Most neurons in language models are "polysemantic" – they respond to multiple unrelated things. For example, one neuron in a small language model activates strongly on academic citations, English dialogue, HTTP requests, Korean text, and others.

→ View original post on X — @anthropicai