© 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 - 131k
    5. 20-GEMMASCOPE-RES-131K
    6. 69536
    Prev
    Next
    INDEX
    Explanations

    phrases that indicate uncertainty or recommendations

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

    check details after To or can"}, {"explanation": "explore relationships"}, {"explanation": "this is not recommended"}, {"explanation": "any games"}, {"explanation": "very rare"}, {"explanation": "check details"}Looking at `MAX_ACTIVATING_TOKENS`: `explore`, `very`, `any`, `To`, `can`, `note`, `not`, `recommended`, `of`, `can`, `check`Looking at `TOKENS_AFTER_MAX_ACTIVATING_TOKEN`: `the`, `I`, `rare`, `games`, `check`, `this`, ``, `play`, `practice`, `a`, `I`, `details`Combinations:- `check` is followed by `details` and `check`.- `To` is followed by `check`.- `can` is followed by `I`, `be`, `check`.- `note` is not explicitly followed by anything here, but `note` is associated with `not recommended`.- `any` is followed by `games`.- `very` is followed by `rare.- `not` is followed by `recommended` and `play`.- `explore` is followed by `the`.`TOP_POSITIVE_LOGITS` include technical terms like `tagHelperRunner`, `WriteTagHelper`, `AddHtmlAttribute`, `HandlerContext`, `ScopeManager`. This suggests the neuron might be related to programming, code, or technical documentation.The combination of `To check` and `check details` from the lists points towards instructions or information retrieval.The `can` token appearing multiple times followed by general words or verbs also suggests capability or possibility.Let's re-evaluate the prompt's emphasis: "finding patterns in lists", "detects or predicts by finding patterns in lists".The pattern seems to be about checking information or details, often presented as instructions or possibilities.Phrases like "To check", "can check", "check details" are prominent.The `TOP_POSITIVE_LOGITS` reinforce a technical context where checking or verifying things is common.How about focusing on the most frequent/strongest patterns?- `check` followed by `details` or other `check`s. (`check`, `check`, `details`)- `To` followed by `check`. (`To`, `check`)- `can` followed by verbs like `check`, `play`. (`can`, `check`)This neuron seems to be most interested in *instructions* on *how to check* or *verify* something.Consider the example text:- "`KinesisRecorder` constructor, call `get`" -> associated with `To check`- "call `get`" -> implies checking something- "You can check details here" -> associated with `can check` and `details`.The combination of "check" and "details" is strong. "To check" is also a strong indicator.Let's try to distill this into a short phrase."to check details" - 3 words. This fits the prompt."checking details or how to" - 5 words."information lookup and verification" - 4 words."instructions for checking details" - 4 words.The phrase should be specific. "Checking details" is quite specific.The logs also include `tagHelperRunner`, `WriteTagHelper`, `AddHtmlAttribute`. These are related to web development and code. In this context, "checking details" might refer to checking attributes, parameters, or logs.Let's see if any of the `MAX_ACTIVATING_TOKENS` combined with their `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` form a more concise pattern.- `explore the`- `very rare`- `any games`- `To check`- `can I`- `can be`- `not recommended`- `not play`- `check this`- `check details`The most actionable and direct patterns relate to `check`."to check details" seems very fitting and direct."checking details" is also good. check details

    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-131k
    Configuration
    google/gemma-scope-9b-it-res/layer_20/width_131k/average_l0_81
    Prompts (Dashboard)
    24,576 prompts, 128 tokens each
    Dataset (Dashboard)
    monology/pile-uncopyrighted
    Features
    131,072
    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
     nahilalakip
    -0.24
    do
    -0.24
    About
    -0.23
    ChildScrollView
    -0.23
     katholischen
    -0.23
     दिया
    -0.22
    इ
    -0.22
    ु
    -0.21
     hitam
    -0.21
    ABOUT
    -0.21
    POSITIVE LOGITS
    tagHelperRunner
    0.75
    WriteTagHelper
    0.72
    AddHtmlAttribute
    0.69
     kaarangay
    0.66
    HandlerContext
    0.65
     témoig
    0.65
     queſto
    0.64
     deſſen
    0.64
    ScopeManager
    0.64
    <unused41>
    0.64
    Activations Density 0.069%

    No Known Activations