For the downvoters: This kind of AI is nothing like generative AI / LLMs. This is the kind of AI that's actually been used for decades and has proven very good at handling large amounts of data. Machine learning excels at ingesting huge amounts of data and doing stuff like pattern analysis and predictions. The advancements in hardware have meant larger models, that can handle more data with more precision and weather prediction is one of the places where these advancements have shown good results.

So don't just downvote because AI bad. I fucking hate genAI and what the handful of AI companies are trying to push onto people. But machine learning in data analysis is actual science and a good thing.

Also: Weather systems are chaotic by nature and by that I mean chaotic like in Chaos theory (https://en.wikipedia.org/wiki/Chaos_theory) often misunderstood and misquoted. But it means very small variations can have huge effects. That's why the prediction van be 90% confident and still be absolutely wrong. That's inherent in how these systems work and cannot be fixed. So better weather prediction does not mean perfect weather prediction.

perfect comment

I hate that they repurposed Ai and started using GenAI.

GenAI is marketing genius (if marketing can be called genius) because its a useful abbreviation for GenerativeAI AND General AI.

Artificial Intelligence also never meant actual intelligence. Just computing tasks typically associated with human intelligence.

AI requires input. Its just like every other mathematical function in existence. You provide data to the function and you get data back. Lots of data is used to build the function. The more complex the function is, the more effort is required to calculate the result.

Its INFURIATING to me that they used terms like "thinking" for LLMS. They don't think, they calculate. Text prompt goes in -> processed into token represented numbers -> numbers are entered into the function -> the algorithm calculates the best response token -> the algorithm continues to calculate the likelihood of subsequent tokens until pre-configured bounds are met.

This is why raw LLMs consistently respond with acknowledgement of the prompt, followed by praise of something in the prompt. The models are trained on (neurotypical) human social interactions.

They can't even be called synthetic intelligence.

We also don't have an agreed upon definition for intelligence either. Most serious attempts at discussion about AI are doomed from the get go as we don't even agree on the meaning of terms.

This is what I want AI to be used for. This is actually useful for humans, and it does something that humans cannot reasonably do themselves. It's perfect.

That's why the prediction van be 90% confident and still be absolutely wrong.

I think that's just how probability works? 90% confidence means there's a 10% chance it'll be wrong. So it'll sometimes be wrong. 1 out of every 10 times (on average), which would be all the time because weather is happening constantly.

Regardless, I'll wait to see how accurate this model is long-term before praising it.