The Arabic script feature would be effectively invisible if we only analyzed the model in terms of neurons.

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

• The familiarity of the text block renders its conventions almost invisible.claim0.793
  Points out how deeply internalized layout norms become naturalized and unnoticed.
• Arabic feature A/1/3450 has 27 neurons with coefficient magnitude ≥0.1 and three largest coefficients are negative; most correlated neuron responds to mixture of non-English languagesfinding0.764
  Demonstrates that the Arabic feature is not aligned to any single neuron
• Arabic script feature A/1/3450 fires on 81% Arabic-script tokens when active, with 98% specificity at high activation levelsfinding0.752
  Demonstrates activation specificity of the Arabic script sparse autoencoder feature
• In some sense, this is the simplest language model we profoundly don't understand. And so it makes a natural target for our paper.quote0.744
  Articulates why a one-layer transformer with MLP is the appropriate starting target for mechanistic interpretability
• An interplay between causal abstraction and feature geometry deepens mechanistic understanding of language modelsclaim0.733
  Methodological claim about the scientific value of combining causal abstraction with representational geometry analysis
• Our SAEs' features are more interpretable than neurons.claim0.731
  Automated interpretability and specificity ratings show SAE features are clearer than MLP neurons.
• If a text attempts to stand alone, it will almost certainly attract commentary or interference.hypothesis0.727
  Predicts the inevitability of dialogic intrusion upon any statement.
• "We should avoid quotes around mental terms because there is no absolute, binary distinction between it knows and it knows—only a difference in the degree to which a model will be useful."concept0.727
  Core assertion against categorical distinction between genuine and metaphorical cognition; justifies continuous gradualist approach.