sae-umap-visualization-tool

IN premisesummaries/2026/08/24/templeton-2024-scaling-monosemanticity-chunk-5.md

Created 2026-08-25T02:58:38+00:00

An interactive UMAP visualization tool (linked as ./umap.html) is provided for exploring feature neighborhoods in the Sonnet SAE feature space.

Summary

A local interactive tool (umap.html) is available that lets you visually explore how the individual features learned by the Sonnet Sparse Autoencoder cluster and relate to one another. This matters because it turns an otherwise abstract, high-dimensional feature space into something you can scroll, zoom, and inspect by hand, making it far easier to spot groups of features that fire together or to test whether the SAE is learning coherent, interpretable representations.