The paddle on the other side of the net is moved by the same real connectome as the Fly Brain Simulation model on this site — 166,700 real neurons from a mapped fruit-fly brain, simulated live in your browser. It hasn't been trained to play Pong; there's no such thing as training this model on a new task. Instead, the side the ball is heading toward simply triggers that model's own "turn left" or "turn right" stimulation, exactly as if you'd clicked those buttons yourself, and the fly's own real turning-rate output is what moves the paddle — standing on it is an actual scanned fruit-fly body (the same specimen used by that model), rendered in 3D on this table rather than a plain block.

That makes the fly a slow, imprecise, occasionally distracted opponent on purpose — its reactions rise and decay over roughly the timescale real recorded neurons actually take, not the crisp instant reflexes of a game AI. It isn't blind to you either: while the ball is still on your side, it pre-positions using where you are standing, leaning whichever way practice has actually found pays off — a plain positioning habit, tuned the same way as everything else here, not a prediction of your actual intent. You can watch exactly why it moves the way it does in the neuron cloud and readouts on the right, and you're free to paint extra neurons yourself in its brain pane mid-rally to see how that changes its play. On top of the real simulation sits one gamified layer that isn't biology: a "mood" (shown under its score) that rises after a real return or a won point and dips — with a genuine stress stimulation on the actual model — after conceding a point, briefly sharpening or dulling how strongly its turning circuit steers the paddle.

The first rally needs a one-time, roughly 76 MB download of the model's neural weights (cached afterward); everything then runs locally on your own device, nothing is uploaded anywhere. The "Train reflexes" button doesn't touch the connectome itself — real recorded neurons can't be retrained — it tunes, in a fast local simulation, only the handful of numbers our own bridge uses (when to stimulate, how hard to weight the result, how early to start slowing down), then checks the outcome on practice rallies it was not allowed to tune against — keeping the new tuning only if it really did beat the old one — and saves it in your browser so the fly starts practiced next time instead of naive.

Fly body: NeuroMechFly v2 via flygym, Apache License 2.0 — meshes and rigging vendored unmodified, credited in this page's own fly-body/NOTICE.