Neuronpedia

EXPLANATION TYPE

oai_token-act-pair

Description

OpenAI's Automated Interpretability from paper "Language models can explain neurons in language models". Modified by Johnny Lin to add new models/context windows.

Author

OpenAI

URL

https://github.com/hijohnnylin/automated-interpretability

Settings

Default prompts from the main branch, strategy TokenActivationPair.

Recent Explanations

The neuron detects programming source code (code blocks and code tokens/structures) within the document.

gpt-5-mini

= 0;↵↵function createGrid(rows,

GEMMA-3-12B-IT

41-GEMMASCOPE-2-RES-65K

INDEX 478

References to the assistant or model itself—questions about its capabilities, access (APIs/downloads), or self-descriptions.

gpt-5-mini

Generate text, translate languages, write different kinds of creative

GEMMA-3-12B-IT

41-GEMMASCOPE-2-RES-65K

INDEX 54

references to the internet and digital communication, such as online platforms, social media, websites, and web activity

gpt-5

anything you want with the click of a computer mouse.

LLAMA3.1-8B-IT

11-RESID-POST-AA

INDEX 21635

the opening of procedural/code blocks in hardware description language snippets (e.g., starts of “begin … end” sections).

gpt-5

if (r) begin↵ y <= a

LLAMA3.1-8B-IT

11-RESID-POST-AA

INDEX 74891

document metadata and structural header elements, such as titles, dates, section labels, categories, and other formatting cues at the start of sections.

gpt-5

November?↵↵Wall Street Journal, Saturday/Sunday, September

LLAMA3.1-8B-IT

11-RESID-POST-AA

INDEX 26646

section and subsection headers or other structural separators (like list/category lines) in wiki-style entries.

gpt-5

, Acta Iranica↵↵Category:Living people↵

LLAMA3.1-8B-IT

11-RESID-POST-AA

INDEX 19223

polite feasibility/how-to question phrasing where the writer states they want to know if there is a way to do something.

gpt-5

↵↵I wanted to know if there is any possible way

LLAMA3.1-8B-IT

11-RESID-POST-AA

INDEX 118410

clause-introducing markers in academic prose that link reporting verbs or assertions to their subordinate content.

gpt-5

. Our central argument is that these hierarchical networks generate and

LLAMA3.1-8B-IT

11-RESID-POST-AA

INDEX 127186

phrases that structure or qualify statements—especially negations, question framers, and other discourse markers indicating evaluation or explanation.

gpt-5

. "I mean, how can you stand the boredom

LLAMA3.1-8B-IT

11-RESID-POST-AA

INDEX 24053

mentions of expert consultation or professional opinion used to verify, assess, or explain something

gpt-5

/RL consulted cinematographers to confirm that the video was

LLAMA3.1-8B-IT

11-RESID-POST-AA

INDEX 78878

in-universe lore or flavor-text passages, especially those describing secret/rumored artifacts or leaked manuals with backstory and mystique.

gpt-5

of the manuals purloined by our LegionLeaks hero

LLAMA3.1-8B-IT

11-RESID-POST-AA

INDEX 54775

indicators of the assistant’s response sections, especially header/meta tokens and structured formatting like tables or lists.

gpt-5

LLAMA3.1-8B-IT

11-RESID-POST-AA

INDEX 26778

mentions of ghosts/hauntings and related paranormal or poltergeist-like phenomena, especially unexplained noises, movements, and apparitions in homes or places.

gpt-5

band felt his back being touched even though no one was

LLAMA3.1-8B-IT

11-RESID-POST-AA

INDEX 36101

terms that are domain-specific or technical (often proper nouns, acronyms, or specialized nouns)

gpt-5

Helicobacter pylori-induced gastric cancer.↵DNA

LLAMA3.1-8B-IT

11-RESID-POST-AA

INDEX 1521

end-of-sentence punctuation, especially periods marking sentence boundaries.

gpt-5

as opposed to electrical signals. There exists a need for

LLAMA3.1-8B-IT

11-RESID-POST-AA

INDEX 101207

meta-instructional directives about how to perform or format a task (e.g., requests to search, generate, or follow a specified response format).

gpt-5

response should follow this format↵Configuration Item: ↵Issue

LLAMA3.1-8B-IT

11-RESID-POST-AA

INDEX 51211

technical acronyms and CamelCase-style identifiers (all-caps abbreviations, versioned tokens, and code-like terms) in scientific or programming contexts.

gpt-5

; 95% CI = 3.7-

LLAMA3.1-8B-IT

11-RESID-POST-AA

INDEX 130538

the assistant speaker/header tag in chat-style metadata indicating assistant responses.

gpt-5