11 Comments

This does look similar to diffusion with all the level granularity.

Diffusion on weather really cool 🖤

Expand full comment
author

Have you seen Diffusion Guided AutoRegression? that's an idea that this reminds me off

Expand full comment

Haven’t seen this in practice.

But the UNet reminded me of diffusion.

But the recent SD3 model does have major tweaks in that direction.

Expand full comment

Thank you for the breakdown! This is such an excellent use case for ML and I didn't even know about it until you sent out this article.

Expand full comment
author

Yeah. Sucks this didn't get any attention from the AI community

Expand full comment

I'm going to suggest that one reason it might be muted is that it might not reflect what politicians want it to reflect. Climate modeling is notoriously bad because it is such a complex and dynamic system. Most models lock down varibles like crazy which isn't how Earth works. If this is accurate, I'd wager it's not showing what people want it to.

Expand full comment
author

If that were the case, why release it at all?

Ajayt even if MS was muting this(unlikely because they published the video and released a special blog for it), nothing would have stopped creators like me from talking about it. I'm probably even more disappointed in the AI evangelists/creators that have heaps to say about the most trivial bit of llm research or AI drama but completely overlooked this.

Expand full comment

That’s what has me puzzled too which makes me wonder what bias it’s not confirming.

Expand full comment
author

I don';t think this is a case of people hiding anything as much as it is simply a case of people not caring about AI.

Doing forecasts some days in advance isn't going to change instituional narratives

Expand full comment

Until it comes to climate models. Then it’ll be interesting. :) These folks have been using AI for decades with the latest statistical and stochastic mathematical models. Not caring about AI doesn’t make sense since they’ve been the model leaders for decades.

Expand full comment
author

I meant not caring about AI that's not LLMs. Everyone talks only about them, even though there's so much cool research going on in another fields.

Also climate scientists are a lot less than advanced in AI than you'd think. I spoke to a group recently that was trying to use ChatGPT for time series forecasting.

Pretty sure it's a case of dedicating PR to the development that will move stock markets. Chatbots help consultants write faster. Climate model helps farmers. Who drives stock price more?

Expand full comment