Neuronpedia maps model features and concepts

This is a research-grade interpretability workbench, not a consumer AI toy, which is why people remember it vaguely but have trouble naming it. It fits the description well: model-specific feature browsing, activation inspection, and concept-level exploration all live in one place.

• –The site centers sparse autoencoders, latents, concepts, and custom vectors, which matches the “vector-ish representation” memory
• –It exposes top activations, top logits, activation density, and live inference testing, so users can inspect what a feature seems to mean
• –Search can be done by explanation text or by inference over text, which makes it feel like a semantic browser for model internals
• –The platform is explicitly open source and API-first, so it doubles as infrastructure for interpretability researchers
• –If the remembered interface was text-heavy and concept-oriented, Neuronpedia is a strong match