Fly Invaders

Everything we call "intelligence" right now is artificial: frozen weight matrices trained on scraped text, running as arithmetic on silicon, with no biological substrate, no ion channels, no membrane dynamics. Impressive, but it is statistics all the way down and nobody can point at a neuron in it. So we went the other direction and made some bio shi: a real connectome (MaleCNS Fruit Fly Brain, ~167k neurons, ~25M edges) simulated as leaky-integrate-and-fire, driven closed-loop at frame rate. Bilinear luma sampled at retinal receptor UVs goes in over a WebSocket; DNp20 (steer) and DNpe017 (fire) firing rates come out, lockstep at 167 x 0.1 ms neural steps per frame (40 ticks/s). Cells actually integrate, spike, and refractory stimuli like real brain does . The demo is a fly brain playing Space Invaders, but the same loop is one wiring change from a drone, where the bio brain goes brr. [DEPLOYMENT IS LAGGY DUE TO CHEAP VPS UNABLE TO LOAD ALL 167K NEURONS RUN LOCALLY OR CHECK VIDEO :D]

What it is? Fly Invaders is a live fruit-fly brain connectome yanked running as a controller. It is not a trained model, a simulation of real biology: Fruit Fly Male (~167k neurons, ~25M edges) as leaky-integrate-and-fire, where cells integrate, spike, and go refractory. One WebSocket in, one action out.

What it does? Sensory input goes in, behavior comes out. A game frame is sampled at the fly's retinal receptors, stepped through neural time, and decoded from real descending-neuron firing: DNp20 steers, DNpe017 fires. Space Invaders is just the harness that makes the loop visible and verifiable, frame-locked, on one vCPU.

The direction? Reflex, not reasoning. A fly brain is the cheapest complete sensorimotor loop nature ships: fast, low-power, robust, and it doesn't need a GPU. Swap the game for a camera and the descenders for motor channels and you have a flight controller where the policy is a connectome rather than an MLP. Plug the fly brain into a drone and let it go brr.