Here's the secret behind AI art prompts: every word and every picture gets turned into a point on a giant invisible map. Things that mean the same sit close together. Once you see the map, you'll understand how typing "dragon" can steer a picture. Let's build it — with a real model.
Drag the word chips anywhere you like — put ones that mean similar things near each other. Then hit 🗺️ Map it for real: a real AI reads each word, turns it into numbers, and slides every chip to where it truly belongs. Nobody programmed these positions — the model learned them by reading.
Type any two words. The model measures the angle between their two points on the map and turns it into a match score. cat vs kitten = super high. cat vs car = low. Try it, then try to surprise yourself.
This is the money idea. CLIP is a real model that maps pictures onto the exact same map as words. Show it a picture and a list of words, and it scores how close the picture is to each word — with no labels, no training on your image. This is why typing "dragon" can pull an image generator toward dragons.
A target word appears with three choices. Pick the one closest in meaning. The real model checks your answer with true cosine similarity — no faking. Nail 3 in a row to win!
Embeddings are maybe the most-used idea in all of AI. Once everything is a point on a map, "find similar" becomes "find nearby" — and that unlocks a ton:
Search by meaning, not exact words. "Money problems" finds a doc that says "budget crisis." Powers modern search bars.
Songs, videos, products get embedded. "Because you liked X" = "here's the nearest point on the map."
Remember Module 3? Chatbots look up facts by embedding your question and grabbing the nearest notes. That's this map, at work.
The map above was drawn by real models: MiniLM for text and CLIP for pictures, both running 100% on your machine. Load them once, then throw any two phrases at it and watch it score their meaning — offline.
First load pulls the model from Hugging Face, then it's cached for offline use. On a 4GB laptop give it a minute. ☕
You found the map under all of AI: words and pictures become points, and close points mean similar things. You measured real cosine similarity, and you saw CLIP put a picture and words on the same map — the reason a text prompt can steer an image. This idea powers search, recommendations, and RAG.