Neuronpedia

INDEX 8666

the neuron detects prominent technical topic keywords — especially acronyms, product/model names, and domain-specific terms.

gpt-5-mini

<start_of_turn>user↵what is macro in programming? Explain it

INDEX 9350

This neuron detects prominent topic words or subject tokens (main nouns/proper nouns) that indicate the central subject of a query or document.

gpt-5-mini

philosophers and scientists about Artificial intelligence (AI) is rapidly

INDEX 7732

This neuron detects named entities/proper nouns (people, brands, place names, model names and other capitalized terms).

gpt-5-mini

то такий 2Pac?**↵↵*

INDEX 5849

It detects prominent named entities and salient topic tokens (titles, product names, and other key content words).

gpt-5-mini

best" tactic in Football Manager 2023

INDEX 141914

tokens in assistant messages that offer help or ask the user for more information (requests to share code/details or invitations to continue).

gpt-5-mini

you have some code already, please share it!↵

INDEX 6171

This neuron detects salient topical content words—domain-specific nouns and named entities that carry the main subject matter of a passage.

gpt-5-mini

"face of the Western propaganda" and a symbol of

INDEX 31655

tokens representing numbers and dates (numerical values like years, months, times, counts, and other numeric tokens).

gpt-5-mini

News:**↵↵* **Israel-Hamas Conflict

INDEX 40948

the neuron detects strong affirmative or positive-response tokens—i.e., when the model is asserting agreement or labeling content as positive.

gpt-5-mini

?<end_of_turn>↵<start_of_turn>model↵Absolutely, a Baywatch

INDEX 50796

It detects long verbatim or block‑delimited quoted passages (text inside quotes/triple‑quotes or other context blocks).

gpt-5-mini

Committee (IAEC).↵"""<end_of_turn>↵<start_of_turn>model

INDEX 3202

The neuron detects the "model" (assistant) speaker token—i.e., the start of model/assistant responses.

gpt-5-mini

chess<end_of_turn>↵<start_of_turn>model↵```c↵#include

INDEX 111414

the presence of anger-related words or strong angry emotion (tokens expressing anger/frustration).

gpt-5-mini

a mixture of sadness, anger, disappointment and numbness,

INDEX 13847

text-structuring tokens (headings, section titles, list/item markers, and other formatting/organization cues).

gpt-5-mini

and convenience features are must-haves?↵↵↵↵**

INDEX 10907

the starts of conditional or hypothetical clauses—tokens that begin “if/imagining” style questions or hypothetical statements.

gpt-5-mini

↵↵**In short: If you're writing for

INDEX 244489

The neuron detects emphatic, declarative claims—strong assertions or superlative statements that stress ability, uniqueness, or certainty.

gpt-5-mini

**↵↵"My superpower? Turning chaos into cuteness

INDEX 200216

The neuron detects salient content-carrying words — important task/topic nouns and verbs (i.e., semantically informative tokens).

gpt-5-mini

passes automated tests is automatically deployed to production.↵*

INDEX 3974

Text that gives explicit instructions or directives to the model—especially prompts to answer questions, assume a persona/alter ego, or perform a specified role.

gpt-5-mini

the night. T answers questions with general statements and does

INDEX 76785

It detects evaluative or summary phrases that state overall assessments or qualifiers (comparisons like "relatively", "are", "all") in explanatory text.

gpt-5-mini

/complexity, but all are relatively accessible):**↵↵

INDEX 249537

the presence of the model/assistant role token that marks the model's generated response.

gpt-5-mini

Dragon?<end_of_turn>↵<start_of_turn>model↵## Helicopters in

INDEX 4568

This neuron detects mentions or references to a child character (especially male child/son) in narrative contexts.

gpt-5-mini

by both people."↵↵Leo scrunched up his