A better method, released free
For decades, weather forecasting has been a physics problem solved on national supercomputers, improving a little each year. Google DeepMind's GraphCast, described in Science, took a different route. It learned from decades of historical weather data, produces a ten-day global forecast in around a minute on ordinary hardware, and was released as open-source weights under a permissive license.
What the numbers do and do not say
Against the leading deterministic system, GraphCast did better on more than 90% of the variable and lead-time combinations tested. That is not every measure. Physics-based models still lead on some, including severe convection and the specifics of tropical cyclone tracks. A follow-up, GenCast, published in Nature, covered probabilistic forecasts.
How the field responded
ECMWF, the European medium-range forecasting organization, did not treat it as a threat to manage. It launched its own operational machine-learning forecasting system, the AIFS, and runs it alongside its numerical model. A serious field was handed a better method and put it to work next to the old one.
Most AI stories at work arrive through a vendor's pitch. This one arrived through a journal article and a download link, which changes who has to decide what to do about it.
The shape, and a test for it
This is a real AI leap, and it looks nothing like a productivity survey. There was no rollout, no license purchase, no change program. What made it possible was the kind of task it works on: years of historical data, an output that can be checked against what actually happened, and an expensive incumbent method.
The conditions are worth naming because they are the reason this case is clean. A forecast is checked against what actually happened every day, so a model's claim to be better can be tested on a scoreboard nobody controls. Few workplace tasks come with one. Where a task has no such scoreboard, a better-sounding method is harder to trust, and that is a separate question from whether the method works.
Ask the same of your own organization. Which task has the data, the measurement and the costly way of doing it today? Most teams will not find one. The value is in having looked.
The check takes one working session: See where your organization stands on this →
Keep reading: The cost-to-check rule: a one-line test for when AI is worth it · AI isn't a Calculator
This is the thinking behind a Trusted AI Culture.
Sources: Google DeepMind, GraphCast (Science) and GenCast (Nature). ECMWF, technical bulletins on the Artificial Intelligence Forecasting System.
Written with AI in the loop: my idea, AI drafted and sharpened, my judgment on the way out. Every word is mine to stand behind. — Darren