© 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. Olmo-3-1125-32B
    3. 32-RES-BATCHTOPK-131K
    4. 115045
    Prev
    Next
    INDEX
    Explanations

    The neuron appears to be triggered by words that often precede specific organizational titles, roles, or events, particularly when followed by another noun or a concluding punctuation. It seems to be identifying contexts related to formal structures, decisions, and the passage of time within those contexts.Given the strong signal of "Senate Intelligence" and other common phrases like "Division of" or "president of", a good explanation might focus on these formal structures. The presence of "decision" and "year" also suggests a focus on processes and timelines within these structures.Thinking about the most salient patterns:- "Senate Intelligence"- "Division of"- "president of"- "decision." (finality after a decision)- "year" (temporal context)- "now as" (current context)The neuron seems to pick up on words that kick off or are part of established structures or significant moments. "Intelligence" appearing after "Senate" is a very strong clue. "Division of" is also very structured.Let's try to combine these: detection of formal titles, organizations, or concluding remarks.Considering the provided examples:- "Magical Object Division of the Faerie Affairs Bureau" -> "Division of"- "Sworn in as president on November 24..." -> "president"- "decision." -> "decision"- "Senate Intelligence Committee" -> "Senate"The pattern "X of Y" is common. "X Intelligence" is also strong. "President" is a role, "Division" is an organization. "Decision" is an event. "Year" is temporal. "Now" is temporal.Perhaps it is about identifying key figures, departments, or official proceedings/conclusions.Let's re-evaluate the prompt: "captures what the neuron detects or predicts by finding patterns in lists."Concise (3-20 words).No "tokens" or "patterns".No "This neuron detects/predicts".Specificity.Looking at TOP_POSITIVE_LOGITS: tonight, recently, promised, fifteen, lately, finally, six, eighth, seven, twelve.These words are mostly time-related (tonight, recently, lately, six, eighth, seven, twelve) or conclusion/finality related (promised, finally, fifteen).official roles and organizational structures

    np_acts-logits-general · gemini-2.5-flash-lite
    New Auto-Interp
    Top Features by Cosine Similarity
    Configuration
    bcywinski/Olmo-3-32B-Base-SAE/saes_allenai_Olmo-3-1125-32B_batch_top_k/resid_post_layer_32
    Prompts (Dashboard)
    16,384 prompts, 128 tokens each
    Dataset (Dashboard)
    monology/pile-uncopyrighted
    No Configuration Found
    Embeds
    IFrame
    Link
    Not in Any Lists

    No Comments

    Negative Logits
    Qi
    -0.10
     Tweets
    -0.09
     aforementioned
    -0.09
    Reviewed
    -0.09
     ca
    -0.09
     tiers
    -0.09
     subsequently
    -0.08
    TR
    -0.08
     later
    -0.08
    AAP
    -0.08
    POSITIVE LOGITS
     tonight
    0.12
     recently
    0.11
     promised
    0.11
     fifteen
    0.11
     lately
    0.11
     finally
    0.10
     six
    0.10
     eighth
    0.10
     seven
    0.10
     twelve
    0.10
    Activations Density 0.221%

    No Known Activations