Great question! We'll follow up with you+other CIRA folks+NHC with more details as part of our research collaboration. I'll say we have some understanding of the causes. We'll also probably push FNv3.1 in 1-2 weeks, which will give some improvements (but not full fixes) in the short lead times.

Finally, we're releasing deepmind.google.com/science/weat..., a preview of our AI models. Huge kudos to @tom-andersson.bsky.social who played a key role improving Weather Lab with expert trusted testers since last Fall.

Here's a current cyclone in China, with our model (blue) being more confident.

According to internal evals, reproduced by the Cooperative Institute for Research in the Atmosphere (thanks to @franklinjamesl.bsky.social!), our experimental model excels at both tracks and intensities. We're looking forward to running real-time evals with CIRA & other worldwide partners.

In my case, the conversion happened with:
1) Kenji Kawaguchi's (my labmate at the time) "Deep Learning without Poor Local Minima" 2016 paper and
2) empirically finding gradient descent highly effective for a very non-smooth problem (modular meta-learning) in 2018.

Oops, I'm new to Bluesky :) I've now enabled DMs and sent you one

Hi Blanka! I'm a research scientist at Google DeepMind working on AI for weather; would love to be part of this community!

Paper: www.nature.com/articles/s41...
Joint work with Ilan Price, Alvaro Sanchez-Gonzalez, @tom-andersson.bsky.social, Andrew El-Kadi, Dominic Masters, Timo Ewalds, Jacklynn Stott, @shakirm.bsky.social, Peter Battaglia, Remi Lam & Matthew Willson

Weather is fundamentally chaotic and with finite sensors and computation, forecasts need to be probabilistic past a few days. In GenCast, we used diffusion models to generate an ensemble of realistic future scenarios, showing skill up to 2 weeks into the future!