© 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. Olmo-3-1025-7B
    3. 16-RES-MATRYOSHKA-65K
    4. 58948
    Prev
    Next
    INDEX
    Explanations

    The user wants me to explain the neuron's behavior using the provided lists.The explanation needs to be a concise phrase (3-20 words), specific, and derived from patterns in the lists.Let's break down the lists:1. **MAX_ACTIVATING_TOKENS**: `!`, `Site`, `:`, `.`, `YS`, `.`, `Retro`, `23`, `.com` * This includes punctuation, domain endings (`.com`), and specific words like `Site`, `Retro`.2. **TOKENS_AFTER_MAX_ACTIVATING_TOKEN**: `Two`, `:`, `Cookies`, `The`, `you`, `.com`, `Ful`, `Raw`, `This`, `–` * This list shows what comes *after* the MAX_ACTIVATING_TOKENS. It includes common words, `.com`, and `Raw`.3. **TOP_POSITIVE_LOGITS**: `Collection`, `collection`, `sheriff`, `featuring`, `snapshot`, `collections`, `Sheriff`, `Collection`, `depicting`, `/view` * This list clearly points towards concepts of "collections," "sheriff," "snapshot," "depicting," and views.4. **TOP_ACTIVATING_TEXTS**: * Looks like a mix of website descriptions and adult content. * "Two big cocks for sweet pantyhosed MILF" * "public anal fucking We picked up a hot ... Niches: Teens, Hardcore, Amateurs, Streets, European, Russians Site: Pickup Fuck" * "Hot vaginas of bff teens get drilled by a crazy dude Tags: Cookies help us deliver our" * "living out your craziest sissy fantasies" * "fantastic angel with a clean-shaven vagina" * "one of his famous giant cum loads direct to the mouth of cum sucker Quinn Blackwood." * "variety of full length porn videos featuring young teens, cheating housewives or mature interracial couples exposed in a hot xxx sextapes." * "Marilyn Jess PornURL: http://marilynjessporn.com/index.html?id=1018" (Mentions Retro Raw) * "Naughty Asian schoolgirl is playing show and ... Japanese Av Model Asian" * "more Teens For Cash scenes than ! Browse through our impressive selection of porn videos in HD quality on any device you own. .com – the best free porn videos on internet, 100% free."**Pattern Identification**:* **TOP_POSITIVE_LOGITS**: "collection", "sheriff", "/view", "depicting"* **MAX_ACTIVATING_TOKENS** and **TOKENS_AFTER_MAX_ACTIVATING_TOKEN**: `Site`, `.com`, `Retro`, `Raw`* **TOP_ACTIVATING_TEXTS**: While explicit, there are references to specific genres or sites like "Retro Raw", "Marilyn Jess PornURL", and domain names (`.com`). The explicit content itself is a strong signal.The combination of "Retro", "Raw", "Site", and ".com" points towards specific types of websites or content. The explicit nature of the texts and the logits like "Collection", "featuring", "depicting" suggest that the neuron is looking for content related to specific genres, possibly adult content, often described with terms like "raw" or found on certain "sites".The term "Retro" appears in both MAX_ACTIVATING_TOKENS and TOP_ACTIVATING_TEXTS ("Retro Raw"). "Raw" also appears in TOKENS_AFTER_MAX_ACTIVATING_TOKEN and TOP_ACTIVATING_TEXTS. The presence of `.com` in both MAX_ACTIVATING_TOKENS and TOKENS_AFTER_MAX_ACTIVATING_TOKEN suggests domain references.The logits like "collection", "featuring", "depicting", "snapshot" could be describing the *type* of content or how it's presented on these sites.Considering the presence of "Retro", "Raw", "Site", ".com", and the nature of the activating texts, the neuron seems to be identifying content associated with specific adult website categorizations. "Retro", "Raw" are common tags.Let's try to combine these:- "retro raw sites" (3 words) - Captures a core pattern.- "retro raw .com sites" (4 words) - More specific.- "content from retro raw sites" (5 words) - Too descriptive.- "retro raw adult content" (4 words) - Fits the theme.- "identifying retro raw adult content" - Too long.The instruction asks to "find a pattern that helps you explain the behavior". The pattern is the *combination* of these elements. "Retro", "Raw", and ".com" seem to be key identifiers for specific types of content, which are clearly adult-oriented given the texts.It's not just "collections" or "sheriff" which are logits, but how they might relate to the overall theme. For example, some sites might offer "collections" of "retro raw" content.Let's re-evaluate the logits: `Collection`, `collection`, `sheriff`, `featuring`, `snapshot`, `collections`, `Sheriff`, `Collection`, `depicting`, `/view`.These suggest a focus on visual media, descriptions of items or scenes.The combination of `Retro` and `Raw` from the tokens, along with the explicit nature of the activating texts, strongly points to a category of adult content. The logits, while seemingly unrelated at first, might be verbs or nouns used in descriptions of such content (e.g., "depicting a collection").A concise explanation derived from "Retro", "Raw", ".com" and the explicit texts would be along the lines of a specific adult content genre."retro raw content" (3 words) is good."retro raw site content" (4 words) is also good."retro raw .com content" (4 words) is another possibility.The instruction says "just say the pattern itself"."Retro Raw .com" captures a specific type of origin or categorization.The explicit nature is implicitly understood from the texts provided.Let's consider if there's any other strong pattern."Site" and "23" in MAX_ACTIVATING_TOKENS. "Two", "you", "Cookies", "Ful", "This" in TOKENS_AFTER_MAX_ACTIVATING_TOKEN.The `Collection` logits are frequent.Is there a pattern like "collections from X"?The texts strongly indicate adult content. "Retro" and "Raw" are common descriptors within adult content sites for specific niches. The ".com" suffix is highly indicative of internet content.The most specific and recurring pattern that ties tokens and texts together is the genre descriptor "Retro Raw" and the context of "sites" or ".com" domains.Final candidates:1. retro raw content2. retro raw site content3. retro raw .com sitesI think "retro raw content" is the most direct and concise interpretation of the *behavior* of detecting this type of material. The "site" or ".com" aspects are *where* it's found, but the *content* itself is what the neuron is likely trying to identify.retro raw content

    np_acts-logits-general · gemini-2.5-flash-lite
    New Auto-Interp
    Top Features by Cosine Similarity
    Configuration
    decoderesearch/olmo-3-saes/olmo-3-1025-7b
    Prompts (Dashboard)
    24,576 prompts, 128 tokens each
    Dataset (Dashboard)
    monology/pile-uncopyrighted
    No Configuration Found
    Embeds
    IFrame
    Link
    Negative Logits
     Advertisement
    -0.09
    füg
    -0.09
     Starbucks
    -0.09
     Sachs
    -0.09
    benh
    -0.09
     enthusi
    -0.09
     strugg
    -0.09
    wagon
    -0.09
     Dod
    -0.08
     /[
    -0.08
    POSITIVE LOGITS
     Collection
    0.11
     collection
    0.11
     sheriff
    0.10
     featuring
    0.10
     snapshot
    0.09
     collections
    0.09
     Sheriff
    0.09
    Collection
    0.09
     depicting
    0.09
    /view
    0.09
    Activations Density 0.051%

    No Known Activations