Glossary · Term

knowledge cutoff

← all terms

Definition

Plain language

The date a model's training data stops, after which it knows nothing about what happened in the world.

As stated in the literature

The point in time past which a model has seen no training data; facts that change after this date become stale, and the model may confidently recall pre-cutoff values as if still current, which is the failure mode temporal-drift probes target.

Also called: training cutoff, cutoff

Why it matters: It matters because users can be misled when a model states outdated facts with full confidence, so knowing the cutoff signals what to double-check.

For example, a model trained through last year may still name last year's champion as the current one, unaware a new season has finished.

Heard on the show

“Before chatting, every participant saw a landing page with a model name, a release year, a knowledge cutoff, and capability scores for reasoning, speed, and creativity.”
Episode 205 — The Same AI, Two Labels: How the Pitch Beat the Product in 162 Sessions

Mentioned in 9 episodes

  1. 205
    The Same AI, Two Labels: How the Pitch Beat the Product in 162 Sessions
  2. 172
    One Bad Token Can Sink a Model's Math, And You Can Delete It
  3. 140
    When a Reasoning Model Says "Let Me Double-Check" After It's Already Decided
  4. 117
    How an Open AI System Verified 672 Hard Math Proofs for Under $300
  5. 076
    Same Model, Organized Differently: How an Agent Architecture Beat Frontier Systems at Research Math
  6. 070
    When Models Know the Answer But Say the Wrong Thing Anyway
  7. 069
    When Smarter Models Forecast Worse: The Hidden Failure Mode in LLM Predictions
  8. 048
    How a 30B Open Model Reached Olympiad Gold With the Right Recipe
  9. 037
    Why Hallucination Detectors Miss Stale Facts: A Geometric Story About What Models Know But Don't Say

Related terms