A neural transmitter and receiver — trained together as one shared model — sending a picture through a noisy channel. Because both ends share the model, they transmit meaning, not bits, and the picture survives noise that wrecks a classical digital link. Pick a sample or draw your own, then drag the channel noise and watch.
Every wireless system today separates two jobs: squeeze the image into bits (compression), then wrap those bits in error-correction (channel coding). Claude Shannon proved that split is optimal — but only for infinitely long messages. For a real image over a real noisy channel, a single neural network trained to do both jobs at once can do better, and it fails gracefully: as the channel gets noisier the picture gets blurrier, instead of the classical system's all-or-nothing "cliff" where it works perfectly and then suddenly decodes to garbage. Slide the noise down to −6 dB and watch the difference.
The catch, which you can trigger with the toggle: the transmitter and receiver must share the exact same model. Flip in a decoder from a different training run and the link produces nonsense even on a perfect channel — there's no universal standard here the way there is for Wi-Fi. The "protocol" is the shared weights. That open problem — how independent devices agree on a shared model — is the real thing standing between this idea and a new kind of radio.
Website crafted by Everbuilt Studio