© 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 EvalsExportsAPI Community BlogPrivacy & TermsContact
    1. Home
    2. Gemma-3-27B-IT
    3. 16-GEMMASCOPE-2-TRANSCODER-262K
    4. 215585
    Prev
    Next
    INDEX
    Explanations

    `an` `a` `the` `new` `single` `visit`.The `MAX_ACTIVATING_TOKENS` include punctuation like `!` and `.`, common words like `new`, `the`, and `visit`.The `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` include `single` and `necessary`.The `TOP_POSITIVE_LOGITS` include articles `an`, `a`, and `Al`, `Find`, `Ab`, `Ed`, `In`.Looking at the `TOP_ACTIVATING_TEXTS`:- "Excited to announce the release of my **new single**" - This directly contains `new single`.- "We even got to **visit an** elephant sanctuary" - Contains `visit an`.- "file to create a cube<end_of_turn>`an` `a` `the` `new` `single` `visit`The `MAX_ACTIVATING_TOKENS` contain punctuation (`!`, `.`), common words (`new`, `the`, `visit`) and an underscore.The `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` include words like `single` and `necessary`.The `TOP_POSITIVE_LOGITS` include articles (`an`, `a`) and capitalized words possibly starting sentences or proper nouns (`Al`, `Find`, `Ab`, `Ed`, `In`).Examining `TOP_ACTIVATING_TEXTS`:- "...release of my **new single**..." - 'new single' relates to the word 'single' found in TOKENS_AFTER_MAX_ACTIVATING_TOKEN. 'new' is in MAX_ACTIVATING_TOKENS.- "...got to **visit an** elephant sanctuary..." - 'visit an' relates to 'visit' in MAX_ACTIVATING_TOKENS and 'an' in TOP_POSITIVE_LOGITS.- "...an 80 year old diabetic male..." - 'an' is in TOP_POSITIVE_LOGITS.- "...It finds a free space in RAM, copies the program's code and data, and sets up the **necessary** memory structures." - 'necessary' is in TOKENS_AFTER_MAX_ACTIVATING_TOKEN.The pattern seems to involve articles (`an`, `a`, `the`) and common words, often preceding or following specific nouns or verbs. The presence of 'new single', 'visit an', and 'necessary' points towards descriptive phrases or clauses.Let's consider specific common sequences. `an` and `a` appear frequently in `TOP_POSITIVE_LOGITS`. `an` is also in `TOKENS_AFTER_MAX_ACTIVATING_TOKEN`. `the` is in `MAX_ACTIVATING_TOKENS`.The neuron might be looking for specific sequences or the continuation of certain word types. Given "new single", "visit an", and "necessary", it might be related to announcing things or describing requirements.Let's re-evaluate `MAX_ACTIVATING_TOKENS`: `!`, `)`, `new`, `the`, `visit`, `.`, `.`, `riv`, `_`.`TOKENS_AFTER_MAX_ACTIVATING_TOKEN`: `tag`, `in`, `single`, `necessary`, `an`, `He`, `ol`, `USE`.`TOP_POSITIVE_LOGITS`: `an`, `de`, `Al`, `Find`, `a`, `Ab`, `Ed`, `In`.The presence of `new` in `MAX_ACTIVATING_TOKENS` and `single` in `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` strongly suggests "new single".The presence of `visit` in `MAX_ACTIVATING_TOKENS` and `an` in `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` suggests "visit an".The presence of `an` in `TOP_POSITIVE_LOGITS` supports this article usage.The presence of `necessary` in `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` could be part of a phrase like "the necessary..." or "setup the necessary...".The neuron seems to pick up on grammatical constructions, specifically articles and common words that lead into more descriptive or functional parts of a sentence.Looking at the specific words: `an`, `a`, `the`, `new`, `single`, `visit`.The combination `an` and `a` from `TOP_POSITIVE_LOGITS` with `an` from `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` and `the` from `MAX_ACTIVATING_TOKENS` points to articles.Coupled with `new` and `single` from `MAX_ACTIVATING_TOKENS` and `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` respectively, this strongly hints at "new single".`visit` from `MAX_ACTIVATING_TOKENS` plus `an` from `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` hints at "visit an".The neuron is detecting specific common word sequences or grammatical patterns involving articles and common descriptive words.A concise explanation could be "new single" or "visit an". However, it needs to be more general if it covers both.The list of `TOP_POSITIVE_LOGITS` has `an`, `a`, `Al`, `Find`, `Ab`, `Ed`, `In`. These are mostly articles or capitalized words that might start a sentence or be proper nouns.The `MAX_ACTIVATING_TOKENS` have `new`, `the`, `visit`.The `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` have `single`, `necessary`, `an`.The most prominent pattern is the sequence "new single" and "visit an". The neuron seems to be triggering on common noun phrases or introductory phrases, particularly those with articles."new single" is a very strong candidate."visit an" is another."an" is a very common token, and it appears in both `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` and `TOP_POSITIVE_LOGITS`. `an` is also often preceded by words like `visit`.Consider the combination of article-preceding-word.`visit` (MAX) -> `an` (AFTER)`new` (MAX) -> `single` (AFTER)The neuron highlights common phrases involving the announcement of new items or visiting places.Let's try to find a common theme.The articles `an` and `a` are central. `new` and `single` are together. `visit` and `an` are together. `necessary` is also there.How about "announcements and common phrases"? Too long."new single" captures one element."visit an" captures another.The prompt says "finding patterns in lists".The pattern of [verb/adjective + article] or [adjective + noun] is present.`visit` (token) `an` (token after)`new` (token) `single` (token after)The `TOP_POSITIVE_LOGITS` are articles and capitalized words.The `MAX_ACTIVATING_TOKENS` include `new`, `the`, `visit`.The `TOKENS_AFTER_MAX_ACTIVATING_TOKEN

    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_16_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
    ⏰
    0.56
    সময়
    0.55
    codon
    0.54
    \},
    0.52
    ৩৩
    0.52
    🔯
    0.52
    当時
    0.51
     pushedAt
    0.51
    ⛄
    0.51
    🔩
    0.51
    POSITIVE LOGITS
     an
    0.61
     de
    0.59
     
    0.57
     Al
    0.55
     Find
    0.55
     a
    0.53
     Ab
    0.52
     Ed
    0.52
    ↵
    0.52
     In
    0.51
    Activations Density 0.000%

    No Known Activations