What an embedding is
An embedding is how AI turns a word — or a sentence, or an image — into a list of numbers that acts like a position in space. The clever part is where each word lands: things with similar meaning are placed close together, and unrelated things end up far apart. So “cat” and “kitten” become near-neighbours, while “cat” and “bank” sit on opposite sides of the map. Once meaning is expressed as a location, a computer can compare ideas just by measuring distance. Try the interactive word-map below to feel how it works.
Words as points on a map
Picture a giant map where every word is a dot. On this map, closeness means similarity of meaning, not spelling. “Dog,” “puppy,” and “hound” cluster in one neighbourhood; “apple,” “banana,” and “grape” cluster in another; “Tuesday” lives somewhere else entirely. In the demo, pick a word and watch the tool list its nearest neighbours — the words AI thinks are most similar. Real embeddings do exactly this, except instead of a flat 2D map they use hundreds or thousands of dimensions, giving them room to capture many shades of meaning at once. The 2D map is just a beginner-friendly shadow of that bigger space.
Why numbers beat words for a computer
Computers cannot compare meanings the way we do, but they are brilliant at arithmetic. By giving every concept a numeric position, the AI can answer “which of these is most similar?” simply by computing distances — no understanding required, just measurement. This is why a search for “cheap flights” can surface a page about “affordable airfare” even though none of the words match: the two phrases land near each other in embedding space. Use the compare feature in the demo to see a similarity score between any two words and notice how it tracks your own sense of relatedness.
Where embeddings show up
Embeddings are everywhere once you know to look. Semantic search uses them to match meaning instead of exact keywords. Recommendation systems use them to find items “near” what you liked. Clustering groups similar documents, duplicate detection spots near-identical content, and RAG (retrieval- augmented generation) uses embeddings to fetch the passages most relevant to your question before the AI answers. In every case the trick is the same: convert fuzzy meaning into precise positions, then let distance do the reasoning.
Tips for building intuition
Three ideas will lock this in. First, distance equals dissimilarity — the farther apart, the less related. Second, the positions are learned from usage, not hand-assigned: the model reads mountains of text and pushes words that share contexts together. Third, the real magic is dimensions — with hundreds of axes, an embedding can place “king” near “queen” along one axis and near “ruler” along another simultaneously, capturing meaning far richer than any flat map shows. Play with the demo until “similar meaning = nearby point” feels obvious; that single intuition explains a huge fraction of modern AI plumbing.