Definition
Plain language
A speed trick for code editing where the model verifies the user's near-copy of the answer instead of regenerating it from scratch.
As stated in the literature
A speculative-decoding variant in which a user-supplied draft (e.g., the original file in a code edit) is verified by the target model in batched passes rather than generated token by token.
Also called: predicted outputs
Why it matters: Verifying a near-correct draft is far faster than generating from scratch, which makes interactive code edits feel instantaneous.
For example, when editing a 500-line file with one line changed, the model verifies the original content as a draft instead of regenerating all 500 lines.
Heard on the show
“Code editing with predicted outputs.”Episode 027 — When AI Agents Build the Serving Stack: A Bet on Bespoke Infrastructure