Glossary · Term

entropy

← all terms

Definition

Plain language

A number that captures how unpredictable or spread out a set of possibilities is.

As stated in the literature

In information theory, the expected negative log probability of a random variable's outcome; in LLM contexts, often a measure of how peaked or diffuse a next-token distribution is.

Why it matters: It gives a single number for how confident or hesitant a model is, which is useful for sampling decisions, uncertainty estimates, and detecting when reasoning is going off the rails.

For example, a model that says the next word is 'the' with 99% probability has very low entropy, while one that splits 20% across five plausible words has much higher entropy.

Heard on the show

“One is the model's entropy at that token — and entropy is just how spread out the model's bets were.”
Episode 172 — One Bad Token Can Sink a Model's Math, And You Can Delete It

Mentioned in 15 episodes

  1. 172
    One Bad Token Can Sink a Model's Math, And You Can Delete It
  2. 162
    The Empty-Lake Proof: Why More Rollouts Stop Helping Reasoning Models
  3. 154
    How a 7B Model Out-Investigates a 72B One by Choosing What to Look At
  4. 141
    How Two Tokens Reopened a Reasoning Method the Field Had Given Up On
  5. 108
    The Reasoning Cliff: Why Thinking Longer Makes Models Worse at Exact Step-by-Step Tasks
  6. 084
    Terminal Agents Get Free Supervision From The Tokens We've Been Throwing Away
  7. 077
    Reading a Model's Confidence Curve to Decide When Chain-of-Thought Is Worth It
  8. 073
    When Three LLMs Talk to Each Other, Their Ideas Quietly Stop Moving
  9. 070
    When Models Know the Answer But Say the Wrong Thing Anyway
  10. 057
    How Uber Caught 206 Leaked Credentials With an LLM-Powered Security Stack
  11. 048
    How a 30B Open Model Reached Olympiad Gold With the Right Recipe
  12. 037
    Why Hallucination Detectors Miss Stale Facts: A Geometric Story About What Models Know But Don't Say
  13. 026
    What RL Actually Does to Language Models, at the Token Level
  14. 011
    When RL Actually Teaches Agents Something New, And When It Doesn't
  15. 010
    When Reward Climbs But Reasoning Goes Generic: Diagnosing Template Collapse in Agentic RL

Related concepts

Related terms