WanderWorld

I wanted to see how much geography a language model knows, and whether I could get exact coordinates just by asking. It works for "where is Paris?" and for "places that smell like candy".

A vibe goes in and numbers come out, numbers a map can use.

LLMs are surprisingly good at geography, even small local ones. The trick is not asking for text at all: ask for a fixed JSON shape.

It even grades its own answers. Take those grades with some salt.

The model answers in a forced JSON schema with names and coordinates, and code turns them into map pins. Bigger models pick more creative places; small ones still get the coordinates right.

• Ask for something weird, like places that smell like candy
• Ask for a list: all the European capitals
• Describe a feeling instead of a place