Gateomatron

We take the live stream of Wikipedia edits and try to predict, for each one, whether a bot or a human wrote it. Doing the predicting: 100 logic gates, each slot a gate picked at random when the grid is built, from AND, NAND, OR, NOR, XOR, XNOR, and the single-input NOT. Each edit goes in as 32 bits (namespace, minor flag, how much text changed, how automated the wiki it landed on usually is, and so on, but not the bot flag and not the username), and the bottom row of gates votes on the answer.

The gate in each slot is fixed for the whole run; all that evolves is the wiring. A mutation rewires one connection at random, and we keep it if the new circuit classifies the last 65,536 edits at least as well as the old one. That runs flat out, thousands of times a minute; the picture just slows down to show you one rewire that stuck. A gate may only read an input bit or an earlier gate, so the wiring stays acyclic. No weights, no gradients, no backpropagation.

The dotted line on the chart is what you would score by always guessing whichever class is commoner. It is there because raw accuracy is a slippery number here: the bot-to-human mix of the feed drifts with the hour, and a circuit that never changes at all still drifts several points with it. Only the gap between the two lines is the circuit’s own doing.

It is a demo of evolvable hardware doing online learning: the stream never ends, so there is no training set and no final model, only a circuit that keeps getting rewired. It is inspired by Adrian Thompson, who in 1996 let evolution rewire a real FPGA until it worked instead of designing it.