WeatherNext 2: Our most advanced weather forecasting model

Original Article ↗ Hacker News Discussion ↗

Perceived Forecast Accuracy & Consumer Experience

  • Several commenters say Google’s consumer forecasts (Search, Pixel, default Android weather) have become noticeably worse in the last 6–12 months, with temperature off by a few degrees or rain predictions clearly wrong.
  • Others report consistently good results from national services (e.g., Norway, US NWS) or apps like Windy, yr.no, and some niche apps.
  • A recurring theme: forecasts have objectively improved over decades (linked long‑term statistics), but users still experience jarring local failures, especially for precipitation and in complex microclimates.

Ensembles, Uncertainty, and Metrics (CRPS)

  • Multiple comments explain that modern forecasting is fundamentally probabilistic: an ensemble of scenarios is run, and “chance of rain” reflects the fraction and spatial coverage of rainy members.
  • Some users want a single “best” forecast; others strongly value explicit uncertainty/variance.
  • WeatherNext 2’s emphasis on many scenarios and CRPS loss is discussed:
    • CRPS encourages sharp, well‑spread probabilistic forecasts, countering the blurring and loss of extremes seen with L2 losses.
    • Noise-driven ensembles (applied to inputs or parameters) plus CRPS help generate diverse but calibrated members without heavy post‑processing.
    • This is framed as a major technical advance versus earlier neural weather models and generative approaches.

Comparison with Traditional NWP Models

  • Several commenters insist the key benchmark is accuracy vs major physics-based models (GFS, ECMWF, ICON), not speed. They note the article gives limited direct skill comparisons.
  • There’s praise for Google’s recent hurricane track performance and criticism that US GFS has had a poor hurricane year.
  • Some meteorology-savvy participants argue that, in this stack, WeatherNext still relies on ECMWF analyses, so it doesn’t yet close the loop with observation targeting or new data assimilation techniques.

High-Resolution & Specialized Use Cases

  • Energy market participants need 5–15 minute forecasts for load and renewable generation; they describe using regional high‑resolution models like HRRR and custom NWP runs.
  • Other specialized needs (e.g., structural engineering wind gust statistics, wildfire or severe thunderstorm behavior) often require reanalyses, regional models, or bespoke simulations; commenters doubt generic AI models yet excel here.

Data Sources, Integration, and Access

  • Discussion of smartphone barometer data: historically proposed, but commenters note privacy, quality-control, and limited benefit; WeatherNext 2 does not use such data.
  • WeatherNext 2 outputs are available via Earth Engine, BigQuery, Vertex AI, and are being integrated into Search, Gemini, Pixel Weather, and Maps/Maps Weather API; no dedicated consumer “WeatherNext” app exists.