What are embeddings?
Embeddings are numeric representations (usually vectors) where tokens or words sit in a space and similar meanings tend to be closer together.
Models learn these vectors during training, then use them inside transformer layers for attention and transformations.
mej7.com includes an interactive embeddings section with illustrative dimensions and simplified spatial relationships.
For how researchers read directions inside a model, see what mechanistic interpretability is and the 3D embeddings toy lab. That lab is an illustration, not real model weights.