Time Neurons

About Time Neurons

Time neurons encode the temporal structure of the world — when things happened, how long they lasted, in what order they occurred, and how frequently they recur. They are the second-most-numerous atom type in most models after Symbol, collectively representing 3–24 percent of classified features. The wide range reflects how strongly model training data composition affects temporal representation: models trained on news and historical corpora (OLMo, Mistral) have far more Time neurons than models trained on more general web text. Time.Era neurons are the single largest element in our entire 56-element corpus — the model's representation of historical periods dwarfs every other temporal sub-type, reflecting the centrality of historical context in language understanding.