© 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. Gemma-3-27B-IT
    3. 27-GEMMASCOPE-2-TRANSCODER-262K
    4. 196520
    Prev
    Next
    INDEX
    Explanations

    The MAX_ACTIVATING_TOKENS list only contains "the".The TOKENS_AFTER_MAX_ACTIVATING_TOKEN list contains words like "Communications", "Privacy", "Investment", "Wireless", "IRS", "policies", "Computer".The TOP_POSITIVE_LOGITS list contains words like "latest", "Privacy", "Provider", "Inquiry", "Interpretation", "Innovative".The TOP_ACTIVATING_TEXTS list contains mentions of laws and regulations: "Criminal Code", "Communications Act", "Privacy Act", "Investment Company Act", "Wireless Communications Act", "Fair Work Act", "IRS", "Computer Fraud and Abuse Act". These texts also mention violations, policies, breaches, and legal offenses.The neuron seems to be activated by discussions of legal statutes, acts, and regulations, often in the context of violations, policies, or specific domains like communications, privacy, or investments. It frequently follows the word "the".Let's look for a pattern:- "the *Criminal Code*"- "the *Communications Act*"- "the *Privacy Act*"- "the *Investment Company Act*"- "the *Wireless Communications Act*"- "the *Fair Work Act*"- "the IRS"- "the *Computer Fraud and Abuse Act*"The pattern is "the" followed by a specific Act, Law, or regulatory body name. The TOP_POSITIVE_LOGITS also reinforce themes of "latest", "privacy", "provider", "inquiry", "interpretation".Combining these, the neuron is looking for instances where "the" is followed by a formal legal or regulatory term, often in contexts of compliance, inquiry, or violation.Possible explanations:- legal acts and regulations- followed by acts and laws- specific laws and acts- identifies names of acts and laws- mentions acts and laws"specific laws and acts" fits the pattern of "the [Specific Law/Act Name]"."identifies names of acts and laws" is a bit too descriptive of the *neuron's function* rather than *what it detects*.The prompt asks for *what the neuron detects or predicts by finding patterns*.The pattern is clearly "the" followed by a legal/regulatory name."specific laws and acts" highlights this."acts and laws" is a bit too broad."names of laws and acts" is better.Let's re-check the word count (3-20 words)."specific laws and acts" is 4 words."names of laws and acts" is 4 words.The TOP_POSITIVE_LOGITS also include "Latest",

    np_acts-logits-general · gemini-2.5-flash-lite
    New Auto-Interp
    Top Features by Cosine Similarity
    Configuration
    google/gemma-scope-2-27b-it/transcoder_all/layer_27_width_262k_l0_small_affine
    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
    FeO
    0.39
     వల్ల
    0.35
    Firing
    0.35
    (",");
    0.34
     नामा
    0.34
    Kamol
    0.34
    ⿲
    0.34
    ButtonGroup
    0.33
     Chakraborty
    0.33
     bolje
    0.32
    POSITIVE LOGITS
    最新
    0.38
    emer
    0.38
     Gifts
    0.37
     Privacy
    0.36
     Provider
    0.35
     Inquiry
    0.35
     ditambah
    0.35
     Latest
    0.34
     Interpretation
    0.34
     Innovative
    0.34
    Activations Density 0.003%

    No Known Activations