynamics remain. Visitor recordings use the same target level, lowered if needed to avoid clipping.

The antenna trace shows modeled vibration strength in 20 ms frames, not a resolved animation of every oscillation. The waveform and input traces use separate scales, shared across recordings. Antennal colour follows that modeled strength. Reader gestures are theatrical.

Credits & privacy

Articulated anatomy and pretrained locomotion controllers: flybody (Vaxenburg and colleagues, 2025). Robot voice: Kokoro af_sarah. Human: Denny Sayers, LibriVox, 2006. William Blake’s poem is in the public domain. The staged human is not a likeness of the reader.

Recordings are processed in your browser. We receive no audio, transcript, or result. Model files may be cached. Example links are shareable; private recordings must be saved locally.