© Neuronpedia 2026
    Privacy & TermsBlogGitHubSlackTwitterContact
    Neuronpedia logo - a computer chip with a rounded viewfinder border around it
    Neuronpedia
    Jacobian LensNEW
    Natural Language
    Autoencoders
    MAY 26
    Assistant AxisJAN 26Circuit TracerAPR 26SteerSAE EvalsExportsinterp-engineNEW Community BlogPrivacy & TermsContact
    1. Home
    2. Qwen3-32B
    3. 32-RESID-BATCHTOPK-65K
    4. 40351
    Prev
    Next
    INDEX
    Explanations

    her stories have been chosen...""...His next book is NORMAL...""...He has published several books...""...Her work has received acclaim...""...Day Delivery Available..."This neuron seems to identify possessive pronouns (Her, His) or the verb "has" followed by a noun that typically describes creative output (stories, book, work, books) or related actions (published, next). It also picks up "Day Delivery".Looking at the MAX_ACTIVATING_TOKENS and TOKENS_AFTER_MAX_ACTIVATING_TOKEN:Her storiesHis nexthas publishedHer workseveral booksDay DeliveryThe pattern seems to be about creative works, accomplishments, or specific items, often introduced by a pronoun or "has". The "Day Delivery" is an outlier, but the preceding tokens are "Day" and the following is "Delivery".The most consistent pattern seems to relate to authorship and publications.Let's look for a phrase that captures this."Possessives followed by creative works" - too long, and not just possessives."Pronouns or 'has' followed by noun" - too general.Let's re-examine the MAX_ACTIVATING_TOKENS and TOKENS_AFTER_MAX_ACTIVATING_TOKEN:Her storiesHis nexthas publishedalso hasHer workis ais theseveral booksis anDay DeliveryThe "is a", "is the", "is an" suggest a descriptor follows."has published", "also has" suggest action or possession."Her stories", "His next", "Her work", "several books" suggest creative output.possessive followed by work/books

    np_acts-logits-general · gemini-2.5-flash-lite
    New Auto-Interp
    Top Features by Cosine Similarity
    Configuration
    adamkarvonen/qwen3-32b-saes/saes_Qwen_Qwen3-32B_batch_top_k/resid_post_layer_32
    Prompts (Dashboard)
    16,384 prompts, 128 tokens each
    Dataset (Dashboard)
    monology/pile-uncopyrighted
    No Configuration Found
    Embeds
    IFrame
    Link
    Negative Logits
    æ»´
    -0.09
    满满
    -0.09
    ä¸į足以
    -0.09
    æīĢèĥ½
    -0.09
    çĶļèĩ³æĺ¯
    -0.09
    åıªä¸įè¿ĩæĺ¯
    -0.09
    åĺĺ
    -0.08
    å¼ĤæĢ§
    -0.08
    ä¸Ģçľĭ
    -0.08
     butt
    -0.08
    POSITIVE LOGITS
    oji
    0.10
    Äļ
    0.09
    ellido
    0.09
     Yapı
    0.08
    请æĤ¨
    0.08
     ÙĬÙĪØ¬Ø¯
    0.08
    åľ¨æŃ¤ä¹ĭåīį
    0.08
     Slate
    0.08
    åĽ½åľŁ
    0.08
     Ãĸzellik
    0.08
    Activations Density 0.107%

    No Known Activations