Amazon interview question

If you will be a tech lead to develop an LLM, which architecture you will choose? (i.e., decoder-only or encoder-decoder) and why?