Jeff AtwoodSep 8, 2025
@thomasfuchs reminder that we predict the weather (quite accurately, nowadays, look it up) using similar statistical techniques. They also don't understand "weather" at all, but they are useful. Citation https://ourworldindata.org/weather-forecasts
Weather forecasts have become much more accurate; we now need to make them available to everyone

A four-day forecast today is as accurate as a one-day forecast 30 years ago.

Our World in Data
Show threadColin JensenSep 8, 2025

@codinghorror @thomasfuchs this is incorrect. Weather models do understand weather and simulate the atmosphere to achieve their predictions. Randomness is sometimes introduced to see different outcomes if the initial measured conditions aren’t quite right

It’s not statistical in the sense of a language model. Also putting in “looking at studies” in your prompts might just be cluing the generator to make a more serious sounding answer that matches you language

Show threadPaul C.

@cjensen @codinghorror @thomasfuchs Just to clarify here, you're saying that "weather models do understand weather" in a way that LLMs don't?

What is the definition of "understanding" that is being used here?

Sep 8, 2025 at 7:10amWeb
Show threadColin JensenSep 8, 2025

@paul @codinghorror @thomasfuchs good question!

Weather models simulate the atmosphere. They do this by subdividing the map into sections, simulating the section, then integrating the result. Each loop of simulation is for a period of time (eg an hour of simulated time). Loop for however many hours into the future you want to predict.

1/

Show threadColin JensenSep 8, 2025

@paul @codinghorror @thomasfuchs

So in this sense the weather model knows what atmosphere is how to simulate it

By contrast, language models don’t understand what paragraphs, sentences, or words mean. They are just statistical models which pick a sequence of words and punctuation that closely match what a likely response to your prompt should look like.

2/

Show threadColin JensenSep 8, 2025

@paul @codinghorror @thomasfuchs

It’s a miracle this works as well as it does.

But knowing how it works, you can see where the “hallucinations” come from: if you ask for a legal brief and there is no precedent in the stats for the argument, it will generate words that statistically best fit the question, including fake cites if that statistically seems to fit

Show threadPaul C.Sep 8, 2025

@cjensen @codinghorror @thomasfuchs

Packing an LLM's context window with a request for a nonexistent legal brief seems more like putting fictional synthetic initial conditions into a weather model. Garbage in, garbage out.

Also, it looks like weather labs are rapidly adopting transformer architectures of one kind or another:

https://www.alcf.anl.gov/news/argonne-develops-new-kind-ai-model-weather-prediction

https://e360.yale.edu/features/artificial-intelligence-weather-forecasting

Argonne develops new kind of AI model for weather prediction | Argonne Leadership Computing Facility

Show threadPeter van BroekhovenSep 8, 2025
@cjensen It’s been my experience that LLMs will never admit to not knowing something. They always try to give you *something* no matter how wrong or nonsensical.