THE SHORT VERSION
Microsoft and MIT separately reported machine-learning forecasting gains in the two-to-six-week window — the range where both physics-based and AI models usually fall apart.
Microsoft says its machine-learning method produces more accurate forecasts two to six weeks ahead. MIT reported a machine-learning approach setting records in the same window. Both landed on 14 October.
That specific range is the point. Short-range forecasting is largely solved and seasonal outlooks are statistical; the two-to-six-week band is where physics-based models lose skill and where AI models have generally lost it too. It is also the window emergency planners actually use — staging resources, pre-positioning crews, deciding whether to call something early.
Separately, the open model HiRO-ACEv2 appeared on Hugging Face, simulating extreme weather at 3 km resolution at a fraction of conventional computing cost and without requiring large supercomputing clusters. Three results in one day in a field where the compute barrier has been the main obstacle.
Sources
MIT News