© 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 EvalsExportsinterp-engineNEWAPI Community BlogPrivacy & TermsContact
    1. Home
    2. Qwen3.5-4B
    3. 15-RES-MATRYOSHKA-65K
    4. 21359
    Prev
    Next
    INDEX
    Explanations

    **Explaining the Neuron's Behavior**The neuron is consistently activated by the token "Me" (or its lowercase variant "me").Looking at `TOKENS_AFTER_MAX_ACTIVATING_TOKEN`, the sequence immediately following "Me" is varied, including punctuation (`,`), conjunctions (`and`), and other words (`too`, `choosing`). This suggests "Me" isn't always followed by a specific word, but rather acts as a self-reference.The `TOP_ACTIVATING_TEXTS` show "Me" appearing in phrases like:* "Me, without an iPhone."* "Me too"* "me and my best programmer buddy"* "Meagan Morris" (a proper noun, potentially less relevant to the core "Me" pattern)* "Me-ow" (a playful usage)* "it was basically me choosing between career and relationship."The pattern is that the neuron activates when "Me" is used as a subject or part of a self-referential phrase, often in direct statements or descriptions.The `TOP_POSITIVE_LOGITS` are less helpful here, showing foreign characters or less common words.The core pattern is the self-reference initiated by "Me".Let's try to phrase this concisely:- "referring to oneself" (3 words) - A bit abstract.- "self-referential language" (2 words) - Good, but maybe I can be more specific on *what kinds* of self-referential language.- "first-person statements" (2 words) - Good.- "phrases starting with me" (4 words) - Accurate.- "statements about self" (3 words) - Good.- "i and me references" (3 words) - "I" is not in MAX_ACTIVATING_TOKENS, so this is not directly supported by the input.Considering the examples like "Me, without...", "Me too...", "me choosing...", it's about identifying contexts where "Me" is the subject or part of a personal declaration. The presence of punctuation and conjunctions *after* "Me" strengthens the idea that it's looking for "Me" as a subject that then leads into a description or action."first-person statements" seems accurate."statements about self" is also good."phrases about me" (3 words) might be the most direct interpretation of the highest activating token.Let's re-evaluate the examples:- "Me, without an iPhone." -> `Me,`- "Me too" -> `Me ` (followed by `too`)- "me and my best programmer buddy" -> `me ` (followed by `and`)- "me choosing between career and relationship." -> `me ` (followed by `choosing`)The neuron seems to capture the use of "Me" as the subject introducing a statement or thought.Possible explanations:- "phrases about me" (3 words)- "self-referential expressions" (2 words)- "first-person self-references" (3 words)The instruction asks for *what it detects or predicts*. It detects "Me" in contexts where the speaker is referring to themselves."phrases about me" captures this well.What about the `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` like "asuring" and "choosing"? These are verbs. So, "Me" followed by a verb might be a sub-pattern. But "Me, without" and "Me too" don't fit that.The overarching pattern is "Me" as a subject.Let's try to be very specific to the examples."Me" followed by punctuation or verbs indicating self-description.Consider "phrases about me". It's concise and descriptive."first-person statements" is also good.Let's check rule "The explanation should be specific. For example, 'unique words' is not a specific enough pattern, nor is 'foreign words'.""phrases about me

    np_acts-logits-general · gemini-2.5-flash-lite
    New Auto-Interp
    Top Features by Cosine Similarity
    Configuration
    decoderesearch/qwen-3.5-saes/qwen-3.5-4b
    Prompts (Dashboard)
    16,384 prompts, 128 tokens each
    Dataset (Dashboard)
    monology/pile-uncopyrighted
    No Configuration Found
    Embeds
    IFrame
    Link
    Not in Any Lists

    No Comments

    Negative Logits
    er
    -0.07
    vale
    -0.06
    oit
    -0.06
    ´
    -0.06
    -esque
    -0.06
    æĢ§åľ°
    -0.05
    UN
    -0.05
    erable
    -0.05
    人
    -0.05
    ع
    -0.05
    POSITIVE LOGITS
    ichi
    0.07
    اÛĮÙĦ
    0.06
    imi
    0.06
    ä½Ĩä¸įéĻIJäºİ
    0.06
    alls
    0.06
     Mojo
    0.06
    ago
    0.06
    ä¸ĢéĹ®
    0.06
    Ú©Ø´
    0.06
    صر
    0.06
    Activations Density 0.013%

    No Known Activations