Generated by Codex with GPT 5.6 Sol XHigh

Techmeme surfaced Victoria Turk’s August 6 WIRED report, “DeepMind Says Its AI Can Predict Hurricanes Earlier Than Everyone Else,” about an AI weather model that could turn one of forecasting’s slowest improvements into a sudden step forward. Google DeepMind’s WeatherNext Cyclones predicts a tropical cyclone’s path, intensity, and wind extent together, and its three-day guidance was, on average, as accurate as leading operational models were at two days. For emergency managers deciding when to evacuate communities, stage supplies, or move crews, an extra day is not an abstract benchmark gain; it is usable preparation time.

The accompanying Nature paper evaluates the model on tropical cyclones from 2023 through 2025. Across forecasts of track, intensity, and wind radii, WeatherNext Cyclones delivered a lead-time advantage averaging a day or more. The researchers describe that improvement as roughly equivalent to a decade of progress in operational cyclone forecasting. The result is especially notable because earlier AI weather systems tended to improve storm tracks more readily than intensity, the quantity that separates a dangerous storm from a catastrophic one.

One model across two scales

Cyclone forecasting traditionally pulls in opposite directions. A storm’s route depends on large atmospheric systems such as fronts and prevailing winds, which global models can represent. Its strength depends on much smaller processes around the storm’s core and the nearby ocean, which favor expensive, high-resolution regional models. WeatherNext Cyclones attempts to learn both scales in a single system.

The model was co-trained on nearly 20 terabytes of global atmospheric data and the IBTrACS archive of almost 5,000 historical storms. It produces forecasts as far as 15 days ahead, including global weather fields and specialized predictions for cyclone location, intensity, and wind structure. Yet it works from atmospheric inputs with cells about 28 kilometers wide—about 100 times coarser than the high-resolution inputs normally considered necessary for intensity forecasting.

That coarse input is not just a cost-saving detail. The Nature results suggest global atmospheric data contains more information about cyclone intensity than meteorologists previously recognized. DeepMind’s researchers do not yet know exactly which signals the system is using. The unexplained performance makes the model scientifically interesting as well as operationally useful: interpreting what it learned could reveal physical relationships that conventional forecasting methods have overlooked.

WeatherNext is also probabilistic. Instead of committing to one future, it generates an ensemble of plausible scenarios so forecasters can see the distribution of outcomes, including low-probability but devastating rapid-intensification cases. Conventional global systems often run about 50 ensemble members. WeatherNext produced 50 during the 2025 hurricane season and has since scaled to 1,000. A single 15-day forecast can run in under a minute on a TPU, making a much broader search through possible futures computationally practical.

The operational test matters

Historical backtests can flatter a model, especially when extreme events are rare. WeatherNext gained credibility by moving into live collaboration with the US National Hurricane Center and other forecasting organizations. WIRED highlights Hurricane Melissa, which developed over the Caribbean in October 2025. Five days before landfall, when conventional models disagreed about whether the system would remain weak or intensify toward Jamaica, WeatherNext assigned an 80 percent probability to a Category 5 landfall there. Melissa did become a Category 5 hurricane and caused severe flooding and landslides.

That example should not be read as proof that the model will win on every storm. The paper’s formal evaluation covers only three seasons, and tropical cyclones vary by basin, structure, and atmospheric setting. A model that performs exceptionally on one season or one dramatic case can still miss the next. The researchers therefore report that adding WeatherNext to a weighted consensus improves the combined forecast; they do not argue that it should replace every other system.

Human forecasters remain central for the same reason. A model estimates where a storm may go and how strong its winds may become. Experts must reconcile conflicting guidance, account for rainfall, storm surge, terrain, infrastructure, and population vulnerability, and translate probabilities into warnings people can act on. Forecast accuracy is a component of public safety, not the whole system.

Opening the model—and the questions

Google is releasing the WeatherNext Cyclones and WeatherNext 2 code and model weights, along with a smaller WeatherNext 2-mini that can run in a free Colab notebook on a single TPU. That gives meteorological agencies and researchers a chance to reproduce the results, adapt the system to regional needs, test it across future seasons, and investigate why low-resolution data worked so well. It also makes the headline claim more falsifiable than a closed demonstration.

The release does not eliminate practical constraints. Local agencies still need reliable observations, computing access, skilled forecasters, and communication systems. Open weights do not automatically produce trustworthy local warnings, and the Nature page notes that the publicly posted manuscript is an early, unedited version accepted shortly before publication. Continued independent evaluation will matter as much as the initial benchmark.

Still, WeatherNext is a concrete example of AI improving a mature scientific workflow rather than merely imitating its output. Its value comes from combining learned global patterns, cyclone-specific observations, cheap large ensembles, and experienced forecasters. The strongest lesson is not that AI has solved hurricanes. It is that faster probabilistic computation may let people see dangerous possibilities earlier—and that even one additional day can change what communities are able to do before a storm arrives.