For every month since July 2023, we took the most-cited AI papers published that month and weighed their topics by citations. Stacked and connected, the months fall like water: streams widen when a topic surges, thin when the field moves on, and occasionally something hits the pool below hard enough to change the course of everything downstream.
Citation counts are a July 2026 snapshot, so recent months run structurally shallower: their citations are still accumulating.
State-space models erupt from nowhere: Mamba and a cascade of variants briefly out-cite everything else in the stream.
The single deepest drop in the data: 11,843 citations. Biology becomes a main channel of the river, not a tributary.
DeepSeek-R1 lands (5,419) and the month totals 8,940: reinforcement learning and reasoning flood the falls.
Kimi K2 and open agentic systems take the current, with interpretability rising along the banks.