© 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. Gemma-3-27B-IT
    3. 2-GEMMASCOPE-2-TRANSCODER-262K
    4. 107047
    Prev
    Next
    INDEX
    Explanations

    The dominant token is `ensively`.Looking at `TOP_ACTIVATING_TEXTS`, I see several instances of "inexpensively" or "expensively", and "extensively".For example:- "can be bought relatively inexpensively"- "produced relatively quickly and inexpensively"- "Stylus is easily and inexpensively replaceable"- "discuss boundaries *extensively*"- "extensively describes the physical interaction"The `MAX_ACTIVATING_TOKENS` list heavily features `ensively`. The `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` list contains words like `replaceable`, `solid`, and `describes`.The `TOP_POSITIVE_LOGITS` list has `ad`, which might relate to "ad campaigns" mentioned in the text.The core pattern seems to be words ending with "-ensively" or relating to cost/degree.Specifically, "inexpensively" and "expensively" appear frequently in the context of cost. "extensively" appears frequently in the context of degree or detail.The neuron seems to activate for words that describe *how* something is done, often relating to cost (inexpensive, expensive) or extent (extensive).Let's check the relationship between `MAX_ACTIVATING_TOKENS` and `TOKENS_AFTER_MAX_ACTIVATING_TOKEN`.- `ensively` + `)`- `ensively` + `(`- `ensively` + `replaceable`- `oodles` + `.` (outlier)- `ensively` + `)`- `ensively` + `*`- `overlap` + ``` ` ` (outlier)- `ensive` + `,"`- `ensively` + `solid`- `ensively` + `describes`The combination `ensively` + `describes` is notable.The combination `ensively` + `solid` (defensively solid).The combination `ensively` + `replaceable` (inexpensively replaceable).The most consistent pattern is the "-ensively" suffix.Words like "inexpensively", "expensively", "extensively", and "defensively" are highlighted.The items in `TOP_POSITIVE_LOGITS` are too generic (`be`, `and`, `in`) or appear as abbreviations (`ad`). The presence of `ad` might link to `ad campaigns`, which were described as "inexpensively" or "quickly".The focus is on adverbs, particularly those ending in `-ensively`, describing manner, cost, or extent.Let's try to find a concise phrase."words ending in -ensively" - too literal."describing cost or extent" - good, but "defensively" doesn't fit perfectly."manner, cost, or extent adverbs" - a bit long.The `MAX_ACTIVATING_TOKENS` list is dominated by `ensively`. This suggests the neuron is very sensitive to this suffix.The `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` show what follows. If `ensively` is the core, then `replaceable`, `solid`, `describes` are contexts.Considering "inexpensively", "expensively", "extensively", "defensively":"cost or manner adverbs""manner adverbs""adverbs of manner, cost, or extent"The presence of `oodles` and `overlap` is a bit strange if the primary focus is `-ensively`. However, `ensively` appears 7 out of 10 times. `ensive` is also present.This strongly points to the `-ensively` suffix.Let's consider the TOP_POSITIVE_LOGITS: `ad` is the most differentiating word. "ad campaigns" were mentioned in top text."inexpensively" and "expensively" relate to cost, especially for advertisements or products."extensively" relates to how much something is done."defensively" relates to a manner of acting.The pattern is primarily about *how* something is done, often relating to cost (inexpensive), duration/degree (extensive), or a mode of action (defensive).What if the neuron fires for specific adverbial phrases or completions?"adverbs modifying action or cost"Let's try to be more specific based on the suffix."describing manner or cost"This captures "defensively", "inexpensively", "expensively". "extensively" also fits manner/degree.Let's re-examine `TOP_POSITIVE_LOGITS`. `ad` could be strong. `w` `t` `ut` `in` `be` `and` are very common. `ز` `AY` `우` are unusual.The unusual tokens might indicate a specific domain or language feature, but without more context for them, focusing on the `-ensively` pattern seems safer.The pattern `-ensively` is very strong.Examples: "inexpensively", "expensively", "extensively", "defensively".These describe *how* something is done.The `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` are quite varied, but `describes` and `solid` point to descriptive adverbs.Concise options:- manner and cost adverbs- adverbs of manner or cost- describing manner or cost- cost or extent adverbs- adverbs ending in -ensively"cost or extent adverbs" seems to capture the essence well. "defensively" falls under manner.The phrasing "describing manner or cost" is also good.Let's consider the most frequent token in `MAX_ACTIVATING_TOKENS`: `ensively`.This strongly suggests adverbs ending in this suffix.The meaning of these adverbs are:- inexpensively: low cost- expensively: high cost- extensively: to a great degree or extent- defensively: in a defensive mannerSo, the neuron detects adverbs describing manner, cost, or extent."manner, cost, or extent adverbs" is a good description. It's 6 words, within the limit."describing manner, cost, or extent" - also good.Let's simplify. "adverbs of manner and cost" covers most. "extensively" fits manner/degree.The presence of `ad` in `TOP_POSITIVE_LOGITS` suggests advertising context, where cost is very relevant."adverbs describing cost or manner""adverbs of cost and manner"I'll go with a phrase that is descriptive and covers the primary patterns. "describing manner or cost" is concise and accurate for "defensively", "inexpensively", "expensively". "extensively" also fits "manner" in the sense of "how much".Final check:- 3 to 20 words: "describing manner or cost" is 4 words. Ok.- captures what the neuron detects/predicts: Yes, it detects adverbs that describe how something is done or its cost.- finding patterns in lists: Yes, the `-ensively` suffix is the primary pattern.- concise explanation: Yes.- specific: Yes, specifies 'manner' and 'cost'.- avoids "tokens", "patterns", etc.- no leading phrases.- not capitalized unless proper noun.Consider "adverbs of manner or cost"

    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_2_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
    ס
    1.27
    า
    1.23
    ا
    1.17
    س
    1.10
    ",
    0.95
     घ्या
    0.94
    daki
    0.93
    ’
    0.92
    daten
    0.88
    dale
    0.86
    POSITIVE LOGITS
     be
    1.45
    and
    1.34
    ut
    1.20
    w
    1.20
    t
    1.18
    ز
    1.18
    in
    1.05
    AY
    1.05
    ad
    1.02
    우
    0.99
    Activations Density 0.000%

    No Known Activations