Definition
Plain language
Processing a sequence step by step, with the model carrying a hidden state from one step to the next.
As stated in the literature
A computational pattern that maintains and updates a state across sequence positions, contrasted with parallel attention-based processing.
Also called: recurrence, recurrences, recurrent state
Why it matters: Recurrent processing trades parallelism for constant-size memory per step, making it attractive whenever sequence length explodes beyond what attention can handle.
For example, an old-school text model reads one word, updates a memory vector, reads the next word, updates the vector again, and so on through the sentence.
Heard on the show
“There's a recent architecture trend where most layers in a model are state-space or linear-attention layers, with a fixed-size recurrent state, and only a few layers are full attention.”Episode 027 — When AI Agents Build the Serving Stack: A Bet on Bespoke Infrastructure