NANDomatron

A grid of NAND gates with no weights and no gradients. Nothing is trained by backpropagation: a genetic algorithm keeps rewiring connections, usually one at a time, and keeps a change whenever it does not increase the loss. It runs on a small server that everyone watches together, and it never stops.

The task is the live stream of Wikipedia edits: from a few features of each edit, predict whether it was a bot or a human. Inputs enter at the top and flow through the gates to the output row at the bottom. Accuracy is the majority vote of those output gates; the loss the evolution minimises is the binary cross-entropy of the output gates against the true label, over a rolling buffer of recent edits. The lineage is Adrian Thompson's evolved FPGA and Cartesian Genetic Programming: let selection, not gradients, find a circuit that works.