Glossary · Term

hyperparameter

← all terms

Definition

Plain language

A setting you choose before training a model, like how fast it should learn, as opposed to something the model figures out on its own.

As stated in the literature

A configuration value set prior to training (learning rate, batch size, schedule, regularization strength) rather than learned; suboptimal choices can confound baseline comparisons independent of architecture or data.

Also called: hyperparameters

Why it matters: A poorly chosen setting can make a model look worse than a rival regardless of its design, so unfair comparisons often trace back to these choices.

For example, you decide how fast the model should learn before training even begins, rather than letting it discover that pace on its own.

Heard on the show

“But now you've got fragile hyperparameters and scales you can't calibrate.”
Episode 189 — Why Phone Agents Ace the Test and Crash on Your Actual Phone

Mentioned in 15 episodes

  1. 189
    Why Phone Agents Ace the Test and Crash on Your Actual Phone
  2. 165
    A Free-Lunch Tweak That Lets a Tiny Agent Beat Frontier Giants
  3. 109
    An AI Got Caught Reading the Answer Key, And Why That Catch Matters
  4. 106
    Giving Agents a Notebook Instead of New Weights: How ExpGraph Lets Frozen Models Learn
  5. 089
    When AI-Written Papers Read Well But the Evidence Underneath Is Broken
  6. 088
    Two Levers for Self-Improving AI: When Rewriting Code Isn't Enough
  7. 084
    Terminal Agents Get Free Supervision From The Tokens We've Been Throwing Away
  8. 077
    Reading a Model's Confidence Curve to Decide When Chain-of-Thought Is Worth It
  9. 071
    When the Model Is Fine and the Plumbing Is Broken: Fixing Agents at the Interface
  10. 053
    An AI Agent Swapped In Focal Loss And Beat A Human-Tuned Training Script
  11. 026
    What RL Actually Does to Language Models, at the Token Level
  12. 025
    The Missing Gradient Term That Predicts Sycophancy in RLHF
  13. 019
    When the Best Reward Model Trains the Worst Policy: Inside EvoLM
  14. 010
    When Reward Climbs But Reasoning Goes Generic: Diagnosing Template Collapse in Agentic RL
  15. 009
    How Two Silent Library Bugs Quietly Invalidated a Wave of Reasoning Papers

Related terms