© Neuronpedia 2026
    Privacy & TermsBlogGitHubSlackTwitterContact
    Neuronpedia logo - a computer chip with a rounded viewfinder border around it

    Neuronpedia

    Natural Language
    Autoencoders
    NEW    Assistant AxisNEWCircuit TracerUPDATESteerSAE EvalsExportsAPI Community BlogPrivacy & TermsContact
    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 fires on mentions of creating directories or folders (e.g., “mkdir,” “folder,” “os.mkdir,” etc.).
    o4-mini
    str):↵    os.mkdir(os.path
    Neuronpedia logo
    GEMMA-2-27B
    34-GEMMASCOPE-RES-131K
    INDEX 2238
    The neuron activates strongly on the word “very,” i.e. it detects the intensifier “very.”
    o4-mini
     eating in Broward County a very long time. Back in
    Neuronpedia logo
    GEMMA-2-27B
    34-GEMMASCOPE-RES-131K
    INDEX 102643
    The neuron activates on and thus identifies numeric characters (digits) wherever they appear in the text.
    o4-mini
    1), "MM-DD
    Neuronpedia logo
    GEMMA-2-27B
    22-GEMMASCOPE-RES-131K
    INDEX 54248
    The neuron fires on words that are part of named entities—especially organization, place, or proper‐name tokens.
    o4-mini
     to 8 this evening in Harold M. Zimmerman and
    Neuronpedia logo
    GEMMA-2-27B
    22-GEMMASCOPE-RES-131K
    INDEX 54229
    The neuron detects the numeric page‐reference (the specific page numbers) in legal reporter citations.
    o4-mini
    E.2d 578↵CAPITOL
    Neuronpedia logo
    GEMMA-2-27B
    22-GEMMASCOPE-RES-131K
    INDEX 54247
    The neuron fires on standalone numeric literals (multi-digit numbers) in code.
    o4-mini
    ↵    .hljs {↵      display: block;
    Neuronpedia logo
    GEMMA-2-27B
    22-GEMMASCOPE-RES-131K
    INDEX 54239
    The neuron activates on numeric tokens (including integers and decimal numbers) in the text.
    o4-mini
    }\↵    [ $(^3)$ Thales Alenia
    Neuronpedia logo
    GEMMA-2-27B
    22-GEMMASCOPE-RES-131K
    INDEX 54244
    This neuron is keyed to code delimiters—especially semicolons that mark the ends of statements.
    o4-mini
    hbase.filter.Filter.ReturnCode;↵import
    Neuronpedia logo
    GEMMA-2-27B
    34-GEMMASCOPE-RES-131K
    INDEX 73589
    The neuron selectively fires on numeric or quantitative tokens (e.g. numbers and measurement‐style entries) in the text.
    o4-mini
    |              | 123child1 | 
    Neuronpedia logo
    GEMMA-2-27B
    22-GEMMASCOPE-RES-131K
    INDEX 54223
    This neuron detects coordinating conjunctions that link paired items in enumerations (e.g. noun-and-noun lists).
    o4-mini
     provider of high technology products and services to industrial companies and
    Neuronpedia logo
    GEMMA-2-27B
    34-GEMMASCOPE-RES-131K
    INDEX 54080
    This neuron flags section‐header or metadata markers in code or documentation (e.g. headings, bracketed labels, double‐underscore tokens).
    o4-mini
    ↵%↵% Options::↵% 'full'
    Neuronpedia logo
    GEMMA-2-27B
    34-GEMMASCOPE-RES-131K
    INDEX 80965
    The neuron strongly activates on code-like tokens (annotations, keywords, identifiers, and numeric literals) rather than ordinary prose.
    o4-mini
    station="second parameter" ↵daysForward="third
    Neuronpedia logo
    GEMMA-2-27B
    22-GEMMASCOPE-RES-131K
    INDEX 54209
    The neuron strongly activates on sequences of uppercase letters (all‐caps words or acronyms).
    o4-mini
    else if (isupper(*s)) digit = ch
    Neuronpedia logo
    GEMMA-2-27B
    22-GEMMASCOPE-RES-131K
    INDEX 54232
    This neuron activates on expressions of elapsed time or duration (e.g. “for some time,” specific month‐ or year‐counts), especially when a numeric span is given.
    o4-mini
     expect to be so for some time to come. We
    Neuronpedia logo
    GEMMA-2-27B
    34-GEMMASCOPE-RES-131K
    INDEX 60643
    the presence of numeric statistical data (numbers, percentages, years, counts) in the text.
    o4-mini
     5% to approximately 15% (Figure
    Neuronpedia logo
    GEMMA-2-27B
    34-GEMMASCOPE-RES-131K
    INDEX 103683
    The neuron is primarily responding to numeric tokens (years, volume and page numbers, reference numbers) in bibliographic or citation contexts.
    o4-mini
    -28, Soc. Manuf. Eng., Dearborn
    Neuronpedia logo
    GEMMA-2-27B
    34-GEMMASCOPE-RES-131K
    INDEX 81475
    The neuron strongly activates on tokens that are part of named entities or numerical identifiers (e.g. proper names, titles, years, place names).
    o4-mini
     (such as “Robin Hood” and “Ironclad
    Neuronpedia logo
    GEMMA-2-27B
    34-GEMMASCOPE-RES-131K
    INDEX 118109
    The neuron strongly activates on the adverb “quite,” detecting that specific intensifier.
    o4-mini
     most dynamic mix, but quite entertaining if you like electric
    Neuronpedia logo
    GEMMA-2-27B
    22-GEMMASCOPE-RES-131K
    INDEX 54236
    The neuron fires on words that denote connecting or bridging between two places or entities.
    o4-mini
     the Bifrost is what anchors Asgard and Midgard
    Neuronpedia logo
    GEMMA-2-27B
    22-GEMMASCOPE-RES-131K
    INDEX 54218
    The neuron strongly activates on fragments of proper names (and other tokens) beginning with “Pal.”
    o4-mini
    .↵↵Samir Palnitkar↵↵President↵↵Serial
    Neuronpedia logo
    GEMMA-2-27B
    34-GEMMASCOPE-RES-131K
    INDEX 107551