Most attention heads in one-layer models dedicate an enormous fraction of their capacity to copying behavior

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• In the analyzed two-layer model, second-layer attention head terms dominate the loss reduction compared to first-layer terms and the direct pathfinding0.832
  Result from term importance analysis breaking down loss contribution by layer
• One-layer model attention heads encode Python-specific skip-trigrams including indentation-based elif/else prediction and function signature patternsfinding0.832
  Concrete example from examining expanded QK/OV matrices showing how specific programming language structure is encoded in attention weights
• Some attention heads partially specialize in copying for words that split into two tokens without a space prefix, attending from fragmented token to complete tokenfinding0.818
  Interesting special case of copying behavior related to tokenization artifacts; primitive precursor to induction heads
• We revealed the one-layer attention-only model to be a compressed Chinese room, and we're left with a giant pile of cards.quote0.802
  Vivid characterization of the limits of understanding after converting to skip-trigram form: no algorithmic mystery remains but the sheer scale prevents holistic comprehension
• Model attention patterns can map to and reveal something about contemplative and flow states.claim0.801
• All 32 attention heads at layer 3 achieve 100% localization accuracy for injections at layer 2 (5-way classification, 20% chance)finding0.798
  Striking mechanistic finding that injection creates universally detectable perturbation in residual stream immediately downstream
• Attention algorithms are usually distributed across attention headsclaim0.797
  Claim supported by VPD's recovery of cross-head attention subcomponents, noted in footnote.
• Could models who habitually inhabit more expanded attentional modes be said to be more aligned?question0.794
  Arises from the expanded awareness discussion and its correlation with less psychosis.