It doesn't come with an example for how to fine-tune it using hf transformers, so it will be way more work to use than the three models I started with. My model selection process was heavily biased by the availability of easy-to-modify examples.

But worth considering for the longer term.

brohan.org/Robot_Rainfa... does this with a trio of mini-AIs.

It's not just the UK - global land temperatures can be independently confirmed - agupubs.onlinelibrary.wiley.com/doi/full/10....

The future of scientific publishing is YouTube, Github, Medium, Substack, ...
I don't see much of that in the agenda.

It is easy to come up with designs for ground-breaking climate models using ML.
The thing that's very hard is to come up with a ML climate model that looks and operates like an NWP model.
If we can find the courage to move on from NWP, then the future is bright indeed.

They only use ERA5 at all because they are doing NWP imitation. Eventually they will realise that to forecast London precip. (say) an AI model doesn’t have to bother with the whole NWP state vector.
Then we’ll start to see real improvements.

The really interesting result is the reduced observational requirement for DA. How far can we push that?

You could build a model without ERA5 - train it on satellite data, in-situ observations, or what you will. There just isn't much point.

While all the AI models use ERA5, they don't require it.
It really is a supercomputer-free forecasting system.

The state vector describing reality has way too many dimensions.
What else can we do?

I won't be following up myself (the data I'm after is almost all in tables). But I do urge you to give it a go - point your browser at aistudio.google.com upload your image and start asking. It could hardle be easier to experiment with.

And as so often it's not perfect, but it's very promising. Definitely worth looking into.

No.

But since you ask.
I've uploaded a page from "A Collection of Voyages Chiefly in the Southern Atlantick Ocean" books.google.co.uk/books?id=sGx...

And asked Gemini a couple of questions:

I was thinking that maybe in a couple of months there would be an upgrade to Gemini that we should experiment with.

I made insufficient allowance for 'AI time'. An update has come out today -
developers.googleblog.com/en/gemini-2-...

I don't know if it makes any difference.

All the code is available at the link, and Gemini's free tier is enough to experiment with.

Go on - give it a try.

" We can say with some confidence that we are better at reading the logs than the original log-keepers were at writing them."

I think we are very close to being able to replace the 'we' with 'AIs'.