© 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
    EXPLANATION TYPE
    np_max-act-logits
    Description
    Attempts to replicate Anthropic's autointerp used for their attribution graphs paper's features.
    Author
    Neuronpedia
    URL
    https://github.com/hijohnnylin/automated-interpretability/blob/4463a9fab7d4828bfd4c33194e64856b95377166/neuron_explainer/explanations/explainer.py#L811-L1135
    Settings
    Activations shown = 24 tokens around max act. Shows top 10 logits. Shows model the max activating token too. Uses top 10 deduplicated activations.
    Recent Explanations
    organic chemistry
    gemini-2.5-flash-lite
    ite®, diluted with ethyl acetate, washed with brine,
    Neuronpedia logo
    GEMMA-3-27B-IT
    11-GEMMASCOPE-2-TRANSCODER-262K
    INDEX 38106
    period
    gemini-2.5-flash-lite
    ↵        node = stack.pop()↵        print
    Neuronpedia logo
    GEMMA-3-27B-IT
    40-GEMMASCOPE-2-TRANSCODER-262K
    INDEX 244443
    switch statements
    gemini-2.5-flash-lite
    nextInt()↵↵            switch (opcion) {
    Neuronpedia logo
    GEMMA-3-27B-IT
    38-GEMMASCOPE-2-TRANSCODER-262K
    INDEX 119433
    geopolitical tensions
    gemini-2.5-flash-lite
    Kosovo Tensions:** Tensions remained high between Serbia
    Neuronpedia logo
    GEMMA-3-27B-IT
    56-GEMMASCOPE-2-TRANSCODER-262K
    INDEX 251109
    premises
    gemini-2.5-flash-lite
    argument is invalid↵↵↵This argument is valid↵↵↵The conclusion
    Neuronpedia logo
    GEMMA-3-27B-IT
    37-GEMMASCOPE-2-TRANSCODER-262K
    INDEX 136881
    passwords
    gemini-2.5-flash-lite
    .hash('password123'),  # Store
    Neuronpedia logo
    GEMMA-3-27B-IT
    32-GEMMASCOPE-2-TRANSCODER-262K
    INDEX 235328
    Riff Method: 4 Reason: The activating texts prominently feature the word "riff" and related concepts in contexts like music (Riffusion, Riffing) and group names (Riffraff)
    gemini-2.5-flash-lite
    , Udio, Riffusion** - Generate music
    Neuronpedia logo
    GEMMA-3-27B-IT
    1-GEMMASCOPE-2-TRANSCODER-262K
    INDEX 106170
    dude
    gemini-2.5-flash-lite
    <bos><start_of_turn>user↵Yo man do you know anything about
    Neuronpedia logo
    GEMMA-3-27B-IT
    31-GEMMASCOPE-2-TRANSCODER-262K
    INDEX 51375
    common words
    gemini-2.5-flash-lite
    , man nimmt das Ganze Wasser auf der Welt und
    Neuronpedia logo
    GEMMA-3-27B-IT
    6-GEMMASCOPE-2-TRANSCODER-262K
    INDEX 132247
    defines parts or processes
    gemini-2.5-flash-lite
    What parts make it up?↵    * **Functions
    Neuronpedia logo
    GEMMA-3-27B-IT
    13-GEMMASCOPE-2-TRANSCODER-262K
    INDEX 43897
    accommodate
    gemini-2.5-flash-lite
    -shaped areas designed to accommodate larger components.  The
    Neuronpedia logo
    GEMMA-3-27B-IT
    13-GEMMASCOPE-2-TRANSCODER-262K
    INDEX 11888
    fine stripes
    gemini-2.5-flash-lite
    is a creature of whimsy and contradiction. Centuries of
    Neuronpedia logo
    GEMMA-3-27B-IT
    2-GEMMASCOPE-2-TRANSCODER-262K
    INDEX 110105
    multilingual or commands. Method used: 4
    gemini-2.5-flash-lite
    иза њега, истежући врат да
    Neuronpedia logo
    GEMMA-3-27B-IT
    50-GEMMASCOPE-2-TRANSCODER-262K
    INDEX 197953
    computing and Arabic
    gemini-2.5-flash-lite
    efficiency even further (tandem cells).↵*   
    Neuronpedia logo
    GEMMA-3-27B-IT
    2-GEMMASCOPE-2-TRANSCODER-262K
    INDEX 159486
    assignment
    gemini-2.5-flash-lite
    = 5↵y = "Hello"↵print
    Neuronpedia logo
    GEMMA-3-27B-IT
    24-GEMMASCOPE-2-TRANSCODER-262K
    INDEX 178045
    currency
    gemini-2.5-flash-lite
    you like - for example, under 18,
    Neuronpedia logo
    GEMMA-3-27B-IT
    10-GEMMASCOPE-2-TRANSCODER-262K
    INDEX 48493
    Rapid heart rate
    gemini-2.5-flash-lite
    deeply conflicted and largely negative. Here's a breakdown
    Neuronpedia logo
    GEMMA-3-27B-IT
    18-GEMMASCOPE-2-TRANSCODER-262K
    INDEX 88529
    throughout
    gemini-2.5-flash-lite
    are spelled the same way throughout a document, that dates
    Neuronpedia logo
    GEMMA-3-27B-IT
    10-GEMMASCOPE-2-TRANSCODER-262K
    INDEX 27171
    Why not
    gemini-2.5-flash-lite
    * **Why don't we dry out?**
    Neuronpedia logo
    GEMMA-3-27B-IT
    20-GEMMASCOPE-2-TRANSCODER-262K
    INDEX 208875
    spite
    gemini-2.5-flash-lite
    the fact that/In spite of the fact that:**
    Neuronpedia logo
    GEMMA-3-27B-IT
    7-GEMMASCOPE-2-TRANSCODER-262K
    INDEX 32829