Google DeepMind x Neuronpedia
    Introducing Gemma Scope, a new tool for understanding the internals of AI models.
    Launch
    Neuronpedia is an open platform for interpretability research.
    Explore, steer, and experiment on AI models.
    Getting Started
    MIT Technology ReviewGoogle DeepMindOpenMOSS, Fudan UniversityApollo ResearchMATSEleutherAI
    Explore Visually
    Best viewed on a tablet or larger screen.
    Each dot in the visualization represents an idea or concept that the model has learned. The dots, or features, are arranged in a fully interactive UMAP graph: you can zoom, pan, and select dots to see details. You can also filter features by specific terms.
    Run Example Search
    Test Instantly
    Neuronpedia lets you search and test models by doing inference over millions of SAE features. You can use your own custom text, and sort and filter results by layers, tokens, or both.
    Infrastructure For All
    Neuronpedia hosts interpretability data, tools, visualizations, and APIs for all researchers. We eliminate the engineering obstacles in order to move interpretability further, faster. Upload your SAEs and we take care of the rest.
    Collaborate and Share
    Every SAE feature has a unique link with link previews, can be compiled into sharable lists, and supports iframe embedding (as demonstrated here).
    Who We Are
    Neuronpedia was created by Johnny Lin, an ex-Apple engineer who previously founded a privacy startup. Neuronpedia is supported by the Long Term Future Fund and AISTOF.
    Citation
    @misc{neuronpedia,
        title = {Neuronpedia: Interactive Reference and Tooling for Analyzing Neural Networks},
        year = {2023},
        note = {Software available from neuronpedia.org},
        url = {https://www.neuronpedia.org},
        author = {Lin, Johnny}
    }