To a computer, the word 'tank' is just a symbol — it has no built-in sense that a tank holds water, or that it's related to a 'heater' or a 'pump'. An embedding fixes that by giving every word a location on a giant map, built by noticing which words tend to show up in similar situations across huge amounts of text. Words used in similar ways end up near each other on that map — 'tank', 'coil', and 'heater' cluster together because they keep appearing in the same kinds of sentences, while a completely unrelated word like 'football' sits far away. That closeness is what lets the model treat 'tank' and 'coil' as related ideas, even though nobody ever explicitly told it what either word means.