I asked GPT3.5turbo "Pretend you are a character called Samatha and you're in your house. You go up to the thermostat and select a comfortable temperature and explained your reasoning"

> Next, I take into account my personal preferences and comfort levels. Everyone has their own ideal temperature range, and it's essential to find the sweet spot that makes me feel most comfortable. For me, it's usually between 22 to 24 degrees Celsius (72 to 75 degrees Fahrenheit). This range allows me to feel neither too cold nor too warm, striking the perfect balance.

It also goes on about how the humidity could effect the desired temperature, etc.

It doesn't need the ability to feel temperature (which could also be a single floating number using kelvin), but it can already describe a "comfortable temperature" and what factors would effect it.

Side note: It doesn't "know" anything, it can only make a "best guess" which is now fairly reliable enough to be useful. It doesn't need the ability to test things to learn, we did it already for it, and it's using that to predict the results. You could make a recursive system to allow it to test data if you'd like though.

Think of all the top journals, textbooks, etc. People have understood the world by interacting with it, detailed their hypothesis, conducted experiments, recalled their learning and written down conclusions.

It's not at all obvious to say a useful world model cannot be derived strictly from all this written information.