This experience uses Ajinkya Gorad’s recovered TI-46 checkpoint: a 5 × 5 × 5 network with 77 cochlear inputs and a spiking digit readout. Neurons and directed connections correspond to the saved matrices. Only drawing coordinates are relaxed into an organic shape; common axon trunks branch into curved synapses and dendritic endings. The branches are schematic, not reconstructed biological anatomy.
Quest 3: open this page in Meta Quest Browser, enter passthrough and use the floating microphone or sound-demo buttons. The waveform, Lyon cochlea and BSA raster remain visible in your room. Point and trigger / pinch to select a neuron or a floating control. Squeeze to grab the reservoir. Use the right thumbstick to scale it; the left thumbstick rotates it. The size buttons scale the whole workspace from 20% to 300%; Recenter places it in front of you. Small strips beside each cell show its last 2.05 seconds of spikes; output strips are larger. Select an output digit to label the next spoken sample for a personal readout, trained and saved on this device. The original reservoir and TI-46 weights remain unchanged.
Timing: simulation runs at 1 ms; BSA waits for its full 33-sample kernel (32 ms lookahead). Synaptic flashes last 140 ms to make real spikes visible. Lyon’s trained lowpass adds smoothing. Animation speed does not change the neural solver.
Digit recognition is experimental. The recovered weights are real. Live microphones, utterance resets, and speech outside TI-46 differ from training. The sound demo uses synthetic tones and makes no digit predictions. Changing recurrent strength disables the trained-model interpretation. Personal training restores original recurrent strength, gain 1× and the CPU solver. Personal guesses require at least two taught labels and only choose among taught labels.
Compute: WebGL2 instancing and shaders render both eyes. The default sparse solver and compiled WebAssembly cochlea run in a worker. Optional WebGPU compute must pass a CPU comparison on this device before it is enabled. Input audio stays on your device.
Research and implementation notes ↗