How does NN create a concept not in its training data? (Does it explore negative idea-space?) What if a concept uses a word not yet invented? How does LLM produce that word and what cosine similarity would such a word have if it's never appeared next to others? How would we know if such a word is useful?