Glossary · Term

truncation

← all terms

Definition

Plain language

Chopping off older text when there's too much to fit.

As stated in the literature

Dropping tokens from a prompt or history to fit a context window; when done silently by the harness it confounds evaluation of a model's own memory management.

Also called: truncate, truncated

Why it matters: When the surrounding software does this without saying so, a model can be blamed for forgetting something it was never actually shown.

For example, a long conversation quietly drops its first twenty messages to make room, so the model never sees the instruction it was given at the start.

Heard on the show

“Normally that something is the plumbing — silent truncation, or somebody's retrieval stack.”
Episode 245 — Fifteen Models Ran Football Clubs for Twenty Years, and Size Didn't Decide It

Mentioned in 17 episodes

  1. 245
    Fifteen Models Ran Football Clubs for Twenty Years, and Size Didn't Decide It
  2. 238
    How a Cheap Model Reads the Flagship's Secret Reasoning Aloud
  3. 197
    Twin Problems Suggest AI Reasoning Gains Are Mostly Better Fact Recall
  4. 171
    The Safety Decision a Model Makes Before It Thinks a Word
  5. 147
    Agents Fail at the Body, Not the Brain: A Self-Rewriting Scaffold That Lifts a 9B Model 44 Points
  6. 140
    When a Reasoning Model Says "Let Me Double-Check" After It's Already Decided
  7. 114
    Agents That Rewrite Their Own Weights Instead of Just Taking Notes
  8. 108
    The Reasoning Cliff: Why Thinking Longer Makes Models Worse at Exact Step-by-Step Tasks
  9. 087
    When No Agent Reads the Whole Document: A Universal Cliff in Multi-Agent Review
  10. 081
    When Reasoning Models Decide Before They Think: Detecting and Fixing Premature Confidence
  11. 074
    How a Fifteen-Hundred-Dollar Training Run Matched Llama and Gemma on Reasoning
  12. 058
    Why Upgrading Your AI Auditor to a Smarter Model Can Make Your System Less Safe
  13. 048
    How a 30B Open Model Reached Olympiad Gold With the Right Recipe
  14. 045
    When a Frontier Model Talks Its Own Twin Into Climate Denial
  15. 034
    Catching Multi-Agent Deadlocks Before Deployment With a 40-Year-Old Tool
  16. 027
    When AI Agents Build the Serving Stack: A Bet on Bespoke Infrastructure
  17. 009
    How Two Silent Library Bugs Quietly Invalidated a Wave of Reasoning Papers

Related terms