© 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 EvalsExportsAPI Community BlogPrivacy & TermsContact
    1. Home
    2. Gemma-3-4B-IT
    3. 22-GEMMASCOPE-2-RES-262K
    4. 3226
    Prev
    Next
    INDEX
    Explanations

    thoughtThe user wants a concise explanation (3-20 words) of a neuron's behavior based on the provided lists.Looking at the lists:- **MAX_ACTIVATING_TOKENS**: `interesting`, `unsettling`, `fascinating`, `captivating`, `disturbing`, `startling`, `surprising`. This list is dominated by words that describe something evoking a strong, often surprising or thought-provoking, reaction. `unsettling` is the most frequent.- **TOKENS_AFTER_MAX_ACTIVATING_TOKEN**: `books`, `realism`, `themes`, `geometry`, `symbol`, `question`, `and`, `ly`, `result`, `concept`. These seem to be the subjects or qualities being described by the activating tokens.- **TOP_POSITIVE_LOGITS**: `ly`, `revelations`, `discoveries`, `feats`, `juxtaposition`, `juxt`, `mente`, `erweise`, `revelation`, `insights`. These support the idea of new information, understanding, or striking observations. `ly` is present here and in the previous list.- **TOP_ACTIVATING_TEXTS**: Examples include "unsettling realism", "unsettling themes", "unsettling geometry", "fascinating symbol", "captivating question", "disturbing thoughts", "startlingly clear", "surprising result", "fascinating... unsettling concept".The neuron seems to focus on words that describe things that are noteworthy, surprising, thought-provoking, or evoke a strong reaction, often followed by nouns describing abstracts, qualities, or outcomes. The overarching theme is about things that are striking or reveal something unexpected.Let's try to synthesize this."unsettling" is very common. The texts often pair it with "realism", "themes", "geometry", "concept"."fascinating", "captivating", "surprising", "startling", "disturbing" are also key.The phrases often describe things that are not mundane, but remarkable or emotionally charged."revelations", "discoveries", "insights" from `TOP_POSITIVE_LOGITS` suggest the neuron is interested in things that lead to understanding or are novel.Considering the words like "unsettling", "disturbing", "surprising", "startling" combined with concepts like "realism", "themes", "geometry", "concepts", and outcomes like "revelations", "discoveries", "insights", the neuron appears to be about things that are remarkably or curiously revealing.Let's refine. "unsettling realism" and "unsettling geometry" are quite specific. "surprising result".The neuron highlights things that are `interesting`, `fascinating`, `captivating`, `unsettling`, `disturbing`, `startling`, `surprising`.These often lead to `revelations`, `discoveries`, `insights`.A possible explanation could be related to remarkable or surprising observations/information."strikingly novel insights" (3 words)"strikingly novel findings" (3 words)"remarkable curiosities and insights" (4 words)"novel insights and surprising truths" (5 words)"curious and thought-provoking discoveries" (4 words)The `ly` suffix in multiple lists is interesting. It suggests adverbs often modifying these startling/surprising terms. e.g., "startlingly clear", "fascinatingly complex". The logits also have `ly`.If the neuron detects adverbs like "startlingly" which describe qualities, or the concepts they describe, then perhaps "adverbially described striking concepts". Too long.Let's focus directly on the concepts.`unsettling``fascinating``surprising``startling``disturbing`These are often descriptors. The things they describe (realism, themes, geometry, concepts, results) are the *subject matter*.The `TOP_POSITIVE_LOGITS` include `revelations`, `discoveries`, `insights`. This suggests the neuron is interested in things that *lead to* these.Maybe the neuron is about intriguing or problematic discoveries/observations."intriguing discoveries" (2 words) - too short."intriguing discoveries and revelations" (4 words)"surprising revelations and insights" (4 words)"striking insights and discoveries" (4 words)"noteworthy observations and revelations" (4 words)The word `unsettling` is the most dominant in `MAX_ACTIVATING_TOKENS`.The texts show "unsettling realism", "unsettling geometry", "unsettling themes", "unsettling concept".This suggests the neuron might be focused on aspects that are unsettling or revealing in an unusual way."unsettling subjects and discoveries" (4 words)"surprising, unsettling observations" (3 words)"noteworthy and unsettling revelations" (4 words)Let's try to be more direct about the pattern. The words themselves are descriptors of how something is perceived."fascinating, disturbing, surprising descriptors" (4 words) - too meta.Let's look at the `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` again: `realism`, `themes`, `geometry`, `symbol`, `question`, `result`, `concept`. These are often abstract nouns or qualities.Combined with `MAX_ACTIVATING_TOKENS`: `interesting`, `unsettling`, `fascinating`, `captivating`, `disturbing`, `startling`, `surprising`.And `TOP_POSITIVE_LOGITS`: `revelations`, `discoveries`, `insights`.The neuron seems

    np_acts-logits-general · gemini-2.5-flash-lite
    New Auto-Interp
    Top Features by Cosine Similarity
    Configuration
    google/gemma-scope-2-4b-it/resid_post/layer_22_width_262k_l0_medium
    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
    ்களில்
    0.85
    Aplic
    0.83
    Excellent
    0.79
    Boxing
    0.78
     foarte
    0.77
     Các
    0.77
     Более
    0.76
    ்கள்
    0.76
    ევე
    0.76
     rất
    0.76
    POSITIVE LOGITS
    ly
    1.77
     revelations
    1.41
     discoveries
    1.25
     feats
    1.21
     juxtaposition
    1.19
     juxt
    1.18
    mente
    1.15
    erweise
    1.14
     revelation
    1.13
     insights
    1.12
    Activations Density 0.283%

    No Known Activations