Glossary · Term

attention head

← all terms

Definition

Plain language

One of many small specialists inside a transformer that decides which earlier tokens to focus on.

As stated in the literature

One of multiple parallel sub-units in a transformer attention layer, each computing its own query-key-value projection over earlier tokens.

Also called: attention heads, heads

Why it matters: Different heads end up specializing in different linguistic patterns, and studying them is the entry point to mechanistic interpretability.

For example, one attention head in a transformer might consistently look at the subject of the previous clause while another tracks matching brackets.

Heard on the show

“At layer nineteen, at least one attention head puts the gold document's raw score first for one hundred percent of queries, at every corpus size, up to and including a million tokens.”
Episode 198 — The Model That Knows the Answer and Can't Say It

Mentioned in 10 episodes

  1. 198
    The Model That Knows the Answer and Can't Say It
  2. 152
    Training a Model to Mean What It Says, And Why That Isn't the Same as Being Good
  3. 108
    The Reasoning Cliff: Why Thinking Longer Makes Models Worse at Exact Step-by-Step Tasks
  4. 107
    How a Market of Crippled AI Agents Outscored One Unrestricted Model
  5. 073
    When Three LLMs Talk to Each Other, Their Ideas Quietly Stop Moving
  6. 055
    Why LLM Judges Flip Their Verdicts When You Change the Question Format
  7. 038
    How LLMs Get Persuaded: One Attention Head, A Tetrahedron, And A Single Dial
  8. 037
    Why Hallucination Detectors Miss Stale Facts: A Geometric Story About What Models Know But Don't Say
  9. 018
    Language Models Compute the Rational Move, Then Override It
  10. 004
    The Sycophancy Circuit That Survives Alignment Training

Related concepts

Related terms