Glossary · Term

next-token prediction

← all terms

Definition

Plain language

The basic training task of guessing the next chunk of text, over and over.

As stated in the literature

The autoregressive objective over token sequences; it supplies no gradient signal for set completeness or for when a generated list should terminate.

Also called: next-token training, next token prediction

Why it matters: Because this objective only ever rewards the next chunk of text, nothing in training teaches a model that a list is incomplete or that it should have kept going.

For example, given "the capital of France is," the model's whole training task is to make "Paris" the most likely next piece of text.

Heard on the show

“And nothing in ordinary next-token training gives a model a calibrated sense of "I have now covered everything," because there's no training signal for the completeness of a set.”
Episode 233 — Why a Model Can Grade an Answer But Not Write the Answer Key

Mentioned in 7 episodes

  1. 233
    Why a Model Can Grade an Answer But Not Write the Answer Key
  2. 213
    A Model Learned to Control a Robot by Watching Video It Never Acted On
  3. 167
    How Teaching an AI to Predict, Not Act, Made It a Better Actor
  4. 114
    Agents That Rewrite Their Own Weights Instead of Just Taking Notes
  5. 073
    When Three LLMs Talk to Each Other, Their Ideas Quietly Stop Moving
  6. 032
    A Sticky-Note for Every Layer: Letting Transformers Remember What They Were Just Thinking
  7. 022
    Training the Model Spec Directly: An Alignment Lever Aimed at the Say-Do Gap

Related terms