Most of this week's AI news has been about systems we can't quite see inside. Here is a corrective: a piece of AI capability that is concrete, measurable, independently scored, and already running in products hundreds of millions of people use.

Google DeepMind and Google Research released WeatherNext 3 this week, the latest in their line of AI weather models, and the headline is resolution. Where the previous version produced forecasts on a 25-kilometre grid every six hours, WeatherNext 3 forecasts down to 5 kilometres and updates every hour — a global weather picture the company says is roughly five times sharper, sharp enough to track how a front behaves over local terrain rather than smearing it across a region.

The numbers, with the caveats

The accuracy gains are real but worth stating precisely, because weather scores are always measured against a particular reference. On precipitation — the hardest thing to forecast and the thing people actually care about — DeepMind reports improvements at early lead times of up to 60% against one satellite benchmark (IMERG), around 30% against another (MRMS), and about 10% against physical rain-gauge measurements. Looking a day or more ahead, it claims up to 50% more accurate precipitation forecasts. In other words: a big jump on the toughest variable, larger against some yardsticks than others.

What's actually new

The more interesting shift is where the model gets its information. Earlier AI weather systems mostly learned from the outputs of traditional numerical weather models — which carry a built-in lag of about six hours. WeatherNext 3 trains more directly on raw observations: live geostationary satellite feeds and weather-station data. That is what lets it update hourly and stay current, rather than waiting on the slower physics pipeline it used to lean on.

Independently checked, and already shipped

Two things make this more than a press release. First, it is externally evaluated: Brightband, which runs independent live leaderboards for weather models, ranks WeatherNext 3 among the top global systems — an outside scoreboard, not just the lab's own numbers. Second, it is not sitting in a research repo. As of this week the forecasts are live in Google Search, the Gemini app and Google Maps, and available to developers through the Maps Platform Weather API, Earth Engine, BigQuery and Cloud Storage.

There is a quieter application tucked in the announcement that may matter more than the umbrella-forecasting: the model predicts wind speeds at 100 metres — turbine height — and includes variables aimed at solar and wind farms. Grid operators live and die by how well they can predict renewable output a day out; a sharper, faster forecast is worth real money and real carbon there.

It is a useful reminder, in a week full of stories about AI that is hard to observe or attribute, that some of the most consequential machine-learning work is the least mysterious. This is a model doing a narrow, well-defined job, measured against public benchmarks by an independent scorer, and deployed where people can actually use it. Not every frontier is a chatbot.