How AI Understands Meaning

Every word gets an address in a city of meaning. Words that mean similar things live in the same neighborhood.

The Analogy

How AI Understands Meaning

Every word gets an address in a city of meaning. Words that mean similar things live in the same neighborhood. If you know a word's address, you know something about what it means.

Use the arrows below, the dots above, or your keyboard arrow keys to move through the stages.

Stage 1 -- Token ID

First, the Word Becomes a Token ID

cat
Token cat
Token ID 9827

A token ID is a number the model uses to look up the token. The model does not understand the word cat directly. It first works with the ID 9827.

Stage 2 -- Vectorization

Then, the Token ID Becomes a Vector

Token ID 9827
Embedding table row 9827

This is vectorization. The token ID points to one row in the embedding table. That row is a list of numbers, called a vector, that helps represent meaning.

Stage 3 -- Visualize the Embedding

Now, See the Vector as a Meaning Address

Each number is one direction in meaning space. Together, the numbers create the token's location.

The vector is not just a list of numbers. It acts like an address. Similar tokens land near each other because their vector addresses are close.

Stage 4 -- The Neighborhood

Words Live in Neighborhoods

Click a word to inspect it
Sample embedding preview
This shows 6 sample dimensions out of a much larger embedding.
Nearest neighbors

Click any word to lock the lines, see distances, and inspect a clear sample embedding. Click the background to reset.

'Cat' and 'dog' live close together. 'Cat' and 'airplane' are far apart. The model knows meaning through proximity, not definitions.

Stage 5 -- The Bias Reveal

Watch the Pattern Form

Stage 6 -- Context Changes Everything

Same Word, Different Address

Same word, different meaning, different address. Context decides where a word lands.

Takeaway

When AI seems biased, it learned associations from training data. Understanding this helps you spot bias and prompt around it.

Every word has an address. But 'bank' near 'money' and 'bank' near 'river' are different places. How does the model figure that out? →