Toolkit / Training function
Meta SAE Dashboard
An interactive dashboard of the meta-SAE decompositions
Credit
Patrick Leask, Bart Bussmann, Michael T Pearce, Joseph Isaac Bloom, Curt Tigges, Noura Al Moubayed, Lee Sharkey, Neel Nanda Licensing
CC-BY-4 Pipeline
| Dataset | Model | Application |
Subproject
Latent mechanistic interpretability
The Meta SAE dashboard is an interface for exploring meta SAE latents (vis-à-vis conventional SAE latents). It accompanies Leask et al.’s paper “Sparse Autoencodeders Do Not Find Canonical Units of Analysis” (2025), which challenges some implicit assumptions of the broader SAE framework. In that context, Meta SAEs demonstrate the non-atomicity of the list of features an SAE learns (see also Bussmann et al. 2024).
Embedded from
https://metasaes.streamlit.app/
