Glossary · Term

inference-masking

← all terms

Definition

Plain language

A trick where slow bookkeeping happens during the time an AI is already waiting on a network call.

As stated in the literature

A DeltaBox technique that overlaps checkpoint work with the LLM inference round-trip the agent is already blocked on, hiding state-management cost behind perceived idle time.

Why it matters: It turns checkpointing from a visible slowdown into something users don't notice, because it happens during time the agent was idle anyway.

For example, while an agent waits hundreds of milliseconds for a remote LLM to reply, the local system uses that gap to snapshot its working filesystem.

Heard on the show

“The inference-masking trick depends on LLM round-trips staying long.”
Episode 068 — The OS Trick That Makes Tree Search Practical for Coding Agents

Mentioned in 1 episode

  1. 068
    The OS Trick That Makes Tree Search Practical for Coding Agents

Related terms