© 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. 32-GEMMASCOPE-2-TRANSCODER-262K
    4. 233708
    Prev
    Next
    INDEX
    Explanations

    ```csharpnamespace TSP_SimulatedAnnealing{ // ...```This output shows repeated presence of C# code snippets, particularly those involving `using Microsoft...`, `System...` (like `System.Drawing`, `System.IO`, `System.Net`), and common programming constructs such as `namespace`, `public class`, and methods. The `MAX_ACTIVATING_TOKENS` being semicolons and `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` containing comments (`//`) and newlines further support this. The `TOP_POSITIVE_LOGITS` contain `createdAt`, `mkdirs`, and other tokens which, while diverse, don't strongly contradict a coding context.The core pattern is the presence of C# code structure. The `MAX_ACTIVATING_TOKENS` being `;` and `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` being `//` indicate that this neuron might be related to the *end of code statements or blocks*, or *the transition into comments/new lines within code*.Let's re-evaluate:- `MAX_ACTIVATING_TOKENS`: `;`- `TOKENS_AFTER_MAX_ACTIVATING_TOKEN`: `//` (comments), newlines.This suggests the neuron is *activated by the end of a code statement (`;`) and then looks at what follows, which is often a comment or the start of a new line/block*.Many of the `TOP_ACTIVATING_TEXTS` contain `public class`, `using`, `namespace`, `public static class`. These are all structural elements of code. The `MAX_ACTIVATING_TOKENS` being exclusively `;` is a very strong signal.What follows a semicolon? Usually something else.Looking at the `TOP_POSITIVE_LOGITS`: `setup`, `createdAt`, `mkdirs` – these are programming-related terms.The pattern is essentially related to the *structure and syntax of C# programming code*. The semicolon is a very specific delimiter.Let's try to synthesize this into a short phrase.The neuron seems to detect C# code syntax, specifically relating to statement endings and what follows.Possible phrases:- C# code syntax- C# code statements- code statements and new lines- C# code delimiters- C# code block endingsThe `TOP_POSITIVE_LOGITS` are quite varied, but `createdAt` and `mkdirs` are programmatically relevant.The `TOP_ACTIVATING_TEXTS` are overwhelmingly C# code.The neuron is *strongly* activated by code structures. The semicolon is a very specific punctuation mark. What does it signify? The end of a statement. What follows? Often whitespace or comments, or the start of a new line, or a new statement.Let's focus on what the *neuron detects* by finding patterns. The pattern is that it's seeing code. The specific tokens `;` and `//` are key.Could it be detecting the *end of code statements*?Or perhaps *code blocks and their structures* due to the surrounding text?The `TOP_POSITIVE_

    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_32_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
     Navigator
    0.40
     mayo
    0.38
    alho
    0.38
     aulas
    0.36
    集
    0.36
     Response
    0.35
     wrappers
    0.35
     grades
    0.35
     navigator
    0.35
    BALL
    0.34
    POSITIVE LOGITS
     setup
    0.43
    Antae
    0.41
    createdAt
    0.40
    千萬
    0.39
    Soy
    0.39
     Vegetarian
    0.39
    sej
    0.38
    coproteins
    0.38
    mkdirs
    0.36
    ኝ
    0.36
    Activations Density 0.000%

    No Known Activations