© 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. Gemma-3-27B-IT
    3. 31-GEMMASCOPE-2-RES-262K
    4. 42587
    Prev
    Next
    INDEX
    Explanations

    claim. You must prove the defendant owed you a duty of care, breached that duty, and that breach caused your injuries. * **Bre - `prove` -> `the`2. * **Outcome:** The project was abandoned due to concerns about security (counterfeiting) and the complexity of the technology. Linnell - `Outcome` -> `: **`3. * **Lunch:** Leftover Baked Salmon and Quinoa - `option` (from `(P) (V option: Tofu scramble)`) -> `:`4. Interception happens at the level of the HTTP client you're using to make API calls. Here are the most common scenarios - `ception` (from `exception` or `reception` based on context) -> `happens`5. 0):** Excellent value. Good range (259-293 miles). Practical and spacious. Discontinued after 202 - `.` -> `**` (This doesn't match `They` from `TOKENS_AFTER_MAX_ACTIVATING_TOKEN`) - `).` -> `**` (This doesn't match `1` from `TOKENS_AFTER_MAX_ACTIVATING_TOKEN`)6. * **Samsung (Galaxy):** Samsung works well with other Samsung devices (watches, earbuds, tablets). They also play nicely with Windows PCs (though not *as* seamlessly as Apple's integration). Samsung is more - `4` (from `after 2020`?) -> `million` (This doesn't match `that` from `TOKENS_AFTER_MAX_ACTIVATING_TOKEN`)7. AI in breaking it. * **RSA Security Relies on Prime Factorization:** The security of RSA rests on the **fact** that factoring large numbers into their prime factors is computationally very hard. If you can factor the modulus (part of the - `fact` -> `that`8. explain *how* it came to be, which is a crucial piece of the puzzle. **Key takeaway from science:** Science explains *how* we exist, providing a detailed account of the processes that led to our emergence. It doesn - `Science` (from `from science`) -> `: ` (This doesn't match `extensive` from `TOKENS_AFTER_MAX_ACTIVATING_TOKEN`)9. * **Challenges:** **Requires** extensive research, strong storytelling, potentially access to archives. * **Advanced DIY/Maker Skills (Beyond Basic Tutorials - `Requires` -> `extensive`Let's look at `TOP_POSITIVE_LOGITS` again. They are mainly German names (Buch, Schmitt, Schilling, Schroeder) and some other tokens that look like they might be from other languages.This suggests the neuron might be related to specific entities or categories.The `MAX_ACTIVATING_TOKENS` shows:- `prove` followed by `the` (legal context)- `Outcome` followed by `: **` (reporting/summarizing context)- `option` followed by `:` (choices/alternatives context)- `ception` (from exception/reception) followed by `happens` (events/processes)- `fact` followed by `that` (logical statements/assertions)- `Requires` followed by `extensive` (conditions/requirements)The `TOP_POSITIVE_LOGITS` are mostly names. The `TOP_ACTIVATING_TEXTS` include examples from law, project management, computer science, product descriptions, and science explanations.The pattern I see in `MAX_ACTIVATING_TOKENS` mixed with `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` is finding specific keywords followed by punctuation or conjunctions that mark a definition, explanation, or consequence.Let's look for a common theme in the `TOP_ACTIVATING_TEXTS`:- "most common basis for a product liability claim"- "project was abandoned due to concerns"- "option: Tofu scramble"- "Interception happens at the level of the HTTP client"- "RSA Security Relies on Prime Factorization"- "Requires extensive research"These suggest a neuron that looks for terms that introduce subsequent explanations, specifications, conditions, or outcomes.Consider the structure often seen:"Term: Explanation""Term: Outcome""Term [preceded by condition]" -> `fact that`"Term [followed by requirement]" -> `Requires extensive`The `TOP_POSITIVE_LOGITS` are still a bit of a wild card if not strongly tied to a semantic pattern. However, names often appear in lists or specific contexts.What do `prove the`, `Outcome:`, `option:`, `fact that`, `Requires extensive` have in common? They often set up a statement, a condition, or an explanation.Let's look at the `TOP_POSITIVE_LOGITS` again - Buch, Schmitt, Schilling, Schroeder. These are German surnames.The `TOP_ACTIVATING_TEXTS` also contain things like "Samsung", "Galaxy", "Apple".Perhaps the neuron is triggered by specific types of lists, structured information, or conditions that lead to further elaboration.The `MAX_ACTIVATING_TOKENS` are often keywords starting bullet points (`Outcome`, `Requires`, `fact`), or introduce a choice (`option`), or indicate an action/state (`prove`, `ception`).The `TOKENS_AFTER_MAX_ACTIVATING_TOKEN` are often punctuation (`:`, `.**`), conjunctions (`that`), or the subsequent word (`the`, `extensive`, `happens`, `Science`). This suggests the

    np_acts-logits-general · gemini-2.5-flash-lite
    New Auto-Interp
    Top Features by Cosine Similarity
    Configuration
    google/gemma-scope-2-27b-it/resid_post/layer_31_width_262k_l0_medium
    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
    verses
    0.46
    leaves
    0.45
    श्यक
    0.43
    一道
    0.43
     জড়িত
    0.43
    arkan
    0.42
    sight
    0.42
    tons
    0.42
    riminating
    0.41
    umatic
    0.41
    POSITIVE LOGITS
    芑
    0.48
     Buch
    0.47
    शुदा
    0.47
     Schmitt
    0.47
    럿
    0.46
    গুলো
    0.46
    이트
    0.46
     Schilling
    0.46
     Connelly
    0.45
     Schroeder
    0.44
    Activations Density 0.000%

    No Known Activations