AnalysisAI ModelsOctober 15, 2026

Researcher argues neural networks are already clean, challenging superposition

Read original source →already-clean.goodfire.com

A post on already-clean.goodfire.com argues the canonical evidence for polysemanticity and superposition in image and language models was misread, and that what looked like arbitrary concept packing was mostly a motif the author calls recursive generalization. The author says networks can be studied directly rather than untangled first.

1 source

More stories today

Open the live feed