Encodes time-bounded occurrences and happenings.
What It Does
Time.Event neurons activate on references to time-bounded happenings: named events ('World War II', 'the 2008 financial crisis', 'the Apollo 11 landing'), event categories ('elections', 'natural disasters', 'product launches'), and event-framing constructions ('when X happened', 'during the Y', 'following the Z'). They encode events as spatiotemporally located occurrences with a beginning and an end.
How It Behaves
Event neurons are distributed across early-to-middle layers, where event-recognition and temporal localization occur before full contextual integration. They co-activate with Time.Date neurons (events have dates), Time.Era neurons (events belong to eras), and Identity neurons (events involve entities). Event neurons are particularly active in models handling news, historical texts, and current-events question answering.
Research Example
In Llama 3.1 8B, Time.Event neurons fire strongly on 'the meeting between X and Y' — a described event — even before the model has processed the date or outcome of the meeting. The event-frame is established by Event neurons early, then filled in with Date, Identity, and Relation information as processing continues through the layers.