Comment by sharemywin

isn't it closer to concept prediction layered over top of text prediction because of the multiple levels? it compresses text into concepts using layers of embeddings and neural encoding then predicts the concept based on multiple areas of attention. then decompresses it to find the correct words to convey the concept.