This paper introduces ClimyNet-Ω, a weather forecasting system that works differently from existing AI weather models. Instead of one big model trying to predict everything, we use two smaller models that work together — a "Thinker" that reasons about local weather patterns, and a "World Simulator" that models global atmospheric dynamics. Together, they produce forecasts with built-in confidence levels, so you always know how much to trust a prediction. What makes this approach unique is that it's designed for personal, hyperlocal weather — the kind of forecast that actually matters at your specific location. The system learns continuously from local observations, runs on phones and even microcontrollers, and gets better over time. We also introduce ten mathematical techniques that are new to weather AI, including Koopman operators for dynamics modeling, conformal prediction for guaranteed uncertainty bands, and Schrödinger bridges for trajectory forecasting. The base model is just 15.6M parameters — small enough to fit on a phone — and grows organically through a companion system called Neural Spawning (detailed in a separate paper).
haruhito (Wed,) studied this question.