© 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. Olmo-3-1125-32B
    3. 32-RES-BATCHTOPK-131K
    4. 111765
    Prev
    Next
    INDEX
    Explanations

    aloneThe word 'alone' appears consistently in `MAX_ACTIVATING_TOKENS`, `TOP_POSITIVE_LOGITS`, and `TOP_ACTIVATING_TEXTS`. The contexts in `TOP_ACTIVATING_TEXTS` revolve around the state of being alone, not being alone, or feeling isolated. This is the dominant pattern.The word 'own' and 'myself' are also present and relate to solitude or self-reliance.The `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` are diverse, so they don't offer a strong secondary pattern.The explanation should capture the essence of 'alone' and related concepts."alone""being alone""solitude""feeling alone""state of being alone"Let's check the constraints:- Concise explanation (3 to 20 words): "being alone" is 2 words, too short. "state of being alone" is 4 words.- Captures what the neuron detects or predicts by finding patterns in lists.- Could be about tokens following or preceding certain tokens.- Could be about words starting with a sequence.- Avoid simply listing all the tokens. Find patterns.- Just say the pattern itself, do not start with "This neuron detects/predicts".- Do not mention "tokens" or "patterns".- Do not capitalize the first letter unless it is a proper noun.- The explanation should be specific.From `MAX_ACTIVATING_TOKENS`: alone, alone, ONE, own, alone, alone, myself, alone, first, unique.From `TOP_POSITIVE_LOGITS`: alone, atop, -alone, amongst, Alone, zan, alone, among, Lor, ech.From `TOP_ACTIVATING_TEXTS`: "grieve alone", "not alone", "not feel ALONE", "shivers now alone", "you aren’t alone", "stadium to myself", "remain alone in first place", "on your own".The word "alone" is the strongest signal. "own" and "myself" are related to individual states. "first" and "unique" are less consistently tied to the core concept of solitude.The phrase should describe the *condition* or *state* that this neuron is sensitive to, using the words found."being alone or singular" - 4 words"individual solitude or uniqueness" - 3 words"alone, oneself, or unique" - 4 words. This feels a bit list-like.The core theme is 'alone'. The modifiers are sometimes 'myself', 'own', 'first', 'unique'.The combination of 'alone' and 'first place' in one text ('remain alone in first place') suggests a connection to being uniquely positioned or isolated in a leading role.'unique' is also in `MAX_ACTIVATING_TOKENS`.'first' is also in `MAX_ACTIVATING_TOKENS`.Let's re-evaluate:The pattern is very strongly "alone".The contexts from `TOP_ACTIVATING_TEXTS` include:- Emotional state: "grieve alone", "you're not alone". This implies the neuron is sensitive to statements about the state of being alone, or the contrast with it.- Physical state/situation: "shivers now alone", "stadium to myself".- State of being the only one: "remain alone in first place". This links "alone" with being singular or leading.- Self-reliance: "on your own".The phrase "alone and oneself" captures 'alone' and 'myself'/'own'."alone or singular" captures 'alone' and 'first'/'unique'."alone and solitary" seems a good fit too.Let's look at TOP_POSITIVE_LOGITS again: alone, atop, -alone, amongst, Alone, zan, alone, among, Lor, ech.'atop', 'amongst', 'among' suggest proximity or relative position.'zan', 'Lor', 'ech' seem like noise or non-English words, or misspellings.Considering the strongest signal is 'alone', and the secondary signals are 'myself', 'own', 'first', 'unique':- 'alone and oneself'- 'alone and singular'- 'alone and unique'Checking constraints again:- 3 to 20 words.- specific. "alone and singular" is specific.- no "tokens", "patterns", "detects".- lower case unless proper noun.The neuron seems to be sensitive to the state of being alone, or being the only one in a certain context."alone and singular" captures both the core meaning of 'alone' and the implication of 'first', 'unique', 'own', 'myself'."alone and oneself" captures 'alone' and 'myself'.Let's see if 'singular' or 'oneself' better covers the range.'singular' covers 'alone', 'first', 'unique'.'oneself' covers 'alone', 'myself', 'own'.The word 'unique' is also present in `MAX_ACTIVATING_TOKENS`.The word 'first' is also present in `MAX_ACTIVATING_TOKENS`.So, the concept is not just 'alone' but also 'unique' or 'first'."alone and unique" is 3 words."alone and singular" is 3 words.The pattern from the texts is often about contrast ("we grieve alone, and yet we are not alone"), or about a state of solo existence."alone" is the primary.Secondary indicators: 'myself', 'own', 'first', 'unique'."alone and standalone" could work."alone and solitary confinement" - nope, too specific and not supported strongly.Let's try to combine the dominant 'alone' with the other standalone concepts.'alone' (solitude)'myself', 'own' (self-reliance, individual)'first', 'unique' (singular, leading)"alone, unique, or own" - 4 words."solitary, unique, or self" - 4 words.The simplest, most direct pattern from `MAX_ACTIVATING_TOKENS` and `TOP_POSITIVE_LOGITS` is 'alone'.The text examples confirm this.Are 'first' and 'unique' just co-occurring or intrinsically linked?"remain alone in first place" links them."unique string in it" links unique to contextually important items.If I focus ONLY on `MAX_ACTIVATING_TOKENS`:alonealoneONE -> can mean unique, or the number 1.ownalonealonemyselfalonefirstuniqueThe concept is definitely about being *one* or *alone*.The tokens `first` and `unique` are strongly associated with this neuron.The `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` are varied, but the most frequent are '.' and 'in'. This doesn't help

    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
     Neu
    -0.10
    haft
    -0.09
     correspond
    -0.09
    enson
    -0.09
     Correspond
    -0.09
     indiscrim
    -0.08
     Nes
    -0.08
    REC
    -0.08
     succession
    -0.08
    oxy
    -0.08
    POSITIVE LOGITS
     alone
    0.11
     atop
    0.11
    -alone
    0.11
     amongst
    0.11
     Alone
    0.10
    zan
    0.10
    alone
    0.10
     among
    0.10
     Lor
    0.10
    ech
    0.09
    Activations Density 0.009%

    No Known Activations