Neuronpedia - open souce interpretability platform for AI Models (www.neuronpedia.org)

Neuronpedia is an open‑source interpretability platform and interactive reference for analyzing neural networks that lets researchers “explore, steer, and experiment” with models’ internal mechanics. The site aggregates runnable releases, reproducible analyses, and tooling—graph visualizations, jump-to-source/feature links, and circuit‑tracing workflows inspired by Anthropic—that make it easy to visualize and trace a model’s internal reasoning on custom prompts. Notable technical artifacts include sparse autoencoders (SAEs) for many residual stream layers, transcoders for fine‑grained circuit analysis, sparse dictionary learning tools, and a growing library of model‑specific analyses (GPT2‑Small, Pythia‑70M, Llama3.1, Gemma‑2, Qwen variants, DeepSeek, etc.). For the AI/ML community this is significant because it packages state‑of‑the‑art interpretability methods into a reproducible, browsable platform that supports open‑weight models and cross‑model comparisons. Researchers can inspect functionally important features, identify misaligned persona features, test steering baselines, and reuse published SAEs and transcoders—accelerating work on alignment, robustness, and mechanistic interpretability. Created by Johnny Lin and supported by Decode Research, Anthropic, and others, Neuronpedia lowers the barrier to rigorous circuit analysis and collaborative model debugging by making artifacts and code openly accessible.