© 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. 127183
    Prev
    Next
    INDEX
    Explanations

    The neuron seems to detect terms related to commands, power calculations (watts, kilowatts), and specific naming conventions (Net, KillNET, Lambert, Ink). The "ow" followed by "atts" strongly suggests a link to power measurements.Let's refine this based on the rules:- Concise (3-20 words).- Specific.- No "tokens", "patterns", "detects", "predicts".- No capitalization unless proper noun.- Find a pattern across majority.Looking at the `MAX_ACTIVATING_TOKENS`:`NET`, `oj`, `Command`, `Lambert`, `ow`, `ow`, `ow`, `ow`, `Ink`, `l`Looking at `TOKENS_AFTER_MAX_ACTIVATING_TOKEN`:`-`, `oules`, `:**`, `)`, `atts`, `atts`, `atts`, `atts`, `:**`, `one`The most prominent pattern is `ow` followed by `atts`, which appears multiple times. This strongly suggests the neuron is activating on mentions of power, specifically "watts" (`atts`). The preceding `ow` is likely part of numbers or units related to power.Other tokens:- `Command` followed by `-` or `one` (Kill Command, exit)- `Lambert` followed by `)` (Kyle Lambert)- `Ink` appears in `TOP_POSITIVE_LOGITS` and `TOP_ACTIVATING_TEXTS` mentions "Kid Ink".- `NET` appears in `TOP_POSITIVE_LOGITS` and `TOP_ACTIVATING_TEXTS` mentions "KILLNET".The clearest and most repeated pattern is `ow atts`. The `ow` is likely part of a numerical value or context preceding "watts".How to express "watts" or power measurement concisely?The `TOP_POSITIVE_LOGITS` include `.`, `outcry`, `in`, `you`, `0`, `ene`, `ประมาณ`, `egy`, `НЫ`. These are a bit mixed and not as clear as the activating texts/tokens.Let's revisit the `TOP_ACTIVATING_TEXTS`:- mentions of "Kill Command"- mentions of "Kyle Lambert"- mentions of "Kilowatts", "Watts" (e.g., "2.886,960 J", "Convert to Kilojoules", "PowerUsageKW DECIMAL", "Brightness:** ~38W", "Watts or Kilowatts", "Power Consumption (in Watts or Kilowatts)")- mentions of "Kid Ink", "KILLNET"The neuron seems to pick up on technical contexts, commands, and specific entities/names, but the most striking recurring element from `MAX_ACTIVATING_TOKENS` + `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` is `ow atts` -> "watts". The `TOP_POSITIVE_LOGITS` also have `0` and `egy` (could be part of energy) and `in`, `outcry` which could relate to usage or flow.Given the strong presence of "Watts", "Kilowatts" in the activating texts and `ow` followed by `atts` in the token lists, "power measurement" or "watts" is a strong candidate.Let's check if there’s a more specific pattern.`ow atts` is the most explicit.What about `NET`? It appears in `TOP_POSITIVE_LOGITS`. In the texts, "KILLNET" is mentioned. `NET` could be related to network or a specific name.`Ink` in `TOP_POSITIVE_LOGITS` and texts "Kid Ink".The repetition of `ow` and `atts` is very strong. It directly points to power consumption.Let's consider the "Kill Command" aspect. "Command" is a MAX_ACTIVATING_TOKEN."NET" and "Ink" are proper nouns from the lists.If I have to pick ONE thing, "watts" seems very evident.The rule is "look at the lists and find a pattern".`ow` followed by `atts` is a clear sequence pattern.How to phrase it?- power calculation- watts consumption- measurements in watts- power units- electricity usageThe prompt says "your response should be a concise explanation (3 to 20 words) that captures what the neuron detects or predicts by finding patterns in lists."The pattern `ow atts` in conjunction with texts about Watts/Kilowatts is the strongest."power measurements" fits the criteria."electricity usage" also fits.Let's try to be more specific to the tokens. `ow atts` directly relates to watts.The `TOP_POSITIVE_LOGITS` include things like `in`, `outcry`, `you`, `0`. This might suggest usage or numbers.The `MAX_ACTIVATING_TOKENS` also has `NET`, `Lambert`, `Ink`, `Command`.These point to specific entities OR commands.Could it be "command usage figures"? or "network power data"?The `ow atts` pattern is very strong. The texts explicitly mention "Watts" and "Kilowatts".Let's consider the combined information.electricity usage

    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
    b
    0.79
    Cellphone
    0.71
    t
    0.64
    profitable
    0.64
    ală
    0.63
    los
    0.63
    finale
    0.62
     Estadística
    0.61
    schedule
    0.61
    election
    0.61
    POSITIVE LOGITS
     in
    0.89
    .
    0.68
     outcry
    0.67
    ใน
    0.63
     you
    0.61
    0
    0.59
    ene
    0.58
     ประมาณ
    0.58
     egy
    0.58
    НЫ
    0.57
    Activations Density 0.000%

    No Known Activations