The best weather model doesn't give you one forecast. It gives you fifty.

Weather forecasting is the oldest serious prediction problem there is, and the benchmark to beat is ENS, the ensemble system run by the European Centre for Medium-Range Weather Forecasts [1][2].

In December 2024, Google DeepMind published GenCast in Nature. Against ENS it was more accurate on 97.2% of 1,320 tested forecast targets, and on 99.8% of them at lead times beyond 36 hours [1].

The interesting part is how it answers. An earlier model, GraphCast, produced a single best estimate of future weather. GenCast instead produces an ensemble of 50 or more predictions, each a possible trajectory — so the output is not "14°C and raining" but a distribution over what might happen [1].

That distribution is what makes it useful for the things people actually need forecasts for: extreme heat, high winds, the track of a tropical cyclone. A decision about whether to evacuate depends on how likely the bad case is, not on the single most likely case [1].

It is also fast. A 15-day global forecast takes about eight minutes on a single TPU, against hours on a supercomputer for the physics-based systems, and it was trained on four decades of ECMWF's own reanalysis archive up to 2018, then evaluated on 2019 [1].

Its authors and outside experts are clear about the limits: it underpredicts the intensity of cyclones, struggles in the upper troposphere, and is trained on a past climate that is not the one coming. One meteorologist's point is worth keeping — humans still distil this into the forecast anyone acts on [2].

This is the shape of AI worth wanting: it beat a system decades of physics and public investment had produced, and it did so while giving MORE information about its own uncertainty rather than less. The fast, cheap ensemble is a genuine public good — a national met service without a supercomputer budget can now run one. And it stayed honest about where it is weak, which is the part that makes the rest usable [1][2].

The danger is downstream, in what happens as the forecast travels. A distribution over 50 outcomes gets collapsed into a single number by the time it reaches an app, a headline or a meeting — the range is the first thing dropped, because ranges are awkward to display and a single number sounds more authoritative. The model did the hard part and said how sure it was; the reporting chain deletes it. Every wrong-sounding forecast you have complained about was probably a 30% chance that happened, presented as a promise that it would not [1][2].