Instrument · live model data
Interactive instrument
Latent Observatory
Search the learned features of real models by concept, inspect what makes each one fire, and follow the geometry to its neighbours. Every number on this page is live data from a real sparse autoencoder trained on a real model; nothing is a mock-up.
Feature data is served live from the open Neuronpedia API (MIT). The SAEs belong to their cited releases. Feature descriptions are auto-generated by language models and are hypotheses about what a feature detects, not ground truth; the activation records below each one are the evidence to check them against.
01 · Search feature space
Try:
02 · Feature dossier
Search on the left, or follow a deep link.
A feature dossier shows its auto-interp explanations, the real text that makes it fire, its output effects, and its nearest neighbours in decoder space.
A feature dossier shows its auto-interp explanations, the real text that makes it fire, its output effects, and its nearest neighbours in decoder space.