Glossary · Term

Gemma

← all terms

Definition

Plain language

Google's family of open-weight smaller language models.

As stated in the literature

Google's series of open-weight language models in the few-billion-parameter range, used widely in interpretability and academic research.

Also called: Gemma-2, Gemma-3, Gemma three, Gemma-3-12B

Why it matters: Open-weight small models like Gemma are what makes most interpretability and academic AI research possible.

For example, researchers studying how transformers store factual knowledge often pick a Gemma 2B model because the weights are open and the size is manageable on a single GPU.

Heard on the show

“Across GPT-4, Llama, Mistral, and Gemma, the hedged prompts reliably got responses that were more readable — that one moves in the women-associated direction in nearly every model-and-document cell.”
Episode 241 — Swapping the Name Did Nothing, But Hedging Moved Every Model

Mentioned in 18 episodes

  1. 241
    Swapping the Name Did Nothing, But Hedging Moved Every Model
  2. 234
    Two Copies of Gemini Cooperated in a Game Where Betrayal Always Pays
  3. 221
    Two Hundred Clean Economics Answers, And a Model That Endorses Race Science
  4. 207
    An AI Graded Its Own Math Test 94 Percent — It Actually Scored 20
  5. 179
    How DeepSeek Made One User Faster Without Slowing Down the Crowd
  6. 173
    The Free Step-Level Grader Hiding in Every RL Training Run
  7. 143
    When a Model Notices You Forged Its Own Words, And Why That Breaks Safety Tests
  8. 140
    When a Reasoning Model Says "Let Me Double-Check" After It's Already Decided
  9. 118
    Why the Best-Aligned AI Models Are the Easiest to Trick Into Producing Harm
  10. 107
    How a Market of Crippled AI Agents Outscored One Unrestricted Model
  11. 074
    How a Fifteen-Hundred-Dollar Training Run Matched Llama and Gemma on Reasoning
  12. 055
    Why LLM Judges Flip Their Verdicts When You Change the Question Format
  13. 038
    How LLMs Get Persuaded: One Attention Head, A Tetrahedron, And A Single Dial
  14. 037
    Why Hallucination Detectors Miss Stale Facts: A Geometric Story About What Models Know But Don't Say
  15. 032
    A Sticky-Note for Every Layer: Letting Transformers Remember What They Were Just Thinking
  16. 013
    Why Search Keeps Rediscovering the Same Workflow, and What That Means
  17. 008
    Why Long-Horizon AI Agents Get Stuck, and a Milestone-Based Fix That Helps
  18. 004
    The Sycophancy Circuit That Survives Alignment Training

Related terms