Nestor G Pestelos Jr · ELI5

The Language of AI

Embeddings

How computers turn words into coordinates on a map where similar ideas sit close together.

01 / The Meaning Map

Every word gets coordinates on a giant concept map.

An embedding model converts every word into a list of numbers, placing similar concepts in the same neighborhood.

Dimension 1 (e.g. Living Being) Dimension 2 (e.g. Size / Scale) ANIMAL CLUSTER "dog" "puppy" "cat" FRUIT CLUSTER "apple" "banana" VEHICLE CLUSTER "truck" "airplane"
How many dimensions does a real map have?

A paper map uses two dimensions (X and Y). Real AI embeddings use between 768 and 3,072 dimensions. Each dimension captures a subtle trait: formality, physical size, sentiment, technical depth, or grammatical role.

02 / Semantic Distance

Distance equals meaning.

The closer two coordinates are in space, the more closely related their concepts are in the real world.

"doctor" "physician" Distance: 0.12 (Near) "doctor" "submarine" Distance: 0.94 (Distant)
How models measure distance in high dimensions

AI systems use cosine similarity. Instead of measuring physical ruler distance, the system calculates the angle between the two coordinate arrows. An angle of zero means identical meaning, an angle of 90 degrees means unrelated, and 180 degrees indicates opposing concepts.

03 / Vector Arithmetic

Math with meaning: the arrow trick.

Because concepts have directions, you can add and subtract words like numbers to solve analogies.

"Man" + Female gender arrow "Woman" "King" + Parallel gender arrow "Queen" King - Man + Woman ≈ Queen
Why analogical directions hold true

In 2013, researchers at Google discovered this property while building Word2Vec. Because the model learned from billions of real sentences, the difference between "man" and "woman" matches the difference between "king" and "queen", "uncle" and "aunt", or "actor" and "actress".

04 / Practical Retrieval

Find what you mean, not just what you type.

Keyword search requires exact matching words. Embeddings match the underlying intent even when every word is completely different.

USER SEARCH QUERY "feline physician" Vector Distance ✓ "Veterinarian who treats cats" 94% Match ✗ "Physician practicing cardiology" 31% Match
How Retrieval-Augmented Generation (RAG) uses embeddings

When you ask an AI a question about your documents, a vector database converts your query into an embedding and retrieves the three to five text passages closest to it in semantic space. The AI reads those retrieved chunks to provide an accurate, grounded response.