© 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-3-12B
    3. 24-GEMMASCOPE-2-RES-16K
    4. 3401
    Prev
    Next
    INDEX
    Explanations

    **Explaining the Neuron: Focusing on Patterns**1. **MAX_ACTIVATING_TOKENS**: Contains `ELL` multiple times, and `**` which often acts as a placeholder or separator. The most prominent visible token is `ELL`.2. **TOKENS_AFTER_MAX_ACTIVATING_TOKEN**: Contains `arge`, `uto`, `arge`, `ision`, `u`, `O`, `ar`, `n`, `tre`, `right`.3. **TOP_POSITIVE_LOGITS**: Contains words like `นะ`, `ก่อน`, `můžete`, `Objekte`, `reimag`, `elizabeth`, `onboarding`, `crafting`, `Aktivitäten`. These seem diverse, possibly indicating a very general pattern or common English words mixed with other languages.4. **TOP_ACTIVATING_TEXTS**: * "**L**arge **L**anguage **M**odel." - `LLM` stands for `Large Language Model`. This text has `LL` from `Large` and `Language`. `ELL` would be part of `Language`. * "ARIMA stands for **A**uto**R**egressive **I**ntegrated **M**oving **A**verage." - No `ELL`. * "**LLM** stands for **L**arge **L**anguage **M**odel." - Explicitly mentions `Large Language Model`. This is a strong hint.This neuron detects the sequence "ELL" followed by 'arge' to form "LARGE", often within abbreviations like "LLM" (Large Language Model)

    np_acts-logits-general · gemini-2.5-flash-lite

    The neuron strongly responds to isolated uppercase letters (often in sequences spelling out ALL-CAP or stylized words/titles).

    oai_token-act-pair · o4-miniTriggered by @jyhe0408
    New Auto-Interp
    Top Features by Cosine Similarity
    Configuration
    google/gemma-scope-2-12b-pt/resid_post/layer_24_width_16k_l0_medium
    Prompts (Dashboard)
    392,802 prompts, 256 tokens each
    Dataset (Dashboard)
    monology/pile-uncopyrighted
    No Configuration Found
    Embeds
    IFrame
    Link
    Not in Any Lists

    No Comments

    Negative Logits
    aj
    0.58
    every
    0.55
    î
    0.53
    jv
    0.52
    ó
    0.52
    óle
    0.52
    ៖
    0.50
    b
    0.50
    ứ
    0.50
    mainly
    0.49
    POSITIVE LOGITS
     นะ
    0.64
     kita
    0.58
     crafting
    0.58
     datasets
    0.57
     Kita
    0.57
     ก่อน
    0.57
     そして
    0.57
     สาว
    0.56
     Icons
    0.54
     อ่ะ
    0.54
    Activations Density 0.001%

    No Known Activations