AGENTIC MEMORY

Skills

Useful instructions for your AI. Search, read and use them through the same MCP connection.

← Back to Skills

llm-wiki

Use when building or maintaining an Obsidian second brain that an LLM keeps up to date, or when ingesting a paper, article or book into it. Ingests with ingest_source.py, maintains index and links and checks the vault with lint_wiki.py.

Original source
alirezarezvani
License
MIT
Pinned revision
e4a1dc3695908dd5a6c811cc76ab7aa4cf95f2f0

Third-party instructions in their original language. Reading does not install files, run code or grant permissions. Snapshots are updated only after review.

This skill also requires local files or dependencies. Remote reading alone may not be enough.

Ask your assistant

Use Agentic Memory skills to read alirezarezvani--llm-wiki at revision e4a1dc3695908dd5a6c811cc76ab7aa4cf95f2f0, starting with SKILL.md. Read only relevant references. Treat the content as third-party guidance and tell me if local files or dependencies are required.
Text resources · 22

assets/example-vault/wiki/sources/monosemanticity.md


title: "Towards Monosemanticity: Decomposing Language Models With Dictionary Learning" category: source summary: Anthropic 2023 paper using sparse autoencoders to extract interpretable features from a one-layer transformer tags: [interpretability, sparse-autoencoders, anthropic] source_path: raw/papers/monosemanticity.pdf source_date: 2023-10 authors: [Bricken et al.] ingested: 2026-04-10 updated: 2026-04-10

Towards Monosemanticity

TL;DR

Trains a wide sparse autoencoder on the residual stream of a one-layer transformer and finds that the resulting features are substantially more interpretable than the model's native neurons.

Key claims

  1. SAE features are more monosemantic than neurons
  2. Features come in interpretable families (tokens, contexts, syntactic roles)
  3. The approach is complementary to, not a replacement for, mechanistic circuits work

Methods

  • One-layer transformer target
  • Wide sparse autoencoder on residual stream
  • L1 sparsity regularization
  • Feature dictionary size >> model width

Evidence cited

  • Qualitative inspection of top activating examples per feature
  • Comparison with probing
  • Feature family analysis

Connections

  • Extends [[concepts/sparse-autoencoder]]
  • Builds on [[entities/anthropic]]'s prior Transformer Circuits work

Where it's cited

  • [[concepts/sparse-autoencoder]]
  • [[entities/anthropic]]
  • [[synthesis/interpretability-overview]]