About this experiment
Every prediction runs a fresh FlyWire neural simulation with fixed internal connections and weights, artificial input, and a trained external readout. The full simulation contains 139,255 neurons. The activity fraction counts active neurons within the 5,273 monitored KC / MBON readout neurons. Active means at least one spike in the four 50 ms observation windows (200 ms total); the denominator is this monitored subset. The body and laser visualize its answer; they are not neural motor control. Game images come from all 10,000 images in the official MNIST test split, shuffled without replacement within each pass. Reloading starts a new shuffle. The source label only places the image in its lane and checks the answer; recognition receives pixels only. Handwriting example buttons use test images too. Session hits and misses are an interactive sample, not a completed full-test accuracy result. Freehand inputs are separate from MNIST. Digit dataset provenance.
Anatomy: FlyBody · Google DeepMind / HHMI Janelia, Apache 2.0. License · Asset provenance. Rendering: Three.js · MIT。