The neuron detects positive subjective evaluation or praise—tokens expressing favorable sentiment or admiration.

This neuron lights up on informal, interactive bits of user comments—especially question marks and small reaction/interjection tokens (e.g. “back,” “wow,” “now?”) that signal a conversational or reactive utterance.

tokens that indicate personal emotional reactions or expressive interjections (worry, excitement, shock, or emphasis).

long, multi-sentence assistant responses or explanatory/system-generated text.

A strong detector for sudden, emphatic exclamations or high-intensity emotional interjections (loud reactions, urgencies, and similar bursty dialogue).

the token "voxel" (references to voxel variables in code).

Finds assistant safety/policy language — refusals, disclaimers, and explanations about prohibited content or why the model can't comply.

The neuron detects key content nouns and technical/quantitative terms (domain-specific entities, measurements, statuses and similar important words).

sentence-initial expletive/subject pronoun "it" (including contractions like "it's") used to start or emphasize clauses.

the neuron detects important content words — the meaningful nouns, verbs, and adjectives that carry the main information in a sentence.

Finds prominent topic or heading tokens — the words that start or label important sections or summary sentences.

the neuron detects tokens that are parts of string literals (text inside quotes) in code/examples.

sentences or tokens containing mathematical notation, equations, or formal math problem statements.

phrases and tokens that mark the start of an assistant’s explanatory or organizing reply (e.g., "Okay", "Here", section headers and similar discourse markers).

The neuron primarily detects numeric tokens and numeric/legal citation-like sequences (i.e., numbers and number-heavy references).

The neuron detects text about performing model inference or instructions for running inference (discussions of carrying out inference).

tokens representing numeric values (numbers/digits) and numeric parameters in code or instructions.

the neuron activates when the model is producing explanatory/corrective output—reformulations, translations, grammar corrections, and related instructional text.

It detects text produced by the model (assistant replies / model-generated turns).

the literal token "model" (appearing in system/meta lines).