This is why “AI” is such a stupid generic term.
This kind of modeling is exactly what machine learning is great at, and this is a perfect use of it. Not fucking chatbots.
Yeah, back when “data centers” were “super computers”, one of the biggest reasons for building them was weather forecasting.
back when it was useful and not some shitty fad
I think the point of the CNN article is that it was a Google AI product that supposedly “outperformed” traditional purpose built machine learning models… They pointed to a single example with mathematically small differences. The rest is just unquantitated AI adoring fluff quotes by Hurricane center director. The whole article seems like a promotional effort by Google
No its about traditional physics based models vs Machine Learning models (or AI). Google’s model isn’t using LLMs, but still uses deep learning models to predict the weather. Per the article:
At that point, many of the physics-based computer models, which rely on mathematical equations describing how the atmosphere works to simulate future weather conditions, were vacillating between different projections but generally forecasting a weaker storm that would take a different track than the DeepMind and other AI models were showing.
The original commentors point is valid in that we just call everything AI right now. While it’s not a terrible term for ML models, the general public largely associates AI with LLMs which is a misnomer.
They weren’t using machine learning models before. They were using physics models which just run simulations on a given set of equations, with a given set of assumptions (which are where the models all differ, those underlying assumptions).
They did not point to a single example though either. They pointed to several. The WeatherNext model has been outperforming the traditional models on everything over the past year. Consistently.
The term is fine - people just don’t know the actual definition of it.
AI is a broad, unspecific term. It’s like the term “plants” - it covers everything from grass to giant redwoods but we have more specific terms for the subcategories for a reason. If we just refer to all life that photosynthesizes as plants then nobody knows which plant you’re talking about. Nowadays when someone speaks of AI it’s relatively safe to assume that they’re talking about one of the multimodal frontier models (that are not only LLMs) but it’s sloppy, imprecise use of language and it will lead to people talking past each other.
In the classical sense AI refers to a system capable of performing any cognitive task typically attributed to a human. Under this definition, even an Atari chess engine is an AI. However nowadays the meaning has shifted a little and the term typically refers to a system that has been trained with a bunch of data.
The good machine learning, versus asking ChatGPT if you’re safe from the hurricane.
Yeah, this seems fine.
- using encoders to create useful sequence data representations in much more information efficient ways
- using transformers to act on sequential data, with attention altering focus dynamically instead odd classical AR style
- training neural networks inside differential equations is very effective
- you can output higher dimensional outputs, including classifications that need less human review
This is machine learning mixed with classical modelling, not “AI” by any definition of AI.
Because more AI = more climate change and thus more hurricanes to justify its existence?
What are they used before then? Wasnt it before not just a lot of if … else … .chains or how the frick did they “predicted” it then before? (Okay maybe per hand, but still )
Physical models use physics based simulations to predict wind paths.
Think
PV=nRT, Navier Stokes, Finite Element Analysis-type stuff. A mix of fundamental physical equations and simplified engineering math based on historical data.Different simulations use different assumptions in the underlying models so they give different output.
The catch is there’s often orders of magnitude more variables at play than physical models are designed to utilize, which is where machine learning comes in super handy - drawing statistical inferences from millions of data points to provide a forecast based on statistics, not physics.
