Tokenization. Pick some text. It splits into tokens, each token has an ID, and each ID picks one row of a learned table: that row is the token’s vector. The IDs and numbers are made up.