key technical or proper terms that stand out in structured text (often emphasized in lists, tables, or quotes)

technical or domain-specific vocabulary and key terminology appearing in academic and professional documentation.

Instances where the text refers to the model's identity or system role (system/instruction messages declaring the assistant/AI).

Instances where the assistant gives a self-referential disclaimer describing itself as an AI language model and stating its capabilities/limitations.

instances of the first-person pronoun "I" (self-references by the assistant).

instances of the assistant's standard self‑referential safety/disclaimer phrasing (e.g., "I'm sorry, but as an AI language model...").

the assistant claiming it was developed by Meta AI (developer attribution statements).

Instances where the assistant self-identifies or gives a disclaimer about being an AI (the model's "As an AI ..." style preface).

mentions that the speaker is an AI (e.g., "AI", "language model", "assistant") or self-identifying phrases about being an AI assistant.

instances where the assistant refers to itself as an AI (e.g., "As an AI language model").

phrases where the speaker self-identifies as an AI (assistant saying it is an AI).

mentions of the assistant identifying itself as an AI (self-referential statements about being an AI).

the word "The" at the beginning of sentences in encyclopedic or technical articles.

capitalized letters that begin proper nouns or brand names.

This neuron appears to be malfunctioning or capturing noise rather than finding a meaningful linguistic pattern. The activations are sporadic, scattered across completely unrelated tokens (proper nouns, fragments, punctuation) in diverse text genres (gambling sites, legal documents, coffee shops, sports articles

The main thing this neuron does is find single letters serving as initials or the first character of capitalized words, acronyms, or specific identifiers.

Mentions of database table operations—especially self-joins and queries combining or comparing table rows.

The neuron detects first-person self-reference (speaker-focused pronouns and constructions indicating "I"/the narrator).

tokens that appear in assistant self-descriptions (mentions of "AI"/"language model", training/knowledge cutoff, updates, system time and related limitations).

This neuron detects first-person self-referential words and role/identity declarations (tokens like "I", "I'm", "am" and similar self-identifying phrases).