Neuronpedia
About
Neuronpedia was founded in summer 2023 by Johnny Lin, an ex-Apple engineer who previously founded Lockdown Privacy, a consumer privacy startup. It began as a reference catalog for neurons in GPT2-Small and rapidly expanded in early 2024, in collaboration with Joseph Bloom (SERI-MATS), into a comprehensive platform for sparse autoencoder (SAE) research. The platform now supports all major forms of mechanistic interpretability research, including probes, transcoders, circuits, and custom steering vectors, not just SAEs. Neuronpedia's infrastructure hosts over 4 terabytes of interpretability datasets covering 11 models, more than 60 million features, 50 million explanations, and 3 billion activations — all freely available. Its tools include interactive feature dashboards, circuit/attribution graph tracing (built on Anthropic's circuit-tracer work), automatic feature interpretation via LLMs, activation steering, semantic search over latent space, and a comprehensive public API. The platform logged over 100,000 API calls per day following the release of its circuit tracing tools in mid-2025. The project went fully open-source on March 31, 2025, releasing both the codebase (MIT license) and all hosted datasets to the public. Neuronpedia has collaborated closely with Anthropic, Google DeepMind (including the Gemma Scope and Gemma Scope 2 projects), EleutherAI, and Goodfire. It operates as Decode Research, a project of Players Philanthropy Fund, Inc. (a 501c3 fiscal sponsor). Funders have included the Long Term Future Fund, Survival and Flourishing Fund, Open Philanthropy, AISTOF, Anthropic, and Manifund. Core team members include Johnny Lin (founder), David Chanin, and Michael Hanna.
Theory of Change
Neuronpedia bets that mechanistic interpretability — understanding what computations AI models are actually performing — is a critical precondition for aligning advanced AI systems. By providing free, high-quality infrastructure for interpretability research (dashboards, APIs, datasets, circuit tracing, steering tools), Neuronpedia accelerates the pace at which researchers worldwide can study model internals, discover safety-relevant features and circuits, and develop alignment techniques. The theory is that reducing the cost and friction of interpretability research multiplies the output of the field, increasing the probability that humans will have reliable tools to understand and correct AI behavior before models become dangerously capable.
Details
- Start Date
- -
- End Date
- -
- Expected Duration
- -
- Funding Raised to Date
- -
- Last Updated
- Apr 3, 2026, 1:23 AM UTC
- Created
- Apr 3, 2026, 1:23 AM UTC