Step inside a transformer FFN layer and see what actually fires. From neuron activation patterns to the key-value memory hypothesis, this visual walkthrough explains how two matrix multiplications and a nonlinearity encode most of what a language model knows.