© 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-TRANSCODER-262K
    4. 78360
    Prev
    Next
    INDEX
    Explanations

    arc(x, y, radius, 0, Math.PI * 2)""ctx.arc(x, y, radius, 0, Math.PI * 2)""arc(x, y, radius, 0, Math.PI * 2)""arc(x,y,r,0,2*PI)""rounded corners with arc.""arc of a circle""circumference=2*pi*r;""arc of radius""the arc of the orbit""helical pattern. ... this tilt rotates continuously along the chain. This spiral arrangement""rounded rectangular""arc"The neuron seems to be activated by code snippets involving drawing circles or arcs, and descriptions of circular or helical patterns, movement, and rotation. The `MAX_ACTIVATING_TOKENS` also include 'arc', 'rotates', and 'cx' (which is often a center coordinate in graphics).Looking at the `TOP_POSITIVE_LOGITS`, 'arcu' and 'articul' are related to curves and articulation. The text examples clearly show `arc(...)` function calls in context of drawing.Considering the prompt's rules:- Concise (3-20 words).- Find patterns, not just list tokens.- Specific.- Avoid certain phrases.- Don't mention "tokens" or "patterns".The most prominent pattern is the use of "arc" in code or descriptions, often related to geometry or graphics. The other keywords like "rotates" and "cx" reinforce a geometric/graphical context involving circles or curves.How can I combine this?- `arc(...)` function calls- circular shapes- rotational descriptions- spiral/helical movementLet's try to distill this into a short phrase.- "drawing arc shapes" - too simple, doesn't capture rotation.- "circular path drawing" - better, but "arc" is key.- "arc drawing and rotation" - good, but "drawing" might be too action-oriented.- "arc and circular motion" - captures the drawing/shape aspect ('arc') and the movement/geometric aspect ('circular motion').- "circular curves and rotation" - similar.- "arc drawing, spirals, rotation" - too long.Revisiting `MAX_ACTIVATING_TOKENS`: `arc`, `cx`, `rotates`, `0`.`TOKENS_AFTER_MAX_ACTIVATING_TOKEN`: `(`, `model`, `**`, `continuously`, `C`.`TOP_POSITIVE_LOGITS`: `arcu`, `articul`, `epiph` (maybe related to epicycle?).`TOP_ACTIVATING_TEXTS`: code `arc(x, y, radius, 0, Math.PI * 2)`, `circle`, `rounded corners`, `spiral`, `rotates continuously`, `helical pattern`.The emphasis is on 'arc' in code and geometrical descriptions involving circles, spirals, and rotation. Therefore, a phrase that covers arcs and related rotations/curves seems most appropriate.Let's consider the phrase "arc drawing and circular patterns". This fits within 3-20 words and covers both the specific token 'arc' and the broader patterns observed.Alternatively, "arc functions and circular shapes".The prompt says "what the neuron detects or predicts by finding patterns in lists."The phrase "arc functions and circular shapes" seems descriptive of what it *detects*.Let's review the sample outputs:- "drawing arcs"- "drawing arcs and circles"- "arc patterns"- "drawing arcs or circles"- "drawing circles"These are very focused on drawing arcs or circles. My current candidates are "arc functions and circular shapes". The text examples contain actual code `ctx.arc(...)`. So "arc functions" is good. "circular shapes" is good.How about just focusing on the most common elements derived from the texts and logits? 'arc', 'circle', 'spiral', 'rotation'."arc, circle, spiral, rotation". This is a list of concepts the neuron is sensitive to.The prompt asks for a phrase that captures *what* it detects/predicts.Let's try to be more action-oriented or conceptual, like the examples seem to suggest."drawing arcs" is a common pattern. Also seeing "circles"."arc path drawing""circular shapes and pathways"Looking at `TOP_POSITIVE_LOGITS`: `arcu`, `articul`, `epiph`, `sacc`. `arcu` points to arcs. `articul` might point to articulation, which involves joints and movement, or perhaps articulation of shapes.The `MAX_ACTIVATING_TOKENS` 'arc' and 'rotates' and 'cx' are very strong indicators.What if it's about drawing round things that might move?"drawing arcs and curves""arc drawing and rotational paths"Let's try to simplify. The most frequent and specific term is "arc". Many examples are `ctx.arc(...)`. This points to drawing. The `TOP_POSITIVE_LOGITS` also contain `arcu`. The `TOP_ACTIVATING_TEXTS` contain `arc(x, y, radius, 0, Math.PI * 2)` multiple times, `circle`, `rounded corners`.The neuron is highly sensitive to `arc` in code or geometric descriptions.What follows `arc`? Often `(x, y, radius, 0, Math.PI * 2)` which is a full circle.What is related to `arc` in `TOP_POSITIVE_LOGITS`? `arcu`.What else is strongly present? Rotation, spirals.Perhaps it's focused on drawing specific graphical primitives that are round or curved."drawing circular shapes and spirals" - 5 words. Looks good."drawing arcs and spirals" - 4 words. Good."drawing curved paths" - 3 words."drawing circular arcs" - 3 words.Let's consider the examples again."arc(x, y, radius, 0, Math.PI * 2)" - drawing a full circle using an arc definition."helical pattern. ... this tilt rotates continuously along the chain. This spiral arrangement" - describes a shape and movement.The neuron is associated with 'arc', 'rotates', 'cx'. It's associated with code like `ctx.arc(...)` and text describing spirals and rotations.The most direct pattern is the use of "arc" combined with elements suggesting circularity or rotation."drawing circular shapes and spirals" seems to cover it well."drawing arcs and spirals" also seems good.Let's try to be more specific to `arc`."drawing arc segments for circles""arc endpoints and rotational paths"The prompt says "finding patterns in lists". The lists clearly show `arc` appearing frequently, often in code, and related concepts like `circle`, `rotates`, `spiral`, `helical`.I'll try to combine the specific element (`arc`) with the common geometric/motion concepts."drawing arcs for wheels or spirals" - too specific

    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_31_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
     ring
    0.53
     rings
    0.52
    rings
    0.52
     circunfer
    0.52
     circumfer
    0.50
    inner
    0.50
    radial
    0.49
     inner
    0.49
     Ring
    0.48
    ring
    0.48
    POSITIVE LOGITS
     Ivan
    0.42
    olulu
    0.42
     arcu
    0.39
     oxalic
    0.39
    üse
    0.38
     পায়ের
    0.38
     articul
    0.36
    Ivan
    0.36
     epiph
    0.36
     sacc
    0.35
    Activations Density 0.027%

    No Known Activations