[AI] Why are most LLMs decoder-only, and not encoder-only or encoder-decoder, in relation to transformer type?