Skip to content

AI

Google's WeatherNext 3 beats the supercomputers at guessing rain

Google's new AI forecaster tops a public benchmark, including against government-grade traditional forecasts. It is heading into Search, Maps and Gemini, so your umbrella excuses are on a deadline.

Cover photo: Dietmar Rabich / CC BY-SA 4.0

TL;DR

  1. WeatherNext 3 ranked most accurate on Operational WeatherBench, per the benchmark results.
  2. It also edged traditional forecasts from the US National Weather Service and ECMWF.
  3. Coming to Search, Maps and Gemini; researchers and enterprises get it on Google Cloud.

What happened

Google has released WeatherNext 3, its latest AI weather forecasting model. According to TechCrunch, it came out on top among the leading contenders tested on Operational WeatherBench, a tool for comparing AI forecasts built by the startup Brightband. The benchmark looks at metrics such as temperature, windspeed and humidity.

The field it beat is not a bunch of pushovers. It includes deep-learning models from Google itself, Microsoft, Nvidia and the European Centre for Medium-Range Weather Forecasts (ECMWF). It also edged out traditional forecasts from the US National Weather Service and the ECMWF.

Those traditional forecasts come from physics equations run on giant supercomputers, the old-school way of predicting weather. The newcomer, trained on patterns, is now ahead of them on this benchmark.

Google says the model will start feeding the weather information you see in Search, Maps and Gemini. It will also be offered to researchers and enterprise users through Google Cloud.

Google senior staff engineer Samier Merchant told TechCrunch it is the first time some of the model's core variables will feed many Google products directly.

Why it matters

Weather is one of the rare AI use cases where being right is easy to check. Tomorrow arrives, and it either rained or it did not. A benchmark win here is less abstract than a chatbot acing a quiz.

Distribution is the bigger story. A better model in a research paper helps researchers. A better model behind the weather card in Search and Maps lands in the pockets of a very large number of people who will never know its name.

One caveat, sapiens: the benchmark result is the headline, and it is a comparison on a specific test built by a third party. Real-world forecasts will be judged one soggy commute at a time.

It is also another sign that deep learning keeps muscling into a field long owned by physics-based simulation. The supercomputers are not retired. They just have a very confident new colleague.

What to do with it

  • Check the weather card.Watch how Search, Maps and Gemini forecasts change as the rollout reaches you, and note when they beat your old app.
  • Cross-check before big plans.Keep a second source handy for weddings, hikes and anything involving ladders.
  • Builders, look at Google Cloud.Researchers and enterprise teams can get the model there, so logistics, energy and agriculture folks should see what it offers.
  • Read the benchmark.Operational WeatherBench is a public comparison tool, so see which metrics matter for your use case.

Ciao, sapiens. Take the umbrella anyway. The model is good, not a deity.