© 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-3-12B
    3. 24-GEMMASCOPE-2-RES-16K
    4. 9857
    Prev
    Next
    INDEX
    Explanations

    The user's previous explanation was too long and conversational. I need to return *only* the concise phrase (3-20 words) that explains the neuron's behavior, without any introductory sentences.Let's re-evaluate the patterns:**MAX_ACTIVATING_TOKENS**:- have- color- money- hiding- actually- includes- Housing- decided- GPU- Driver**TOKENS_AFTER_MAX_ACTIVATING_TOKEN**:- plans (after 'have')- like (after 'color')- so (after 'money')- your (after 'hiding')- copied (after 'actually')- three (after 'includes')- Provider (after 'Housing')- to (after 'decided')- architecture (after 'GPU')- driver (after 'driver')This list clearly shows tokens that often appear *after* the corresponding MAX_ACTIVATING_TOKEN. This is a "what comes next" pattern.**TOP_POSITIVE_LOGITS**:- ing- ode- iodo- elle- hol- uer- williams- ibilit- egenes- busThese are word suffixes or partial words, indicating a focus on morphology or word structure towards the end of words.**TOP_ACTIVATING_TEXTS**:- "Worried about the economy? Want to stop spending money on dues and taxes" -> money on- "two option color sets for tips card vol.6. The color start with the spring tone color like green" -> color like- "buys up the next number of shares. It includes three options in its execution logic." -> includes three- "Supported Housing Provider services." -> Housing Provider- "customer ultimately decided to return the monitor" -> decided to- "older GPU architecture (Kepler)." -> GPU architecture- "AndroidDriver driver; // Your Appium driver instance" -> Driver driverThe most prominent pattern across MAX_ACTIVATING_TOKENS and TOKENS_AFTER_MAX_ACTIVATING_TOKEN is identifying what tokens *follow* a specific input token. The TOP_POSITIVE_LOGITS suggest it might be related to word endings or specific word structures.Combining these, the neuron seems to predict or associate words that follow a given word, possibly with a lean towards certain word types based on suffixes.Let's try to phrase this concisely:- "predicting words that follow given words" (5 words)- "associates words with their subsequent neighbors" (6 words)- "identifies token sequences" (3 words) - but "tokens" is not allowed.- "identifies word sequences" (3 words)- "what comes next in word sequences" (6 words)- "detecting sequential word associations" (4 words)- "following words" (2 words) - too short.- "sequential word completion" (3 words)The pattern here is strongly about predicting the *next* word in a sequence. The "ing", "ode", "elle" etc. from TOP_POSITIVE_LOGITS might hint at a general tendency, but the most concrete pattern is the token-after-token relationship.sequential word completion

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

    The neuron strongly fires on tokens that introduce or carry quantitative or measurement information (e.g. fees, totals, durations, amounts).

    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
     worded
    0.77
     doubted
    0.75
     まず
    0.75
     Ер
    0.75
     ridiculed
    0.74
     Standing
    0.73
     predictors
    0.73
     मोती
    0.73
     criticized
    0.72
     OUR
    0.72
    POSITIVE LOGITS
    ing
    0.78
    ode
    0.71
    hol
    0.71
    iodo
    0.70
    elle
    0.70
     williams
    0.69
    uria
    0.68
    SafeArea
    0.68
     eigenes
    0.68
    ibilit
    0.67
    Activations Density 0.000%

    No Known Activations