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

    Neuronpedia

    Jacobian LensNEW
    Natural Language
    Autoencoders
    NEW
    Assistant AxisNEWCircuit TracerUPDATESteerSAE EvalsExportsinterp-engineNEWAPI Community BlogPrivacy & TermsContact
    1. Home
    2. Gemma-3-27B-IT
    3. 14-GEMMASCOPE-2-TRANSCODER-262K
    4. 98078
    Prev
    Next
    INDEX
    Explanations

    The neuron detects concepts like "Effectiveness", "memorable experience", and "Impact", often appearing after capitalized words or phrases that signal important themes, and it activates on discussions involving complex or nuanced situations.Let's try again, focusing on the commonality.MAX_ACTIVATING_TOKENS: s, at, disasters, ., the, ,, ', a, **, sTOKENS_AFTER_MAX_ACTIVATING_TOKEN: Effectiveness, me, ., Core, end, a, d, memorable, ImpactTOP_POSITIVE_LOGITS: ... (various non-English words)TOP_ACTIVATING_TEXTS: mentions "Propaganda's Effectiveness", "devastating consequences", "disasters", "loaded and aggressive set of questions", "specific historical and political context", "memorable experience", "AI's Impact", "opportunities".The TOP_POSITIVE_LOGITS list is predominantly non-English. This suggests the neuron might be sensitive to languages other than English, or perhaps it's a general semantic neuron where the English representation is less direct.However, the prompt emphasizes finding patterns *in lists*.MAX_ACTIVATING_TOKENS has "disasters". TOKENS_AFTER_MAX_ACTIVATING_TOKEN has "Effectiveness", "memorable", "Impact".TOP_ACTIVATING_TEXTS has:- "Propaganda's Effectiveness"- "devastating consequences" (related to disasters)- "disasters"- "historical and political context"- "memorable experience"- "AI's Impact"There's a strong theme of impact, consequences, effectiveness, and things that are memorable or significant, often in a context of history, politics, or significant events.Let's consider the non-English tokens. If the neuron activates on these languages, it's capturing something broad. But the activation texts are English.Let's look at the MAX_ACTIVATING_TOKENS again: 's', 'at', 'disasters', '.', 'the', ',', "'", 'a', '**', 's'.And TOKENS_AFTER_MAX_ACTIVATING_TOKEN: 'Effectiveness', 'me', '.', 'Core', 'end', 'a', 'd', 'memorable', 'Impact'.'disasters' followed by something.The neuron detects significant impacts and consequences

    np_acts-logits-general · gemini-2.5-flash-lite
    New Auto-Interp
    Top Features by Cosine Similarity
    Configuration
    google/gemma-scope-2-27b-it/transcoder_all/layer_14_width_262k_l0_small_affine
    Prompts (Dashboard)
    238,145 prompts, 512 tokens each
    Dataset (Dashboard)
    lmsys + oasst1
    No Configuration Found
    Embeds
    IFrame
    Link
    Not in Any Lists

    No Comments

    Negative Logits
     Renew
    0.49
    aran
    0.48
     New
    0.47
     roma
    0.46
     Auch
    0.46
    iasi
    0.46
     arra
    0.46
    Work
    0.45
     appliance
    0.45
     Ad
    0.44
    POSITIVE LOGITS
     кандида
    0.58
     የሚ
    0.54
    ogeneous
    0.54
     নিহত
    0.53
    ーンズ
    0.53
    ။
    0.52
     возможностей
    0.52
    の花
    0.50
    騙
    0.50
    ние
    0.50
    Activations Density 0.000%

    No Known Activations