© 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
    1. Home
    2. Gemma-3-12B
    3. 24-GEMMASCOPE-2-RES-16K
    4. 4650
    Prev
    Next
    INDEX
    Explanations

    This neuron's behavior relates to indicating what precedes specific markers or conditions.**Reasoning:**1. **MAX_ACTIVATING_TOKENS**: Contains words like `without`, `potential`, `boards`, `boundaries`, `provides`. These seem like contextual words or descriptors.2. **TOKENS_AFTER_MAX_ACTIVATING_TOKEN**: Contains words like `first` (after `without`), `to` (after `potential`), `for` (after `boards`). This suggests a pattern where a word from MAX_ACTIVATING_TOKENS is followed by a specific token.3. **TOP_ACTIVATING_TEXTS**: * "...`without` first explicitly separating..." - Confirms `without` followed by `first`. * "...`potential` to capture people's attention..." - Confirms `potential` followed by `to`. * "...leaderboards for collaborative quizzes…" - Confirms `boards` followed by `for`. * "within ethical `boundaries`: 1. `Architecturally`..." - Confirms `boundaries` followed by a colon or a number/word indicating a list item. * "...`provides` their cultural, their fear..." - Shows `provides` and `their`. The `their` could be a pronoun, and `fear` is in MAX_ACTIVATING_TOKENS, but the sequence isn't as clear as the others.The dominant pattern is a word from `MAX_ACTIVATING_TOKENS` being closely followed by a specific word or punctuation (like `first`, `to`, `:`, `1`). This suggests the neuron is identifying a specific linguistic construction or a state that is then qualified or elaborated upon.Phrases considered:- "conditions followed by details" (too generic)- "precedes elaborations" (getting closer)- "identifies specific continuations" (too

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

    The neuron consistently lights up on common “glue” tokens—i.e. punctuation and high-frequency function words (commas, conjunctions, prepositions, auxiliary verbs) that link clauses.

    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
     đầu
    0.79
    والفقار
    0.74
     součástí
    0.73
     tiến
    0.72
     lắm
    0.72
     Thủ
    0.71
     Giáo
    0.70
     CHIKV
    0.70
     tử
    0.69
     triệu
    0.69
    POSITIVE LOGITS
    ii
    0.86
    くれます
    0.79
    browser
    0.74
    iy
    0.71
    your
    0.71
     charming
    0.70
    ffen
    0.69
     ballet
    0.68
     slogans
    0.68
     понятно
    0.68
    Activations Density 0.016%

    No Known Activations