Interactive Lab · How models represent meaning
A model can't do math on the word "king." So it turns every word into a vectora point in space, placed so that things that mean similar things sit close together. Once meaning is geometry, two magical things become true: distance = similarity, and direction = relationship. Click any word to see its nearest neighbors, then run the famous analogy.
🎬 Watch the 55-second explainer, then try it yourself below ↓
▶ The meaning map, click any word
Click a word, say dog or king, and watch the map draw lines to its closest neighbors. Similar meanings live in the same neighborhood.
Embeddings are the currency of modern AI infrastructure: every RAG system, semantic search, recommendation engine, and vector database (Pinecone, pgvector, FAISS) stores and compares exactly these vectors. The neighbors-and-analogies geometry you just explored is what “similarity” means computationally everywhere.
The code that's actually running
Click lion, which words light up as neighbors, and why those and not apple? Now hit ✨ Show analogy: before it lands, predict where king − man + woman should point. Same trick powers real systems, if apple and orange sit close, what do you think a search engine does when you type "orange"? (Hint: it looks up the nearest neighbors of your query's vector.)
The takeaway