© Neuronpedia 2026
    Privacy & TermsBlogGitHubSlackTwitterContact
    Neuronpedia logo - a computer chip with a rounded viewfinder border around it
    Neuronpedia
    Jacobian LensNEW
    Natural Language
    Autoencoders
    MAY 26
    Assistant AxisJAN 26Circuit TracerAPR 26SteerSAE EvalsExportsinterp-engineNEW Community BlogPrivacy & TermsContact
    1. Home
    2. Qwen3.5-4B
    3. 15-RES-MATRYOSHKA-65K
    4. 16837
    Prev
    Next
    INDEX
    Explanations

    The neuron detects sequences like "come and" or "follow me". It seems to be associated with directions or instructions.The provided lists are:- **MAX_ACTIVATING_TOKENS**: down, here, omen, follow, come, hands, and, meet, sit- **TOKENS_AFTER_MAX_ACTIVATING_TOKEN**: in, in, om, me, and, ,, dec, give, him, upLet's identify patterns:1. **"come" -> "me"**: The context from TOP_ACTIVATING_TEXTS shows "follow me—we must find Number Three." and "come and look I wanted to show her something". Also, the instruction "Please come by my desk and give me your travel arrangements." implies "come" followed by an action or recipient.2. **"follow" -> "me"**: As seen in "follow me—we must find Number Three.".3.The neuron detects sequences indicating upcoming actions or destinations, often involving "come" or "follow" followed by a pronoun or directive.- **MAX_ACTIVATING_TOKENS**: `down`, `here`, `omen`, `follow`, `come`, `hands`, `and`, `meet`, `sit`- **TOKENS_AFTER_MAX_ACTIVATING_TOKEN**: `in`, `in`, `om`, `me`, `and`, `,`, `dec`, `give`, `him`, `up`Let's analyze the pairs:- `here` -> `in` (e.g., "gather here in silent meeting")- `come` -> `me` (e.g., "come and look", "give me your travel arrangements" - `me` follows `give` which is often preceded by `come by me desk AND GIVE ME`)- `follow` -> `me` (e.g., "follow me—we must find Number Three.")- `meet` -> `give` (e.g., "to meet him at the Training Grounds ... give me your travel arrangements")The pattern seems to be action verbs or location indicators followed by a direct object, pronoun, or prepositional phrase indicating target or continuation. Many of these are related to giving instructions or making requests.Let's look at TOP_POSITIVE_LOGITS as well:- `apel`- `ufe`- `întâ`- `uest`- `åĬłå·¥è®¾å¤Ĩ`- `åıĹæ½®`- `漸`- `.tools`- `ä¿¡æģ¯åĴĮ`- `åĪ«`These are mostly non-English characters or short strings. They might indicate a specific context where these ASCII-like or garbled tokens appear, but it's hard to derive a semantic meaning for a general neuron explanation. The main semantic clues come from MAX_ACTIVATING_TOKENS, TOKENS_AFTER_MAX_ACTIVATING_TOKEN, and TOP_ACTIVATING_TEXTS.From TOP_ACTIVATING_TEXTS, we see:- " Lie down in the Fire"- "gather here in silent meeting"- "come and look"- "follow me—we must find Number Three."- "come by my desk and give me your travel arrangements."- "meet him"- "hand, and helped her sit up."The strongest patterns are:- `here` -> `in`- `follow` -> `me`- `come` -> `and` or `me` (in context of `come by desk AND GIVE ME`)- `sit` -> `up`The consistent theme is about movement, location, and subsequent actions or direct recipients.follow or come with directions

    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
    æĢ§æĪĸ
    -0.06
    æİ
    -0.05
    bare
    -0.05
    ئÙĬ
    -0.05
    mano
    -0.05
    é»
    -0.05
    ¾
    -0.05
    iog
    -0.05
    alone
    -0.05
     Flesh
    -0.05
    POSITIVE LOGITS
    apel
    0.07
    ufe
    0.06
     întâ
    0.06
    uest
    0.06
    åĬłå·¥è®¾å¤ĩ
    0.06
    åıĹæ½®
    0.05
    漸
    0.05
    .tools
    0.05
    ä¿¡æģ¯åĴĮ
    0.05
    åĪ«
    0.05
    Activations Density 0.006%

    No Known Activations