The other half of a transformer block, sitting after attention. Attention decides which words to look at; this part does the actual processing of what was gathered. It is where most of the model’s parameters live, and increasingly where interpretability researchers think factual knowledge is stored.
Technical view
A position-wise two-layer network, expanding to roughly four times the model dimension and projecting back, with a non-linearity between. Modern models use gated variants such as SwiGLU. It accounts for approximately two thirds of the parameters in a dense transformer, which is why mixture-of-experts designs replace precisely this component with routed experts.
Related
Part of AI Terminology Index, section: How the Machinery Works.