This neuron activates on programming keywords, type identifiers, and technical terms in source code, particularly those that define or reference code structures, classes, functions, and data types.

mentions of racism, harmful/discriminatory content, or policy-style refusals explaining why hateful content can't be provided.

This neuron never activates—it doesn’t detect any meaningful pattern (a “dead” neuron).

This neuron detects mentions of dialog slot names together with their values (e.g. “area = east”, “people = 3”).

The neuron activates strongly on words that describe taking away or withholding possessions (e.g. “confiscation,” “footwear,” “defending,” “preoccupied”).

function words that serve as grammatical glue—especially prepositions and auxiliary verb contractions linking actions or clauses

imperative action commands in code or technical instructions, especially those manipulating windows or performing drawing/movement operations.

uncommon subword fragments, especially question-word stems and isolated letter-like tokens.

the model's self-identification — it activates on mentions of the assistant's name / self-introduction.

named entities, especially distinctive proper nouns like company, brand, platform, or person names.

titles/headings and salient domain-specific terms or proper nouns that signal the main topic of a passage.

tokens used in headings or emphasized/important document structure (bold markers, section numbers, dates, and other emphasis/heading tokens).

chat turn-taking structure and the assistant’s opening response markers (role tokens and initial affirmations).

markdown-style section headers and subheadings in outlines, especially bolded headings that end with a colon.

words describing prohibited content types and policy violations on online platforms.

the character sequence "thro" inside tokens (a common subword in medical/biological terms).

Mentions of running external processes or using subprocess/shell commands to execute and capture program input/output.

the neuron responds to technical or scientific content—terms, measurements, and data-heavy/highly specific words found in experimental or domain-specific descriptions.

text written in a robotic/AI persona with formal, protocol-driven technical phrasing, structured acknowledgments, and system-style markers (often including numeric designations).

tokens that occur at the start of a sentence or turn (beginning-of-sentence/turn tokens).