Imagine an interpreter in a huge hall who hears not only the person next to them but also a whisper in the far corner, and from those fragments reconstructs the entire conversation. That's how the CAS-Canglong neural network works: it picks up connections between pressure, wind, and heat across the whole planet and uses them to predict what the ocean will look like three months from now.

Previously, forecasts ranging from two weeks to a season were meteorology's weak spot. The atmosphere quickly forgets what happened yesterday, while the ocean hasn't yet built up enough momentum to dictate the weather. Physics-based models require supercomputers and still make mistakes, especially away from the tropics. The new neural network trained on 51 years of weather records and now delivers a forecast in seconds on an ordinary graphics card.

In the southern oceans, the correlation between forecast and reality reaches 0.9 one month ahead and 0.7 three months ahead. This means that if the neural network says the water is warming off the coast of Australia, it's highly likely to happen. The error in the area of the warm pool in the Indo-Pacific region was only 3.8%.

Previously, physics-based models produced large errors at such lead times, especially at extratropical latitudes. Now the neural network predicts the global ocean temperature field more accurately than all 22 physics-based models. The secret is that it doesn't try to reproduce the physics; instead, it learns from historical data, like an interpreter who has mastered a language through thousands of conversations.

Such forecasts can be integrated into early warning systems for droughts, floods, and heatwaves. A farmer would learn about a poor harvest months in advance, energy companies could prepare for peak loads, and insurance companies could assess risks more precisely. The interpreter who hears the entire hall is finally translating not just individual words but the whole conversation of the planet.