Towards Monosemanticity: Decomposing Language Models With Dictionary Learning
Date : 2023-10-04
Description
In this very impressive work, Anthropic tackle interpretability of Large Language Models. Working around the problem of superposition causing polysemanticity, they use a weak dictionary learning algorithm called a sparse autoencoder to generate learned features from a trained model that offer a more monosemantic unit of analysis than the model's neurons themselves.
Read article here
Recently on :
Artificial Intelligence
Research
PITTI - 2026-03-05
Scaling Trust : a Missing Piece in Multi-Agent Worlds
Humanity’s ability to build complex civilizations relies on an "invisible infrastructure" - the shared culture, institutions, a...
PITTI - 2026-01-14
Cultural, Ideological and Political Bias in LLMs
Transcription of a talk given during the work sessions organized by Technoréalisme on December 9, 2025, in Paris. The talk pres...
WEB - 2025-11-13
Measuring political bias in Claude
Anthropic gives insights into their evaluation methods to measure political bias in models.
WEB - 2025-10-09
Defining and evaluating political bias in LLMs
OpenAI created a political bias evaluation that mirrors real-world usage to stress-test their models’ ability to remain objecti...
WEB - 2025-07-23
Preventing Woke AI In Federal Government
Citing concerns that ideological agendas like Diversity, Equity, and Inclusion (DEI) are compromising accuracy, this executive ...