Word embeddings (that 1000-dimension vector you mention) are not new. No comment on the rest of your comment, but that aspect of LLMs is "old" tech - word2vec was published 11 years ago.