🗺️ What are 'embeddings'? (How AI understands meaning)
Turning words into coordinates in space, so similar meanings sit close together.
3 min read
Computers are great with numbers and terrible with meaning. An 'embedding' is the bridge: it turns a word (or sentence, or image) into a list of numbers — a coordinate in a high-dimensional space — where similar meanings end up close together.
Imagine plotting words on a map where 'king' and 'queen' sit near each other, far from 'banana'. Real embeddings use hundreds of dimensions, not two, but the idea is the same: the position encodes meaning, and distance encodes similarity.
The magic is that this falls out of training. The model notices that words used in similar contexts have similar meanings, and arranges them accordingly. It even captures relationships — famously, 'king' minus 'man' plus 'woman' lands near 'queen'.
Embeddings are the quiet workhorse of AI. Search, recommendations, and 'related items' all lean on them. When ChatGPT retrieves relevant info, it's often comparing embeddings — finding text whose meaning is close to your question, not just matching keywords.
Words as points in space
This is what an "embedding" looks like (squashed to 2D). Words with similar meaning sit close together. Click a word to see its nearest meaning-neighbours.
This is a simplified illustration. Real embeddings use hundreds of dimensions, which is what lets them capture meaning so precisely.
Explore a tiny 2D word-map: notice how words with similar meanings cluster together.
🕒 Try it in 60 seconds
- 1Play with the word-map demo below — see which words sit close together.
- 2Ask an AI: 'How are embeddings used in search engines?'
- 3Then ask: 'What does it mean that king - man + woman = queen?'
Want the full picture?
Get the free 5-day "AI in your job" email course — jargon-free, five minutes a day.