There's a really interesting thing happening with generative AI and weather forecasting: The genAI is forecasting better than the physics models.
Some people are worried that it will fail at the edges, namely when forecasting extreme (read: uncommon) weather events. And yet it seems to be better at forecasting hurricane paths, which are the most difficult things to predict, and which are so poorly determined by the physics models that we run something like 20 of the things to try to get an idea of what's going on.
Why, though? Why is genAI doing so well at weather forecasting? I have some thoughts.
Weather forecasting is a game of statistics. We ask, what's the most likely thing to happen, given what we know has already happened? You know what is programmed specifically to answer that question? Generative AI.
This is so, so fascinating to me.
If you forecast the weather by simply predicting that things will be the same tomorrow as they are today, your forecast accuracy will be something like 60 to 80 percent, depending on location. That's bonkers, right? That a simple heuristic based on almost nothing will render a better forecast than a coin toss. If you just always predicted the most dominant weather for a location, you'd get something like 40 to 70 percent. For instance, forecasting that Houston will have temperatures in the 90s, 100% humidity, and 30% chance of afternoon showers is going to end up being correct more than half the days of most years. Again, a ridiculous heuristic can produce a better forecast than a coin toss.
No one forecasts that way, of course, because no one cares very much what the typical weather is. They want to know when it will rain, so they can make plans. They want to know when wind is going to blow their possessions away, so they can secure them. They want to know when flooding will happen, and where, and how severe it will be. So weather forecasting has been far more sophisticated than any of that since humans became human.
Preindustrial forecasters utilized what we referred to as rules of thumb. If the sky is green at dawn, this will happen. If the rock in the yard has dew at this hour, that will happen. These are highly location specific, and utilize purely empirical data. With such methods, people could predict the next 24 hours or so with accuracy better than 80%.
At some point, people started working together to compile temperature and wind data across large areas. This data sharing improved and extended forecasts. Then someone invented weather balloons, and this further improved forecasting. Weather satellites improved things yet again. All of these tools were observation based, and provided layers of information about what had already happened. It became possible to watch a storm travel across the country, instead of just seeing the edge on the horizon the day before. Forecast reliability reached 4 to 5 days. Weather forecasting is one of the most amazing results of cooperative human activity ever. I always get a little awe-struck when I think about it.
The next big players in weather forecasting were the physics models. Now forecasters could look at likely future scenarios, in addition to what had already happened. Using models to inform forecasts, we reached about 10 days of forecast with 95% accuracy. That is so bananas, when you think about it. We can predict the future, y'all. It sounds like some sci-fi shit, but it's real.
So why would the genAI models outperform the physics models? This is the question I've been brooding about, and I think it comes down to two things. First, the physics models require so much estimation, rounding, and smoothing, that they hold an inescapable error. It is not possible to model the real world perfectly with a computational model. Secondly, the genAI models are basically looking at the weather patterns, and then finding that same scenario in the past, and thus predicting that what happened before will happen again. This works because the weather is consistent in its behavior. It's the same concept as saying that tomorrow will be the same as today, except that the genAI program can go digging through mountains of data and find a day in the past that mimics today perfectly, and then see what happened the next day. I'm oversimplifying this somewhat, obviously, but that's the hand-wavy concept of what's going on. IT'S SO FUCKING COOL.
So while I despise genAI language and art models on principle, and I cannot fathom why anyone would use one to do math, there is one great use for the fuckers: Predicting the weather.
(Oh, and for the people concerned about resource usage, genAI weather forecasting uses less energy and fewer computational resources than the physics models do.)
