© 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-3-4B-IT
    3. 3-GEMMASCOPE-2-TRANSCODER-262K
    4. 6400
    Prev
    Next
    INDEX
    Explanations

    The explanation should focus on mathematical terms, especially those related to algebra and quadratic equations.- `MAX_ACTIVATING_TOKENS` contains `px`, `ax`, `bx`, `ix`, `mx`. These are standard mathematical notations for coefficients and variables.- `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` shows `+`, `^`, `²`. This strongly suggests operations like addition and squaring (`ax^2`, `bx`).- `TOP_ACTIVATING_TEXTS` examples: - "cubic equation in the form x³ + px + q = 0" - "$(ax^2+b)^2$" - "$\mathbf{ax² + bx + c = 0}$" - "quadratic equation of the form ax^2 + bx + c = 0" - "$x = ax + by + c$" (this one is a bit different, involves systems of equations or transformations) - "quadratic formula to solve the equation $ax^2 + bx + c = 0$" - "general form of a quadratic equation is: **ax² + bx + c = 0**" - "quadratic equation of the form ax² + bx + c = 0"- `TOP_POSITIVE_LOGITS` shows some non-English words, but also `toep` (part of 'coefficient' perhaps?), `undirected`, `oldest`. The non-English words obscure the pattern but the presence of mathematical terms in the other lists is strong.The overwhelming pattern is the detection of forms closely related to quadratic equations (`ax^2 + bx + c = 0`) and potentially cubic equations. The tokens `ax`, `bx`, `px` followed by `+` or operators like `^2` are strong indicators.A good concise explanation would capture this mathematical context.Possible explanations:* quadratic equations* algebraic coefficients* ax bx c form* polynomial forms* mathematical equations"quadratic equations" is quite specific and covers the most prominent pattern."ax bx c form" is also good and very literal to the tokens."polynomial forms" is slightly more general than quadratic, but still fitting.Considering the prompt: "finding patterns in lists" and "concise explanation (3 to 20 words)", "specific", and "not just listing tokens"."quadratic equations" - 2 words."ax bx c form" - 3 words.Let's re-examine `MAX_ACTIVATING_TOKENS`: `px`, `ax`, `ax`, `bx`, `ey`, `ax`, `bx`, `ix`, `mx`, `bx`. This is heavily weighted towards coefficients like `a`, `b`, `p`, `m`. The ones that appear frequently are `ax`, `bx`.`TOKENS_AFTER_MAX_ACTIVATING_TOKEN`: `+`, `^`, `²`, `+`, `+`, `^`, `+`, `+`, `+`, `+`. This strongly suggests `ax + b`, `ax^2`, etc.The `TOP_ACTIVATING_TEXTS` are almost exclusively about quadratic equations (`ax^2 + bx + c = 0`). It also mentions cubic equations.Therefore, "quadratic equations" is a very good fit. "polynomial coefficients" could also work. "algebraic terms" fits too.The prompt asks for "what the neuron detects or predicts by finding patterns in lists."The pattern is the structure of algebraic expressions, particularly quadratic ones.Let's try to be as specific as the data allows without listing.`ax`, `bx`, `c` are coefficients. `px` is also coefficient-like.The operators `+`, `^2` strongly link these coefficients into equation forms."quadratic equation forms" - 3 words."algebraic equation structures" - 3 words."quadratic and cubic forms" - 4 words."ax bx + c" - 4 words. This is very literal."ax^2 + bx + c" - 5 words. This is also very literal.Given `MAX_ACTIVATING_TOKENS = ['px', 'ax', 'ax', 'bx', 'ey', 'ax', 'bx', 'ix', 'mx', 'bx']`, and `TOKENS_AFTER_MAX_ACTIVATING_TOKEN = ['+', '^', '²', '+', '+', '^', '+', '+', '+', '+']`.The presence of `ax`, `bx` followed by `+` or `^2` or `²` is extremely strong.The `TOP_ACTIVATING_TEXTS` cement this by repeatedly mentioning `ax^2 + bx + c = 0`.I think "quadratic equation structures" or "quadratic equation forms" is appropriate."quadratic equation structures" - is 3 words."quadratic equation forms" - is 3 words."ax^2 + bx + c" - is 5 words. This is very specific to the form."algebraic equation terms" - 3 words.Let's consider the prompt's example: "unique words" is not specific enough. "foreign words" is not specific.Here, "quadratic equations" is specific.Let's check the word count: 3 to 20 words.The explanation should be specific.The TOP_POSITIVE_LOGITS are a distraction if they are in other languages or unrelated. The core pattern is in the tokens and texts.The neuron is firing for tokens that are coefficients (`a`, `b`, `p`, `m` etc.) and when these are followed by operators (`+`, `^2`) that form algebraic equations, especially quadratic ones, as seen in the sample texts."quadratic equation forms" seems like a good balance of specificity and conciseness."algebraic equation components" - 3 words.What if we focus on the tokens themselves? `ax`, `bx`, `px`, `ix`, `mx`.And the following `+`, `^`, `²`.This implies terms like `ax`, `bx`, `ax^2`, `bx^2`, `ax + b`.Let's stick with the overall mathematical context."quadratic equation structure" as a potential answer.quadratic equation structure

    np_acts-logits-general · gemini-2.5-flash-lite
    New Auto-Interp
    Top Features by Cosine Similarity
    Configuration
    google/gemma-scope-2-4b-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
    س
    3.92
    ie
    3.02
    ્સ
    2.98
     teg
    2.66
    aş
    2.66
     dakika
    2.65
    رف
    2.64
    ná
    2.63
    Ě
    2.62
    ü
    2.56
    POSITIVE LOGITS
    אם
    2.81
     лиц
    2.78
    ئے
    2.51
     toep
    2.49
     undirected
    2.43
     oldest
    2.43
    ়
    2.42
    ुक्त
    2.41
     अमाव
    2.40
    чения
    2.36
    Activations Density 0.037%

    No Known Activations