Definition
Plain language
Generating one word at a time, with each word depending on what came before.
As stated in the literature
A generation regime where the model predicts each token conditional on the sequence of previously generated tokens.
Also called: autoregressively
Why it matters: It's the generation pattern behind every mainstream LLM and shapes both their strengths — coherent text — and their weaknesses, like compounding errors.
For example, an autoregressive model writes a sentence one word at a time, with each new word chosen given everything it's written so far.
Heard on the show
“… Imagine you want to deploy a model — let's say one of those new multimodal ones that interleaves autoregressive text and diffusion image steps in a single forward pass — and you want to run it on a MacBook. …”Episode 027 — When AI Agents Build the Serving Stack: A Bet on Bespoke Infrastructure