assets/example-vault/wiki/synthesis/interpretability-overview.md
title: Interpretability Overview category: synthesis summary: Current synthesis on mechanistic interpretability in large language models tags: [interpretability, overview] sources: 1 updated: 2026-04-10
Interpretability Overview
Thesis
(early — only one source ingested) Mechanistic interpretability is shifting from direct neuron inspection (limited by polysemanticity) toward sparse-autoencoder-based feature decomposition. The empirical bet is that overcomplete sparse dictionaries recover the "true" feature basis of trained models.
The landscape
- Sparse-autoencoder line — pursued by [[entities/anthropic]]; see [[sources/monosemanticity]]
- Circuits work — (no sources ingested yet)
- Probing — (no sources ingested yet)
Key concepts
- [[concepts/sparse-autoencoder]] — primary method under investigation
Key sources
- [[sources/monosemanticity]] — foundational SAE paper
Current open problems
- Scaling SAEs beyond toy models
- Whether features are truly monosemantic vs merely "more" monosemantic
- How SAE features relate to circuits-based analysis
How this synthesis has changed
- 2026-04-10 — initial synthesis after ingesting [[sources/monosemanticity]]. Only one data point; thesis is intentionally tentative.