© 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. 21765
    Prev
    Next
    INDEX
    Explanations

    materialsfamilysourcesformatvehicleselementsformatsrom1resourcesTOKENS_AFTER_MAX_ACTIVATING_TOKEN:intoreturnswere,arefromandolecules(inTOP_POSITIVE_LOGITS:(ï¼Īï¼Ī(>ãĢįï¼Ī(=(++å°½(--ãĢıï¼ĪTOP_ACTIVATING_TEXTS:anical modeling of biological cells in microinjection. Microinjection is an effective technique to introduce foreign materials into a biological cell. Although some semi-automatic and fully-automatic microinjection systems have been developed, a full understanding of: In which case PHP sort function returns FALSE? PHP sort() function and other functions from this family returns true on success or false on failure. In which case sort function returns false? What is possible source of failureThe antigenic structure and the drug resistance pattern of 65.4% of the E coli isolates from different sources were used as epidemiologic markers to trace the source of the infection. The predominant E coli serotypes involved in infections ofto your calendar:Browse to the WebStore Events Calendar, select your course, select the calendar symbol , select your calendar format, open the file, andthe rate at which overworn or worn-out rubber tires (hereinafter referred to simply as tires) of land vehicles are being discarded. Of these tires, one portion is being reutilized as retreaded tires, but most oflittle or no structural support for sealing the window pane with which they are associated. The door seals are designed to prevent outside elements from entering the passenger compartment and to allow the window pane to move translationally without undue resistanceBundesverband Musikindustrie (Federal Association of the German Music Industry, BVMI) announced that revenues from the sale of physical formats and downloads, as well as income from music streaming increased by 1.2 percent to 1.45 billionand blood-brain barrier selectively restrict the entry of most compounds into cells to < 500 Da, delivering macromolecules into cells was, until recently, little more than a goal. However, with significant effort to capitalise on therapeutic00 were selected. There were 103 patients who had ASD. All 64 patients in group 1 (surgery) had a successful operation, with only 2 patients with a mild residual shdownlink data transmission to User Equipments (UEs) in one or more cells. The BS needs to allocate communication resources in the time-frequency domain to each UE when the BS transmits downlink data to a specific UE since a plurality ofThe neuron appears to be detecting terms related to categories, types, or components, often followed by words indicating their properties, origins, or states.Let's look for specific patterns.- **materials**, **family**, **sources**, **format**, **vehicles**, **elements**, **formats**, **resources** are all nouns that describe types, categories, or components.- The tokens after are common: `into`, `,` `are`, `from`, `and`, `in`. These suggest descriptions or relationships.Consider the texts:- "introduce foreign **materials into**"- "functions from this **family returns**" / "source of **failure**" (after 'sources')- "select your calendar **format, open**"- "tires) of land **vehicles are**"- "prevent outside **elements from**"- "physical **formats and**"- "allocate communication **resources in**"The pattern seems to revolve around identifying specific types of things (materials, elements, sources, formats, etc.) and the context around them (what they are used for, what they do, what comes from them).The TOP_POSITIVE_LOGITS are mostly non-standard characters, which can sometimes indicate structural or formatting elements within text, or perhaps specific ways of encoding information. However, the MAX_ACTIVATING_TOKENS are very clear English words.Let's focus on the MAX_ACTIVATING_TOKENS and TOP_ACTIVATING_TEXTS.The keywords like `materials`, `family`, `sources`, `format`, `vehicles`, `elements`, `resources` all point to categories or types of things.The context around them usually describes their properties, functions, or relationships.For example:- materials -> introduce foreign materials into- family -> functions from this family returns- sources -> isolates from different sources were- format -> select your calendar format, open- vehicles -> tires) of land vehicles are- elements -> prevent outside elements from- resources -> resources inThe neuron seems to identify specific types of nouns or categories and the descriptions or actions associated with them.The common tokens after suggest these categories are often being described, listed, or related to other things.Let's try to synthesize this into a short phrase."types of things and their descriptions" - a bit long."categories and their context" - closer."categories and related words" - okay.Let's re-examine TOP_POSITIVE_LOGITS. They are mostly `(` and `)` related symbols, and some other unicode characters. This might suggest that the neuron is sensitive to specific structural elements or delimiters that often appear *around* these categories, or perhaps when these categories are presented in a structured list or code-like format. However, without more context for these logits, it's hard to be certain.The MAX_ACTIVATING_TOKENS are quite diverse categories. The rule is "find patterns."- materials- family- sources- format- vehicles- elements- formats- rom- 1- resourcesThe commonality is that these are all nouns that represent distinct items, types, or groups. They are often followed by verbs or prepositions that describe their state, origin, or use.Consider the phrase: "types of things followed by description". This is 5 words.Let's try to be more specific by looking at the *kind* of types."materials", "elements", "sources", "vehicles" suggest physical or abstract components/origins."family", "format", "resources" are similar category words."rom", "1" are more specific.What if the neuron detects specific categories and then common follow-up words that further define them?"categories followed by context""types and their usage""elements and their role"Let's check the constraints: 3 to 20 words. Not start with "This neuron detects/predicts". No "tokens" or "patterns". No capitalization unless proper noun. Be specific."materials", "family", "sources", "format", "vehicles", "elements", "formats", "resources" all sound like lists of items or categories. The tokens after are often verbs (`returns`, `are`), prepositions (`into`, `from`, `in`), or conjunctions (`and`), which tend to describe or connect these items.What if the neuron detects lists of items, or categories, and their immediate context?"categories and surrounding words""item types and descriptions""generic categories and their attributes"Let's focus on the

    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
    avant
    -0.06
     rasse
    -0.06
     entretien
    -0.06
    agrada
    -0.05
    ubahan
    -0.05
     аÑĥдиÑĤ
    -0.05
    =localhost
    -0.05
     Ente
    -0.05
     bred
    -0.05
    anta
    -0.05
    POSITIVE LOGITS
     (
    0.09
    ï¼Ī
    0.07
     ï¼Ī
    0.07
     (>
    0.07
    ãĢįï¼Ī
    0.07
     (=
    0.06
     (++
    0.06
    å°½
    0.06
     (--
    0.06
    ãĢıï¼Ī
    0.06
    Activations Density 0.076%

    No Known Activations