> So embeddings are bits of binary code that are associated with the words and word-snippets that the computer ‘reads’. These embeddings never change.

I'm not entirely sure but I think this definition of embedding is wrong: first they are not binary, they are floats (as the other parameters/weights) and they do change, as the error backpropagates through them. They are simply "swappable parameters", explicitly corresponding to each word, thus it's possible to detach and reuse them for other purposes after training, which is not necessarily easy (or meaningful) with any other weight matrix in a model.