© 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. 3-GEMMASCOPE-2-TRANSCODER-262K
    4. 86454
    Prev
    Next
    INDEX
    Explanations

    The neuron appears to be sensitive to Latinate suffixes and prefixes, particularly those found in words related to educational roles and academic contexts.Based on the provided lists, here's an attempt to find a pattern:* **MAX_ACTIVATING_TOKENS**: 'ento', 'ister'* **TOKENS_AFTER_MAX_ACTIVATING_TOKEN**: 'Mechan', 'optim', 'ium', 'ial'* **TOP_POSITIVE_LOGITS**: 'il', 's', 'ological', 'ado', 'ia', 'c', 'al', 'rom'* **TOP_ACTIVATING_TEXTS**: contain terms like "magento", "Magister Mechanicus", "magistral", "magister optimus", "Magisterium".The recurring themes point towards Latin or words derived from Latin, specifically related to titles, learning, or systems.* 'ento' appears in words like 'magento' (a platform, derived from the Spanish word for magnet).* 'ister' strongly suggests 'magister' (master, teacher).* 'Mechan' suggests 'mechanicus'.* 'optim' suggests 'optimus'.* 'ological' suggests academic subjects.* 'Magisterium' is a central theme.The most prominent and consistent theme seems to revolve around the word "Magister" and related Latinate terms.Concise explanation: **magister and related Latinate terms**Let's try to be more specific to the *tokens* provided.'ento' + 'omechanical'/'ento' + 'logy''ister' + 'Mechanicus' / 'ister' + 'logy' / 'ister' + 'al'Considering the TOP_POSITIVE_LOGITS also contains 'il', 's', 'ological', 'ado', 'ia', 'c', 'al', 'rom', these are common suffixes and parts of words.The specific sequence 'ister' followed by 'Mechan' is strongly indicative of 'Magister Mechanicus'. The word 'magento' (derived from 'magnet') also appears.If I have to choose just one pattern or concept:The neuron seems to activate for terms related to 'Magister' and specific types of academic/technical roles or systems.Let's look at the `MAX_ACTIVATING_TOKENS` again: `ento`, `ister`.And `TOKENS_AFTER_MAX_ACTIVATING_TOKEN`: `Mechan`, `optim`, `ium`, `ial`.The combination of `ister` + `Mechan` -> `Magister Mechanicus`.The combination of `ister` -> `magister`, `magistral`, `magisterium`.The `ento` might relate to `magento` or suffixes.The `TOP_POSITIVE_LOGITS` like `ological` reinforce a learned/academic context.A broad interpretation could be "academic teacher/master roles".But looking for specific *token patterns* as requested:"ister" followed by "Mechan" or completions leading to academic titles.Let's try to find a more concrete, unified pattern.'ento' appears in 'magento'.'ister' appears around 'Magister', 'Magisterium', 'Mechanicus'.The pattern could be "words beginning with 'magister' or featuring 'ento'".This is still a bit broad.Let's focus on `MAX_ACTIVATING_TOKENS` and `TOKENS_AFTER`:`ister` + `Mechan` is a very strong hit for `Magister Mechanicus`.`ister` + `ial` -> `magistral``ister` + `um` (`ium` in list) -> `magisterium`The `ento` part:`ento` itself is not common without a prefix like `mag`.The text snippet "Hire magento 2 developers" has 'magento'.The `TOP_POSITIVE_LOGITS` like `ological`, `al`, `il`, `c` are common word endings in Latin/English.The neuron seems to be looking for specific Greek/Latin root-based words, particularly those involving 'magister' and technical/academic domains.Let's consider what makes the neuron *activate*.The presence of 'ister' is a strong signal. The tokens following 'ister' in the examples are 'Mechan', 'ial', 'ium'. These lead to 'Magister Mechanicus', 'magistral', 'magisterium'.The `ento` token in `MAX_ACTIVATING_TOKENS` is a bit less clear from the direct `TOKENS_AFTER` list, but it appears in "magento" in the `TOP_ACTIVATING_TEXTS`.If the neuron captures the concept of "master/teacher/ruler" or related entities:"magister and related terms""academic/technical master roles"Let's re-read the rules:- Concise (3 to 20 words)- Single word, phrase, or pattern.- About tokens following or preceding.- About words starting with a sequence.- Avoid simply listing. Find patterns.- Just say the pattern itself.- No "This neuron detects/predicts".- No "tokens" or "patterns" in explanation.- Not ALL texts/tokens have to match.- Specific.The most specific recurring element is "Magister".The tokens `ister` and `Mechan` strongly point to `Magister Mechanicus`.The `TOP_POSITIVE_LOGITS` like `ological` and `al` suggest it's not just `Magister` but related concepts or suffixes.The neuron might be detecting phrases related to "Master/Teacher" in specific contexts.Consider the examples:"Magister Mechanicus" - a specific title."magister optimus" - teacher best."Magisterium" - a governing body (often religious or academic)."magento" - commercial platform.The common denominator is "master" or "teacher" related words, often with a technical or organizational aspect.If I have to find a pattern in *tokens*:- `ister` + `Mechan`- `ister` + `ial`- `ister` + `sapiens` (implied from 'optimus' and quote context)And the `ento` token from `MAX_ACTIVATING_TOKENS` with its presence in `magento`.This is a bit of a blend.How about focusing on the root and its common associations?The root "magister" (master, teacher) is present in multiple examples."Mechanicus" (mechanic, engineer) is associated."magento" is present.Let's try to find a short phrase that captures these."magister, mechanicus, magento" - lists specific words, not ideal."master, teacher, engineer, platform" - too generic.Let's re-examine the `MAX_ACTIVATING_TOKENS`:`ento``ento``uro` (appears in, e.g.,

    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_3_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
    า
    1.01
    ை
    0.89
    ೆ
    0.78
    ಿ
    0.75
    ो
    0.71
    IS
    0.70
    াৰ
    0.68
    ো
    0.67
    い
    0.66
    </u>
    0.65
    POSITIVE LOGITS
    il
    1.11
    s
    0.89
    ological
    0.71
    ado
    0.69
    ية
    0.69
    ς
    0.68
    al
    0.68
    c
    0.68
    and
    0.67
    rom
    0.67
    Activations Density 0.001%

    No Known Activations