“This is a completely different approach to what people have done before. The writing’s on the wall that this is going to transform things, it’s going to be the new way of doing forecasting,” Turner said. He said the model would eventually be able to produce accurate eight-day forecasts, compared with five-day forecast at present, as well as hyper-localised predictions.
Dr Scott Hosking, the director of science and innovation for environment and sustainability at the Alan Turing Institute, said the breakthrough could “democratise forecasting” by making powerful technologies available to developing nations around the world, as well as assisting policymakers, emergency planners and industries that rely on accurate weather forecasts.
Not to undermine their work, but didn’t Microsoft not already release Aurora Forecasting (a 1.5b model, which compared to models for text is rather small - those tend to start at 3B - which makes sense because there is also a lot less data to build a model on).
Anyway, I am happy to see competitors popping up, because NWP is hard enough already.
One of the few good uses of AI, and actual real use of generative AI. Predicting the next frames of a storm seeing things that previous algorithms and humans could not. This will honestly probably save lives with earlier detection.
paper (link without AWS tracking or “email attachment” url): https://arxiv.org/abs/2404.00411
GitHub repo (no content yet): https://github.com/annavaughan/aardvark-weather-public/tree/main
This repo will contain code and weights used to run the Aardvark Weather model. This will be released to reviewers during the review period and will be publicly available on the completion of peer review. If you would like to be notified when the codebase becomes available please email […]
I’d settle for detailed short term local rain forecast. It’s such a huge application and as far as I can tell (please correct me if I’m wrong!) nobody does it at all well.
- Often ensembles are only run every 3 or 6 hours, so the predictions are needlessly out of date.
- The ensembles have a very small number of runs (like less than 10) so you can’t get a good estimate of probabilities.
- Usually you can’t get access to the raw data anyway and user facing sites dumb things down to a single number, so e.g. “it’s going to drizzle all day” and “it’s going to tip it down for half an hour at some point” are presented exactly the same. As are “it’s either going to rain loads or not at all” and “it’s definitely going to rain a bit”.
