© 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. Qwen3.5-9B Base
    3. 15-QWENSCOPE-RES-64K
    4. 26907
    Prev
    Next
    INDEX
    Explanations

    made the day after the U.S. election, were directed at the President-elect Barack Obama and his new" - "...for how many days I spend in the office, or how many days I spend doing other things. \"But if you think I am going to do a diary for you in order to satisfy some populist pursuit of somebody you didn" - "In this retaliatory-discharge case a principal issue is whether a state insurance-department subrogation waiver override the insurer's statutory right to reimbursement." - "...just after we went to prom. Next up, me, Abby-René, Beth, and my old friend Elden. This one was also shot at" - "Seu servidor deve \"escutar\" na porta provida a você pelo primeiro argumento para seu programa." (Portuguese: "Your server must 'listen' on the port provided to you by the first argument to your program.") - "Abuse is also common in homes where pathological gambling is present" - "...aftermath of a 9-3 victory over the Dodgers in the home opener at Petco Park. The Padres on Tuesday were patient, they were aggressive. They were resilient. They were clutch. “Everything we didn" - "Both servers, that is the library / email and web site / forum servers were attacked. AVSIM is totally offline at" - "...in hopes of spreading Japanese soul food to different parts of Asia. Since then, Sama has branched out to 10 stores in Japan, infiltrated Hong Kong’s culinary realm," - "Also, is there an easy way to perform this conversion in Ada, without having to manipulate the bits?"Looking at the `TOP_ACTIVATING_TEXTS` and `TOP_POSITIVE_LOGITS`, there seems to be a mix of meanings.However, `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` contains common introductory words like "These", "But", "In", "Next", "The", "Both", "Since", "Also". This suggests the neuron might be related to the start of clauses or sentences, or sequence markers.The `TOP_POSITIVE_LOGITS` has some less common terms and some potential English words like "mania", "strikes", "summary", and Chinese "pursuit".The provided `MAX_ACTIVATING_TOKENS` is empty, which is a significant constraint. Usually, I would look for patterns immediately following the max activating token.Given the `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` list contains words that often start a new thought or transition (e.g., "Next", "Since", "Also", "But", "In", "The", "Both", "These"), and the `TOP_ACTIVATING_TEXTS` show transitions like "day after", "after we went", "aftermath of", "Since then", this neuron might be involved in identifying or predicting sequences, transitions, or the start of new information.The presence of "pursuit" (追究) and "mania" (gambling context) could point towards something related to goals or intense activities, but the other terms are too varied.Let's re-evaluate.The empty `MAX_ACTIVATING_TOKENS` is critical. If it were, for example, a word like "after", then "Next", "Since", "Also", "The", "Both", "These" would be strong candidates for something related to temporal or sequential events.The `TOP_POSITIVE_LOGITS` includes words that could be part of entities or concepts: "mania", "ken", "sey", "summary".The texts mention specific entities (Barack Obama, NATO allies, Dodgers, Petco Park), actions (deployment, election, requests, sell, shoot, gambling, attacked, spreading, infiltrated), and circumstances (after, since, next).Consider the possibility that the neuron activates on transitional phrases or sentence beginnings.The most common pattern in `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` is words that start a clause or sentence.Let's look for tokens in `TOP_POSITIVE_LOGITS` that could precede these."mania", "ken", "sey", "strikes", "summary".The phrase "made the day after the U.S. election" suggests a temporal marker."just after we went to prom" also suggests temporal."in the immediate aftermath of a 9-3 victory" temporal."Since then" temporal.The items in `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` are very common sentence starters.If the neuron activates, it could be signalling the *start* of something new, or a transition.However, the constraint is "find patterns that helps you explain the behavior."Given the empty `MAX_ACTIVATING_TOKENS`, and the `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` are common introductory words, the neuron could be responsible for recognizing the beginning of a new sentence, paragraph, or thought."after the U.S. election", "after we went to prom", "aftermath of", "Since then". These all suggest the idea of "after" or "since" a point in time, or a transition.The list of `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` seems to be the most structured clue. Words like "Next", "Since", "Also", "The", "Both", "These", "In", "But" are often used to introduce elements or connect ideas.Let's consider the possibility of it detecting temporal markers or sequential markers.However, the `TOP_POSITIVE_LOGITS` are quite varied and many are gibberish.What if the neuron detects specific word types that precede these common starters?The text samples show phrases like "made the day after...", "just after we went...", "in the immediate aftermath...", "Since then...".The key word "after" or "since" is present in the source text, and then the `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` appear.If the neuron is activated by "after" or "since" and then predicts the subsequent word from the `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` list, what would that mean?It could mean detecting temporal shifts or sequential information.Let's revisit the `TOP_POSITIVE_LOGITS`:- mania (could be a noun)- 追究 (pursuit - noun)- ken (part of names, or other words)- sey (part of names, or other words)- strikes (verb/noun)- summary (noun)These are mostly nouns or verbs. The `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` are mostly determiners, prepositions or adverbs.Let's consider the "after" pattern again, as it appears multiple times in the `TOP_ACTIVATING_TEXTS`."after the U.S. election""after we went to prom""aftermath of""Since then"If "after" or "since" (or related concepts) were the antecedent, the `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` provide words often following such transitions.However, without

    np_acts-logits-general · gemini-2.5-flash-lite
    New Auto-Interp
    Top Features by Cosine Similarity
    Configuration
    Prompts (Dashboard)
    24,576 prompts, 128 tokens each
    Dataset (Dashboard)
    monology/pile-uncopyrighted
    No Configuration Found
    Embeds
    IFrame
    Link
    Not in Any Lists

    No Comments

    Negative Logits
     folks
    -0.07
    æĥij
    -0.07
    å¬
    -0.07
    tach
    -0.07
    à®Łà®¿
    -0.07
    utch
    -0.07
    elter
    -0.06
    Æł
    -0.06
    ãĥ¼ãĤ¸ãĥ£
    -0.06
    éĺħè¯»æľ¬æĸĩåīį
    -0.06
    POSITIVE LOGITS
    mania
    0.07
    ÑĢна
    0.07
    追究
    0.07
    ken
    0.06
    æķħäºĭæĥħ
    0.06
    sey
    0.06
    rikes
    0.06
     اÙĦذاتÙĬØ©
    0.06
    summary
    0.06
    ìĭľë¡ľ
    0.06
    Activations Density 0.081%

    No Known Activations