Neuronpedia logo - a computer chip with a rounded viewfinder border around it

    Neuronpedia

    APICircuit TracerNEWSteerSAE EvalsBlogSlackPrivacy & TermsContact
    © Neuronpedia 2025
    Privacy & TermsBlogGitHubSlackTwitterContact
    EXPLANATION TYPE
    eleuther_acts_top20
    Description
    Eleuther's "Default" Explainer, which shows the auto-interp model a sample from activating texts (with max activations highlighted) and asks the model to think through possible patterns, and then provide the explanation. This is an alternate version that doesn't use quantiles.
    Author
    EleutherAI
    URL
    https://github.com/EleutherAI/sae-auto-interp
    Settings
    Default prompts from the main branch. The model is shown top 20 examples, with a threshold of 60% of the max activation to consider highlighting. Temperature is set to 0.7.
    Recent Explanations
    Discussions related to consciousness, perception, and the mind from scientific and philosophical perspectives, often discussing the influence of the environment and biological factors. The examples come from technical discussions.
    gemini-2.0-flash
    the typically metaphysical distinction between what is inside the box
    Neuronpedia logo
    LLAMA3.1-8B
    21-LLAMASCOPE-RES-32K
    INDEX 17571
    Technical discourse about consciousness, mind, perception, and cognitive processes, often in philosophical or scientific contexts.
    claude-3-7-sonnet-20250219
    the typically metaphysical distinction between what is inside the box
    Neuronpedia logo
    LLAMA3.1-8B
    21-LLAMASCOPE-RES-32K
    INDEX 17571
    Frequent occurrences of technical and philosophical discourse around consciousness, perception, mind, and environmental influences, often with detailed linguistic analysis of these abstract concepts.
    claude-3-5-haiku-20241022
    the typically metaphysical distinction between what is inside the box
    Neuronpedia logo
    LLAMA3.1-8B
    21-LLAMASCOPE-RES-32K
    INDEX 17571
    Programming language syntax elements and control structures in code snippets, including function parameters, loop constructs, and variable references.
    claude-3-7-sonnet-20250219
     of Light, making their mana regen come mostly from Illumination
    Neuronpedia logo
    GEMMA-2-2B
    19-GEMMASCOPE-TRANSCODER-16K
    INDEX 10165
    Various linguistic patterns observed including: idioms, comparative word endings, possessive nouns, programming code tokens, grammatical markers, and computational/mathematical terminology.
    claude-3-5-haiku-20241022
     of Light, making their mana regen come mostly from Illumination
    Neuronpedia logo
    GEMMA-2-2B
    19-GEMMASCOPE-TRANSCODER-16K
    INDEX 10165
    Common code patterns for a boss AI in a game, including timers, spell casting, and instance management. Uses pointer syntax and specific naming conventions for game-related functions and variables.
    claude-3-5-sonnet-20240620
     of Light, making their mana regen come mostly from Illumination
    Neuronpedia logo
    GEMMA-2-2B
    20-GEMMASCOPE-TRANSCODER-16K
    INDEX 2124
    Video game terminology related to RPGs and MMORPGs, including character attributes, abilities, game mechanics, social structures, and content features.
    claude-3-7-sonnet-20250219
     he’s the only character who does this. And
    Neuronpedia logo
    GEMMA-2-2B
    0-GEMMASCOPE-TRANSCODER-16K
    INDEX 14521
    References to "cats" as a singular or plural noun, often in the context of general statements or specific animal-related discussions.
    claude-3-5-haiku-20241022
    <bos>I like cats
    Neuronpedia logo
    GEMMA-2-2B
    12-GEMMASCOPE-RES-16K
    INDEX 11195
    The highlighted elements are variable or identifier names in UI-related code, typically associated with graphical components such as buttons, text fields, labels, and views. These identifiers often follow naming conventions that reflect the component type (e.g., "btn" for button, "lbl" for label, "txt" for text field), and are used in contexts involving event handling or UI initialization.
    gpt-4o
    BtnAceptar);↵ lblMensaje = (TextView)
    Neuronpedia logo
    GEMMA-2-2B
    12-GEMMASCOPE-RES-16K__L0-22
    INDEX 0
    The pattern involves important words that are past participles or adjectives derived from verbs, often indicating an action related to publishing, providing, reproducing, making, supporting, or translating.
    gpt-4.1-nano
    paradoxa is published in print twice a year
    Neuronpedia logo
    GPT2-SMALL
    8-RES-JB
    INDEX 56
    Suffixes include "urers", "urer", "ers", or "ious" are found at the end of words. "Urers" and "urer" designate a job function. "Ious" indicates a strong emotional state.
    gemini-2.0-flash
    America, the Can Manufacturers Institute, Glass Packaging
    Neuronpedia logo
    GPT2-SMALL
    0-RES-JB
    INDEX 14059
    The suffix "-ers" attached to nouns, frequently denoting a group of people or things performing a similar action or profession (e.g., manufacturers, lecturers, insurers). There are also instances of "-er" suffixes attached to nouns modifying the base noun (e.g., torturer, perjurer).
    gemini-1.5-flash
    America, the Can Manufacturers Institute, Glass Packaging
    Neuronpedia logo
    GPT2-SMALL
    0-RES-JB
    INDEX 14059
    The adverb "closer" is frequently used to describe a process of gradual approximation or progress towards a goal. It often appears in contexts involving spatial proximity, relationship development, or achieving a target.
    gemini-1.5-flash
    a feat that gets us closer to understanding how human speech
    Neuronpedia logo
    GPT2-SMALL
    10-RES-JB
    INDEX 10
    Text segments preceding or following quotation marks, often containing proper nouns, technical terms, or specialized vocabulary related to specific topics or industries.
    claude-3-5-sonnet-20240620
    <|endoftext|>Feeling so grateful for
    Neuronpedia logo
    GPT2-SMALL
    8-RES_FS6144-JB
    INDEX 1812
    Tokens that introduce or modify clauses, often connecting ideas or providing context in complex sentences. These include conjunctions, prepositions, and relative pronouns that help structure the flow of information in the text.
    claude-3-5-sonnet-20240620
    <|endoftext|>On the guards��
    Neuronpedia logo
    GPT2-SMALL
    8-RES_FS6144-JB
    INDEX 5541
    The token "," before "because" in sentences, connecting clauses or expanding reasoning.
    deepseek-v3
    a bank heist that, marvels a quoted police
    Neuronpedia logo
    GPT2-SMALL
    8-RES_FS6144-JB
    INDEX 4518
    Explanation could not be parsed.
    deepseek-r1
    a bank heist that, marvels a quoted police
    Neuronpedia logo
    GPT2-SMALL
    8-RES_FS6144-JB
    INDEX 4518
    Explanation could not be parsed.
    o3-mini
    a bank heist that, marvels a quoted police
    Neuronpedia logo
    GPT2-SMALL
    8-RES_FS6144-JB
    INDEX 4518
    Commas used to introduce additional information, often following a conjunction or preceding a quote, particularly in complex sentences discussing social, political, or cultural topics.
    claude-3-5-sonnet-20240620
    a bank heist that, marvels a quoted police
    Neuronpedia logo
    GPT2-SMALL
    8-RES_FS6144-JB
    INDEX 4518
    Sequences of related terms, often categories or options, separated by slashes. These frequently represent choices, versions, or specifications within a system or context.
    gemini-1.5-flash
    a Race/Class/Gender combo. WeâĢĻ
    Neuronpedia logo
    GPT2-SMALL
    8-RES-JB
    INDEX 41