© 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. Gemma-3-27B-IT
    3. 37-GEMMASCOPE-2-TRANSCODER-262K
    4. 139372
    Prev
    Next
    INDEX
    Explanations

    *- **Keywords:**" to quickly summarize. I've checked for availability as of *right now* (Oct 26, 2023) but things change *fast* on Instagram, so double-check before committing! **Ibroad.*Vibe Check:* Intelligent, cutting-edge, slightly abstract.) (Good Availability) * **AlgorithmicAesthetics:** (Clearly states the focus. *Vibe Check:* Academic, thoughtful, precise.).* **The Static Hour:** A bit unsettling, good for art/moody content.* **Echo & Veil:** Mysterious and evocative.* **Lunar Codex:** Suggests secrets and knowledge(YouTube):** Another great resource with a lot of beginner-friendly content: [https://www.youtube.com/@SquashSkills](https://www.youtube.com/@SquashSkills) **II. Footwork (Highit's *immediately* taken, but you *must* verify on Instagram itself. **Important:** I'm checking availability using Namecheckr *right now*, but Instagram's database updates constantly. **You *need**Figures** (Common, check) * **@PastGreats** (Decent, might be available) * **@HistorysHeroes** (Check, could be taken) * **@GreatMindsPastKids' Activities, Wedding Decor, Budgeting Tools, etc.) The name should hint at this if possible.* **Target Audience:** Who are you selling to? (Moms, Brides, Students, Professionals, etc.)---Let's find a pattern.* **MAX_ACTIVATING_TOKENS**: Starts with '@' or '/' or "'" or empty. This looks like it's looking for delimiters or specific token types.* **TOKENS_AFTER_MAX_ACTIVATING_TOKEN**: This list shows tokens like 'Sc', '\*', '2', '\*\*', 'Squ', 'm'. This is highly varied and doesn't immediately suggest a strong pattern with the other lists.* **TOP_POSITIVE_LOGITS**: These are a mix of languages (cki, अंतर, 중요, лек, േണ്ട, криптовалю) and English terms (PersonalInfo, AppComponent, Trat, COMPARE). This suggests the neuron might be sensitive to different languages or specific technical/business terms.* **TOP_ACTIVATING_TEXTS**: This text is very diverse. * It includes social media handles like "@Scandinaviansoccer". * It has descriptions of content and "Vibe Check" sections. * It mentions "Important Considerations", "Keep it short & memorable", "Brand Consistency", "Keywords", "Target Audience". * It mentions checking availability for names/handles. * It includes links like `https://www.youtube.com/@SquashSkills`. * It mentions checking availability for Instagram. * It lists historical figures/handles like "@PastGreats", "@HistorysHeroes".The presence of "@username", links with "/", and the focus on "availability", "names", "handles", "keywords", "brand consistency", and "target audience" strongly suggests the neuron is related to identifying or processing online handles, usernames, or perhaps specific types of URLs/references. The `@` in `MAX_ACTIVATING_TOKENS` is a strong clue. The diversity in `TOP_POSITIVE_LOGITS` might indicate a general sensitivity to identifying named entities or structured data across different contexts, potentially including different languages or specific technical components.The texts are about checking for *availability* of *names/handles/identifiers* which are often formatted with '@' or '/'.Let's try to form a concise explanation."online handles" is good but maybe too short."checking availability of online handles" is better."identifying and checking availability of online handles" is more descriptive but might be too long.Looking at the `@` and `/` in `MAX_ACTIVATING_TOKENS`, and phrases like `@Scandinaviansoccer`, `twitter.com`, `youtube.com/@SquashSkills`, and "checking availability" in `TOP_ACTIVATING_TEXTS`.The neuron seems to be finding tokens that are part of usernames/handles or links, and the surrounding text often discusses checking their availability."online handles and availability" - 4 words. This might be too general."usernames and checking availability" - 4 words."social media handles and URLs" - 4 words."handles, usernames, and website links" - 5 words.Let's re-examine the `MAX_ACTIVATING_TOKENS` and `TOKENS_AFTER_MAX_ACTIVATING_TOKEN`.MAX_ACTIVATING_TOKENS: `@`, `/`, `@`, `'`, `@`, `/`, `@`, `'`, `'`TOKENS_AFTER_MAX_ACTIVATING_TOKEN: `Sc`, `*`, `*`, `2`, `**`, `**`, `Squ`, `m`, `*`, `*`The pattern of `@` and `/` repeated in MAX_ACTIVATING_TOKENS is very strong. This strongly points to detecting online identifiers. The texts reinforce this by discussing handles (`@Scandinaviansoccer`, `@PastGreats`), URLs (`twitter.com`, `youtube.com/@SquashSkills`), and the *process* of checking their availability.So, the neuron detects elements related to online identifiers.Consider:- "online usernames and web addresses" - 5 words- "social media handles and links" - 4 words- "identifying online handles and URLs" - 5 words. This states *what* it does.The rule says "what the neuron detects or predicts."The neuron detects tokens like `@` and `/` which are common in online handles and URLs. The context is about availability of these.So, it's related to identifying these specific types of textual elements."social media handles and links" is concise and covers the main observation."online handles and links" is also good.Let's consider the TOP_POSITIVE_LOGITS again:`скі`, ` अंतर`, `PersonalInfo`, `AppComponent`, `중요`, ` лек`, `േണ്ട`, `криптовалю`, `Trat`, `COMPARE`These might indicate that the neuron is not *just* about handles/links but perhaps about *categories* or *metadata* associated with them, or that it generalizes from specific handle patterns to broader concepts. `AppComponent` and `PersonalInfo` suggest technical or data-related contexts. `COMPARE` suggests comparing options. `криптовалю` and other foreign words suggest it might be multilingual or handle diverse data.However, the prompt emphasizes finding a pattern in the lists and being specific. The most concrete and consistent pattern is the `@` and `/` delimiters leading to considerations about online handles and availability.Let's refine:The MAX_ACTIVATING_TOKENS are `@`, `/`. The texts frequently mention `@username` and URLs. The context is about checking availability for branding.This strongly points to identifying online identifiers."online handles and website addresses" - 5 words."social media handles and links" - 4 words. This seems strong and specific enough.Let's check the rules:- 3 to 20 words: "social media handles and links" is 4 words. OK.- Captures what it detects/

    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_37_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
     guy
    0.50
     chica
    0.45
     warrior
    0.45
     ragazzo
    0.45
     wandering
    0.45
    wolf
    0.44
    guy
    0.44
     redhead
    0.43
     Mama
    0.43
     blogger
    0.42
    POSITIVE LOGITS
    цкі
    0.44
     अंतर
    0.43
    PersonalInfo
    0.43
    AppComponent
    0.40
     중요
    0.38
    letek
    0.38
    േണ്ട
    0.38
     криптовалю
    0.38
     Trat
    0.37
     COMPARE
    0.37
    Activations Density 0.003%

    No Known Activations