© 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. 593
    Prev
    Next
    INDEX
    Explanations

    - Mentions "deliver them from what they were describing" (Quranic quote)- Mentions "abolishing", "Civil War", "Gettysburg Address", "speeches" (Historical text about Lincoln)- Mentions "Levy War", "conclude Peace", "contract Alliances", "establish Commerce", "do all other Acts and Things" (Declaration of Independence)- Mentions "jejak-i hwan-gyeongjeok-eoji" (Korean text, potentially about environmental issues)**Pattern Analysis:**1. **`spoken` / `spoken;`**: Appears in MAX_ACTIVATING_TOKENS and TOKENS_AFTER_MAX_ACTIVATING_TOKEN (indirectly via "spoken"). Seen in a biblical text about David.2. **`his` / `his weapon`**: `his` is in MAX_ACTIVATING_TOKENS. `weapon` appears after `his` in a top activating text (Nehemiah).3. **`mount`**: Appears in MAX_ACTIVATING_TOKENS and a top activating text (Isaiah).4. **`3`**: Appears in MAX_ACTIVATING_TOKENS and TOKENS_AFTER_MAX_ACTIVATING_TOKEN. Also visible in text snippets like `*14*(10), e0223331.` and `f4.bp.blogspot.com%2f-PD4rsMwvX3M%2fU5yikZrxEtI%2fAAAAAAAAAB8%2faudZj6-yzVA%2fs`. This suggests involvement with numerical sequences, possibly references or identifiers.5. **`they` / `they were describing` / `Our` (from "Они")**: `they` is in MAX_ACTIVATING_TOKENS. `Они` (Russian "they") is a top positive logit. This suggests it recognizes pronouns or references to groups.6. **`Acts` / `Acts and Things`**: `Acts` is in MAX_ACTIVATING_TOKENS. "Acts and Things" appears in the Declaration of Independence text. This points towards actions or legal/declarative statements.7. **`टी` / `सब`**: These are Devanagari script characters. They appear in MAX_ACTIVATING_TOKENS (`टी`) and TOKENS_AFTER_MAX_ACTIVATING_TOKEN (`सब`). One top activating text is in Hindi, discussing slavery and Civil War.8. **`TOP_POSITIVE_LOGITS`**: Includes brackets (`⸩`, `⸨`), Cyrillic (`Ⲃ`, `Фі`), Turkish name (`Timurtaş`), common English words (`Their`, `Merged`), UI element (`SearchBar`), and Russian (`Они`). This mix suggests a broad understanding of different languages and possibly UI contexts.9. **`TOP_ACTIVATING_TEXTS`**: Covers religious texts (Quran, Bible, Isaiah), historical/political documents (Declaration of Independence, Lincoln's speech), and academic/media references (Facebook, PLoS ONE).**Synthesizing a Pattern:**The neuron seems to activate on texts that often involve:- **References to power, authority, or declarations**: "what Allah has given you", "Lord", "exalt my throne", "highest places", "Political Campaigns & Elections", "full Power to levy War, conclude Peace", "all other Acts and Things".- **Pronouns indicating groups or entities**: "you", "they", "their", "Они".- **Scripture or historical documents**: Bible, Quran, Declaration of Independence.- **Action-oriented phrases**: "slay him", "built it", "fought with the other", "levy War", "conclude Peace", "establish Commerce", "do all other Acts".- **Non-English characters/words**: Devanagari, Cyrillic, Turkish. This indicates a multilingual capability, especially when combined with English.The presence of `spoken`, `weapon`, `mount`, `Acts`, `3`, and mixed scripts points towards a neuron that might identify specific types of historical, religious, or political pronouncements and actions, potentially involving numerical references or quotes. The `Acts and Things` phrase is particularly strong. The presence of `3` and numerical fragments suggests it might be sensitive to specific formatting or reference styles in these documents.Considering the mix of religious (Quran, Bible), political (Declaration of Independence, speeches), and action-oriented terms ('weapon', 'slay', 'war', 'peace', 'acts'), a good unifying theme is declarative statements of power, action, or rights, often found in historical or canonical texts. The numerical token `3` could relate to specific citation formats or quantities.Let's try to find a concise phrase."Declarations of power and rights" - too long."Historical pronouncements and actions" - good."Religious and political declarations" - also good."Acts, rights, and declarations" - 5 words. This seems specific enough. It captures `Acts`, references to rights/power from the Declaration of Independence, and the declarative nature of religious/historical texts.Let's re-evaluate:- `spoken` + biblical text- `his weapon` + biblical text- `mount` + biblical text- `they` + Quranic text- `3` + numerical references in academic text- `Acts` + Declaration of Independence- Devanagari/Cyrillic + Multilingual context, religious/historical texts.The neuron seems to be sensitive to phrases that *declare* something significant, often related to authority, rights, actions, or prophecies, and it can handle various forms of script and context."Formal declarations of rights" feels a bit too narrow."Historical and religious declarations" is good but misses "Acts"."Acts and declarations of power" - 5 words."Declarations of power, rights, and acts" - 6 words."Acts, rights, and historical declarations" - 5 words.The token `3` and the numerical context in the academic text might suggest references."Acts, rights, and numbered declarations" - 5 words."Acts, rights, and written declarations" - 5 words.Let's look at the top positive logits again: `SearchBar`, `Timurtaş`, `Merged`, `Their`, `Они`. `SearchBar` suggests some interaction or lookup. `Their` and `Они` are pronouns. `Merged` and `Timurtaş` suggest specific entities or merging concepts. This is a bit confusing with the other strong patterns.However, the dominant theme from MAX_ACTIVATING_TOKENS and TOP_ACTIVATING_TEXTS is more about the content and style of pronouncements."Acts, pronouncements, and rights" - 4 words.This covers:- `Acts` (from MAX_ACTIVATING_TOKENS and Declaration)- `pron

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

    The neuron fires strongly on first‐person, reflective or appreciative commentary—personal, subjective expressions (e.g. “I simply wanted to say thanks…,” “I’m not certain what I would’ve created…”) rather than neutral or technical text.

    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
     
    0.47
     November
    0.41
     uptake
    0.39
     ৬৬
    0.39
     part
    0.38
     ২৮
    0.38
     fi
    0.38
     ৫৭
    0.38
     ৭৭
    0.37
     ৩৮
    0.37
    POSITIVE LOGITS
    Ⲃ
    0.45
     Hesap
    0.42
    liquidacion
    0.41
    кугӀ
    0.40
     Они
    0.40
    SearchBar
    0.40
     Бы
    0.39
    SanPham
    0.39
     Timurtaş
    0.39
    separatedBy
    0.39
    Activations Density 0.413%

    No Known Activations