Thank you for this article. As a now-retired federal government employee, my division would occasionally be contacted by companies pushing their new AI weather forecasts, saying they were superior to the physics-based models we used. But when pressed they 1) wouldn't show us the data they used; 2) would only provide vague statistics on basic items like temperature; and 3) stammered and finally admitted that AI couldn't yet predict the things we really needed like thunderstorms, turbulence, icing, winter precipitation, etc. AI has lots of promise but beware of the overselling going on right now. Good job pointing that out.
Thanks for this post, Aaron. We can't stop the AI train, so it really does pay to understand how it works in the area of atmospheric science. I guess my concern is that weather prediction will become totally privatized, when history has shown that the best way to approach forecasting is by collaboration: private, public, military, and even amateur weather watchers. I hope that doesn't go away, but we'll see...
I was a huge fan of AIWP/AIFS models when they first came out. After reading this article I have new doubts to add to my old one. My old hesitation was based on the fact that a LLM is only as good as its training data and the latter is date-stamped. It's like a new car that becomes a used car as soon as it rolls off the car lot. In an era of very rapid global warming I would guess that atmospheric dynamics are changing quickly. Can the AI models keep up?
Thank you for this article. As a now-retired federal government employee, my division would occasionally be contacted by companies pushing their new AI weather forecasts, saying they were superior to the physics-based models we used. But when pressed they 1) wouldn't show us the data they used; 2) would only provide vague statistics on basic items like temperature; and 3) stammered and finally admitted that AI couldn't yet predict the things we really needed like thunderstorms, turbulence, icing, winter precipitation, etc. AI has lots of promise but beware of the overselling going on right now. Good job pointing that out.
Thanks for this post, Aaron. We can't stop the AI train, so it really does pay to understand how it works in the area of atmospheric science. I guess my concern is that weather prediction will become totally privatized, when history has shown that the best way to approach forecasting is by collaboration: private, public, military, and even amateur weather watchers. I hope that doesn't go away, but we'll see...
I was a huge fan of AIWP/AIFS models when they first came out. After reading this article I have new doubts to add to my old one. My old hesitation was based on the fact that a LLM is only as good as its training data and the latter is date-stamped. It's like a new car that becomes a used car as soon as it rolls off the car lot. In an era of very rapid global warming I would guess that atmospheric dynamics are changing quickly. Can the AI models keep up?