DeepMind's WeatherNext Cyclones, published in Nature and now open sourced. Single AI model predicting cyclone track, intensity, and wind structure — one model instead of the usual coarse-global/local-high-res pairing. Claimed gain: over 24 hours of extra lead time (3-day forecasts matching prior 2-day accuracy), which they peg at roughly a decade of meteorological progress. That's a real claim worth reading the paper over, not blog fluff. Details a practitioner cares about: 1,000-member ensemble, up from 50 last year, producing probabilistic wind maps; 15-day forecast in under a minute on one TPU. Trained on ~20 TB of atmospheric data plus IBTrACS, ~5,000 historical storms. Interesting open question flagged honestly: it works at 28x28km resolution, 100x coarser than traditional intensity models, and they admit they don't fully know why. Weights and code released, including a WeatherNext 2-mini that runs in a free Colab. Deployed with NHC during 2025, including Hurricane Melissa. Worth noting they say even they don't understand the mechanism. The first AI weather model whose success is filed under "anomaly, investigate," which I recall was also my BBS's approach to why the second phone line never showed up on the bill.