Ω · Oneironexus · Blog

Fifteen Days Ahead

A weather forecast that reaches two weeks out, runs in minutes on one chip, and beats the best operational system on 97 percent of what it was tested on. What that gives a farmer, a ferry and a festival.

14 October 2026 · Aaron Michael Weis

Day 1. A farmer in the San Luis Valley looks at the sky and then at her phone. Cutting hay is a bet on three dry days in a row. Today the bet has better odds than it has ever had.

Day 4. A ferry operator on a northern coast is choosing which crossings to schedule next week. The swell forecast reaches far enough now that next week is a plan, not a hope.

Day 9. A festival organizer is deciding whether the main stage goes under a roof. Nine days out used to be a coin flip. It is a forecast now.

Day 15. A city water manager is looking at a reservoir and a two-week outlook, and setting releases with a number she trusts.

Here is what changed underneath those four days.

The machine that learned the atmosphere

For decades the best weather forecasts came from giant physics simulations: divide the atmosphere into boxes, write down the equations, and march forward in time on some of the largest computers on Earth. It is one of the great achievements of applied science, and it is slow and expensive by design.

In 2023 a team published GraphCast, a machine-learning model trained on four decades of past weather. Given the state of the atmosphere now and six hours ago, it produces a ten-day global forecast at a quarter-degree resolution in under a minute. Tested against the most accurate operational deterministic system in the world, it was more accurate on 90 percent of 1,380 verification targets — different variables, heights and lead times, scored one by one.

In December 2024 the same group published GenCast, which goes further in two ways. It forecasts fifteen days instead of ten. And it forecasts probabilities — an ensemble of fifty possible futures, which is what a decision-maker actually needs when the question is "how likely is the rain." Against the European Centre's operational ensemble, the gold standard for probabilistic forecasts, GenCast was more skilful on 97.2 percent of 1,320 targets. One fifteen-day forecast takes about eight minutes on a single chip.

Both results are peer-reviewed, in Science and in Nature. Both come from Google DeepMind, which is worth saying plainly — and both were scored against the public, operational systems the whole world already uses, on targets everyone can inspect.

What minutes instead of hours buys

Speed sounds like a convenience. In forecasting it is a capability.

When a forecast takes hours on a supercomputer, you run it a few times a day. When it takes minutes on one chip, you can run it fifty times for the ensemble, re-run it as new observations arrive, and run it in places that never had a supercomputer at all. A regional weather service with a modest budget can produce a fifteen-day outlook that, a few years ago, only three or four organizations on Earth could make.

That is the quiet part of the story. The headline is accuracy. The gift is access.

The physics stays in the room

The learned models did not push the physics simulations aside; they learned from them. Forty years of reanalysis — the physics models' own best reconstruction of what the weather was — is the textbook GraphCast and GenCast studied. The two approaches now run side by side in forecasting centres, each checking the other. The science that took a century to build is the reason the learned model has anything to learn.

Day 16

The farmer cut her hay on days two, three and four. The ferry ran every crossing it scheduled. The main stage stayed under the sky. The reservoir is where the manager wanted it.

None of them thinks about the model. They think about the plan. That is what a good forecast is for.

Where the numbers come from