The second half of every transformer block: a widening projection, a nonlinearity, and a projection back down.
Broad technical consensus on what this term means. Disagreements that invoke it are about the world, not the word.
After attention has moved information between positions, each position is processed independently by a small network that projects it up to several times its width, applies a simple elementwise function, and projects it back. Most of a model's parameters sit in these two matrices.
Attention decides what to look at. The feed-forward layers are where a great deal of what the model knows appears to be stored.