Our research reveals how AI models understand and connect places through language. The network visualization shows how different locations, objects, and concepts are interlinked in the model's understanding.
Words highlighted in green show how the model breaks down language into subword tokens. Notice how "georgetown" becomes "george##town", revealing the model's basic building blocks of understanding.
In the Paris cluster, we see strong connections between "Arc de Triomphe" and "Olympic". The model consistently associates these landmarks with the upcoming 2024 games, showing how temporal context influences place understanding.
The Stabroek Market cluster reveals local transportation patterns, with "minibus" and "license plate" frequently appearing together, indicating the model's grasp of daily urban life in Georgetown.
Examining connections between Paris and Georgetown reveals how the model understands different types of urban spaces, from tourist landmarks to local markets.