Definition
Plain language
When a single internal unit in a model fires for a jumble of unrelated things at once.
As stated in the literature
The phenomenon where individual neurons respond to incoherent mixtures of concepts (e.g., citations, Korean text, and suspicion together), explained by the superposition hypothesis and motivating dictionary-learning approaches to recover monosemantic features.
Also called: polysemantic
Why it matters: It's why raw model internals are so hard to read, and explaining it is what motivates the methods that try to untangle concepts.
For example, a single neuron might light up for academic citations, Korean text, and expressions of suspicion all at the same time.
Heard on the show
“The field calls this polysemanticity, "many meanings," and it was the central roadblock.”Episode 098 — Finding Millions of Readable Concepts Inside a Real, Deployed AI Model