© 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. Google DeepMind · Exploring Gemma 2 with Gemma Scope
    3. Gemma-2-9B-IT
    4. Residual Stream - 16k
    5. 20-GEMMASCOPE-RES-16K
    6. 4589
    Prev
    Next
    INDEX
    Explanations

    expressions related to societal perceptions and stereotypes

    oai_token-act-pair · gpt-4o-miniTriggered by @bot

    Follows a word indicating uncertainty or possibility

    np_acts-logits-general · gemini-2.0-flash

    thoughtThe user wants me to explain the behavior of a neuron with a short phrase (3-20 words) based on four lists.The explanation should capture patterns in the lists and avoid generic phrasing.Let's analyze the lists:1. **MAX_ACTIVATING_TOKENS**: `are`, `some`, `impossible`, `role`, `merely`, `decay`, `relevant`, `be`, `möglich`, `who`, `that`, `biggest`2. **TOKENS_AFTER_MAX_ACTIVATING_TOKEN**: `companies`, `intimidating`, `simply`, `in`, `support`, `and`, `to`, `insignificant`, `sei`, `bore`, `this`, `jerk`3. **TOP_POSITIVE_LOGITS**: `незавершена`, `LLocation`, `enumi`, `ніципалі`, `Voci`, `Билгалдахарш`, `ofluor`, `Смотрите`, `ilaire`, `drm` *These are mostly non-English words. They might indicate a focus on specific entities, concepts, or languages, but without context, they are less directly useful for a simple phrase than the English tokens and texts.*4. **TOP_ACTIVATING_TEXTS**: * "Or are companies voluntarily tying one hand behind their backs?" (Connects 'are' to 'companies') * "...assuming that such commonality is impossible simply because that person couldn’t imagine such a thing..." (Connects 'impossible' to 'simply') * "...it has received relatively little attention. However, we know that waveform has a significant impact on ... Often thought to merely support normal bowel function..." (Connects 'merely' to 'support') * "...cities will decay and die as wealthy people move out..." (Connects 'decay' to 'die' - not directly from the list, but shows a negative outcome association) * "...digital twins only relevant to physical machines..." (Connects 'relevant' to 'to') * "...dass es nicht möglich sei, einen korrekten Satz mit mehr als vier aufeinanderfolgenden Verben zu formulieren..." (German: "that it is not possible to formulate a correct sentence with more than four consecutive verbs" - again 'impossible' and related negation/qualification) * "...it is thought to be insignificant and expensive to measure..." (Connects 'be' to something negative like 'insignificant' or 'expensive') * "...he thinks that House is the biggest jerk but is absolutely addicted to the show." (Connects 'biggest' to 'jerk')Let's look for patterns by combining MAX_ACTIVATING_TOKENS and TOKENS_AFTER_MAX_ACTIVATING_TOKEN:* `impossible` -> `simply`* `role` -> `in`* `merely` -> `support`* `decay` -> `and` (followed by `die`)* `relevant` -> `to`* `be` -> `insignificant` (or something akin to negative assessment)* `biggest` -> `jerk`The `TOP_POSITIVE_LOGITS` contain non-English words, which might hint at a broader scope beyond just English, or specific technical terms not easily translated. However, the English tokens and texts provide a clearer pattern.The pattern seems to be words that are often used to **qualify, contrast, or assess** something, often followed by another word that helps define that qualification or assessment.Let's look at the *type* of words:* `impossible`, `merely`, `decay`, `relevant`, `be`, `biggest` are descriptive or relational.* The subsequent words like `simply`, `in`, `support`, `and`, `to`, `insignificant`, `jerk` often complete a thought or provide context.Consider the overall theme from the texts:* Sustainability strategy as a strategic advantage or hindrance.* Philosophy as intimidating vs. accessible.* Commonality in relationships being impossible or assumed.* Cell propulsion, bowel function (biological/functional descriptions).* Cities decaying, digital twins.* Rainfall studies, insignificance, expense.* Grammatical rules (impossible sentences).* Risks of having children vs. modern medical understanding.* Addiction to a show despite disliking a character.There's a recurring theme of **negation, qualification, or questioning reality/possibility/significance**.* `impossible`* `merely` (implies less than)* `decay` (negative outcome)* `relevant` (questioning scope)* `insignificant`* `nicht möglich` (not possible)* `jerk` (negative assessment)The neuron seems to activate when discussing things that are difficult, not possible, not significant, or evaluated negatively/qualifiedly. It focuses on the *limits*, *qualifications*, or *negations* of concepts.Let's try to synthesize this into a short phrase:- "Difficulties and limitations" (4 words) - A bit too generic for "impossible", "merely", "decay", "relevant", "biggest".- "Negation and qualification" (3 words) - Good, captures `impossible`, `merely

    np_acts-logits-general · gemini-2.5-flash-lite
    New Auto-Interp
    Top Features by Cosine Similarity
    Comparing With GEMMA-2-9B-IT @ 20-gemmascope-res-16k
    Configuration
    google/gemma-scope-9b-it-res/layer_20/width_16k/average_l0_91
    Prompts (Dashboard)
    24,576 prompts, 128 tokens each
    Dataset (Dashboard)
    monology/pile-uncopyrighted
    Features
    16,384
    Data Type
    float32
    Hook Name
    blocks.20.hook_resid_post
    Hook Layer
    20
    Architecture
    jumprelu
    Context Size
    1,024
    Dataset
    monology/pile-uncopyrighted
    Activation Function
    relu
    Embeds
    IFrame
    Link
    Not in Any Lists

    No Comments

    Negative Logits
    CloseOperation
    -0.55
    MessageTagHelper
    -0.52
    WireFormat
    -0.43
    AddTagHelper
    -0.42
     sumpay
    -0.40
     estu
    -0.40
     Reſ
    -0.40
     interv
    -0.40
     împre
    -0.39
     ſta
    -0.38
    POSITIVE LOGITS
     незавершена
    0.54
    LLocation
    0.47
    enumi
    0.45
    ніципалі
    0.44
    Voci
    0.44
    Билгалдахарш
    0.44
    ofluor
    0.43
    Смотрите
    0.42
    ilaire
    0.41
    drm
    0.40
    Activations Density 0.343%

    No Known Activations