December 19, 2025

Cloudy with a chance of hot takes

NOAA deploys new generation of AI-driven global weather models

NOAA’s new AI forecasts spark hype, side-eyes, and stormy comment wars

TLDR: NOAA launched AI-driven global models that deliver faster forecasts using far less computing power. Fans cheered speed, skeptics flagged GraphCast roots, ECMWF comparisons, energy tradeoffs, and warned: please no LLM steering hurricanes.

NOAA just dropped a trio of AI-powered weather models promising faster, cheaper, and more accurate forecasts—think “same forecast, way less computer time.” Cue the comment section going full thunderstorm. One camp is begging for sanity: not all “AI” is chatbots. As one skeptic pleads, please don’t put a talky robot in charge of hurricanes; an LLM (large language model) writes sentences, not forecasts. Another thread dives into model lineage drama: the new system reportedly leans on Google’s 2023 GraphCast, not the shiny WeatherNext 2, prompting a chorus of “cool, but not bleeding edge.” Then the global scoreboard came out—users asked if it now rivals Europe’s gold-standard ECMWF (that’s the European Centre for Medium-Range Weather Forecasts). Early NOAA verification here has folks cautiously optimistic, while others want clearer numbers and resolution details. The energy nerds chimed in with a twist: training is pricey, but day-to-day AI forecasts might actually save power compared to traditional physics simulations. Meanwhile, accessibility sparked mini-drama: some tools show up behind the WeatherBell paywall, while Tropical Tidbits offers a taste for free—cue the “forecast FOMO.” And yes, people joked about “Alexa, stop the rain” and “Skynet tracking storms,” even as NOAA admits hurricane intensity still needs work.

Key Points

  • NOAA launched operational AI-driven global weather models: AIGFS, AIGEFS, and HGEFS.
  • AIGFS delivers comparable or improved forecasts to GFS using only 0.3% of its compute and finishing a 16-day run in ~40 minutes.
  • AIGEFS is a 31-member AI ensemble requiring 9% of GEFS compute; early tests suggest extended forecast skill by 18–24 hours, with overall skill comparable to GEFS.
  • HGEFS combines 31 GEFS and 31 AIGEFS members into a 62-member hybrid ensemble that outperforms both AI-only and physics-only systems.
  • NOAA reports first-of-its-kind operational use of a hybrid physical-AI ensemble, with ongoing work to improve hurricane intensity forecasts.

Hottest takes

"not that they have actually integrated an LLM into a weather model" — margalabargala
"it's based on a model from 2023 (GraphCast) rather than the WeatherNext 2 model" — Workaccount2
"weather is an interesting contradiction of the AI-energy-waste narrative" — apawloski
Made with <3 by @siedrix and @shesho from CDMX. Powered by Forge&Hive.